text
stringlengths 4
2.78M
| meta
dict |
|---|---|
---
author:
- |
Kurt Mehlhorn[^1]\
Max-Planck-Institut f[ü]{}r Informatik,\
Saarbr[ü]{}cken,\
Germany
- |
Sanjeev Saxena[^2]\
Computer Science and Engineering,\
Indian Institute of Technology,\
Kanpur, INDIA-208 016
title: 'A Still Simpler Way of Introducing Interior-Point Method for Linear Programming'
---
{#section .unnumbered}
Linear programming is now included in algorithm undergraduate and postgraduate courses for computer science majors. We give a self-contained treatment of an interior-point method which is particularly tailored to the typical mathematical background of CS students. In particular, only limited knowledge of linear algebra and calculus is assumed.
Introduction
============
Terlaky [@terlaky] and Lesaja [@lesaja] have suggested simple ways to teach interior-point methods. In this paper, we suggest an alternative and maybe still simpler way which is particularly tailored to the typical mathematical background of CS students. In particular, only limited knowledge of linear algebra and calculus is assumed. We have selected most of the material from popular textbooks [@roos; @van; @BT; @karloff; @ye; @saigal] to assemble a self-contained presentation of an interior point method– little of this material is new.
The canonical *linear programming problem* is to $$\label{primal problem}
\text{minimize $c^Tx$ subject to $Ax=b$ and $x\geq 0$.}$$ Here, $A$ is an $m\times n$ matrix, $c$ and $x$ are $n$-dimensional, and $b$ is an $m$-dimensional vector. A *feasible solution* is any vector $x$ with $Ax = b$ and $x \ge 0$. The problem is *feasible* if there is a feasible solution, and *infeasible* otherwise. A feasible problem is *unbounded* (or more precisely the corresponding objective function is unbounded) if for every real $z$, there is a feasible $x$ with $c^T x \le z$, and *bounded* otherwise.
In our presentation, we first assume that feasible solutions to the primal and the corresponding dual LP satisfying a certain set of properties (properties (I1) to (I3) in Section \[sec:invariants\]) are available. We then show how to iteratively improve these solutions in Sections \[sec:iteration\] and \[sec:invariants\]. In each iteration the gap between the primal and the dual objective value is reduced by a factor $1 - O(1/\sqrt{n})$, where $n$ is the number of variables. The iterative improvement scheme leads to solutions that are arbitrarily close to optimality. In Sections \[sec:init\] and \[sec:extraction\] we discuss how to find the appropriate initial solutions and how to extract an optimal solution from a sufficiently good solution by rounding. Either or both these sections may be skipped in a first course.
It is easy to deal with maximization instead of minimization and with inequality constraints. Indeed, maximize $c^Tx$ is equivalent to minimize $-c^Tx$. Constraints of type $\alpha_1x_1+{\ldots}
+\alpha_nx_n\leq \beta$ can be replaced by $\alpha_1x_1+{\ldots}
+\alpha_nx_n+\gamma = \beta$ with a new (slack) variable $\gamma\geq 0$. Similarly, constraints of type $\alpha_1x_1+{\ldots} +\alpha_nx_n\geq
\beta$ can be replaced by $\alpha_1x_1+{\ldots} +\alpha_nx_n-\gamma =
\beta$ with a (surplus) variable $\gamma\geq 0$.
We consider another problem, the *dual problem*, which is $$\label{dual problem} \text{
maximize $b^T y$, subject to $A^Ty+s=c$, with variables $s\geq 0$ and
unconstrained variables $y$.}$$
The vector $y$ has $m$ components and the vector $s$ has $n$ components. We will call the original problem the *primal problem*.
\[first\]\[weak duality\] If $x$ is a solution of $Ax=b$ with $x\geq 0$ and $(y,s)$ is a solution of $A^Ty+s=c$ with $s\geq 0$, then
$x^Ts=c^Tx-b^Ty$, and
$b^Ty\leq c^Tx$, with equality if and only if $s_ix_i=0$ for all $i$s.
We multiply $s=c-A^Ty$ with $x^T$ from the left and obtain $$x^Ts=x^Tc- x^T(A^Ty)=c^Tx-(x^TA^T)y= c^T x - (A x)^T y = c^Tx-b^Ty .$$ As $x,s\geq 0$, we have $x^Ts\geq 0$, and hence, $c^Tx\geq b^Ty$.
Equality will hold if $x^Ts=0$, or equivalently, $\sum_i s_ix_i=0$. Since $s_i,x_i\geq 0$, $\sum_i s_ix_i=0$ if and only if $s_ix_i=0$ for all $i$.
If $x$ is a feasible solution of the primal and $(y,s)$ is a feasible solution of the dual, the difference $c^T x - b^T y$ is called the *objective value gap* of the solution pair. Thus, if the objective values of a primal feasible and a dual feasible solution are the same, then both solutions are optimal. Actually, from the Strong Duality Theorem, if both primal and dual solutions are optimal, then the equality will hold. We will prove the Strong Duality Theorem in Section \[part\] (Corollary \[strong:dual\]).
If the primal and the dual are both feasible, neither of them can be unbounded as by Claim \[first\], the objective value of all dual feasible solutions are less than or equal to the objective values of any primal feasible solution. As a consequence: If the primal and the dual are feasible, both are bounded. If the primal is unbounded, the dual is infeasible, and if the dual is unbounded, the primal is infeasible. It may happen that both problems are infeasible. It is also true, that if the primal is feasible and bounded, the dual is feasible and bounded, and vice versa. This is a consequence of strong duality.
*We will proceed under the assumption that the primal as well as the dual problem are bounded and feasible.* This allows us to concentrate on the core of the interior point method, the iterative improvement scheme. We come back to this point in Section \[sec:init\].
Claim \[first\] implies, that if we are able to find a solution to the following system of equations and inequalities $$Ax=b,\ A^Ty+s=c,\ x_i
s_i = 0 \text{ for all $i$},\ x\geq 0,\ s\geq 0,$$ we will get optimal solutions of both the original primal and the dual problem. Notice that the constraints $x_i s_i = 0$ are nonlinear and hence it is not clear whether we have made a step towards the solution of our problem. The idea is now to relax the conditions $x_i s_i = 0$ to the conditions $x_i s_i
\approx \mu$ (with the exact form of this equation derived in the next section), where $\mu \ge 0$ is a parameter. We obtain $$(P_\mu) \quad
Ax=b,\ A^Ty+s=c,\ x_i s_i \approx \mu \text{ for all $i$},\ x > 0,\ s >
0.$$ We will show:
(initial solution) For a suitable $\mu$, it is easy to find a solution to the problem $P_\mu$. This will be the subject of Section \[sec:init\].
(iterative improvement) Given a solution $(x,y,\mu)$ to $P_\mu$, one can find a solution $(x',y',s')$ to $P_{\mu'}$, where $\mu'$ is substantially smaller than $\mu$. This will be the subject of Sections \[sec:iteration\] and \[sec:invariants\]. Applying this step repeatedly, we can make $\mu$ arbitrarily small.
(final rounding) Given a solution $(x,y,\mu)$ to $P_\mu$ for sufficiently small $\mu$, one can extract an exact solutions for the primal and the dual problem. This will be the subject of Section \[part\].
![\[central path\] The interior of the polygon comprises all points $(x,y,s)$ satisfying $Ax = b$ and $A^T y + s = c$, $x > 0$, and $s
> 0$. The blue (bold) line consists of all points in this polygon with $x_i s_i = \mu$ for all $i$ and some $\mu > 0$. These points trace a line inside the polygon that ends in an optimal point. The optimal solution lies on the boundary of the polygon (in the figure, the optimal point is a vertex of the polygon) and satisfies $x_i s_i = 0$ for all $i$. The red (dashed) line illustrates the steps of the algorithm. It follows the blue (bold) line in discrete steps. The close-up shows the situation near the optimal solution. The algorithm stops tracing the blue curve and rounds to the near-optimal red solution obtained at this point of time to an optimal solution. ](centralpath-1a.pdf){width="60.00000%"}
For the iterative improvement, it is important that $x > 0$ and $s > 0$. For this reason, we replace the constraints $x \ge 0$ and $s \ge 0$ by $x
> 0$ and $s > 0$ when defining problem $P_\mu$ (see Figure \[central path\]).
Note that $x_i s_i \approx \mu$ for all $i$ implies $b^Ty-c^Tx \approx
n\mu$ by Claim \[first\]. Thus, repeated application of iterative improvement will make the gap between the primal and dual objective values arbitrarily small.
*Throughout the paper we assume that the rows of $A$ are linearly independent and that $n > m$*, i.e., we have more variables than constraints.[^3]
Iterative Improvement: Use of the Newton-Raphson Method {#sec:iteration}
=======================================================
This section and the next follow Roos et al [@roos] (see also Vishnoi [@vishnoi]).
Let us assume that we have a solution $(x,y,s)$ to $$Ax=b \mbox{ and } A^Ty+s=c \text{ and } x > 0 \text{ and } s > 0.$$ We will use the Newton-Raphson Method [@roos] to get a “better” solution. Let us choose the next values as $x' = x+h$, $y' = y+k$, and $s' = s+f$. We can think of the steps $h$, $k$, and $f$ as small values. Then we want, ignoring the positivity constraints for $x'$ and $s'$ for the moment:
1. $Ax' = A(x+h)=b$, or equivalently, $Ax+Ah=b$. Since $Ax=b$, this is tantamount to $Ah=0$.
2. $A^T y' + s' = A^T(y+k)+(s+f)=c$. Since $A^Ty+s=c$, we get $A^Tk+f=c-A^Ty-s=0$.
3. $x'_i s'_i = (x_i+h_i)(s_i+f_i)\approx \mu'$, or equivalently, $x_is_i+h_is_i+f_ix_i+h_if_i\approx \mu'$. We drop the quadratic term $h_i f_i$ (if the steps $h_i$ and $f_i$ are small, the quadratic term $h_i f_i$ will be very small) and turn the approximate equality into an equality, i.e., we require $x_is_i+h_is_i+f_ix_i=\mu'$ for all $i$.
Thus, we have a system of linear equations for $h_i,k_i,f_i$, namely, $$\begin{aligned}
&&A h &= 0\\
\text{system (S)}&& A^T k + f &= 0\\
&&h_i s_i+f_i x_i &=\mu' - x_i s_i \quad \text{for all $i$}\end{aligned}$$ We show in Theorem \[theorem\] that system (S) can be solved by “inverting” a matrix. Note that there are $n$ variables $h_i$, $m$ variables $k_j$, and $n$ variables $f_i$ for a total of $2n + m$ unknowns. Also note that $Ah = 0$ constitutes $m$ equations, $A^T k + f =
0$ constitutes $n$ equations, and $h_i s_i+f_i x_i =\mu' - x_i s_i$ for all $i$ comprises $n$ equations. So we have $2n + m$ equations and the same number of unknowns. Also note that the $x_i$ and $s_i$ are *not* variables in this system, but fixed values.
Before we show that the system has a unique solution, we make some simple observations. From the third group of equations, we conclude
\[obs:lp:remark:1\] $(x_i+h_i)(s_i+f_i) = \mu'+h_if_i$, and $(x + h)^T (s + f) = n \mu' + h^T
f$.
From the third group of equations, we obtain $$(x_i+h_i)(s_i+f_i) = x_is_i+h_is_i+f_ix_i+h_if_i = \mu' + h_i f_i.$$ Summation over $i$ yields $$(x + h)^T (s + f) = \sum_i
(x_i+h_i)(s_i+f_i) = \sum_i \left(\mu' + h_i f_i\right) = n \mu' + h^T f.
\vspace{-5ex}$$
\[claim:hifi:1\] $h^Tf=f^Th= \sum_i h_i f_i = 0$, i.e., the vectors $h$ and $f$ are orthogonal to each other.
Multiplying $A^Tk+f=0$ by $h^T$ from the left, we obtain $h^TA^Tk+h^Tf=0$. Since $h^TA^T=(Ah)^T=0$, the equality $h^Tf=0$ follows.
\[second\] $c^T (x + h) - b^T (y + k)= (x+h)^T(s+f)=n\mu'$.
From Claims \[obs:lp:remark:1\] and \[claim:hifi:1\], $(x+h)^T(s+f)= n \mu' + h^T f = n \mu'$. Also, applying Claim \[first\] to the primal solution $x' = x+h$ and to the dual solution $(y',s') =
(y+k,s+f)$ yields $c^T (x + h) - b^T (y + k) = (x + h)^T (s + f)$.
Note that $n \mu'$ is the objective value gap of the updated solution.
\[theorem\] The system (S) has a unique solution.
We will follow Vanderbei [@van] and use capital letters (e.g. $X$) in this proof (only) to denote a diagonal matrix with entries of the corresponding row vector (e.g. $X$ has the diagonal entries $x_1,x_2,{\ldots} ,x_n$). We will also use $e$ to denote a column vector of all ones (usually of length $n$).
Then, in the new notation, the last group of equations becomes $$S h+Xf=\mu' e-XSe.$$ Let us look at this equation in more detail. $$\begin{aligned}
Sh+Xf &=\mu' e-XSe\\
h+ S^{-1}Xf &= S^{-1}\mu' e- S^{-1}XSe &\mbox{ pre-multiply by } S^{-1}\\
h+S^{-1}Xf &= \mu' S^{-1}e -X {S^{-1}}{S} e &\mbox{ diagonal
matrices commute}\\
h+S^{-1}Xf &= \mu' S^{-1}e -x &\mbox{ as } Xe=x\\
Ah+ AS^{-1}Xf &= \mu' AS^{-1}e-Ax &\mbox{ pre-multiply by } A\\
AS^{-1}Xf &= \mu' AS^{-1}e-b &\mbox{ since } Ax=b \mbox{ and } Ah=0\\
-AS^{-1}XA^Tk &= \mu' AS^{-1}e-b &\mbox{ using } f=-A^Tk\\
b-\mu' AS^{-1}e &= (AS^{-1}XA^T)k\end{aligned}$$ As $XS^{-1}$ is diagonal with positive items, the matrix $W=\sqrt{XS^{-1}}$ is well-defined. Note that the diagonal terms are $\sqrt{x_i/s_i}$; since $x > 0$ and $s > 0$, we have $x_i/s_i > 0$ for all $i$. Thus, $AS^{-1}XA^T = A W^2 A^T = (AW)(AW)^T$. Since $A$ has full rank, $(AW)(AW)^T$, and hence $AS^{-1}XA^T$, is invertible (see Appendix). Thus, $$k=(AS^{-1}XA^T)^{-1}\left(b-\mu' AS^{-1}e\right).$$ Then, we can find $f$ from $f=-A^Tk$. And to get $h$, we use the equation: $h+S^{-1}Xf = \mu' S^{-1}e -x$, i.e., $$h= -XS^{-1}f+\mu' S^{-1}e-x.$$ Thus, system $(S)$ has a unique solution.
What have we achieved at this point? Given feasible solutions $(x,y,s)$ to the primal and dual problem, we can compute a solution $(x',y',s') =
(x + h, y + k, s + f)$ to $Ax' = b$ and $A^T y' + s' = c$ that also satisfies $h^T f = 0$ and $ x'^T s = n \mu'$ for any prescribed parameter $\mu'$. Why do we not simply choose $\mu' = 0$ and be done? It is because we have ignored that we want $x' > 0$ and $s' > 0$. We will attend to these constraints in the next section.
Invariants in each Iteration {#sec:invariants}
============================
Recall that we want to construct solutions $(x,y,s)$ to $P_\mu$ for smaller and smaller values of $\mu$. The solution to $P_\mu$ will satisfy the following invariants. The first two invariants state that $x$ is a positive solution to the primal and $(y,s)$ is a solution to the dual with positive $s$. The third invariant formalizes the condition $x_i
s_i \approx \mu$ for all $i$.
(I1)
: (primal feasibility) $Ax=b$ with $x>0$ (strict inequality).
(I2)
: (dual feasibility) $A^Ty+s=c$ with $s>0$ (strict inequality).
(I3)
: $\sigma^2 {\mathrel{:=}}\sum_i \left(\frac{x_i
s_i}{\mu}-1\right)^2 \le \frac{1}{4}$.
Even though the variance of ${x_is_i}$ is $\frac{1}{n} \sum_i
\left(x_i s_i-\mu \right)^2$, we still use the notation $\sigma^2$.
We need to show $$x' > 0 \text{ and } s' > 0 \text{ and }
\sigma'^2 {\mathrel{:=}}\sum_ i \left(\frac{x'_i s'_i}{\mu'} - 1 \right)^2 \le
\frac{1}{4}.$$ We will do so for $\mu' = (1 - \delta) \mu$ and $\delta=\Theta\left(\frac{1}{\sqrt{n}}\right)$. Claim \[obs:lp:remark:1\] gives us an alternative expression for $\sigma'^2$, namely, $$\label{sigma-prime-alternative}
\sigma'^2=\sum_i
\left(\frac{(x_i+h_i)( s_i+f_i)}{\mu'}-1\right)^2=\sum_i
\left(\frac{h_if_i}{\mu'}\right)^2$$ We first show that the positivity invariants hold if $\sigma'$ is less than one.
\[fact:int:1:1\] If $\sigma'<1$, then $x' >0$, and $s' >0$.
We first show that if $\sigma'<1$ then each product $x'_i s'_i =
(x_i+h_i)(s_i+f_i) = \mu' + h_i f_i$ is positive. From $\sigma'<1$, we get $\sigma'^2< 1$. Since $\sigma'^2 = \sum_i \left({h_if_i}/{\mu'}
\right)^2$, each term of the summation must be less than one, and hence, $-\mu'< h_if_i<\mu'$. In particular, $\mu'+h_if_i>0$ for every $i$. Thus, each product $(x_i+h)(s_i+f)$ is positive.
Assume for the sake of a contradiction that both $x_i+h_i<0$ and $s_i+f_i<0$. But as $s_i>0$ and $x_i>0$, this implies $s_i(x_i+h_i)+x_i(s_i+f_i)<0$, or equivalently, $\mu' +x_is_i<0$, which is impossible because $\mu',x_i,s_i$ are all non-negative. This is a contradiction.
We next show $\sigma' \le 1/2$. We first establish
\[helper\] $\frac{\mu}{x_is_i} \le \frac{1}{1 -
\sigma}$ for all $i$ and $\sum_i {\left\vert \hspace{\minispacing} 1 - \frac{x_i s_i}{\mu} \hspace{\minispacing} \right\vert} \le
\sqrt{n} \cdot \sigma$.
As $\sigma^2 = \sum_i \left(1 - {x_i s_i}/{\mu}\right)^2$, each individual term in the sum is at most $\sigma^2$. Thus, ${\left\vert \hspace{\minispacing} 1 -
{x_i s_i}/{\mu} \hspace{\minispacing} \right\vert} \le \sigma$, and hence, ${x_is_i}/{\mu} \ge 1 - \sigma$, and further, ${\mu}/{x_is_i} \le {1}/(1 - \sigma)$.
For the second claim, we have to work harder. Consider any $n$ reals $z_1$ to $z_n$. Then $(\sum_i {\left\vert \hspace{\minispacing} z_i \hspace{\minispacing} \right\vert})^2 \le n \sum_i z_i^2$; this is the frequently used inequality between the one-norm and the two-norm of a vector[^4]. We apply the inequality with $z_i = 1 - {x_i s_i}/{\mu}$ and obtain the second claim.
Let us define two new quantities $$H_i = h_i\sqrt{\frac{s_i}{x_i\mu'}} \quad\text{and}\quad F_i =
f_i\sqrt{\frac{x_i}{s_i\mu'}}.$$
Observe that $\sum_i H_iF_i=\sum \frac{h_if_i}{\mu'}=0$ (from Claim \[claim:hifi:1\]) and $\sum_i (H_i F_i)^2 = \sum_i
\left(\frac{h_i f_i}{\mu'}\right)^2 = \sigma'^2$. Also,
$$\begin{aligned}
H_i + F_i &= \sqrt{ \frac{1}{x_i s_i \mu'}} \left(h_i s_i + f_i x_i\right)
= \sqrt{ \frac{1}{x_i s_i \mu'}} \left(\mu' - \mu + \mu - x_i s_i\right)
\nonumber\\
&= \sqrt{ \frac{\mu}{x_i s_i}\frac{\mu}{\mu'}} \left(\frac{\mu'}{\mu} -1
+ 1 - \frac{x_i s_i}{\mu}\right) = \sqrt{\frac{\mu}{x_is_i (1 - \delta)}}
\left(-\delta + 1 - \frac{x_i s_i}{\mu}\right). \label{Hi + Fi}\end{aligned}$$
Finally, $$\begin{aligned}
\sigma'^2 &= \sum_i (H_i F_i)^2 = \frac{1}{4} \left(\sum_i (H_i^2 + F_i^2)^2 - \sum_i (H_i^2 - F_i^2)^2\right)\\
&\le \frac{1}{4} \sum_i (H_i^2 + F_i^2)^2 &\text{since $\sum_i (H_i^2 - F_i^2)^2 \ge 0$}\\
&\le \frac{1}{4} \left(\sum_i (H_i^2 + F_i^2) \right)^2 &\text{more positive terms}\\
&= \frac{1}{4} \left(\sum_i (H_i + F_i)^2 \right)^2 &\text{ since $H^T F = 0$}\\
&=\frac{1}{4} \left( \sum_i \frac{\mu}{x_is_i (1 - \delta)} \left(-\delta + 1 - \frac{x_i s_i}{\mu}\right)^2 \right)^2 &\text{by {(\ref{Hi + Fi})}}\\
&\le \frac{1}{4 (1 - \delta)^2 (1 - \sigma)^2} \left( \sum_i \left(-\delta + 1 - \frac{x_i s_i}{\mu}\right)^2 \right)^2 &\text{since $\mu/(x_is_i) \le 1/(1 - \sigma)$}\\
&\le \frac{1}{4 (1 - \delta)^2 (1 - \sigma)^2} \left( n \delta^2 -
2\delta \sum_i \left(1 - \frac{x_is_i}{\mu}\right) + \sum_i \left(1 -\frac{x_i s_i}{\mu}\right)^2 \right)^2 &\text{remove inner square}\\
&\le \frac{1}{4 (1 - \delta)^2 (1 - \sigma)^2} \left( n \delta^2 +
2\delta \sum_i {\left\vert \hspace{\minispacing} 1 - \frac{x_is_i}{\mu} \hspace{\minispacing} \right\vert} +
\sum_i \left(1 - \frac{x_i s_i}{\mu}\right)^2 \right)^2\\
&\le \frac{1}{4 (1 - \delta)^2 (1 - \sigma)^2} \left( n \delta^2 + 2\delta \sqrt{n} \cdot \sigma + \sigma^2 \right)^2 &\text{by Claim~\ref{helper}}\\
&= \frac{1}{4 (1 - \delta)^2 (1 - \sigma)^2} \left( \left( \sqrt{n} \delta + \sigma\right)^2 \right)^2, &\text{forming inner square}\end{aligned}$$ and hence, $$\label{defining inequality}
\sigma' \le \frac{\left(\sqrt{n} \delta + \sigma\right)^2}{2 (1 -
\sigma)(1 - \delta)} \le \frac{\left(\sqrt{n} \delta + 1/2 \right)^2}{2
(1 - 1/2) (1 - \delta)} \stackrel{!}{\le} \frac{1}{2},$$ where the second inequality holds since the bound for $\sigma'$ is increasing in $\sigma$, and $\sigma \le 1/2$. We need to choose $\delta$ such that the last inequality holds. This is why we put an exclamation mark on top of the $\le$-sign. Setting $\delta = {c}/\!\sqrt{n}$ for some to be determined constant $c$ yields the requirement $$\frac{\left( c +
1/2 \right)^2}{(1 - \delta)} \stackrel{!}{\le} \frac{1}{2}, \quad\text{or
equivalently,}\quad \left( 2c + 1 \right)^2 \stackrel{!}{\le} 2 \left(1 -
\frac{c}{\sqrt{n}}\right).$$ This holds true for $c = 1/8$ and all $n \ge
1$. Thus, $\delta = 1/(8 \sqrt{n})$.
Why do we require $\sigma \le 1/2$ in the invariant? Let us formulate the bound as $\sigma \le \sigma_0$ for some to be determined $\sigma_0$. Then, the inequality [(\[defining inequality\])]{} becomes $$\frac{\left(\sqrt{n} \delta + \sigma_0 \right)^2}{2 (1 - \sigma_0) (1 -
\delta)} \stackrel{!}{\le} \sigma_0.$$ We want this to hold for $\delta =
\frac{c}{\sqrt{n}}$ and some $c > 0$. In order for the inequality to hold for $c = 0$, we need $\sigma_0 \le 2(1 - \sigma_0)$, or equivalently, $\sigma_0 \le 2/3$. Since we want it to hold for some positive $c$, we need to choose a smaller $\sigma_0$; $1/2$ is a nice number smaller than $2/3$.
#### An Alternative Proof for Invariant (I3) (provided by Andreas Karrenbauer)
Andreas Karrenbauer derived an alternative proof for invariant (I3) that avoids introduction of the quantities $H$ and $F$ and is more compact than the above.
\[alternative proof\] Assume $\delta \le 1/6$. Then $\sigma \le \delta$ implies $\sigma' \le
\delta$.
As $\sigma^2=\sum
\left(\frac{x_is_i}{\mu}-1\right)^2\le \delta^2$, each individual term must be bounded by $\delta^2$. Thus, $\sigma \le \delta$ implies $\left|\frac{x_is_i}{\mu}-1\right|\le \delta$, or $-\delta\leq
\frac{x_is_i}{\mu}-1$ or $x_i s_i \ge (1 - \delta) \mu$.
We define $${{\| \sigma' \|}_1} = \sum_i {\left\vert \hspace{\minispacing} \frac{(x_i + h_i)(s_i +f_i)}{\mu'} -
1 \hspace{\minispacing} \right\vert}$$ Then from the definition of $\mu'$ and triangle inequality, $${{\| \sigma' \|}_1} = \sum_i {\left\vert \hspace{\minispacing} \frac{(x_i + h_i)(s_i +f_i)}{\mu'} -
1 \hspace{\minispacing} \right\vert}\le \sum_i {\left\vert \hspace{\minispacing} \frac{{x_i s_i + x_i f_i + h_i
s_i}}{{\mu'}} - 1 \hspace{\minispacing} \right\vert} + \sum_i
{\left\vert \hspace{\minispacing} \frac{ h_i f_i}{\mu'} \hspace{\minispacing} \right\vert} = \sum_i {\left\vert \hspace{\minispacing} \frac{ h_i f_i}{\mu'} \hspace{\minispacing} \right\vert}.$$ Again from $x_i f_i + s_i h_i = \mu' - x_i s_i$, we obtain (by squaring) $$\label{hifi}
h_i f_i = \frac{1}{2 x_i s_i} \left[ (\mu' - x_i s_i)^2 - (h_i s_i)^2 -
(x_i f_i)^2\right]$$ Summing over $i$ and using the fact that from Claim \[claim:hifi:1\], $f^T h = 0$ we obtain $$\label{sum}
\sum_i (\mu' - x_i s_i)^2 = \sum_i \left( (h_i s_i)^2 + (x_i f_i)^2
\right).$$
Assume that $\mu' = (1 - \tau) \mu$ for a $\tau$ to be fixed later. Then $$\begin{aligned}
{{\| \sigma' \|}_1}
&\le \sum_i {\left\vert \hspace{\minispacing} \frac{ h_i f_i}{\mu'} \hspace{\minispacing} \right\vert} \\
&\le \sum_i \frac{1}{2 x_i s_i
\mu'} \left[ (\mu' - x_i s_i)^2 + (h_i s_i)^2 + (x_i f_i)^2\right]
&\text{by~{(\ref{hifi})}} \\
&= \sum_i \frac{ (\mu' - x_i s_i)^2}{x_i s_i \mu'}
&\text{by~{(\ref{sum})}}\\
&\le \sum_i \frac{\left(\frac{\mu'}{\mu}
- \frac{x_i s_i}{\mu}\right)^2}{(1 - \delta) (1 - \tau)}
&\parbox{3cm}{\text{as } $x_i s_i \ge (1 - \delta) \mu$\\
\text{and } $\mu'=(1-\tau)\mu$}\\
&= \sum_i \frac{ \left(\tau - \left(\frac{x_is_i}{\mu}-1\right)\right)^2}{(1
- \delta) (1 - \tau)}\\
&= \frac{n \tau^2 -2 \tau \sum \left(\frac{x_is_i}{\mu}-1\right)+ \sigma^2}{(1
-\delta)(1 - \tau)}\\
&\le
\frac{n \tau^2 + 2 \tau {{\| \sigma \|}_1} + \sigma^2}{(1 -\delta)(1 -
\tau)}\\
&\le \frac{(\sqrt{n} \tau + \delta)^2}{(1 - \delta)(1 - \tau)}
&\text{since ${{\| \sigma \|}_1} \le \sqrt{n} \delta$ (Claim~\ref{helper})}\\
&\le \frac{4 \delta^2}{(1 - \delta)(1 - \delta/\sqrt{n})} &\text{for the choice $\tau = \delta/\sqrt{n}$}\\
&\le \delta &\text{for $\delta \le 1/6$}\end{aligned}$$ The claim follows as the two norm is always less than the one norm[^5], $\sigma'={{\| \sigma' \|}_2} \le
{{\| \sigma' \|}_1}$.
Initial Solution {#sec:init}
================
This section follows Bertsimas and Tsitsiklis [@BT p430]; see also Karloff [@karloff p128-129]. We have to deal with three problems:
1. how to make sure that we are dealing with a bounded problem
2. how to make sure that the problem is feasible and if the problem is feasible, then how to find an initial solution
3. how to guarantee condition (I3) for the initial solution.
A standard solution for the second problem is the *big M method*. Let $x_0 \ge 0$ be an arbitrary nonnegative column vector of length $n$. We introduce a new variable $z \ge 0$, change $Ax = b$ into $Ax + (b-A
x_0) z = b$ and the objective into “minimize $c^T x + Mz$”, where $M$ is a big number. Note that $x = x_0$ and $z = 1$ is a feasible solution to the modified problem. We solve the modified problem. If $z^*= 0$ in an optimal solution, we have also found an optimal solution to the original problem. If $z^* > 0$ in an optimal solution and $M$ was chosen big enough, the original problem is infeasible.
There are several other methods of dealing with the problem of getting a starting solution. These include self-dual method [@terlaky; @van] and the infeasible interior point method [@MKD; @ziang].
*We assume for the remainder of the presentation that $A$, $b$, and $c$ are integral and that $U$ is an integer with $U \ge {\left\vert \hspace{\minispacing} a_{ij} \hspace{\minispacing} \right\vert},
{\left\vert \hspace{\minispacing} b_i \hspace{\minispacing} \right\vert}, {\left\vert \hspace{\minispacing} c_j \hspace{\minispacing} \right\vert}$ for all $i$ and $j$.*
We need the following Fact which we will prove in Section \[bounds\].
\[bounded by W\] Let $W = (mU)^m$. If [(\[primal problem\])]{} is feasible, there is a feasible solution with all coordinates bounded by $W$. If, in addition the problem is bounded, there is an optimal solution with this property.
We now give the details. We add the constraint $e^T x+z \le (n+2) W$. If the problem was feasible, it will stay feasible. If the problem was bounded, the additional constraint does not change the optimal objective value. If the problem was unbounded, the additional constraint makes it bounded. Using an additional slack variable $x_{n+1}$ we get the equality $e^Tx+x_{n+1}+z= (n+2)W$. If we use “normalized variables” $x'_i=\frac{x_i}{W}$, drop the primes and use $x_{n+2}$ for $z$, we obtain the following auxiliary primal problem. $$\begin{aligned}
\label{artificial primal}
\begin{array}{r r c c c c c l }
\text{minimize $c^Tx +Mx_{n+2}$, subject to}\quad & Ax & & & + & \rho
x_{n+2} & = & d\\
& e^T x & + & x_{n+1} & + & x_{n+2} & = & n+2\\
& \multicolumn{2}{l}{x \ge 0} & \multicolumn{2}{l}{x_{n+1} \ge
0}&\multicolumn{2}{l}{x_{n+2} \ge 0},
\end{array}\end{aligned}$$ where $d = \frac{1}{W} b$, $\rho = d - Ae$, and we show later in this section, that $M$ can be chosen as $M = 4nU/{R}$, where ${R}=
\frac{1}{W^2} \cdot \frac{1}{2n \left((m+1) U\right)^{3(m+1)}}$. In matrix form, the auxiliary primal is $$A' \left( \begin{matrix} x \\ x_{n+1} \\ x_{n+2} \end{matrix}
\right) = b', \quad\text{where}\quad A' = \left(\begin{matrix}A & 0 & \rho \\
e^T & 1 & 1 \end{matrix} \right)
\text{ and } b' = \left(
\begin{matrix} d \\ n+2
\end{matrix}\right)$$ We make the following observations.
As $x_i = 1$ for $1 \le i \le n+2$ is a feasible solution, [(\[artificial primal\])]{} is feasible. The feasible region is a polytope contained in the cube defined by $0 \le x_i \le n+2$ for all $i$. The following Fact is shown in Section \[integer\].
\[nonzero coordinates\] The nonzero coordinates of the vertices of this polytope are at least ${R}$.
As $0 \le x_i \le n+2$ and $c_i \ge -U$ for all $i$, the objective value is at least $-U(n+2)$ Thus, [(\[artificial primal\])]{} is bounded.
If $x$ is feasible solution to [(\[primal problem\])]{} with $x_i \le W$ for $1\leq i\leq n$ then $(\frac{1}{W} x, (n+2) - \frac{1}{W}
e^T x, 0)$ is a feasible solution to [(\[artificial primal\])]{} with objective value $\frac{1}{W} c^T x$.
In particular, if [(\[primal problem\])]{} is feasible, then [(\[artificial primal\])]{} has a solution with objective value less than or equal to $nU$. This follows from $x_i/W \le 1$ and $c_i \le U$ for $1\le
i \le n$.
We next show that if [(\[artificial primal\])]{} has an optimal solution $(x^*, x^*_{n+1},x^*_{n+2})$ with $x^*_{n+2} = 0$ then [(\[primal problem\])]{} is feasible. Indeed, $A W x^* = W A x^* = W d = b$ and hence $W x^*$ is feasible for [(\[primal problem\])]{}. If, in addition, [(\[primal problem\])]{} is bounded, $W x^*$ is an optimal solution of [(\[primal problem\])]{}. Note that if [(\[primal problem\])]{} is bounded, it has an optimal solution $x$ with $x_i \le W$ by Fact \[bounded by W\]. This solution induces a solution of [(\[artificial primal\])]{} with objective value $\frac{1}{W} c^T x$ by the preceding item. The optimality of $(x^*, x^*_{n+1},x^*_{n+2})$ implies $c^T x^* \ge \frac{1}{W} c^T x$.
We finally show that if [(\[artificial primal\])]{} has an optimal solution with $x^*_{n+2} > 0$, [(\[primal problem\])]{} is infeasible. Indeed, then there must be an optimal vertex solution of [(\[artificial primal\])]{}. For this vertex, $x^*_{n+2} \ge {R}$. The objective value of this solution is at least $M\cdot {R}- (n+2) U = 2nU$. On the other hand, if [(\[primal problem\])]{} is feasible, [(\[artificial primal\])]{} has a solution with objective value at most $nU$. Any value of $M$ for which $M\cdot {R}- (n+2) U > nU$ would work for this argument. $M = 4 U/R$ is one such value. This explains the choice of $M$.
We summarize: Our original problem is feasible if and only if $x^*_{n+2}
= 0$ in every optimal solution to [(\[artificial primal\])]{} if and only if $x^*_{n+2} = 0$ in some optimal solution to [(\[artificial primal\])]{}. Moreover, if $x^*_{n+2} = 0$, and [(\[primal problem\])]{} is bounded, $\frac{1}{W} x^*$ is an optimal solution of [(\[primal problem\])]{}.
By the above, our original problem is feasible if and only if $x^*_{n+2} = 0$ in an optimal solution to [(\[artificial primal\])]{}. So we can distinguish feasible and infeasible problems. How can we distinguish bounded and unbounded problems? Note that the primal is unbounded if it is feasible and the problem “minimize 0 subject to $c^T
x = -1$, $Ax = 0$, and $x \ge 0$” is feasible. So the test for unboundedness reduces to two feasibility tests.
The dual problem (with new dual variables $y_{m+1}, s_{n+1}$ and $s_{n+2}$) is $$\begin{aligned}
\label{artificial dual}
\text{maximize } & d^Ty+(n+2)y_{m+1}, \text{ subject to } & A^Ty+e
y_{m+1}+s &=c, \\ \nonumber&&\rho^T y+y_{m+1}+s_{n+2}& =M\\ \nonumber
&&y_{m+1}+s_{n+1}&=0 \end{aligned}$$
with slack variables $s\geq 0, s_{n+1} \ge 0,
s_{n+2} \ge 0$ and unconstrained variables $y$.
Which initial solution should we choose? Recall that we also need to satisfy (I3) for some choice of $\mu$, i.e., $\sum_{1 \le i \le n+2} (x_i
s_i/\mu - 1)^2 \le 1/4$. Also, recall that we set $x_i$ to $1$ for all $i$. As $x_{n+1}=1$, we choose $s_{n+1}= {\mu}/{x_{n+1}}=\mu$. Then, from the last equation, $y_{m+1}=-s_{n+1}=-\mu$. The simplest choice for $y$ is $y=0$. Then, from the first equation, $s=c+e\mu$, and from the second equation $s_{n+2}=M-y_{m+1}=M+\mu$. Observe that all slack variables are positive (provided $\mu$ is large enough). For this choice, $$\begin{aligned}
\frac {x_i s_i}{\mu}-1&=\frac{c_i}{\mu} &\text{for $i \le n$}\\
\frac{x_{n+1}s_{n+1}}{\mu} - 1 &=0\\
\frac{x_{n+2}s_{n+2}}{\mu}-1&=\frac{M}{\mu}.\end{aligned}$$ Thus, $\sigma^2=\left(M^2+\sum c^2_i\right)/\mu^2$. We can make $\sigma^2
\le {1}/{4}$ by choosing $$\label{definition of mu0}
\mu^2= 4\left(M^2+\sum c^2_i\right).$$
#### Summary:
Let us summarize what we have achieved.
For the auxiliary primal problem and its dual, we have constructed solutions $(x^{(0)},y^{(0)},s^{(0)})$ that satisfy the invariants for $\mu^{(0)} = 2\left(M^2+\sum c^2_i\right)^{1/2}$.
From the initial solution, we can construct a sequence of solutions $(x^{(t)},y^{(t)},s^{(t)})$ and corresponding $\mu^{(t)}$ such that
$x^{(t)}$ is a solution to the auxiliary primal,
$(y^{(t)},s^{(t)})$ is a solution to its dual,
$\mu^{(t)} = (1 - \delta) \cdot \mu^{(t-1)} = (1 - \delta)^t \cdot
\mu^{(0)}$, and $\sum_j \left(x_j^{(t)}s_j^{(t)}/\mu^{(t)} - 1\right)^2
\le 1/4$.
For $t \ge 1$, the difference between the primal and the dual objective value is exactly $(n+2) \mu^{(t)}$ (Claim \[second\]). The gap decreases by a factor $1 - \delta = 1 - 1/(8\sqrt{n+2})$ in each iteration, and hence, can be made arbitrarily small.
In the next section, we will exploit this fact and show how to extract the optimal solution. Before doing so, we show the existence of an optimal solution.
[Existence of an Optimal Solution:]{} This paragraph requires some knowledge of calculus, namely continuity and accumulation point. Our sequence $(x^{(t)},y^{(t)},s^{(t)})$ has an accumulation point (this is clear for the sequence of $x^{i}$ since the $x$-variables all lie between $0$ and $n+2$ and we ask the reader to accept it for the others). Then there is a converging subsequence. Let $(x^*,y^*,s^*)$ be its limit point. Then $x^*$ and $(y^*,s^*)$ are feasible solutions of the artificial primal and its dual respectively, and $x_i^* s_i^* = 0$ for all $i$ by continuity.
Extracting an Optimal Solution {#part}
==============================
\[sec:extraction\]
We will show how to round an approximate solution for the auxiliary problems for a sufficiently small $\mu$ to an optimal solution. This section is similar to [@ye Theorem 5.3] and to the approach in [@roos Section 3.3]. See also [@greenberg]. The auxiliary problem has $m + 1$ constraints in $n+2$ variables. The auxiliary dual problem has $n + 2$ constraints in $m + 1 + n + 2$ variables. We use $x$ to denote the variables of the auxiliary primal including $x_{n+1}$ and $x_{n+2}$, and $y$ and $s$ for the variable vectors of the dual (including the additional variables). Moreover, we use $A$ for the entire constraint matrix and $b$ for the full right hand side. So $A$ is $(m+1)
\times (n+2)$, $b$ is a $(m+2)$-vector and $c$ is a $(n+2)$-vector.
Consider an iterate $(x,y,s,\mu)$. We will first show that $x_i \ge x^*_i/(4(n+2))$ and $s_i \ge s_i^*/(4(n+2))$ for all optimal solutions $x^*$ and $(y^*,s^*)$ (Lemma \[conclude zero\]), i.e., if $x_i^* > 0$ ($s_i^* > 0$) for some $i$, then $x_i$ ($s_i)$ cannot become arbitrarily small. However, since $x_i s_i \le 2\mu$ always and $\mu$ decreases exponentially, at least one of $x_i$ or $s_i$ has to become arbitrarily small. We use this observation to conclude that if $x_i$ is sufficiently small (Lemma \[B and N\] quantifies what sufficiently small means) then $x^*_i = 0$ in every optimal primal solution. Similarly, if $s_i$ is sufficiently small, then $s^*_i = 0$ in every optimal dual solution.
Let $N$ be the set of indices for which we can conclude $x^*_i = 0$ and let $B$ be the set of indices for which we can conclude $s^*_i = 0$. We show $B \cup N = {\{ 1,\ldots,n \}}$ and $B \cap N = \emptyset$. We split our last iterate ${\bar{x}}$ into two parts ${\bar{x}}_B$ and ${\bar{x}}_N$ accordingly, round the $N$-part to zero and recompute the $B$-part. Since the coordinates in the $N$-part are tiny, this has little effect on the $B$-part and hence the solution stays feasible. It stays optimal because of complementary slackness.
\[conclude zero\] Let $(x,y,s,\mu)$ satisfy (I1) to (I3).
1. For all $i \in {\{ 1,\ldots,n \}}$: $x_i \ge x^*_i/(4(n+2))$ for every optimal solution $x^*$ of the auxiliary primal.
2. For all $i \in {\{ 1,\ldots,n \}}$: $s_i \ge s^*_i/(4(n+2))$ for every optimal solution $(y^*,s^*)$ of the auxiliary dual.
By (I1) and (I2), $x$ is a feasible solution of the auxiliary primal and $(y,s)$ a feasible solution of the auxiliary dual. By (I3), we have $\sigma^2 = \sum_i (\frac{x_is_i}{\mu} - 1)^2 \le
\frac{1}{4}$. Thus, $(\frac{x_is_i}{\mu} - 1)^2 \le \frac{1}{4}$, and hence, $\mu/2 \le x_i s_i \le 3\mu/2 < 2\mu$ for all $i$. Further, $x^T s
= \sum_i x_i s_i < 2(n+2)\mu$.
Let $x^*$ be any optimal solution of the primal. Then $c^T x \ge c^T
x^*$. We apply Claim \[weak duality\] first to the solution pair $x$ and $(y,s)$ and then to the pair $x^*$ and $(y,s)$ to obtain $$x^T s = c^T x - b^T y \ge c^T x^* - b^T y = (x^*)^T s.$$ Consider any $i \in {\{ 1,\ldots,n+2 \}}$ and assume $x_i <
x^*_i/(4(n+2))$. Since $x_i s_i \ge \mu/2$, we have $s_i \ge \mu/(2x_i) >
2 (n+2) \mu/x^*_i$, and hence $$(x^*)^T s \ge x_i^* s_i > 2(n+2) \mu \ge x^T s \ge (x^*)^T s,$$ a contradiction.
Let $(y^*,s^*)$ be any optimal solution of the dual. Then $b^T y^* \ge
b^T y$. We apply Claim \[weak duality\] first to the solution pair $x$ and $(y,s)$ and then to the pair $x$ and $(y^*,s^*)$ to obtain $$x^T s = c^T x - b^T y \ge c^T x - b^T y^* = x^T s^*.$$ Consider any $i
\in {\{ 1,\ldots,n+2 \}}$ and assume $s_i < s^*_i/(4(n+2))$. Since $x_i
s_i \ge \mu/2$, we have $x_i \ge \mu/(2s_i) > 2 (n+2) \mu/s^*_i$, and hence $$x^T s^* \ge x_i s^*_i > 2(n+2)\mu \ge x^T s \ge x^T s^*,$$ a contradiction.
The preceding Lemma implies strong duality, one of the cornerstones of linear programming theory.
\[strong:dual\] For each $i$, either $x^*_i=0$ in every optimal solution or $s^*_i=0$ in every optimal solution. Thus, $c^T x^* - b^T y^* = (x^*)^T s^* = 0$.
Let $x^*$ and $(y^*,s^*)$ be any pair of optimal solutions. Assume that there is an $i$ such that $x^*_i s^*_i > 0$. Let $(x,y,s,\mu)$ satisfy the invariants (I1) to (I3). Then $x_i \ge
x^*_i/(4(n+2))$ and $s_i \ge s^*_i(4(n+2))$ by Lemma \[conclude zero\]. Thus $2\mu > x_i s_i \ge x^*_i s^*_i /(16(n+2)^2)$. For $\mu < x^*_i
s^*_i /(32(n+2)^2)$, this is a contradiction.
We leave it to the reader to derive strong duality for the original primal and dual from this.
By the Strict Complementarity Theorem (see e.g. [@saigal pp 77-78] or [@ye pp 20-21]), there are optimal solutions $x^*$ and $(y^*,s^*)$ in which $x^*_i>0$ or $s^*_i>0$ for every $i$. A Quantitative version of strict complementarity is next stated in Fact \[strict complementarity\] (the proof is in Section 7).
\[strict complementarity\] Let $Q = R/(n+2)$. Then there are optimal solutions $x^*$ and $(y^*,s^*)$ such that for all $i$ either $x^*_i \ge Q$ and $s_i^* = 0$ or $s^*_i \ge Q$ and $x^*_i = 0$.
#### The Rounding Procedure:
Throughout this section $x^*$ and $(y^*,s^*)$ denote optimal solutions as in Fact \[strict complementarity\]. We run the iterative improvement algorithm until $$\label{muf}
\mu < \mu_f {\mathrel{:=}}R \cdot Q/(64 (n+2)^2 ((m+1) U)^{m+2}.$$ Let $({\bar{x}}, {\bar{y}}, {\bar{s}}, {\bar{\mu}})$ be the last iterate. Let $$B = {\{ \hspace{\minispacing} i \, | \,{\bar{s}}_i < Q/(4(n+2)) \hspace{\minispacing}\}} \quad\text{and}\quad
N = {\{ \hspace{\minispacing} i \, | \,{\bar{x}}_i < Q/(4(n+2)) \hspace{\minispacing}\}}.$$
\[B and N\] $B \cup N = {\{ 1,\ldots,n \}}$, $B \cap N = \emptyset$, $x^*_i = 0$ and ${\bar{x}}_i < 8 {\bar{\mu}}/ Q$ for every $i \in N$ and $s^*_i = 0$ and ${\bar{s}}_i < 8
{\bar{\mu}}/Q$ for every $i \in B$.
Since $x_i s_i < 2\mu$ and $\mu \le Q^2/(32 n^2)$, we have either ${\bar{x}}_i < Q/(4(n+2))$ or ${\bar{s}}_i < Q/(4(n+2))$. Thus $B \cup N =
{\{ 1,\ldots,n \}}$. Since ${\bar{x}}_i \ge x^*_i/(4(n+2))$ and ${\bar{x}}_i \ge
s^*_i/(4(n+2))$ and either $x^* \ge Q$ or $s^*_i \ge Q$, we have $B \cap
N = \emptyset$. Consider any $i \in B$. Then ${\bar{s}}_i < Q/(4(n+2))$ and hence $s^*_i < Q$. Thus $s^*_i = 0$. Similarly, $i \in N$ implies $x^*_i
= 0$. Finally, since ${\bar{x}}_i {\bar{s}}_i < 2 {\bar{\mu}}$, we either have ${\bar{x}}_i \ge
Q/(4(n+2))$ and ${\bar{s}}_i < 8 {\bar{\mu}}/Q$ or ${\bar{s}}_i \ge Q/(4(n+2))$ and ${\bar{x}}_i <
8 {\bar{\mu}}/Q$.
We split the variables $x$ into $x_B$ and $x_N$ and the matrix $A$ into $A_B$ and $A_N$. Then our primal constraint system (ignoring the non-negativity constraints) becomes $$A_B x_B + A_N x_N = b .$$ $(x^*_B, x^*_N)$ and $({\bar{x}}_B,{\bar{x}}_N)$ are solutions of this system, and $x^*_N = 0$ by Lemma \[B and N\]. Thus $A_B x^*_B = b$.
Let us concentrate on the equation $A_B x_B = b$. If it has a unique solution, call it ${\hat{x}}_B$, then ${\hat{x}}_B = x^*_B$. We can find ${\hat{x}}_B$ by Gaussian elimination and $({\hat{x}}_B,0)$ will be the optimal solution and we are done.
What can we do if $A_B x_B = b$ has an entire solution set? Then the rank of the matrix $A_B$ is smaller than the cardinality of $B$. Let $B_1
\subseteq B$ be such that the rank of the matrix $A_{B_1}$ is equal to the cardinality of $B_1$ and let $B_2 = B \setminus B_1$. We can find $B_1$ by Gaussian elimination. Then our system becomes $$A_{B_1} x_{B_1} + A_{B_2} x_{B_2} + A_N x_N = b.$$ For every choice of $x_{B_2}$ and $x_N$ this system has a unique solution[^6] for $x_{B_1}$. Let ${\hat{x}}_{B_1}$ be the solution of
$$A_{B_1} {\hat{x}}_{B_1} + A_{B_2} {\bar{x}}_{B_2} = b \quad (\text{$x_N$ is set to
zero and $x_{B_2}$ is set to ${\bar{x}}_{B_2}$}).$$ Subtracting this equation from $A_{B_1} {\bar{x}}_{B_1} + A_{B_2} {\bar{x}}_{B_2} +
A_N {\bar{x}}_N = b$ yields $$A_{B_1} ({\bar{x}}_{B_1} - {\hat{x}}_{B_1}) + A_N {\bar{x}}_N = 0.$$ The coordinates of ${\bar{x}}_N$ are bounded by $8 {\bar{\mu}}/Q$ and hence the coordinates of $A_N {\bar{x}}_N$ are bounded by $8 (n+2) U {\bar{\mu}}/ Q = R/(8 (n+2)
((m + 1 ) U)^{m + 1})$ in absolute value. By the remark after Lemma \[solutions to a linear system\] of Section 7, all coordinates of ${\bar{x}}_{B_1} - {\hat{x}}_{B_1}$ are bounded by $((m+1) U)^{m+1}$ times this number in absolute value, i.e., are bounded by $R/(8(n+2))$ in absolute value. Since ${\bar{x}}_i \ge {R}/(4(n+2))$ for every $i \in N$, we have ${\hat{x}}_{B_1} \ge 0$. Thus $\tilde{x} = ({\hat{x}}_{B_1}, {\bar{x}}_{B_2}, 0)$ is a feasible solution of \[artificial primal\]. Since $\tilde{x}^T s^* =
\sum_{i \in B} \tilde{x}_i s^*_i + \sum_{i \in N} \tilde{x}_i s^*_i = 0 +
0 = 0$, $\tilde{x}$ is an optimal solution to \[artificial primal\].
Complexity
==========
Let us assume that the initial value of $\mu$ is $\mu_0$ and that we want to decrease $\mu$ to $\mu_f$. Since every iteration decreases $\mu$ by the factor $(1 - \delta)$, we have $\mu = (1-\delta)^r\mu_0$ after $r$ iterations. The smallest $r$ such that $(1 - \delta)^r \le \mu_f$ is given by $$\ln \frac{\mu_0}{\mu_f} = -r \ln(1-\delta)\approx -r
(-\delta),$$ or equivalently, $$r=O\left(\frac{1}{\delta}\log \frac{\mu_0}{\mu_f}\right)=
O\left(\sqrt{n}\log \frac{\mu_0}{\mu_f}\right).$$ In [(\[definition of mu0\])]{}, we defined $$\mu_0^2 = 4 \left(M^2+\sum c_i^2\right)\leq
4 \left( \frac{16n^2 U^2 }{{R}^2} + n U^2 \right) \le 68 \frac{n^2
U^2}{{R}^2}.$$ In [(\[muf\])]{}, we defined $\mu_f$. Thus, the number of iterations will be $$\begin{aligned}
r &=O\left(\sqrt{n}\log \frac{\mu_0}{\mu_f}\right)=
O\left(\sqrt{n}\log \frac{n^2U^2/{R}^2}{RQ/(64 (n+2)^2 ((m+1)
U)^{m+2})}\right)\\
&=O\left(\sqrt{n}(\log n+ m\log (mU) + \log \frac{1}{{R}} \right) \\
&= O\left(\sqrt{n}(\log n+m \left(\log (mU)\right) \right), \end{aligned}$$ as $\log \frac{1}{{R}} = O(\log n + m (\log (mU)))$.
The Proofs of Facts 1 to 3 {#integer}
==========================
\[bounds\]
In the previous sections, we used upper bounds on the components of an optimal solution and lower bounds on the nonzero components of an optimal solution. In this section, we derive these bounds. In this section, we assumes more knowledge of linear algebra, namely, determinants and Cramer’s rule, and some knowledge of geometry. Unless stated otherwise, we assume that all entries of $A$ and $b$ are integers bounded by $U$ in absolute value.
The determinant of a $k \times k$ matrix $G$ is a sum of $k!$ terms, namely, $$\det G =\sum_\pi {{\mathit sign}}(\pi) \cdot g_{1\pi(1)}g_{2\pi(2)}{\ldots}
g_{k\pi(k)}.$$ The summation is over all permutations $\pi$ of $k$ elements, ${{\mathit sign}}(\pi)
\in {\{ -1,1 \}}$, and the product corresponding to a permutation $\pi$ selects the $\pi(i)$-th element in row $i$ for each $i$. Each product is at most $U^k$. As there are $k!$ summands, we have ${\left\vert \hspace{\minispacing} \det G \hspace{\minispacing} \right\vert} \le k!
U^k \le (kU)^k$; see [@BT pp 373-374], [@karloff p75] or [@saigal pp 43-44].
Cramer’s rule states that the solution of the equation $Gz = g$ (for a $k\times k$ non-singular matrix $G$) is $z_i=(\det G_i)/\det G$, where $G_i$ is obtained by replacing the $i$th column of $G$ by $g$.
\[solutions to a linear system\] Let $G z = g$ be a linear system in $k$ variables with a unique solution. Let $z^*$ be the solution of the system. If all entries of $G$ and $g$ are integers bounded by $U$ in absolute value then ${\left\vert \hspace{\minispacing} z^*_i \hspace{\minispacing} \right\vert} \le (kU)^k$ for all $i$ and $z^*_i \not= 0$ implies ${\left\vert \hspace{\minispacing} z^*_i \hspace{\minispacing} \right\vert} \ge
1/(kU)^k$.
Since the system has a unique solution there is a subsystem $G' z = g'$ consisting of $k$ equations such that $G'$ is non-singular and $G' z^* = g'$. Then $z^*_i = (\det G'_i)/\det G'$, where $G'_i$ is obtained from $G'$ by replacing the $i$th column of $G$ by $g'$. Since all entries of $G$ and $g$ are integral, $\det G'$ is at least one in absolute value, $\det G'_i$ is at least one in absolute value if nonzero, and $\det G'_i \le (kU)^k$. The bounds follow.
If the entries of the right-hand side $g$ are bounded by $U'$ instead of $U$, the upper bound becomes $k^k U^{k-1} U'$.
\[F and O\] Assume that [(\[primal problem\])]{} is feasible. Let $x$ be a feasible solution with the maximum number of zero coordinates (equivalently the minimum number of nonzero coordinates).[^7] Let $B$ be the set of indices for which $x_i \not= 0$, and let $A_B$ be the submatrix of $A$ formed by the columns indexed by $B$. Then $A_B z = b$ has a unique solution, where the dimension of $z$ is equal to the number of columns of $A_B$.
If, in addition [(\[primal problem\])]{} is bounded, the same claim holds for an optimal solution with a maximum number of zero coordinates.
Let $x_B$ be the restriction of $x$ to the indices in $B$. Then $A_B x_B = b$. Assume there is a second solution $x'_B$ of $A_B z =
b$ with $x'_B \not= x_B$. Then all points $z(\lambda) = x_B +
\lambda(x'_B - x_B)$, $\lambda \in {\mathbb{R}}$, satisfy $A_B z = b$. These points form a line. Consider the intersection $z^*$ closest to $x_B$ of this line with one of the coordinate planes $z_i = 0$; if there are several with the same distance choose one of them. Then $z^* \ge 0$ because we consider an intersection closest to $x_B$ and $z^*_i = 0$ for at least one $i \in B$. Thus $z^*$ is a feasible solution to [(\[primal problem\])]{} with one more zero coordinate, a contradiction to the definition of $x$.
If [(\[primal problem\])]{} is bounded, there is an optimal solution. Let $x$ be an optimum solution with a maximum number of zero coordinates. Define $x_B$, $x_B'$, and $z(\lambda)$ as above. Since $x_B > 0$, the $z(\lambda)$ is feasible for small enough ${\left\vert \hspace{\minispacing} \lambda \hspace{\minispacing} \right\vert}$. Also $c_B^T z(\lambda) = c_B^T x_B
+ \lambda (c_B^T x'_B - c_B^T x_B)$; here $c_B$ is the restriction of $c$ to the indices in $B$. Since $\lambda$ may be positive or negative, we must have $c_B^T x'_B = c_B^T x_B$ and hence $z(\lambda)$ is feasible and optimal as long as $z(\lambda) \ge 0$. The proof is now completed as in the preceding paragraph.
We can now give the proof of Facts \[bounded by W\], \[nonzero coordinates\], and \[strict complementarity\].
(Fact \[bounded by W\]) Consider a feasible (optimal) solution $x$ of [(\[primal problem\])]{} with a maximum number of zero coordinates. Then $x$ is of the form $x = (x_B, x_N)$ with $x_N = 0$ and $x_B$ being the unique solution to the system $A_B x_B = b$. Thus the coordinates of $x_B$ are bounded by $(m U)^m$.
(Fact \[nonzero coordinates\]) Let $x^*$ be an optimal vertex of the artificial primal [(\[artificial primal\])]{}. How small can a nonzero coordinate of $x^*$ be? The constraint system is $$\begin{array}{r c c c c c l } Ax & & & +
& (\frac{1}{W} b - Ae) x_{n+2} & = & \frac{1}{W} b\\ e^T x & + & x_{n+1}
& + & x_{n+2} & = & (n+2).
\end{array}$$ Let $B$ be the index set of the nonzero coordinates of $x^*$. Then $x^*_B$ is the solution to a subsystem formed by ${\left\vert \hspace{\minispacing} B \hspace{\minispacing} \right\vert}$ columns of the above and this subsystem has a unique solution. For $i \in B$, $x^*_i =
\det G_i/\det G$, where $G$ is a nonsingular square matrix and $G_i$ is obtained from $G$ by replacing the $i$th column by the corresponding entries of the right hand side. In the system above, the entries in the column corresponding to $x_{n+2}$ are bounded by $(n+1) U$, and all other entries are bounded by $U$. Since any product in the determinant formula for $G$ can contain only one value of the column for $x_{n+2}$, we have ${\left\vert \hspace{\minispacing} \det G \hspace{\minispacing} \right\vert} \le (m+1)! (n+1)U^{m+1}$. Consider next $\det G_i$. We need to lower bound ${\left\vert \hspace{\minispacing} \det G_i \hspace{\minispacing} \right\vert}$. The matrix $G_i$ may contain two columns with fractional values. If we multiply these columns with $W$, we obtain an integer matrix. Thus, ${\left\vert \hspace{\minispacing} \det G_i \hspace{\minispacing} \right\vert} \ge 1/W^2$ if nonzero. Thus $$\label{definition of sep}
x_i^* \ge \frac{1}{W^2} \cdot \frac{1}{2n \left((m+1) U\right)^{m+1}} \ge
\frac{1}{2 n \left((m+1) U\right)^{3(m+1)}}.$$
(Fact \[strict complementarity\]) We prove the fact for the auxiliary primal. Let ${{\cal O}}$ be a smallest set of optimal vertices with the property that if for some $i$ there is an optimal solution with $x^*_i > 0$, then ${{\cal O}}$ contains an optimal vertex with this property. Then ${\left\vert \hspace{\minispacing} {{\cal O}}\hspace{\minispacing} \right\vert} \le n + 2$. Let $x^{**} = \frac{1}{{\left\vert \hspace{\minispacing} {{\cal O}}\hspace{\minispacing} \right\vert}} \sum_{x^*
\in {{\cal O}}} x^*$ be the center of gravity of the vertices in ${{\cal O}}$. Then $x^{**}_i \ge x^*_i/(n+2)$ for every $x^* \in {{\cal O}}$. Thus $Q = R/(n+2)$ works.
#### Beyond the Integral Case
If the entries of $A$ and $b$ are rational numbers, we write the entries in each column (or row) with a common denominator. Pulling them out brings us back to the integral case. For example, $$\left| \begin{array}{cc} {2}/{3} & {4}/{5} \\
{1}/{3} & {6}/{5} \end{array} \right| = \frac{1}{15} \left| \begin{array}{cc} {2} & {4}\\
{1} & {6} \end{array}\right|.$$ Thus, if the determinant is nonzero, it is at least $1/15$.
Acknowledgments {#acknowledgments .unnumbered}
---------------
The first author thanks Andreas Karrenbauer and Ruben Becker for intensive discussions and Andreas Karrenbauer for teaching from an earlier draft of this note. The work of the second author was inspired by an informal lecture given by Nisheeth Vishnoi at IIT Kanpur. The second author also thanks the students of CS602 (2014-15 and 2015-2016 batches) for their helpful comments and questions. Thanks also to Romesh Saigal for very prompt replies to queries.
The authors also thank Andreas Karrenbauer for providing the proof of Lemma \[alternative proof\].
[10]{}
D. Bertsimas and J.N. Tsitsiklis, Introduction to linear optimization, Athena Scientific, 1997.
H.J. Greenberg, The use of the optimal partition in a linear programming solution for postoptimal analysis, Operations Research Letters, 15 (1994), 179-185.
H. Karloff, Linear Programming, Birkhauser, 1991.
G. Lesaja, Introducing Interior-Point Methods for Introductory Operations Research Courses and/or Linear Programming Courses, The Open Operational Research Journal, 2009, 3, 1-12.
S. Mizuno, M. Kojima, M.J. Todd, Infeasible-interior-point primal-dual potential-reduction algorithms for linear programming, SIAM J. Optim. 5, 52–67 (1995)
C. Roos, T. Terlaky and J-P. Vial, Interior Point Methods for Linear Optimization, 2nd Ed, 2006, Springer.
R. Saigal, Linear Programming, A Modern Integrated Analysis, Kluwer, 1995.
T. Terlaky, An easy way to teach interior-point methods, European J of Operational Research, 130 (2001), 1-19
R.J. Vanderbei, Linear Programming: Foundations and Extension, 1st Ed: Kluwer Academic Publishers, 1997 (3rd Ed: Springer).
N.K. Vishnoi, Convex Optimization Notes, <http://theory.epfl.ch/vishnoi/Nisheeth-VishnoiFall2014-ConvexOptimization.pdf>
Y. Ye, Interior Point Algorithms, Wiley, 1997
Y. Zhang, On the convergence of a class of infeasible interior-point methods for the horizontal linear complementarity problem, SIAM J. Optim. 5, 208–227 (1994).
Appendix: Result from Algebra {#appendix-result-from-algebra .unnumbered}
-----------------------------
Assume that $A$ is $m\times n$ matrix and the rank of $A$ is $m$, with $m<n$. Then, all $m$ rows of $A$ are linearly independent. Or, $\alpha_1A_1+\alpha_2A_2+{\ldots} +\alpha_mA_m=0$ ($0$ here being a row vector of size $n$) has only one solution $\alpha_i=0$. Thus, if $x$ is any $m \times 1$ matrix (a column vector of size $m$), then $x^TA=0$ implies $x=0$. Note that $(x^T A)^T = A^T x$. Thus, $A^T x = 0$ implies $x = 0$.
As $A$ is $m\times n$ matrix, $A^T$ will be $n\times m$ matrix. The product $AA^T$ will be an $m\times m$ square matrix.
Consider the equation $(AA^T)x=0$. Pre-multiplying by $x^T$ we get $x^TAA^Tx =0$ or $(A^T x)^T (A^T x) =0$. Now, $(A^T x)^T (A^T x)$ is the squared length of the vector $A^T x$. If a vector has length zero, all its coordinates must be zero. Thus, $A^T x = 0$, and hence, $x = 0$ by the preceding paragraph.
Thus, the matrix $AA^T$ has rank $m$ and is invertible.
Also observe that if $X$ is a diagonal matrix (with all diagonal entries non-zero) and if $A$ has full row-rank, then $AX$ will also have full row-rank. Basically, if the entries of $X$ are $x_1,x_2,{\ldots} ,x_m$ then the matrix $AX$ will have rows as $x_1A_1,x_2A_2,{\ldots} ,x_mA_m$ (i.e., $i$th row of $A$ gets scaled by $x_i$). If rows of $AX$ are not independent, then there are $\beta$s (not all zero) such that $\beta_1x_1A_1+\beta_2x_2A_2+{\ldots} +\beta_mx_mA_m=0$, or there are $\alpha$s (not all zero) such that $\alpha_1A_1+\alpha_2A_2+{\ldots}
+\alpha_mA_m=0$ with $\alpha_i=\beta_ix_i$.
[^1]: E-mail: [email protected]
[^2]: E-mail: [email protected]
[^3]: Indeed, we can use Gaussian elimination to remove superfluous constraints and to make the rows of $A$ independent. Assume first that $A$ contains a row $i$ in which all entries are equal to zero. If $b_i$ is also zero, we simply delete the row. If $b_i$ is nonzero, the system of equations has no solution, and we declare the problem infeasible and stop. Now, every row of $A$ contains a nonzero entry, in particular, the first row. We may assume that $a_{11}$ is nonzero. Otherwise, we interchange two columns. We multiply the $i$th equation by $-\frac{a_{11}}{a_{i1}}$ and subtract the first equation. In this way, the first entry of all equations but the first becomes zero. If any row of $A$ becomes equal to the all zero vector, we either delete the equation or declare the problem infeasible. We now proceed in the same way with the second equation. We first make sure that $a_{22}$ is nonzero by interchanging columns if necessary. Then we multiply the $i$th equation (for $i > 2$) by $-\frac{a_{22}}{a_{21}}$ and subtract the second equation. And so on. In the end, all remaining equations will be linearly independent. Equivalently, the resulting matrix will have full row-rank.
We now have $m$ constraints in $n$ variables with $n \ge m$. If $n = m$, the system $Ax =b$ has a unique solution (recalling that $A$ has full row-rank and is hence invertible). We check whether this solution is non-negative. If so, we have solved the problem. Otherwise, we declare the problem infeasible. So, we may from now on assume $n > m$ (more variables than constraints).
[^4]: Indeed, $$n \sum_i z_i^2 - \left(\sum_i z_i\right)^2 = n
\sum_i z_i^2 - \sum_i z_i^2 - 2\sum_{i < j} z_i z_j = (n - 1) \sum_i
z_i^2 - 2\sum_{i < j} z_i z_j = \sum_{i < j} (z_i - z_j)^2 \ge 0.$$
[^5]: If $\alpha=(\alpha_1,{\ldots},\alpha_n)$ then $\left(\sum
|\alpha_i|\right)^2=\sum |\alpha_i|^2+2\sum_{i<j}|\alpha_i||\alpha_i|\geq
\sum |\alpha_i|^2=\sum \alpha_i^2$
[^6]: Let $m' \le m$ be the rank of $A_B$. By row operations and permutation of columns, we can transform the system $A_B x_B + A_N
x_N = b$ into $$\begin{array}{ccccccc} I x_{B_1} &+ & A'_{B_2} x_{B_2} & + & A'_N x_N
& = & b' \\ 0 &+& 0 &+& A''_N x_N & = & b'',
\end{array}$$ where $I$ is a $m' \times m'$ identity matrix, $A'_{B_2}$, $A'_N$, and $b'$ have $m'$ rows, and $A''_N$ and $b''$ have $m - m'$ rows. Since $(x^*_B, x^*_N)$ is a solution to this system and $x^*_N = 0$, we have $b'' = 0$. Since $({\bar{x}}_B,{\bar{x}}_N)$ is a solution to this system, we have further $A''_N {\bar{x}}_N = 0$. Thus for every choice of $x_{B_2}$ and $x_N$ this system has a unique solution for $x_{B_1}$.
[^7]: Consider minimize 0 subject to $x_1 + x_2 = 1$, $x_1 \ge 0$ and $x_2 \ge 0$. The feasible solutions $(0,1)$ and $(1,0)$ have one nonzero coordinate. The feasible solutions $(x_1,x_2)$ with $x_1
> 0$ and $x_2 > 0$ and $x_1 + x_2 = 1$ have two nonzero coordinates.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We investigate what powers hyperluminous infrared galaxies (HyLIRGs; $L_{\rm IR, 8-1000\mu m}>$10$^{13}$L$_{\odot}$) at z$\sim$1-2, by examining the behaviour of the infrared AGN luminosity function in relation to the infrared galaxy luminosity function. The former corresponds to emission from AGN-heated dust only, whereas the latter includes emission from dust heated by stars and AGN. Our results show that the two luminosity functions are substantially different below 10$^{13}$L$_{\odot}$ but converge in the HyLIRG regime. We find that the fraction of AGN dominated sources increases with total infrared luminosity and at $L_{\rm IR}>10^{13.5}\, \rm L_{\odot}$ AGN can account for the entire infrared emission. We conclude that the bright end of the $1<z<2$ infrared galaxy luminosity function is shaped by AGN rather than star-forming galaxies.'
bibliography:
- 'references.bib'
date: 'Accepted Received; in original form'
nocite:
- '[@RR91]'
- '[@Soifer94]'
- '[@SM96]'
- '[@Sanders03]'
- '[@RR00]'
- '[@RRW10]'
- '[@Sanders88]'
- '[@Ivison98]'
- '[@Hines99]'
- '[@Farrah02a]'
- '[@Farrah02b]'
- '[@RCP07]'
- '[@Ruiz10]'
- '[@Ruiz13]'
- '[@Hines95]'
- '[@GDF96]'
- '[@RR00]'
- '[@RRW10]'
- '[@Verma02]'
- '[@Franceschini00]'
- '[@Wilman03]'
- '[@RCP07]'
- '[@Stern14]'
- '[@Hines95]'
- '[@Granato96]'
- '[@YS98]'
- '[@Efstathiou13]'
- '[@Marconi04]'
- '[@MD14]'
- '[@Bethermin11]'
- '[@Toba15]'
- '[@Gruppioni13]'
- '[@Aird15]'
- '[@Richards06]'
- '[@Ross13]'
- '[@Maiolino07]'
- '[@Symeonidis16]'
- '[@DP90]'
- '[@BSP88]'
- '[@Page97]'
- '[@LGT98]'
- '[@York00]'
- '[@Boyle00]'
- '[@Ross13]'
- '[@RRW10]'
- '[@Goto05]'
- '[@Kartaltepe10]'
- '[@Yuan10]'
- '[@Goto11a]'
- '[@RR00]'
- '[@Farrah02a]'
- '[@Farrah02b]'
- '[@Ruiz10]'
- '[@Ruiz13]'
- '[@Calderon16]'
- '[@TN16]'
- '[@RCP07]'
- '[@Ruiz10]'
- '[@Nardini]'
- '[@Kennicutt98]'
- '[@RRW10]'
- '[@RR16]'
- '[@Banerji17]'
- '[@Stacey18]'
- '[@RR00]'
- '[@Farrah02b]'
- '[@Dai12]'
- '[@Leipski13]'
- '[@Leipski14]'
- '[@Fan16]'
- '[@Duras17]'
- '[@Symeonidis17]'
- '[@YS98]'
- '[@NI07]'
- '[@MQT05]'
- '[@Baugh05]'
- '[@Swinbank08]'
- '[@Dave10]'
- '[@Secrest15]'
- '[@TN16]'
- '[@Lawrence91]'
- '[@HB07]'
- '[@LE10]'
- '[@Mateos17]'
- '[@Akylas12]'
- '[@Buchner15]'
- '[@Akylas16]'
- '[@BU12]'
- '[@Vignali14]'
- '[@Mullaney11]'
- '[@Netzer07]'
title: 'What powers Hyperluminous infrared galaxies at $z$$\sim$1–2?'
---
\[firstpage\]
galaxies: general galaxies: high-redshift infrared: galaxies galaxies:active
Introduction {#sec:introduction}
============
Hyperluminous infrared galaxies (HyLIRGs) are defined as galaxies with $8-1000$ $\mu$m infrared luminosities (hereafter $L_{\rm IR}$) exceeding $10^{13}$L$_{\odot}$. They were first discovered in the *IRAS* all-sky survey (Rowan-Robinson et al. 1991; Soifer et al. 1994; Sanders $\&$ Mirabel 1996) and the origin of their extreme luminosities immediately became a topic of much debate. HyLIRGs are rare; none have been found in the local Universe ($z<0.1$; e.g. Sanders et al. 2003) and it is estimated that there are about 200 *IRAS*-detected hyperluminous galaxies over the whole sky (Rowan-Robinson 2000; Rowan-Robinson $\&$ Wang 2010).
Follow-up studies of HyLIRGs found that their morphologies show signs of interactions or QSO-like appearance (e.g. Sanders et al. 1988; Ivison et al. 1998; Hines et al. 1999; Farrah et al. 2002a) and virtually all host AGN (e.g. Farrah et al. 2002b; Ruiz, Carrera $\&$ Panessa 2007; Ruiz et al. 2010; 2013), many of which are found to be buried (Hines et al. 1995; Granato et al. 1996). The observation that both AGN and starburst components are often necessary to fit their broadband spectral energy distributions (SEDs) has generated differing views on what the dominant energy source is in HyLIRGs: on one hand it has been suggested that emission longwards of 50$\mu$m is primarily attributed to star-formation, with star-formation rates that were predicted to be in excess of 1000M$_{\odot}$/yr (e.g. Rowan-Robinson et al. 2000; Rowan-Robinson $\&$ Wang 2010; Verma et al. 2002), whereas on the other hand, the detection of buried, even Compton-thick, AGN in many HyLIRGs (e.g. Franceschini et al. 2000; Wilman et al. 2003; Ruiz, Carrera $\&$ Panessa 2007; Stern et al. 2014) has given rise to the idea that all HyLIRGs could potentially host AGN and be powered by AGN (e.g. Hines et al. 1995; Granato et al. 1996; Yun $\&$ Scoville 1998; Efstathiou et al. 2013).
In this letter, we investigate what is the dominant power source in HyLIRGs, using the infrared galaxy luminosity function and the X-ray AGN luminosity function at $z\sim$1–2. This redshift range corresponds to the peak of the cosmic star-formation history and supermassive black hole accretion rate history (e.g. Marconi et al. 2004; Madau $\&$ Dickinson 2014). Moreover, the contribution to the cosmic energy budget from HyLIRGs is seen to undergo rapid growth from $z=0$ to $z=2$, amounting to at least a few percent of the total comoving infrared luminosity density at $z\sim2$ (e.g. B[é]{}thermin et al. 2011; Toba et al. 2015).
This letter is laid out as follows: in section \[sec:method\] we describe our method and our results are presented in section \[sec:results\]. A discussion of our findings and conclusions are presented in sections \[sec:discussion\] and \[sec:conclusions\] respectively. Throughout, we adopt a concordance cosmology of H$_0$=70kms$^{-1}$Mpc$^{-1}$, $\Omega_{\rm M}$=1-$\Omega_{\rm \Lambda}$=0.3.
Method {#sec:method}
======
Our investigation rests on the comparison of the infrared galaxy luminosity function and the AGN luminosity function. For the infrared galaxy luminosity function ($\phi_{\rm IR}$) we use the one presented in Gruppioni et al. (2013; hereafter G13). $\phi_{\rm IR}$ is a function of $L_{\rm IR}$, which includes the total dust-reprocessed emission from stars and AGN. The uncertainties on $\phi_{\rm IR}$ from G13 are a combination of Poisson errors and photometric redshift uncertainties derived through Monte Carlo simulations.
For the AGN luminosity function we use the absorption-corrected hard X-ray (2-10keV) AGN luminosity function from Aird et al. (2015; hereafter A15). This is significantly more complete than the optical AGN luminosity function (e.g. Richards et al. 2006; Ross et al. 2013) since it includes absorbed AGN. The X-ray AGN luminosity function is translated to an infrared AGN luminosity function ($\phi_{\rm IR, AGN}$) as follows: first, hard X-ray luminosity is converted to optical luminosity at 5100$\AA$ ($\nu L_{\nu, 5100}$), adopting the relation from Maiolino et al. (2007). Subsequently, to convert from $\nu L_{\nu, 5100}$ to infrared luminosity in the 8–1000$\mu$m range ($L_{\rm IR, AGN}$) we use the intrinsic AGN SED of Symeonidis et al. (2016; hereafter S16). The errors on $\phi_{\rm IR, AGN}$ are Poisson, as are the original errors on the X-ray AGN luminosity function from A15. $L_{\rm IR, AGN}$ is the intrinsic IR luminosity of the AGN, i.e. it does not include the contribution of dust heated by starlight. $\phi_{\rm IR}$ and $\phi_{\rm IR, AGN}$ are monotonically decreasing functions of $L_{\rm IR}$ and $L_{\rm IR, AGN}$ respectively, over the luminosity range considered in this letter and $\phi_{\rm IR} (L_{\rm IR}) \geq \phi_{\rm IR, AGN}(L_{\rm IR, AGN})$.
Results {#sec:results}
=======
The data and functional forms of $\phi_{\rm IR, AGN}$ and $\phi_{\rm IR}$ are shown in Fig. \[fig:LFs\] in two redshift bins within the $1<z<2$ interval, corresponding to the peak in comoving infrared luminosity density. $z>1$ is where HyLIRGs first appear in the data of both luminosity functions, and we chose $z\sim$2 as a conservative upper limit relating to the reliability of photometric redshifts, because the bright end of the luminosity function is very sensitive to redshift mis-identification. $\phi_{\rm IR}$ is shown in the $1.2<z<1.7$, $1.7<z<2$ redshift bins; G13 fit the data with the function from Saunders et al. (1990), behaving as a power-law for $L<L_{\star}$ and as a Gaussian for $L>L_{\star}$ (see G13 for more details). $\phi_{\rm IR, AGN}$ is shown in the $1.2<z<1.5$, $1.5<z<2$ redshift bins and the data are fit with a double power-law model, whose parameters are themselves functions of redshift evaluated at the centre of the relevant bin (see A15 for more details); for the redshift bins of interest here, these are $z=1.35$ and $z=1.75$. As the G13 and A15 redshift bins are not precisely the same, we also evaluated the A15 LF model at the the centre of the G13 bins, namely $z=1.45$ and $z=1.85$. We found the mean shift in the A15 LF to be only about 0.1dex at the bright end, so we use the original redshift bins for $\phi_{\rm IR, AGN}$ in Fig. \[fig:LFs\], as the AGN luminosity densities were calculated in those bins in A15.
Fig. \[fig:LFs\] shows that for $L_{IR}<10^{12}$ L$_{\odot}$, $\phi_{\rm IR}$ exceeds $\phi_{\rm IR, AGN}$ by more than 1dex. However, this difference decreases with increasing luminosity because $\phi_{\rm IR}$ declines faster than $\phi_{\rm IR, AGN}$ and at $L_{\rm IR} > 10^{13}$L$_{\odot}$ the two luminosity functions converge, with the measurements of $\phi_{\rm IR}$ and $\phi_{\rm IR, AGN}$ consistent at 1 $\sigma$. The models of $\phi_{\rm IR}$ and $\phi_{\rm IR, AGN}$ cross at log ($L_{\rm IR}$/L$_{\odot}$) $\sim$13.6 in both redshift ranges.
The ratio of $\phi_{\rm IR, AGN}$ to $\phi_{\rm IR}$ provides a simple estimate of the fraction of AGN-dominated sources as a function of $L_{\rm IR}$. This ratio, as derived from the parametric model luminosity functions, is shown in Fig. \[fig:AGN\]. In contrast to Fig. \[fig:LFs\], here we use the functional form of the A15 LF at the the centre of the G13 redshift bins. Fig. \[fig:AGN\] shows that the contribution of the AGN to the total infrared luminosity and hence the fraction of AGN-dominated sources is small ($<$10 per cent) until the HyLIRG regime, where it undergoes a rapid increase and at log($L_{\rm IR}/\rm L_{\odot})>13.5$ the galaxy population becomes AGN dominated.
Discussion {#sec:discussion}
==========
Implications for the infrared luminosity function
-------------------------------------------------
In this work, we have used the AGN and galaxy infrared luminosity functions to explore what powers HyLIRGs at the peak of the cosmic star-formation history at $1<z<2$. We believe that this is the first time that the AGN luminosity function has been expressed as a function of *intrinsic* infrared AGN power, enabling an estimate of the fraction of AGN dominated sources per luminosity bin. Our results show that the galaxy and AGN luminosity functions are initially offset in the $10^{11} - 10^{13}$L$_{\odot}$ luminosity range, but converge in the HyLIRG regime with the model luminosity functions crossing at about $10^{13.6}$L$_{\odot}$. By definition (see Section \[sec:method\]) $\phi_{\rm IR}$ cannot be less than $\phi_{\rm IR, AGN}$, implying that beyond the crossover ($>10^{13.6}$L$_{\odot}$), the shape of one or both of the luminosity functions must be different to that predicted by the parametric models.
The relatively small areas covered by the deep infrared surveys with suitable redshift information are a hindrance to adequately sampling the bright tail of the infrared luminosity function; Fig. \[fig:LFs\] shows that the only HyLIRG data point in $\phi_{\rm IR}$ from G13 is at log$L_{\rm IR}/ \rm L_{\odot} <$13.5. On the other hand, the shape of the AGN luminosity function is well-established as a double power-law at all wavelengths (e.g. Dunlop $\&$ Peacock 1990, Boyle, Shanks & Peterson 1988, Page et al. 1997) and the availability of large surveys such as the 2dF (Lewis, Glazebrook $\&$ Taylor 1998) and SDSS (York et al. 2000) have enabled precise measurements of the bright tail of the AGN luminosity function, e.g. Boyle et al. (2000), Ross et al. (2013). The most natural resolution of the apparent crossing of $\phi_{\rm IR}$ and $\phi_{\rm IR, AGN}$, would therefore be for $\phi_{\rm IR}$ to flatten above the highest luminosities probed by G13 to match the shape of $\phi_{\rm IR, AGN}$. Thus the IR emission from log($L_{\rm IR}$/L$_{\odot}$)$>$13.5 galaxies is dominated by emission from AGN, and the luminosity function of galaxies in this regime is shaped by AGN rather than star-forming galaxies.
We are able to test the prediction that $\phi_{\rm IR}$ must flatten at $L_{\rm IR} > 10^{13.5}$ with the $1<z<2$ *IRAS* HyLIRG luminosity function presented in Rowan-Robinson $\&$ Wang (2010; hereafter RRW10). This is shown in Fig \[fig:LFs\_RRW10\], where we combine the two redshift bins for $\phi_{\rm IR}$ and $\phi_{\rm IR, AGN}$ by averaging the luminosity functions, to give us $\phi_{\rm IR}$ and $\phi_{\rm IR, AGN}$ at $1.2<z<2$. To compare with the RRW10 results, we convert the abscissa of the RRW10 luminosity function from 60$\mu$m to total infrared luminosity using the S16 intrinsic AGN SED, a reasonable assumption given the SED shapes of HyLIRGs presented in RRW10, and that the AGN incidence at these luminosities is 100 per cent. Fig \[fig:LFs\_RRW10\] shows that the space density of $1<z<2$ *IRAS* HyLIRGs does indeed appear to be consistent with the space density of the extrapolated $\phi_{\rm IR, AGN}$. RRW10 themselves noted the apparent flattening of the luminosity function implied by their data at the highest luminosities, but were unable to identify the mechanism responsible. Here it is explained naturally by the transition to an entirely AGN-dominated population.
On the nature of HyLIRGs
------------------------
Our study of the HyLIRG luminosity function indicates that in the HyLIRG regime, the dominant power source in the galaxy population transitions from star-formation to AGN: figs. \[fig:LFs\] and \[fig:AGN\] show that the total infrared emission of the most luminous HyLIRGs can be accounted for entirely by AGN. An implication of this result, and a condition necessary for it to hold, is that at $L_{\rm IR}>10^{13.5} \rm
L_{\odot}$ almost all HyLIRGs must host AGN. Observations of HyLIRGs readily satisfy this condition: the incidence of AGN in infrared galaxies rises as a function of infrared luminosity (e.g. Goto et al. 2005; Kartaltepe et al. 2010; Yuan et al. 2010; Goto et al. 2011a), plateauing at 100 per cent in the HyLIRG regime. Indeed, as far as can be ascertained, all known HyLIRGs host AGN (e.g. Rowan-Robinson 2000; Farrah et al. 2002a; 2002b; Ruiz et al. 2010; 2013; Calder[ó]{}n et al. 2016; Toba $\&$ Nagao 2016).
Although the number of HyLIRGs studied in detail so far is small, it seems that the transition from star-formation to AGN power within the HyLIRG regime is consistent with the variation of HyLIRG spectral properties with luminosity. The most luminous HyLIRGs show, on average, more prominent AGN signatures: optical and X-ray spectra characteristic of AGN and SEDs that are AGN-dominated from the optical to (at least) the mid-IR (e.g. Ruiz, Carrera $\&$ Panessa 2007; Ruiz et al. 2010).
We note that although the transition from SF to AGN dominated sources occurs in the HyLIRG regime at $z=1-2$, this could be different at other redshifts as the knee of the luminosity function evolves with redshift. It is plausible that at lower redshift the transition takes place at lower luminosities — for example Nardini et al. (2010) suggest that, for $z<0.35$ ULIRGs with log$[L_{\rm IR}/\rm L_{\odot}]>12.7$, the fraction of AGN-dominated sources is about 40 per cent.
Implications for galaxy star formation rates {#sec:sfrs}
--------------------------------------------
We have found that AGN can account for a significant part of the infrared emission in $13< \rm log (L_{\rm IR}/ L_{\odot})<13.5$ sources and the entire infrared emission at $L_{\rm IR}>10^{13.5}\, \rm L_{\odot}$. Assuming that the total infrared luminosity is linked to star-formation, $L_{\rm IR}=10^{13}\, \rm L_{\odot}$ and $L_{\rm IR}=10^{13.5}\, \rm L_{\odot}$ would translate to an SFR of 1700M$_{\odot}$/yr and 5400M$_{\odot}$/yr respectively, using the Kennicutt (1998) SFR-$L_{\rm IR}$ relation. Our results show that, at least at $1<z<2$, SFRs claimed to be greater than 5400M$_{\odot}$/yr (e.g. Rowan-Robinson $\&$ Wang 2010; Rowan-Robinson et al. 2016; Banerji et al. 2017; Stacey et al. 2018) should be scrutinised carefully because such objects are likely to be entirely powered by AGN. The dominance of AGN at those luminosities does not imply that SFRs$>$5400M$_{\odot}$/yr are implausible, rather than the existence of such SFRs cannot be established from the mere existence of galaxies in this luminosity range.
Even sources with lower reported SFRs but still in the HyLIRG regime are likely to require a large correction due to the AGN contribution at infrared wavelengths. Although most authors use multi-component SED fitting to estimate HyLIRG star-formation rates (e.g. Rowan-Robinson et al. 2000; Farrah et al. 2002b; Dai et al. 2012; Leipski et al. 2013; 2014; Fan et al. 2016; Duras et al. 2017), the far-IR has always been attributed to dust heated by star-formation. In Symeonidis et al. (2016) and Symeonidis (2017) however, we showed that for galaxies hosting powerful AGN, the far-IR emission is likely to originate from kpc-scale dust heated by the AGN.
We believe that it is imperative to consider diverse star-formation rate indicators for HyLIRGs, not associated with broadband photometry. Some studies which have examined gas masses, mid-IR spectral features, or extended soft X-ray emission in such sources (e.g. Yun $\&$ Scoville 1998; Nandra $\&$ Iwasawa 2007), have concluded that the implied SFRs are not sufficient to power the infrared emission. Our proposal that the most luminous infrared galaxies (at least at $1<z<2$) have lower SFRs than often suggested, could be an important step towards bridging the discrepancies between models and observations. High SFRs of the order of a few thousand M$_{\odot}$/yr or more have been difficult to explain; for example they are much higher than the estimated Eddington limited SFR (in the optically thick limit) of $\sim$1400M$_{\odot}$/yr for typical high redshift ULIRGs (e.g. Murray, Quataert $\&$ Thompson 2005). Moreover, it has been challenging for the $\Lambda$ Cold Dark Matter ($\Lambda$CDM) galaxy formation models to reproduce the most luminous dusty galaxies at high redshift without evoking a flat or top heavy initial mass function (IMF; e.g. Baugh et al. 2005, Swinbank et al. 2008; Davé et al. 2010).
Implications for galaxy surveys
-------------------------------
Since the discovery of HyLIRGs from the *IRAS* survey, the increased sensitivity of the *WISE* survey has facilitated the discovery of an unprecedented number of hyperluminous sources (e.g. Secrest et al. 2015; Toba $\&$ Nagao 2016). Here we have shown that the most luminous sources discovered by *WISE*, or any other all-sky survey (at least at $1<z<2$), are likely to be AGN-powered. Since AGN completely dominate the emission in $L_{\rm IR}>10^{13.5}\,\rm L_{\odot}$ galaxies, a sample of HyLIRGs in this luminosity range may be perfect for finding out the obscured fraction of AGN at the highest luminosities and discriminating between the various torus models that have been put forward (e.g. Lawrence 1991; [H[ö]{}nig]{} $\&$ T. Beckert 2007; Lawrence $\&$ Elvis 2010; Mateos et al. 2017). Moreover, identifying the most luminous IR galaxies is a means of discovering AGN even if these are Compton thick, a significant step towards quantifying the contribution of the most heavily obscured AGN to the X-ray background (e.g. Akylas et al. 2012; Buchner et al. 2015).
The robustness of our computed fraction of AGN-dominated sources
----------------------------------------------------------------
### The contribution of compton thick AGN
The A15 luminosity function includes a component from Compton thick (CT) AGN, which affects its normalisation. Although the CT AGN fraction is uncertain, the value quoted in A15 is 34 per cent of the absorbed AGN, lower than other CT fractions quoted in the literature which range from 35 to 50 per cent of the whole AGN population. For example, Akylas et al. (2016) argue that for a CT fraction that evolves as a function of redshift, a value as high as 50 per cent at $z\sim1$ would be consistent with the 2-10keV number counts. An evolving CT fraction is in agreement with the results of Brightman $\&$ Ueda (2012), who find that the CT fraction increases with redshift, measuring 42$\pm7$ per cent per cent at $z\sim2$ (see also Vignali et al. 2014). To find how our results would change, we used the upper end of the quoted CT fraction range, i.e. 50 per cent, to re-compute the fraction of AGN-dominated sources as a function of $L_{\rm IR}$. We find that the point at which the galaxy population becomes AGN-dominated shifts to log\[$L_{\rm IR}/L_{\odot}] > 13.4$, 0.1dex lower than the value quoted in section \[sec:results\], but within the 1$\sigma$ confidence interval outlined in Fig \[fig:AGN\].
### The effect of SED choice
The S16 intrinsic AGN SED we used to convert the X-ray AGN luminosity function to an IR AGN luminosity function (see section \[sec:method\]) is arguably the most appropriate choice because it includes emission from AGN-heated dust in the torus as well as further afield, at kpc scales. S16 present a detailed comparison between the S16 SED and other AGN SEDs, finding that they all fall short of the S16 SED in the far-IR, be it because their derivation inherently assumes that AGN do not heat dust at kpc scales, or because their formulation only treats dust emission from the AGN torus. Nevertheless, we examine the effect of SED choice on our results, by re-computing the IR AGN luminosity function using the high luminosity AGN SED from Mullaney et al. (2011), which has the least far-IR emission out of the three SEDs they present and is similar to the Netzer et al. (2007) AGN SED (see S16). We find that the AGN luminosity function is shifted leftward by 0.1dex which translates to a rightward shift in the curve of the fraction of AGN-dominated sources by about 0.07dex, within the 1$\sigma$ confidence interval shown in Fig \[fig:AGN\].
Conclusions {#sec:conclusions}
===========
We have investigated what powers HyLIRGs at z$\sim$1-2, by examining the behaviour of the infrared AGN luminosity function in relation to the infrared galaxy luminosity function. The former corresponds to emission from AGN-heated dust only, whereas the latter includes emission from dust heated by AGN and starlight. At the peak of cosmic SFR history, $z\sim 1-2$:
- The AGN and galaxy luminosity functions converge in the HyLIRG regime, where the measured space densities of galaxies and AGN become comparable. This implies a transition between galaxies primarily powered by star-formation to galaxies primarily powered by AGN, the transition locus being at log\[$L_{\rm IR}/\rm L_{\odot}] \sim$13.5
- The shape of the infrared galaxy luminosity function at the most extreme luminosities flattens to match that of the AGN luminosity function, implying that at the bright end the infrared galaxy luminosity function is shaped by AGN, not by star formation.
- HyLIRG SFRs calculated using broadband photometry are likely overestimated and require a correction due to the AGN contribution. In particular, galaxies which have infrared luminosities that would translate to SFRs of more than a few thousand M$_{\odot}$/yr are instead likely to be primarily powered by the AGN.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'An unsteady three-dimensional boundary element method is developed to provide fast calculations of biological and bio-inspired self-propelled locomotion. The approach uniquely combines an unsteady three-dimensional boundary element method, a boundary layer solver and self-propelled equations of motion. This novel implementation allows for the self-propelled speed, power, efficiency and economy to be accurately calculated. A Dirichlet formulation is used with a combination of constant strength source and doublet elements to represent a deforming body with a nonlinearly deforming wake. The wake elements are desingularized to numerically stabilize the evolution of the wake vorticity. Weak coupling is used in solving the equations of motion and in the boundary layer solution. The boundary layer solver models both laminar and turbulent behavior along the deforming body to estimate the total skin friction drag acting on the body. The results from the method are validated with analytical solutions, computations and experiments. Finally, a bio-inspired self-propelled undulatory fin is modeled. The computed self-propelled speeds and wake structures agree well with previous experiments. The computations go beyond the experiments to gain further insight into the propulsive efficiency for self-propelled undulating fins. It is found that the undulating fin produces a time-averaged momentum jet at 76% of the span that accelerates fluid in the streamwise direction and in turn generates thrust. Additionally, it is discovered that high amplitude motions suppress the formation of a bifurcating momentum jet and instead form a single core jet. Consequently, this maximizes the amount of streamwise momentum compared to the amount of wasted lateral momentum and leads to a propulsive efficiency of 78% during self-propelled locomotion.'
address: 'Department of Mechanical Engineering and Mechanics, Lehigh University, Bethlehem, PA 18015, USA'
author:
- 'K. W. Moored'
bibliography:
- 'BEM\_Validation\_Literature.bib'
title: 'Unsteady Three-Dimensional Boundary Element Method for Self-Propelled Bio-Inspired Locomotion'
---
Bio-inspired propulsion ,boundary element method ,panel method ,self-propelled swimming ,unsteady flows ,vortex dynamics
Introduction
============
Boundary element methods (BEMs) are a class of numerical methods used to solve boundary value problems throughout physics from electromagnetics [@Kagami1984] and fracture mechanics [@Portela1992] to fluid flows at both low [@Pozrikidis2001] and high Reynolds numbers [@Basu1978]. In high Reynolds number flows they are classically described as panel methods and have been well established in the study of aerodynamics over several decades [@Hess1972; @Maskew1987; @Katz2001]. High Reynolds number BEMs assume that a fluid flow is incompressible, irrotational (except at singular elements) and inviscid, that is, a potential flow. This leads to simplified forms of the continuity and momentum equations that govern the fluid flow. Yet, unsteady BEM solutions are still rich with flow physics [@Quinn2014] and give accurate solutions at computational times that are several orders of magnitude faster than Navier-Stokes solvers [@Pan2012; @Willis2006].
Unsteady three-dimensional BEM computations have been used by many researchers to explore both biological and bio-inspired propulsion. The flight performance of birds [@Smith1996] and the swimming performance of fin whales [@Liu1997] and fish [@Cheng2001] have been examined to reveal features of high efficiency locomotion. For example, Zhu et al. [@Zhu2002] found that constructive or destructive interactions can occur between the shed vorticity from finlet structures and the caudal fin of tuna and giant danio. This can lead to enhanced thrust production or efficiency, respectively, with maximum efficiencies of 75% being calculated. More recently, Zhu [@Zhu2007] showed that spanwise and chordwise flexibility can enhance both thrust production and efficiency of a flapping wing. The benefit of flexibility was also found to be highly dependent upon the mass ratio between the wing and the surrounding fluid environment. Additionally, Zhu [@Zhu2008], Shoele & Zhu [@Shoele2009] and Shoele & Zhu [@Shoele2010] determined that the flexibility of ray-finned fish caudal and pectoral fins also improved their efficiency performance and reduced the time-varying lateral forces acting on the fish. Importantly, none of these previous studies have examined the locomotion of *self-propelled* swimmers nor the *free-flight* of flyers, yet these conditions are a critical feature of bio-inspired locomotion.
One complicating factor is that an inviscid BEM does not inherently calculate viscous drag. This gives no opposing force to balance the thrust production, which leaves out a necessary ingredient for calculating a steady-state self-propelled speed. However, viscous drag has been estimated in several other BEM studies by using a boundary layer momentum-integral approach on streamwise strips [@Robinson1988; @Liu1999; @Voutsinas2006]. Even with a viscous drag estimate included these studies focused on fixed freestream velocity conditions.
This work describes a novel implementation for computing the self-propelled performance of biological and bio-inspired propulsors within a BEM framework. There are three main components that must be combined to model self-propelled swimming: (1) a three-dimensional BEM fluid solver, (2) a boundary layer solver, and (3) an equations of motion solver. These components to the method are described in section \[sec:methods\]. Validation with several analytical, numerical and experimental solutions are presented in section \[sec:validation\]. Finally, comparison of the BEM solution with a three-dimensional self-propelled undulating fin experiment is presented in section \[sec:results\]. The free-swimming performance and wake structures are shown to agree well with the experiments. Additionally, the self-propelled performance of cases that extend beyond the previous experiments are examined to provide novel physical insight into the self-propulsion of three-dimensional ray-inspired fins.
Computational Methods {#sec:methods}
=====================
Governing Equations and Boundary Conditions
-------------------------------------------
To model a high Reynolds number fluid flow around an self-propelled bio-inspired device or animal an unsteady three-dimensional boundary element method is employed. The flow field is modeled as an incompressible, irrotational and inviscid flow, that is, a potential flow. For the self-propelled problem we define the problem in an inertial frame of reference that is attached to the undisturbed fluid (denoted by $(X,Y,Z)$ is Figure \[BEM\]). As such the velocity field, ${\mathbf}{u}$, may be defined everywhere as the gradient of a scalar velocity potential, $$\label{VelPot}
{\mathbf}{u} = {\boldsymbol}{\nabla} \Phi^*,$$
where $\Phi^*$ is defined in the inertial frame of reference and it is known as the perturbation potential. The pressure field within this fluid can be calculated from the unsteady Bernoulli equation, $$\label{BernI}
P(X,Y,Z,t) = -\rho \frac{\partial \Phi^*}{\partial t}\bigg|_{inertial} - \rho \frac{\left({\boldsymbol{\nabla}}\Phi^* \right)^2}{2},$$
which is formulated here in the inertial frame where the reference pressure $P_\infty = 0$ and the perturbation potential at infinity is zero. The time derivative of the perturbation potential for a point on the surface of the body is then calculated by using a body-fixed Lagrangian frame (denoted by $(x,\, y,\, z)$ in Figure \[BEM\]) [@Cheng2001; @Willis2007; @Pan2012], that is, $$\begin{aligned}
\label{BernBody}
P(x,y,z,t) = -\rho \frac{\partial \Phi^*}{\partial t}\bigg|_{body} + \rho \left({\mathbf}{u_{rel} + {\mathbf}{U_0}} \right) \cdot {\boldsymbol{\nabla}}\Phi^*- \rho \frac{\left({\boldsymbol{\nabla}}\Phi^* \right)^2}{2}.\end{aligned}$$
![The inertial reference frame fixed to the undisturbed fluid is denoted by $(X,Y,Z)$ while the body-fixed reference frame is denoted by $(x,y,z)$. The local normal, streamwise and cross-stream unit vectors are denoted by $\mathbf{\hat{n}}$, $\mathbf{\hat{s}}$, and $\mathbf{\hat{c}}$, respectively. The body surface, ${\mathcal}{S}_b $, is layered with distributions of doublet elements of strength $\mu$ and source elements of strength $\sigma$. The wake surface, ${\mathcal}{S}_w$, is layered with distributions of doublet elements of strength $\mu_w$.[]{data-label="BEM"}](figure_01.pdf){width="85.00000%"}
The translational velocity of a body-fixed frame of reference is ${\mathbf}{U_0}$ while the relative velocity of a point on the surface of the body to the body-fixed reference frame is ${\mathbf}{u_{rel}}$. Once the perturbation potential is known, then the pressure on the body surface may be found and the forces can be calculated by integrating the pressure and shear stress acting on the body. $$\label{Forces}
{\mathbf}{F}\left(x,y,z,t \right) = \int_{{\mathcal}{S}_b } \left(-P\; {\mathbf}{\hat{n}} + \tau \; {\mathbf}{\hat{s}} \right)\; d{\mathcal}{S}$$
The body surface is denoted as ${\mathcal}{S}_b $, the outward normal vector from the body surface is ${\mathbf}{\hat{n}}$ and the tangential vector along the body surface in the streamwise direction is ${\mathbf}{\hat{s}}$. This inviscid formulation is coupled to a viscous boundary layer solver described in section \[BLSolve\], which estimates the shear stress, $\tau$, acting on the body in the streamwise direction produced by the outer potential flow. Note that the shear stress acting in the cross-stream direction is not accounted for in the viscous boundary layer solver and is therefore not present in equation (\[Forces\]). The problem is then reduced to solving for the perturbation potential throughout the fluid, which is governed by Laplace’s equation, $$\label{LaplaceI}
\nabla^2 \Phi^* = 0.$$
The boundary conditions that must be satisfied for an inviscid fluid are that there is no fluid flux through the body surface and that the flow disturbances caused by the body must decay far away, $$\begin{aligned}
\label{NoFlux}
&{\mathbf}{n} \cdot {\boldsymbol}{\nabla} \Phi^* = {\mathbf}{n} \cdot \left( {\mathbf}{u_{rel} + {\mathbf}{U_0} } \right) \quad \quad \mbox{on } {\mathcal}{S}_b \\
&{\boldsymbol}{\nabla} \Phi^* \bigg|_{|{\mathbf}{x}| \rightarrow \infty} = 0 \quad \quad \mbox{on } {\mathcal}{S}_\infty\end{aligned}$$
where ${\mathcal}{S}_\infty$ is the surface at infinity bounding the fluid and ${\mathbf}{x} = \left[ x, \, y, \, z\right]^T$ is measured from the body-fixed frame of reference.
Boundary Integral Equation
--------------------------
A general solution to Laplace’s equation for the potential anywhere within the fluid domain, $\mathcal{V}$, can be determined. This is done by considering the potential response at a point ${\mathbf}{x} = [x,\, y, \, z ]^T$ due to a source located at ${\mathbf}{x_0} = [x_0,\, y_0, \, z_0 ]^T$, that is, the infinite space Green’s function which satisfies the Poisson equation with homogeneous far-field boundary conditions [@Haberman2004]. The infinite space Green’s function in three dimensions is then, $$\begin{aligned}
G({\mathbf}{x};{\mathbf}{x_0}) = - \frac{1}{4 \pi r}, \quad \quad \mbox{where } r = |{\mathbf}{x} - {\mathbf}{x_0} |.\end{aligned}$$
By invoking Green’s formula twice with respect to the perturbation potential within the fluid volume, the internal perturbation potential (the volume enclosed by ${\mathcal}{S}_b$) and the Green’s function and then adding the results, the following boundary integral equations (BIEs) for the internal or external perturbation potential are derived, $$\begin{aligned}
\label{IntPertPot}
\Phi_i^*({\mathbf}{x}) & = \oiint_{{\mathcal}{S}_b} \left[ \sigma({\mathbf}{x_0}) \; G({\mathbf}{x};{\mathbf}{x_0}) - \mu({\mathbf}{x_0}) \; {\mathbf}{\hat{n}} \cdot {\boldsymbol}{\nabla} G({\mathbf}{x};{\mathbf}{x_0})\right] \; d{\mathcal}{S}_0 -\oiint_{{\mathcal}{S}_w} \mu_w({\mathbf}{x_0}) \; {\mathbf}{\hat{n}} \cdot {\boldsymbol}{\nabla} G({\mathbf}{x};{\mathbf}{x_0}) \; d{\mathcal}{S}_0 \\
\Phi^*({\mathbf}{x}) & = \oiint_{{\mathcal}{S}_b} \left[ \sigma({\mathbf}{x_0}) \; G({\mathbf}{x};{\mathbf}{x_0}) - \mu({\mathbf}{x_0}) \; {\mathbf}{\hat{n}} \cdot {\boldsymbol}{\nabla} G({\mathbf}{x};{\mathbf}{x_0})\right] \; d{\mathcal}{S}_0 -\oiint_{{\mathcal}{S}_w} \mu_w({\mathbf}{x_0}) \; {\mathbf}{\hat{n}} \cdot {\boldsymbol}{\nabla} G({\mathbf}{x};{\mathbf}{x_0}) \; d{\mathcal}{S}_0 \\
& \mbox{where,} \nonumber \\
\sigma({\mathbf}{x_0}) & = {\mathbf}{\hat{n}} \cdot {\boldsymbol}{\nabla} \left(\Phi^* - \Phi_i^*\right) \label{Source} \\
-\mu({\mathbf}{x_0}) & = \Phi^* - \Phi_i^* \\
-\mu_w({\mathbf}{x_0}) & = \Phi_+^* - \Phi_-^*\end{aligned}$$
During the derivation of the boundary integral equation the source and observation locations switch roles. Note that when this happens $G({\mathbf}{x_0};{\mathbf}{x}) = G({\mathbf}{x};{\mathbf}{x_0})$ due to the reciprocity of the Green’s function, however, ${\boldsymbol}{\nabla} G({\mathbf}{x_0};{\mathbf}{x}) = -{\boldsymbol}{\nabla} G({\mathbf}{x};{\mathbf}{x_0})$. The surface boundary integral is broken up into an integral over the body boundary, the wake boundary and the far-field boundary. To formulate the problem in an inertial frame of reference attached to the undisturbed fluid the contribution to the potential from the farfield boundary is set to zero, that is $\Phi_\infty = 0$. The potential jump $\mu({\mathbf}{x_0})$ represents the strength of a dipole or doublet while the jump in the normal derivative of the potential $\sigma({\mathbf}{x_0})$ represents the strength of a source. The wake boundary in the limit as it becomes infinitesimally thin has a continuous normal derivative of the potential so it does not support a source distribution. The local potential jump between the top and bottom surface of the wake is represented by $\mu_w({\mathbf}{x_0})$, where $\Phi^*_+$ is the potential above the wake and $\Phi^*_-$ is the potential below the wake.
The general solution to the potential flow problem in a fluid domain ${\mathcal}{V}$ is now reduced to finding a distribution of sources and doublets on the boundaries ${\mathcal}{S}_b$ and ${\mathcal}{S}_w$ that satisfy the boundary conditions. Since the continuity and thus Laplace’s equation is time-independent, all of the time-dependency comes from the unsteady Bernoulli equation and the time-dependent boundary conditions.
Enforcement of the Boundary Conditions
--------------------------------------
The elementary solutions of the doublet and source both implicitly satisfy the far-field boundary condition. In this paper, the no-flux condition on the body is satisfied with an indirect Dirichlet formulation since it saves computational effort as compared to the Neumann formulation [@Katz2001]. For the Dirichlet problem, we set the internal potential to a constant, which in our case is selected to be zero $$\begin{aligned}
\label{Dirichlet}
\Phi_i^* = 0.\end{aligned}$$
To enforce this condition the BIE must be satisfied with the left-hand side equal to zero. This condition also leads to the simplification of the source and doublet strength equations, which become $$\begin{aligned}
-\mu & = \Phi^* \\
\sigma & = {\boldsymbol}{\nabla} \Phi^* \cdot {\mathbf}{n} = \left( {\mathbf}{u_{rel} + {\mathbf}{U_0} } \right) \cdot {\mathbf}{n}.\end{aligned}$$
The perturbation velocity on the surface of the body is simplified and can be found by a local differentiation of the perturbation velocity potential, $$\begin{aligned}
{\mathbf}{u_b} = {\boldsymbol{\nabla}}\Phi_b^* = \frac{\partial \Phi*}{\partial s} \; {\mathbf}{\hat{s}} + \frac{\partial \Phi*}{\partial c} \; {\mathbf}{\hat{c}} + \frac{\partial \Phi*}{\partial n} \; {\mathbf}{\hat{n}} = - \frac{\partial \mu}{\partial s} \; {\mathbf}{\hat{s}} - \frac{\partial \mu}{\partial c} \; {\mathbf}{\hat{c}} + \sigma \; {\mathbf}{\hat{n}},\end{aligned}$$
where ${\mathbf}{\hat{c}}$ is the tangential vector along the surface in the cross-stream direction. The pressure over the body can also be found in terms of the boundary element strengths, $$\begin{aligned}
\label{BernBody}
P_b(x,y,z,t) = \rho \frac{\partial \mu}{\partial t}\bigg|_{body} + \rho \left({\mathbf}{u_{rel} + {\mathbf}{U_0}} \right) \cdot {\mathbf}{u_b} - \rho \frac{{\mathbf}{u_b} ^2}{2}.\end{aligned}$$
Now, the problem is reduced to finding a distribution of sources and doublets that solve the BIE when the Dirichlet condition is applied.
Numerical Solution
------------------
To numerically solve this problem, the boundaries are discretized into constant strength quadrilateral boundary elements over the body and wake. Then the boundary integral equation (eq. (\[IntPertPot\])) with the Dirichlet condition substituted (eq. (\[Dirichlet\])) can be discretized into the summation over a series of boundary elements, $$\begin{aligned}
\label{Discretized}
\sum^{N_b}_{j = 1} B_{ij} \sigma_j & +
\sum^{N_b}_{j = 1} C_{ij} \mu_j
+ \sum^{N_w}_{k = 1} C_{w,ik} \mu_{w,k} = 0 \\
& \mbox{with,} \nonumber \\
B_{ij} & = -\frac{ 1}{4 \pi} \int_{ele} \frac{1} {| {\mathbf}{r_{ij}} |} \; \mbox{d}{\mathcal}{S}_0 \\
C_{ij} & = -\frac{ 1}{4 \pi} \int_{ele}\frac{{\mathbf}{\hat{n}} \cdot {\mathbf}{r_{ij}} }{|{\mathbf}{r_{ij}}|^3} \; \mbox{d}{\mathcal}{S}_0 \\
C_{w,ik} & = -\frac{ 1}{4 \pi} \int_{ele}\frac{{\mathbf}{\hat{n}} \cdot {\mathbf}{r_{ik}} }{|{\mathbf}{r_{ik}}|^3} \; \mbox{d}{\mathcal}{S}_0 \\
\mbox{and} \quad {\mathbf}{r_{ij}} =& {\mathbf}{x_i} - {\mathbf}{x_{0,j}}, \quad {\mathbf}{r_{ik}} = {\mathbf}{x_i} - {\mathbf}{x_{0,k}} \end{aligned}$$
where $N_b$ is the number of body elements, $N_w$ is the number of wake elements, $d\mathcal{S}_0$ is the differential area of a boundary element, ${\mathbf}{x_i}$ is the vector denoting the position of the $i^{th}$ collocation point, ${\mathbf}{x_{0,j}}$ is the vector denoting the position of a differential area of the $j^{th}$ element and ${\mathbf}{x_{0,k}}$ is the vector denoting the position of a differential area of the $k^{th}$ element. Explicit solutions for the influence integrals over the elements can be found in Katz [@Katz2001]. Equations (\[Discretized\]) must be satisfied at every point within the boundary ${\mathcal}{S}_b$, which in the discretized form are satisfied at $N_b$ collocation points. In the current method, the collocation points were located at the center of the elements but moved into the body by 15% of the distance of the half-thickness of the body at that point, along the element normal vector. The system of equations (\[Discretized\]), however, needs to be modified by an explicit or implicit Kutta condition to allow the support of bound circulation. Also, for time-stepping, a wake shedding procedure must be used to satisfy Kelvin’s condition.
Wake Model
----------
In this work an explicit Kutta condition is chosen for its simplicity. A trailing-edge doublet element is the first element in the set of wake elements and it is used to enforce the Kutta condition of finite velocity at the trailing edge. This occurs by the trailing-edge element having a strength that cancels the vorticity at the trailing-edge. The strength of the trailing-edge element is found at each time step from the difference in strengths between the top and bottom trailing-edge doublet elements, i.e., $$\begin{aligned}
\label{Kutta}
\mu_{w,TE} = \mu_{t,TE} - \mu_{b,TE}.\end{aligned}$$
The orientation of the element is set to be along a line that bisects the angle of the trailing-edge, which is typical for steady flow conditions [@Katz2001]. Traditionally, the trailing-edge element length is set to $0.3$ – $0.5 \, U_0 \Delta t$ [@Willis2007], where $\Delta t = 1/ (f N_{step})$ and $N_{step}$ is the number of time steps per oscillation cycle. Here a length of $0.4 \,U_0 \Delta t$ was used since it gave good solution convergence while maintaining solution accuracy with the validation cases.
During each time step the trailing-edge element from the previous time step is ‘shed’ a distance $U_0 \Delta t$ downstream and it becomes the second wake element. The strength of that wake element is the same as the strength of the previous trailing-edge element and it remains constant for all subsequent time steps. The trailing-edge element’s strength can be re-written in terms of the unknown body element strengths using equation (\[Kutta\]) and the discretized BIE is then modified, that is, $$\begin{aligned}
\label{BIE_Kutta}
\sum^{N_b}_{j = 1} A_{ij} \mu_j & = - \sum^{N_b}_{j = 1} B_{ij} \sigma_j - \sum^{N_w}_{k = 2} C_{w,ik} \mu_{w,k}\\
& \mbox{with,} \nonumber \\
A_{ij} & =
\begin{cases}
C_{ij} - C_{w,i1}, & \quad \mbox{j = bottom element} \\
C_{ij} + C_{w,i1}, & \quad \mbox{j = top element} \\
C_{ij}, & \quad \mbox{otherwise}
\end{cases} \nonumber\end{aligned}$$
Now, the body source element strengths and wake element strengths are known. The linear set of equations may be solved at each time step for the body doublet strengths, $\mu_j$, by an inversion of matrix $A_{ij}$. The trailing-edge element strengths are then directly calculated from the body doublet element strengths. In the current method, the number of unknown body doublet element strengths is reduced in half by assuming left-right symmetry for the problem and using mirror image elements to represent the left-half of the body and the wake.
Nonlinear Wake Deformation {#Wake}
--------------------------
The wake elements that are shed at each time step model the shedding of vorticity from the body into its wake. These elements cannot support loads so they must be free to advect with the local velocity field. At each time step the induced velocity at the corner points of each wake element, ${\mathbf}{u}_{w}$, is determined. The wake element corner points are then displaced by $\Delta {\mathbf}{d} = {\mathbf}{u}_{w} \Delta t$. Calculating the induced velocity at the wake element corner points will lead to a numerically unstable solution if the doublet elements are not desingularized. Here we take advantage of the equivalence of a constant strength doublet element and a vortex ring around the edge of that element by using the desingularized Biot-Savart law, $$\begin{aligned}
{\mathbf}{u(x)} = \frac{\Gamma}{4 \pi} \oint \frac{{\mathbf}{s} \times {\mathbf}{r}}{r^3 + \delta^3} \; ds,\end{aligned}$$
to calculate the induced velocity field [@Krasny1986]. Here, the circulation of an element is $\Gamma = -\mu$ and the desingularization parameter, $\delta$, is a constant and a free-parameter for the method. Provided that $\delta$ is large enough, the transfer of energy to high wavenumbers is minimized thereby preventing solution breakdown [@Zhu2002]. The desingularization parameter mimics the effect of viscosity in a real fluid by giving each vortex ring element a core radius directly related to $\delta$.
Lumped Wake Elements {#Lump}
--------------------
A lumped wake element model is used to restrict the growth of the problem size as the number of wake elements increases with every time step. The lumped elements conserve the net circulation in the wake such that Kelvin’s condition still holds. There is one lumped wake element for every trailing-edge element, with the trailing-edge elements acting as the wake element generators and the lumped elements acting as wake element absorbers in the far-field. The strength of the lumped elements at the $n^{th}$ time step is the summation of the circulation of the lump elements at the previous time step and the absorbed elements at the $n^{th}$ time step. $$\begin{aligned}
\label{LumpedStrength}
\Gamma_{lump}^{n} &= \Gamma_{lump}^{n-1}+ \Gamma_{w,absorbed}^{n} \end{aligned}$$
The lumped element corner point locations, ${\mathbf}{p}_{lump}$, are at the weighted-average locations of the absorbed elements where the weights are based on the magnitude of the increment of circulation added to the lumped element compared to the total absorbed magnitude of circulation, $$\begin{aligned}
{\mathbf}{p}_{lump}^{n}&= \left[\frac{\Gamma_{mag}^{n-1}}{\Gamma_{mag}^{n}}\right] {\mathbf}{p}_{lump}^{n-1}+ \left[\frac{|\Gamma_{w,absorbed}^{n}|}{\Gamma_{mag}^{n}}\right] {\mathbf}{p}_{absorbed}^{n} \\
& \mbox{with,} \nonumber \\
\Gamma_{mag}^{n} &= \Gamma_{mag}^{n-1}+ |\Gamma_{w,absorbed}^{n}| \nonumber\end{aligned}$$
The lumped elements absorb wake elements such that the most recent $N_l$ oscillation cycles of wake elements remain to fully-resolve the near wake. It is found that by measuring the time-averaged forces on the body, the lumped wake solution is within 1% of the fully-resolved solution if $N_l \geq 4$ for two-dimensional flows and $N_l \geq 2$ for three-dimensional flows. The lumped wake elements are included in the third term of equation (\[BIE\_Kutta\]).
Equations of Motion {#EoM}
-------------------
To be able to calculate the self-propelled body dynamics and performance of a bio-inspired device or an animal, the equations of motion for the body must be solved. For the results in this work, we only allow streamwise translation to be unconstrained while the other degrees of freedom undergo fully prescribed motions. To further simplify the implementation of the unconstrained body dynamics, a loose or one-way coupling with the fluid solution is used. That is, the forces from the solution of the BEM fluid problem are used as the driving forces in the equations of motion. The body velocity and position is then explicitly determined without sub-iterations between the fluid solver and the body dynamics solver. Following [@Borazjani2008], the loose-coupling scheme uses the body position and velocity at the current time step to explicitly solve for the position and velocity at the subsequent time step, $$\begin{aligned}
\label{EoM}
x_b^{n+1} &= x_b^{n} + \frac{1}{2}\left( U_0^{n+1} + U_0^{n} \right) \Delta t \\
U_0^{n+1} &= U_0^{n} + \frac{F_x^n}{M} \Delta t\end{aligned}$$
Viscous Boundary Layer Solver {#BLSolve}
-----------------------------
An estimate of skin friction drag is critical to computing self-propelled body dynamics and performance as this is the main source of drag for streamlined swimmers. Using the potential outer flow solution, the boundary layer properties and the skin friction can be estimated without an assumption on the form of a drag law acting on a body. To calculate the boundary layer properties a two-dimensional momentum integral analysis is performed over streamwise strips along the body. Some three-dimensionality is embedded into the analysis by using an effective outer flow velocity at each chordwise station, which is the magnitude of the streamwise and cross-stream tangential velocities at those locations, $$\begin{aligned}
|q_{outer}| = \sqrt{q_{s}^2 + q_{c}^2}\end{aligned}$$
This analysis can be used to approximate the boundary layer displacement thickness and the skin friction drag coefficient over each boundary element based on the outer flow alone. The outer flow is coupled to the boundary layer solution in a loose coupling, that is, there are no sub-iterations between the outer flow solution and the shape of the body based on the displacement thickness of the boundary layer. From the outer flow solution the stagnation point on the body is determined for each strip. Then based on the stagnation point the surface is divided into an upper and lower part where each boundary layer begins growing. First, the method of Thwaites is used to solve for the laminar boundary layer properties, next Michel’s criterion is used to determine the transition to turbulence location and finally a one-seventh-power law is used to calculate the turbulent boundary layer properties [@Robinson1988; @White2006].
Numerical Method Validation {#sec:validation}
===========================
To validate the current boundary element method implementation a series of analytical, numerical and experimental results are used to analyze the accuracy of the current numerical method. The validation cases include a two-dimensional steady and unsteady flow case, a three-dimensional steady and unsteady flow case, a skin friction drag case, a self-propelled biological propulsion case and a self-propelled bio-inspired undulating fin case.
Two-Dimensional Steady Flow
---------------------------
The steady pressure distribution over a two dimensional van de Vooren airfoil is first examined. The van de Vooren airfoil is chosen since there is an exact analytical solution for its pressure distribution [@VandeVooren1969] and since it has a finite-angle trailing edge as opposed to the cusped trailing edge of a Joukowski airfoil. Cusped trailing edges are numerically difficult to obtain accurate solutions when using a Dirichlet formulation.
![(a) Schematic of a 15% thick van de Vooren airfoil at an angle of attack, $\alpha$, in a steady flow, $U_\infty$. The coefficient of pressure over the top and bottom surface of the airfoil for three angles of attack: (b) $\alpha = 0^0$, (c) $\alpha = 5^0$ and (d) $\alpha = 10^0$. The analytical solution for the top surface is represented by the solid red line while the numerical solution is represented by the red triangles. For the bottom surface, the analytical solution is represented by the solid blue line and the numerical solution is represented by the blue squares.[]{data-label="fig:CoefPres"}](figure_02.pdf){width="80.00000%"}
The current unsteady three-dimensional formulation is used to solve this two-dimensional steady flow problem by solving for the flow over a van de Vooren wing with a rectangular planform shape and an aspect ratio of $1000$. The wing is discretized into $30$ chordwise boundary elements for both the top and bottom surfaces and $50$ spanwise boundary elements for a total of $3000$ body elements. The computation is discretized into $10$ timesteps with the elapsed time for each step being $\Delta t = 100$ s. The starting vortex present in the unsteady numerical solution is $1000$ chord lengths downstream of the wing since the total elapsed time for the simulations is $\Delta t_{total} = 1000$ s, the free-stream velocity is $U_\infty = 1$ m/s and the chord length of the wing is $c = 1$ m. This is necessary since the analytical solution assumes that the starting vortex is infinitely far away. Also, to conform with the analytical solution the wake is ‘frozen’, that is, it is not allowed to roll-up or advect with the local velocity field.
The coefficient of pressure over the wing is calculated for the section nearest the symmetry plane of the wing where the flow is nearly two-dimensional. Figure \[fig:CoefPres\]a shows a schematic of the physical problem where a 15% thick van de Vooren wing is placed in a steady flow of $U_\infty$ and an angle of attack, $\alpha$. Figure \[fig:CoefPres\]b–d compares the exact solutions (solid lines) to the numerical BEM solutions (square and triangle markers). For $\alpha = 0^o$, the pressure coefficient is observed to stagnate at the leading edge ($C_p = 1$). Subsequently, the pressure drops as flow is accelerated symmetrically over the top and bottom surfaces until a minimum pressure of $C_p = -0.5$ is reached around $x/c = 0.2$ near the maximum thickness. After the maximum thickness location there is a gradual pressure recovery of the flow to the trailing edge. For $\alpha > 0^o$, a prominent leading-edge suction peak can be observed that grows with increasing $\alpha$. Importantly, there is excellent agreement between the analytical and numerical solutions over a range of angles of attack from $\alpha = 0$ – $10^o$.
Two-Dimensional Unsteady Flow
-----------------------------
The unsteady lift produced by a two-dimensional thin airfoil heaving with small amplitude motion is examined next. Theodorsen [@Theodorsen1935] provided an analytical solution to this problem and a schematic detailing the problem is shown in Figure \[Theo\]a. The Theodorsen model for a pure heaving motion can be cast into a non-dimensional form, $$\begin{aligned}
c_l &= -2 \pi^2 \; St \; |C(k)| \; cos(2 \pi \tau + \phi) + \pi^2St\; k \; sin(2 \pi \tau) \\
& \mbox{where,} \nonumber \\
c_l & = \frac{L}{1/2 \; \rho c U_\infty^2} \quad h = h_0 \; sin (2 \pi \tau) \quad St = \frac{2 h_0 f}{U_\infty} \quad k = \frac{\pi f c}{U_\infty} \quad \tau = ft \nonumber\end{aligned}$$
Here, $h$ is the time varying heave, $|C(k)|$ is the magnitude and $\phi$ is the phase angle of the lift deficiency factor, which is a complex number. A 0.1% thick symmetric NACA airfoil was used to simulate an infinitesimally thin plate. The parameters used were $c = 1$ m, $U_\infty = 1$ m/s, $h_0/c = 0.001$ with two cases of $St = 0.001$, $k = \pi/2$ and $St = 0.01$, $k = 5 \pi$. The three-dimensional formulation is used to solve this two-dimensional unsteady problem by solving for the flow over airfoil with a rectangular planform shape and an aspect ratio of $1000$. The lift coefficient from the numerical solution is then normalized by the planform area instead of the chord length as in the Theodorsen model. The wing was discretized into $50$ chordwise boundary elements for both the top and bottom surfaces and $40$ spanwise boundary elements for a total of $4000$ body elements. The computation was discretized into $50$ timesteps per oscillation cycle for $4$ cycles. The force calcuations were taken from the fourth oscillation cycle. A frozen wake was used to simulate the assumptions in the Theodorsen model. Figure \[Theo\] shows excellent agreement between the time-varying lift coefficients of the analytical and numerical solutions. The BEM solution of the time-varying lift coefficient for a Strouhal number of $St = 0.01$ has a slightly lower peak lift coefficient of $C_L = 1.505$ compared to the Theodorsen model of $c_l = 1.545$, which is also seen in other studies [@Willis2007].
![Comparison of the numerical solution marked by circles (o) with the Theodorsen model denoted by the solid line (–). The parameters of $c = 1$, $U_\infty = 1$, $h/c = 0.001$ were used for both graphs while the left graph used $St = 0.001$ and $k = \pi/2$ and the right graph used $St = 0.01$ and $k = 5 \pi$.[]{data-label="Theo"}](figure_03.pdf){width="80.00000%"}
Three-Dimensional Steady Flow
-----------------------------
The steady flow over a three-dimensional elliptical wing was examined to help validate the three-dimensional aspects of the current formulation. A teardrop shaped airfoil with a maximum thickness-to-chord ratio of $t/c = 0.02$ was used to simulate a flat plate. A 2% thickness was chosen to be sufficiently large so that there is a well-defined leading-edge to avoid numerical sensitivity in the induced drag calculation. There were 70 chordwise elements on the top and bottom surfaces as well as 50 spanwise elements for a total of $7000$ body elements. The computation used $4$ timesteps with $\Delta t = 250$ s. A “frozen” wake was used in the force calculations of this validation. Figure \[fig:EllpWing\] shows good agreement between the analytical and numerical calculations. It is expected that the numerical and analytical solutions would deviate at lower aspect ratios. This is where the finite-wing theory solution begins to breakdown. An example of the wake structure for a “free" wake is shown in Figure \[fig:EllpWing\]c. In this calculation the wing has an aspect ratio of $AR = 5$, an angle of attack of $\alpha = 15^o$. The airfoil thickness-to-chord ratio is $t/c = 0.2$. There are $70$ timesteps with $\Delta t = 0.07$ s between them. The wake is allowed to rollup showing the characteristic horseshoe vortex system connecting the tip vortex system to the starting vortex. On the body the colormap represents the pressure coefficient. There is stagnation pressure near the leading and trailing-edges while there is a minimum in pressure near the leading-edge responsible for leading-edge suction and a majority of the lift production.
![Calculated coefficient of lift and induced drag of steady flow over an elliptical wing at varying angles of attack $\alpha$. The numerical solution (circles) compare well with the analytical solution (solid line). There are four aspect ratios used in the calculations: $AR = 5, \: 10, \: 15,$ and $20$.[]{data-label="fig:EllpWing"}](figure_04.pdf){width="99.00000%"}
Three-Dimensional Unsteady Flow
-------------------------------
The unsteady flow problem of the lift on a oscillating three-dimensional wing is examined next. There is a numerical solution to this problem presented by Katz [@Katz1985] where he calculated the lift generated by a heaving three-dimensional rectangular wing at a negative angle of incidence using a boundary element method. For this case the rectangular wing has an aspect ratio of $AR = 4$, an angle of incidence $\alpha = -5^o$ and a heave amplitude of $h = 0.1c$, where $c$ is the chord length. The reduced frequency $k = \pi f c/ U_\infty$ is varied for three cases $k = 0.1, \: 0.3,$ and $0.5$ where $f$ is the oscillation frequency and $U_\infty$ is the free-stream velocity. In the current numerical implementation a $2$% thick tear-drop shaped airfoil is used to simulate a infinitely-thin flat plate. The top and bottom surface of the wing is discretized into $30$ chordwise and $50$ spanwise elements for a total of $3000$ body elements. The computation is discretized into $50$ timesteps per cycle with $\Delta t = [0.63, 0.21, 0.125]$ s for each of the reduced frequency cases respectively. A total of $4$ oscillating cycles are computed. Figure \[fig:UnsRectWing\]a shows the lift coefficient as a function of time over the fourth oscillation cycle. Excellent agreement can be seen between the current numerical implementation (solid line) and the numerical solution of [@Katz1985] (open markers). As the reduced frequency is increased the magnitude of the unsteady lift is seen to grow and there is a negative phase-shift in the lift oscillation as well. Figure \[fig:UnsRectWing\]b shows the rollup of the wake elements that are free to advect with the local velocity field for the case where $k = 0.5$. As the airfoil heaves upward, the circulation from the unsteady motion adds to the circulation due to the angle of incidence and subsequently the tip vortices increase their strength. When the airfoil heaves downward the circulation due to the incidence angle and the downward motion counteract each other to produce weak tip vortices. The net effect is to produce oscillating regions of strong and weak upwash.
![Wing oscillating under unsteady flow conditions. (a) Lift coefficient calculated with the current numerical method (solid lines) compared to a previous numerical calculation (open markers) by Katz [@Katz1985]. (b) Boundary element wake deformation.[]{data-label="fig:UnsRectWing"}](figure_05.pdf){width="90.00000%"}
Viscous Boundary Layer Solution
-------------------------------
The viscous drag over a NACA 0012 airfoil was examined next by using the viscous boundary layer solver coupled with the outer potential flow solution from the boundary element method. Figure \[fig:ViscDrag\]a shows a wing with $AR = 4$ at an angle of attack $\alpha = 3^o$ and a Reynolds number of $Re = 1\times 10^6$. The boundary elements are colored with the value for the shear stress acting on each element. When the wing is examined up close, the shear stress is seen to be high at the minimum pressure location on the wing as flow is accelerated around the leading-edge. Then the shear stress decreases downstream of the minimum pressure point until the boundary layer transitions to a turbulent boundary layer. At this location, the shear stress is seen to discontinuously increase and then slowly decrease toward the trailing edge.
![(a) Shear stress on the surface of a NACA 0012 wing in steady flow conditions. Drag coefficient calculated numerically (solid line) compared to experimental data (circles) from Abbott and von Doenhoff [@Abbott1959]: (b) $Re = 3 \times 10^6$, (c) $Re = 6 \times 10^6$, and (d) $Re = 9 \times 10^6$[]{data-label="fig:ViscDrag"}](figure_06.pdf){width="90.00000%"}
A wing with an $AR = 1000$ was used to compare the calculated drag with experimental measurements published in Abbott and von Doenhoff [@Abbott1959]. The calculations spanned a range of angles of attack between $-8^o$ and $8^o$ and three Reynolds numbers of $Re = [3\times 10^6, \; 6\times 10^6, \; 9\times 10^6]$. The top and bottom surface of the wing is discretized into $50$ chordwise and $50$ spanwise elements for a total of $5000$ body elements. The computation is discretized into $5$ timesteps with $\Delta t = 200$ s. The wake is ‘frozen’ in the computations.
Excellent agreement can be seen between the calculated drag coefficient (solid line) and the experimentally measured values (open circles) for low angles of attack less than $\pm 4^o$. For higher angles of attack, the drag coefficient is under-predicted by the numerical solution. In the worst case examined here the calculated drag is $22$% lower than the measured value at $\alpha = 8^o$. This discrepancy could be due to mild or moderate trailing-edge separation that may occur in the experiments, but not modeled in the current numerical solution. Future work of incorporating a separation model may improve the drag calculations.
Self-Propelled Biological Propulsion
------------------------------------
Recent work has examined in detail the forces and flow structures exhibited by a numerical manta ray swimmer by using the code presented in the current study [@Fish2016]. In the previous work the numerical method was not shown to be extensively validated as is the case and the focus of the current work. From extensive biological data gathered in aquaria and in the field, the manta ray swimmer (figure \[manta\]a) was modeled to be geometrically similar to real mantas and to oscillate its fins with the same kinematic motions of the rays (see [@Fish2016] for further details). The swimming motions of the manta are three-dimensionally complex because of the flexibility of the propulsive pectoral fins. The manta propels itself by vertically moving its enlarged pectoral fins in a flapping motion, combined with waves moving through the fins in both the spanwise and chordwise directions. The motion is thus both oscillatory, and shape-changing (undulatory), although the undulatory component in the chordwise direction is small (the wavelength of the undulation is greater than the chord length of the pectoral fin).
![Reproduced from [@Fish2016]. (a) Vortex wake of a manta ray swimmer identified by isosurfaces of the $\lambda_2$ criteria. (b) Numerical calculations of the self-propelled speed as a function of frequency (solid lines with open markers) compared to biological data (filled markers). The left vertical axis shows the mean speed normalized by the body length while the right vertical axis shows the mean speed normalized by the frequency and body length.[]{data-label="manta"}](figure_07.pdf){width="99.00000%"}
Figure \[manta\]a shows the vortex wake produced by a free-swimming manta ray. The vortices are marked as isosurfaces of the $\lambda_2$ criteria, which distinguishes pressure minima in a plane after discarding unsteady straining and viscous effects [@Jeong1995a]. The manta is seen to shed a series of interlocked vortex rings with one set originating from each pectoral fin.
The left axis line in Figure \[manta\]b shows the mean swimming speed normalized by the body length as a function of oscillation frequency for the numerical manta ray. The right axis line further normalizes the mean speed by the oscillation frequency to highlight the near linear dependence upon the oscillation frequency. The small blue points represent biological data of the swimming speed of real manta rays as a function of their oscillation frequency. Importantly, the present numerical method shows excellent agreement with the biological data indicating that the current method can predict the performance of real biological self-propelled swimmers.
Results {#sec:results}
=======
To investigate the capabilities and limitations of the current BEM implementation in predicting the self-propelled performance and wake structure of a bio-inspired device, the undulatory elliptical fin experiments of Moored *et. al* [@Moored2011b] are modeled. In these experiments a fin having an elliptical planform with an aspect ratio $\AR = 1.6$ and a NACA 0020 cross-section was fabricated. The root chord length of the fin was $c =0.254$ m. The fin had four actuating spars that were embedded into a PVC polymer and were actuated in such a manner as to produce a chordwise traveling wave (see [@Moored2011b] for the experimental details). The ratio of the traveling-wave wavelength to root chord in the experiments was varied from $\lambda/c = 3-12$. In the computations, the two cases of $\lambda/c = 6$ and $12$ were modeled where $\lambda/c = 6$ was previously found to be the most efficient case [@Clark2006; @Moored2012]. The fin had a linearly increasing amplitude from root to tip in the spanwise direction with a peak-to-peak tip amplitude of $A_{pp} = 0.05$ m. The computational model of the undulatory fin can be seen in Figure \[fig:Wake1\]a. The experimental fin was placed into a free-swimming water tank facility where it was attached to a carriage that was supported by low-friction carts on a rail system (Article No.: ME-9454; PASCO scientific). This setup allowed for unconstrained motion in the streamwise direction. The root of the fin was located at the free-surface where an acrylic plate was placed to dampen free-surface waves generated by the unsteady undulations of the fin.
To properly model this experiment several details must be considered. The free-surface and acrylic plate introduce a no-flux boundary condition that is modeled by using a mirror image fin connected at the root section. At the same time, the free-surface can produce waves from the streamwise motion of the plate, however, the wave drag from this motion is neglected in the computations. The acrylic plate though has another drag source, that is, skin friction drag from the production of its boundary layer. In the computations, the plate geometry is modeled such that the boundary layer solver calculates the skin friction drag from the plate as well as the submerged body. The forces acting on the mirror image body and mirror image plate are neglected in the free-swimming calculations. The low-friction carts actually have a finite rolling friction coefficient, which has been measured to be $\mu_r = 0.0065 \pm 0.0002$ [@Mungan2012]. The friction coefficient was found to be relatively independent of the speed, that is, it changed by less than 10% for speeds between $0$–$2$ m/s. This speed range covers the swimming speeds attained during the experiment, so the friction force resisting the motion, $F_{fr} = \mu_r W$, where $W$ is the net weight of carriage fin system. This rolling frictional force was included in some of the computations. The PVC polymer used to fabricate the fin is less dense than water and produces a buoyant force that counteracts the dry weight of the carriage and fin. The net weight of the whole system when the fin is submerged is estimated at $W \approx 15$ N.
Self-Propelled Performance Metrics
----------------------------------
There are several self-propelled performance metrics that are reported in this study. Time-averaged quantities are marked by an overbar and are always averaged over the last cycle of the computations. The time-averaged thrust coefficient, net thrust coefficient and power coefficient are defined as $$\begin{aligned}
C_T = \frac{\overline{T}}{1/2 \, \rho \overline{U}^2 S}, \quad \quad C_{T,net} = \frac{\overline{T}_{net}}{1/2 \, \rho \overline{U}^2 S}, \quad \quad C_{P} = \frac{\overline{P}}{1/2 \, \rho \overline{U}^3 S},\end{aligned}$$
respectively. The time-averaged swimming speed is $\overline{U}$, the time-averaged power input to the fluid is $\overline{P}$, the planform area is $S$ and the fluid density is $\rho$. The time-averaged thrust force, $\overline{T}$, is the streamwise component of the force from pressure forces alone and the force is considered positive when it is acting on the body in the $-x$ direction. The time-averaged net thrust force, $\overline{T}_{net}$, is the difference between the thrust force and the total drag force, which includes all of the sources of drag that are present in a given computation, that is, the skin friction drag over the fin and the wave-suppression plate, and the rolling friction drag of the rail-carriage system.
In addition, the Strouhal number, reduced frequency, non-dimensional speed, propulsive efficiency and the swimming economy are reported as $$\begin{aligned}
St = \frac{f A}{\overline{U}}, \quad \quad k = \frac{2 \pi f c}{\overline{U}}, \quad \quad U^* = \frac{\overline{U}}{f c}, \quad \quad \eta = \frac{\overline{T U}}{\overline{P}}, \quad \quad \xi = \frac{\overline{U}}{\overline{P}},\end{aligned}$$
respectively, where the frequency of oscillation is $f$, the tip amplitude is $A = A_{pp}/2$ and $\overline{T U}$ is the time-averaged useful power output. The non-dimensional speed, $U^*$, represents the distance traveled in chord lengths over a period of oscillation and is effectively the inverse of the reduced frequency. The propulsive efficiency is the ratio of useful power output to power input, while the swimming economy represents the distance that can be traveled with a unit of energy. Both energetic swimming metrics are reported since the economy is more readily measured in self-propelled swimming experiments, while the efficiency is non-dimensional and easily comparable between different systems.
Discretization Independence
---------------------------
Convergence of the time-averaged swimming speed, $\overline{U}$, and the swimming economy, $\xi$, was tracked as the number of body elements and the number of time steps per oscillation cycle were varied (Figure \[fig:converge\]). The convergence computations are executed for 3 oscillation cycles. On the right axes in Figure \[fig:converge\], good convergence of the swimming speed and swimming economy are shown for both variations in the number of body elements and the number of time steps per oscillation cycle. The left axes show the percent change (%$\Delta$) in either the swimming speed or economy when the number of body elements or time steps per cycle are doubled. It is determined that the free-swimming solution changes by $\mathcal{O}(1\%)$ when $N = 3,200$[^1] is doubled and by $\mathcal{O}(2\%)$ when $N_{step} = 80$ is doubled. These values for $N$ and $N_{step}$ are fixed for all of the following results.
![Right axes: Convergence of the time-averaged swimming speed and the swimming economy for varying numbers of the body elements and time steps per oscillation cycle. Left axes: percent change (%$\Delta$) in the swimming speed or economy when the number of body elements or time steps per cycle are doubled. For (a) and (b), the number of time steps per oscillation cycle is fixed at $N_{step} = 50$. For (c) and (d), the number of body elements is fixed at $N = 800$.[]{data-label="fig:converge"}](figure_08.pdf){width="95.00000%"}
Wake Structure and Self-Propelled Performance
---------------------------------------------
The self-propelled performance of an undulating elliptical fin is examined in detail to offer a direct comparison of the self-propelled three-dimensional unsteady boundary element implementation formulated in this study with previously published experimental data. Additionally, the self-propelled performance of other cases that extend beyond the previous experiments are examined to provide novel physical insight into the self-propulsion of three-dimensional bio-inspired fins. The undulatory fin model examined by the computations more generally represents the fin motions observed in stingrays and pectoral fin swimmers [@Rosenberger2001]. The simulation parameters used for the four numerical cases and the previous experiment data [@Moored2011b] are summarized in table \[tab:parameters\]. The first numerical case models the exact experimental conditions and parameters used in [@Moored2011b]. Numerical cases 2–4 examine different parameter combinations of amplitude, $A$, and nondimensional wavelength, $\lambda/c$, under no plate drag and no carriage friction conditions. Plate drag and carriage friction were necessary conditions imposed by the experimental apparatus, however, these conditions are not present in a self-propelled bio-inspired device. By eliminating these conditions in the numerical simulations, the actual performance of a self-propelled fin can be directly examined. All of the self-propelled simulations have reached a steady-state condition, that is, the number of oscillation cycles was increased until the cycle-averaged net thrust coefficient was $C_{T,net} = \mathcal{O}(10^{-4})$.
<span style="font-variant:small-caps;">Plate Drag</span> <span style="font-variant:small-caps;">Carriage Friction</span> $A$ (m) $\lambda/c$
------------------ -- ---------------------------------------------------------- ----------------------------------------------------------------- --------- -------------
Experiments Yes Yes 0.025 6
Numerical Case 1 Yes Yes 0.025 6
Numerical Case 2 No No 0.025 6
Numerical Case 3 No No 0.025 12
Numerical Case 4 No No 0.05 6
: Simulation parameters used in the present study and experimental parameters used in [@Moored2011b].[]{data-label="tab:parameters"}
![ (a) Boundary element discretization of the wake. The fin surface colormap represents the pressure coefficient. (b) Vortex wake identified by isosurfaces of the $\lambda_2$ criteria. The fin surface colormap represents the skin friction coefficient. (c)–(d) Isosurfaces of the time-averaged $x$-component of the velocity. The pink, red and purple isosurfaces are 5%, 7% and 10% above the mean swimming speed.[]{data-label="fig:Wake1"}](figure_09.pdf){width="99.00000%"}
Figure \[fig:Wake1\] presents the wake structure of the ellipitcal bio-inspired fin for numerical case 1 (see Table \[tab:parameters\]). Figure \[fig:Wake1\]a shows the wake doublet elements rolling up into coherent vortex structures, which is made evident with the $\lambda_2$ isosurfaces highlighted in figure \[fig:Wake1\]b. It is clear that the elliptical fin is shedding a series of interlocked vortex rings through which fluid is accelerated. In fact, the time-average of the streamwise velocity is shown for isosurfaces of 5%, 7% and 10% above the free-stream velocity for the pink, red and purple surfaces, respectively (figure \[fig:Wake1\]c and \[fig:Wake1\]d). Note that the isosurface calculation begins at nearly the half-chord of the fin. The time-averaged jet of fluid accelerated by the vortex rings is observed to have its peak momentum flux at 76% of the span (Figure \[fig:Wake1\]d). This is in good agreement with previous experiments [@Dewey2011] that found the peak momentum flux to occur at about 80% of the span. The surface of the fin is colored with the pressure coefficient in figure \[fig:Wake1\]a, which is defined as $C_p = P/\, (1/2\, \rho \overline{U}^2$). The pressure distribution shows a clear leading-edge suction signature occurring near the fin tip that is coincident with the region of the highest momentum flux generated by the fin. The shear stress distribution computed from the boundary layer solver is shown on the surface of the fin in figure \[fig:Wake1\]b and is nondimensionalized as a skin friction coefficient, that is $C_f = \tau/ \, (1/2\, \rho \overline{U}^2 )$. High shear stress can be seen near the leading edge where the flow is accelerated, but the boundary layer remains laminar. The shear stress rises to high values near the quarter-chord of the fin due to a transition in the boundary layer to turbulence. This indicates that a majority of the skin friction acting on the fin is coming from a turbulent boundary layer. All of the other cases show the same salient features of the wake as well as the pressure and shear stress distributions.
![Time-averaged $x$-velocity (left) and $z$-velocity (right) in an $x$-$z$ plane at 76% of the span for the four numerical cases.[]{data-label="fig:Wake2"}](figure_10.pdf){width="99.00000%"}
Although all of the cases produce time-averaged momentum jets, the structure of the momentum jets are quite different. Figure \[fig:Wake2\] presents the time-averaged streamwise and cross-stream velocities in the $x$-$z$ plane that cuts through the core of momentum jets at 76% of the span for all four numerical cases. The mean streamwise velocity for cases 1–3 show bifurcating jets where the jet splits into two branches. Bifurcating jets have been implicated in poor efficiency performance [@Dong2005; @Dewey2011]. In fact, the time-averaged lateral velocity (figure \[fig:Wake2\] right) highlights that there is excess lateral momentum such that some of the momentum added to the fluid does not perform useful work. In contrast, case 4 shows a single, non-bifurcating momentum jet and minute amounts of excess lateral momentum. By considering the magnitude of the lateral momentum relative to the streamwise momentum, the efficiency of each case could be predicted to be ordered from the least efficient to the most efficient as case 1, 3, 2 and 4, respectively. Indeed, as can be observed later in figure \[fig:xi\_np\] this is exactly the ordering of the cases based on their efficiency. It is further observed that case 4 has twice the amplitude of motion of the other cases, which leads to the suppression of the jet bifurcation and in turn high efficiency locomotion.
Figure \[fig:U\_St\_exp\]a presents the self-propelled swimming speed of the elliptical fin for a range of frequencies of $f = 1.2-2$ Hz. The experimental data [@Moored2011b] shows that the swimming speed of the undulating fin increases nearly linearly with an increasing frequency of motion. Similarly, the numerical simulations show a nearly linearly increasing swimming speed with increasing frequency, however, with a slightly lower slope. Regardless of this discrepancy, there is still excellent agreement between the numerical simulations and the experimental data further validating the self-propelled numerical implementation presented in this study. It is expected that viscous effects in the experiments such as small regions of separated flow near the trailing-edge, the three-dimensional boundary layers, or boundary layer thinning from the unsteady motions [@Ehrenstein2013; @Ehrenstein] can account for the differences among these data.
![(a) Self-propelled swimming speed as a function of frequency. (b) Strouhal number as a function of self-propelled swimming speed. All subsequent experimental and numerical cases marker and line styles follow the legend in (a).[]{data-label="fig:U_St_exp"}](figure_11.pdf){width="99.00000%"}
As expected, when the experimental conditions of plate drag and carriage friction are eliminated (cases 2–4) the undulating fin is observed to have a higher self-propelled swimming speed than case 1 at the same frequency of motion (Figure \[fig:U\_St\_exp\]a). When the nondimensional wavelength of motion is doubled from $\lambda/c = 6$ to $\lambda/c = 12$ there is a slight increase in the slope of the swimming speed curve, but otherwise a negligible effect on the swimming speed. However, when the amplitude of motion is doubled from $A = 0.025$ m to $A = 0.05$ m, the swimming speed is nearly doubled for all frequencies examined (percent increase of 93–97%).
Figure \[fig:U\_St\_exp\]b presents the Strouhal number as a function of the swimming speed. Again, the numerical data (case 1) is observed to show excellent agreement with the experimentally measured Strouhal number and both curves show decreasing $St$ with an increase in the frequency of motion and in turn the swimming speed. Cases 2–4 exhibit a nearly constant $St$, however, they show a slightly decreasing trend with an increase in the swimming speed. In fact for cases 2–4 the swimming speed does indeed increase proportionally with the frequency leading to a nearly constant $St$. In contrast, the additional friction from the carriage and additional drag from the plate modify the swimming speed to be a nonlinear function of the frequency that is not obvious in figure \[fig:U\_St\_exp\]a, but is made evident by the nonlinear relationship between the $St$ and the swimming speed.
![(a) Swimming economy as a function of the self-propelled swimming speed. (b) Propulsive efficiency as a function of the nondimensional swimming speed.[]{data-label="fig:xi_np"}](figure_12.pdf){width="99.00000%"}
Figure \[fig:xi\_np\]a presents the swimming economy as a function of the swimming speed. For all cases, as the swimming speed increases there is an inverse relationship with the economy, that is, when swimming faster the distance that can be travelled for a unit of energy is smaller. This trend has been observed in the self-propelled swimming of manta rays [@Fish2016] and in the self-propelled swimming of heaving flexible panels [@Quinn2014b]. However, this trend contradicts the measurements of swimming economy observed in [@Moored2011b] and indicates that there was likely an error in those power measurements given the excellent agreement in swimming speed between the numerics and experiments. Intuitively, one would expect the economy to, in general, have an inverse trend with swimming speed as observed in the current study since the reverse would imply that it would take little energy to swim a given distance for the fastest swimming speeds, which seems unlikely. To support this intuition, when the economy relationship is examined, $$\begin{aligned}
\xi = \frac{\overline{U}}{\overline{P}} = \frac{1}{1/2 \, \rho \overline{U}^2 S} \left(\frac{1}{C_P}\right),\end{aligned}$$ it can be observed that the economy has an inverse relationship with the swimming speed and the power coefficient, which is also a function of the swimming speed. Furthermore, experimental power measurements have been reported in [@Clark2006] for the same undulating fin as in [@Moored2011b], except that the fin was held in a water channel at a fixed velocity instead of operating in a self-propelled swimming state. If these power measurements are comparable, then [@Clark2006] indicates a 5 fold drop in power when the Strouhal number drops from $St = 0.225$ to $St = 0.1$, precisely the $St$ range observed in [@Moored2011b] that in turn corresponds to the speed range of $0.134 \leq \overline{U} \leq 0.503$. Now, the highest speed, $\overline{U}_H$, in [@Moored2011b] relates to the lowest speed, $\overline{U}_L$, as $\overline{U}_H = 3.75\, \overline{U}_L$ and the subsequent power coefficients are related by $C_{P,H} = 1/5\, C_{P,L}$. By forming a ratio of the economy at the highest swimming speed with the economy at the lowest swimming speed the trend in the economy can be estimated as, $$\begin{aligned}
\mathcal{R} = \frac{\xi_H}{\xi_L} = \frac{\overline{U}_L^2 C_{P,L}}{\overline{U}_H^2 C_{P,H}} = \frac{5}{3.75^2} = 0.3556\end{aligned}$$
When the ratio $\mathcal{R} < 1$ the economy is estimated as having an inverse trend with swimming speed and when the $\mathcal{R} > 1$ *vice versa*. The ratio based on the previous power measurements from [@Clark2006] is $\mathcal{R} = 0.3556$ estimating that the economy should drop by nearly one-third in the experiments over the tested speed range. This estimate further supports the findings of the current numerical study and the idea that the previous experiments have an error in their power measurements. In fact, the ratio of economies using the numerical data (case 1) from figure \[fig:xi\_np\] is $\mathcal{R} = 0.4725$, which is in good agreement with the expected ratio of economies from the experiments.
Figure \[fig:xi\_np\]b presents the propulsive efficiency as a function of the non-dimensional swimming speed. Even though the economy is low for the high swimming speeds and for the high amplitude case (case 4), the efficiency is the highest for these cases with a peak efficiency for case 4 of $\eta \approx 78\%$. Also, the peak efficiency for case 1 is $\eta \approx 50\%$, which is in good agreement with the experimentally measured peak efficiency being $\eta \approx 55\%$ [@Clark2006]. Interestingly, if the experiments were fully self-propelled without the additional plate drag and carriage friction the numerical results indicate that a peak efficiency of $\eta \approx 61\%$ could be attained. Now, it is also clear that even though the higher nondimensional wavelength ($\lambda/c = 12$, case 3) did not produce significantly different swimming speeds than case 2, it did reduce the propulsive efficiency to a peak value of $\eta \approx 54\%$.
![(a) Thrust coefficient as a function of the nondimensional swimming speed. (b) Power coefficient as a function of the nondimensional swimming speed.[]{data-label="fig:Ct_Cp"}](figure_13.pdf){width="99.00000%"}
The trends in propulsive efficiency can be further decomposed by examining the thrust (figure \[fig:Ct\_Cp\]a) and power coefficients (figure \[fig:Ct\_Cp\]b) since $\eta = C_T/C_P$. It is observed that the thrust coefficient for $\lambda/c = 12$ (case 3) is actually higher than $\lambda/c = 6$ (case 2) indicating that the swimming speed is not linearly proportional to changes in $C_T$, but instead is proportional to the frequency and amplitude of motion as proposed in [@Bainbridge1957]. It is clear though that the power coefficient rise is greater than the thrust coefficient rise leading to a lower efficiency. Additionally, the higher amplitude motion (case 4) is observed to slightly reduce the thrust coefficient as compared to case 2. At the same time the power is also significantly reduced leading to higher efficiency. Case 2 shows both a higher thrust coefficient and a lower power coefficient than case 1 leading to an increase the propulsive efficiency.
Conclusions
===========
A novel boundary element method implementation is developed and presented to examine problems of biological and bio-inspired self-propelled locomotion. The formulation uniquely combines an unsteady three-dimensional boundary element fluid solver, a boundary layer solver and an equation of motion solver. The method is validated through a series of analytical, numerical and experimental data that include comparisons with two-dimensional and three-dimensional steady flow solutions, two-dimensional and three-dimensional unsteady flow solutions, viscous drag measurements, self-propelled biological field measurements and self-propelled bio-inspired laboratory measurements. The method is then employed to extend previous experimental results on a ray-inspired model of a self-propelled undulating fin. It is discovered that high propulsive efficiencies of 78% can be obtained when large amplitude motion is used. Additionally, these motions are shown to suppress the formation of a time-averaged bifurcating jet and instead form a single core jet with negligible time-averaged lateral velocity. This maximizes the amount of streamwise momentum compared to the amount of wasted lateral momentum and in turn leads to high propulsive efficiency.
Acknowledgments {#acknowledgments .unnumbered}
---------------
This work was supported by the Office of Naval Research under Program Director Dr. Robert Brizzolara, MURI grant number N00014-08-1-0642. I would like to acknowledge the helpful discussions with Dan Quinn and Peter Dewey on the design of the experimental undulating elliptical fin.
[^1]: More specifically, there are 40 chordwise elements for the top surface, 40 chordwise elements for the bottom surface and 40 spanwise elements for the combined body of the fin and its image
| {
"pile_set_name": "ArXiv"
} |
[**Gravitational theories coupled to matter as invariant theories under Kac-Moody algebras[^1]**]{}
Nassiba Tabti[^2]
[*Service de Physique Théorique et Mathématique, Université Libre de Bruxelles,\
Campus Plaine C.P. 231\
Boulevard du Triomphe, B-1050 Bruxelles, Belgium*]{}
[*The International Solvay Institutes,\
Campus Plaine C.P. 231\
Boulevard du Triomphe, B-1050 Bruxelles, Belgium*]{}\
[[email protected]]{}
**Abstract**
Many recent researches indicate that several gravitational $D$-dimensional theories suitably coupled to some matter fields (including in particular pure gravity in $D$ dimensions, the low energy effective actions of the bosonic string and the bosonic sector of M-theory) would be characterized by infinite dimensional Kac-Moody algebras $\mathcal{G}^{++}$ and $\mathcal{G}^{+++}$. The possible existence of these extended symmetries motivates a development of a new description of gravitational theories based on these symmetries. The importance of Kac-Moody algebras and the link between the $\mathcal{G}^{+++}$-invariant theories and the uncompactified space-time covariant theories are discussed.
Introduction
============
The best candidate for the description and unification of all fundamental interactions is M-Theory in eleven dimensions.Very little is known for sure about M-Theory but it is thought to encompass all superstring theories and in particular, in low energy limit, the eleven dimensional supergravity. Essential elements are lacking in the quest for a unified theory of quantized gravity and matter. In this context, the study of hidden symmetries, exhibited by dimensional reduction, would allow a better understanding of the structure of the unified theory and could lead to a new formulation of gravitational interactions. These hopes have been encouraged by some developments from recent years that certain types of Kac-Moody algebras occur in several $D$-dimensional theories of gravity suitably coupled to dilatons and matter fields associated to $n$-forms, whose lagrangian is: $$\begin{aligned}
\label{actionmax}
\mathcal{L}_{D}= \sqrt{-g}\,\bigg( R -
\frac{1}{2}\sum_{u=1}^{q}\partial_{M}\Phi^{u}
\partial^{M}\Phi^{u} -\sum_{n}\, \frac{1}{2n!}\, e^{\sum_{u}a_{n}^{u}\Phi^{u}}
F^{2}_{(n)} \bigg) \, .\end{aligned}$$ The possible existence of these extended symmetries motivates a development of an original formulation of gravitational theories based on these Kac-Moody algebras.
Dimensional reduction and coset symmetries
==========================================
In this section, we will explain how Kac-Moody hidden symmetries are exhibited. Performing first a compactification on a torus $T^{D-3}$ of a $D$-dimensional gravitational theory given by Eq.(\[actionmax\]), we find in $3$ dimensions a lagrangian containing only scalars coupled to gravity: $$\begin{aligned}
\label{lagred}
\mathcal{L}_{3}=
\sqrt{-g} \, \bigg(R - \frac{1}{2}\,
\partial_{\mu}\bar{\varphi} \cdot \partial^{\mu}\bar{\varphi} -
\frac{1}{2}\,\sum_{\bar{\alpha}}\, e^{\sqrt{2}\,\bar{\alpha}\cdot
\bar{\varphi}}\, \partial_{\mu} \chi_{\bar{\alpha}} \,
\partial^{\mu} \chi_{\bar{\alpha}} \bigg)\, .\end{aligned}$$ One expects that the symmetry of this reduced lagrangian will be $GL(D-3,\mathbb{R})$ which is the symmetry group on the $D-3$ torus. But for some very specific theories, this symmetry is much larger. In fact, under certain conditions[^3], the scalar part of the reduced lagrangian $\mathcal{L}_{3}$ can be identified to a coset lagrangian $\mathcal{L}_{\g/\mathcal{K}}$ invariant under transformations $\g/\mathcal{K}$ where $\g $ is a simple Lie group and $\mathcal{K}$ is the maximal compact subgroup of $\g$ [@Cremmer1979up; @Marcus1983hb].\
For instance, upon dimensional reduction down to $3$ dimensions, the bosonic part of the $11$-dimensional supergravity whose lagrangian is written as $ \mathcal{L}_{11} = \sqrt{-g}\, ( R- \frac{1}{2.4!} \
F_{\mu\nu\rho\sigma}F^{\mu \nu \rho \sigma}+ \ C.S. \ )$ exhibits the simple Lie group $\g=$ $E_8$. The Dynkin diagram of $E_8$ (see Fig.1) characterizes completely this simple Lie algebra. The vertices on the horizontal line define the gravity line. They represent simple roots related to fields coming from the metric $g_{\mu \nu}$. The vertex not belonging to the gravity line is related to a field resulting from the $4$-form $F_{\mu \nu \rho \sigma}$.\
It is interesting to consider the dimensional reduction beyond $3$ dimensions. That would give Dynkin diagrams of Kac-Moody algebras. Roughly speaking, Kac-Moody algebras are infinite dimensional generalization of finite dimensional Lie algebras $\g$. It has been showed that the reduced theory to $2$ dimensions are connected to a infinite dimensional symmetry $\gup$ ( affine extension of $\g$) [@Nicolai1987vy] obtained by adding one root to the Dynkin diagram of $\g$. Motivated by the dimensional reduction, it has been argued that the Kac-Moody algebra $\gdp$ (overextension of $\g$) can play a role in the compactification to $1$ dimension [@Julia1982gx]. Finally, when all the dimensions are compactified, it is obvious to extend the algebra $\gdp$ to $\G$ (triple extension of $\g$) by adding a third vertex. Such $\G$ symmetries were first conjectured in the aforementioned cases [@West2001as; @Lambert2001gk] and the extension to all $\G$ was proposed in [@Englert2003zs]. So this construction motivates the fact that the eleven dimensional supergravity could have the symmetry $\G=E_8^{+++}$ (see Fig.1).\
Similarly other well known theories would also have symmetries under Kac-Moody algebras. Indeed the pure gravity in $D$ dimensions and the effective action of the bosonic string in $26$ dimensions would possess respectively a very-extended symmetry $A_{D-3}^{+++}$ and $D_{24}^{+++}$. More generally all simple maximally non-compact Lie groups $\g$ could be generated from the reduction down to $3$ dimensions of suitably chosen actions [@Cremmer1999du] and it was conjectured that these actions possess the very-extended Kac-Moody symmetries $\G$ [@Englert2003zs]. $\G$ algebras are defined by the Dynkin diagrams depicted in Fig.1, obtained from those of $\g$ by adding three nodes.
$\G$- invariant action
======================
To make the Kac-Moody symmetries manifest, it is interesting to construct an action explicitly invariant under $\G$: $\mathcal{S}_{\G}$ [@Englert2003py]. This action is defined in a reparametrisation invariant way on a world-line, a priori unrelated to space-time, in terms of an infinity of fields $\varphi(\xi)$ where $\xi$ spans the world-line. The fields $\varphi(\xi)$ live in a coset space $\G/\mathcal{K}^{+++}$ where the sub-algebra $\mathcal{K}^{+++}$ is invariant under a ‘temporal involution’ $\Omega$ which ensures that the action is $SO(1,D-1)$ invariant. We can write an element of the coset by exponentiating the Borel sub-algebra of $\G$ (which contains the Cartan and the positive root generators) as: $$\begin{aligned}
\label{corep}
\mathcal{V} (\xi)= exp^{\mathcal{B}^{a} \phi_a(\xi) }\, .
\end{aligned}$$ To each Borel generator $\mathcal{B}^a$, we associate a field $\phi_a(\xi)$. As there is an infinity of generators $\mathcal{B}^a$, there is an infinite number of fields $\phi_a(\xi)$. So we have to organize this summation on the infinity of fields in such a way that a recursive approach has a sense. In this context, we introduce a level decomposition with respect to the finite sub-algebra $A_{D-1}$ where $D$ is interpreted as the space-time dimension. Each $\G$ contains indeed a sub-algebra $GL(D)$ such that $SL(D)(=A_{D-1}) \subset GL(D) \subset
\G$.\
In the particular case of $E_8^{+++}$, the level $l$ counts the number of times the simple root $\alpha_{11}$ (not contained in the gravity line) appears in irreducible representation of $A_{10}$ [@bi10]: $ \alpha = l \alpha_{11} + \sum_{i=1}^{10} a_i \alpha_i \, .$ Using this mechanism, we find a nice decomposition of the infinite number of generators. At each level we have a finite number of generators and fields associated to it. Exploring the decomposition of $E_8^{+++}$ into representations of $A_{10}$, the positive root generators at level 0, 1, 2 and 3 are respectively, $K^a_{\ b}$ ($GL(11)$ generators), $R^{a_1a_2a_3} $ , $R^{a_1a_2a_3a_4a_5a_6}$ and $R^{a_1a_2a_3a_4a_5a_6a_7a_8,b}$. The level $1$ and $2$ tensors are antisymmetric and the level $3$ tensor is associated to a mixed Young tableau (with the constraint $R^{[a_1a_2a_3a_4a_5a_6a_7a_8,b]}=0$). The fields associated to these generators are respectively $h_a^b(\xi)$, $A_{a_1a_2a_3}(\xi)$, $A_{a_1a_2a_3a_4a_5a_6}(\xi)$ and $A_{a_1a_2a_3a_4a_5a_6a_7a_8,b}(\xi)$. At higher levels ($> 3$), there is an infinite number of representations characterized by some Young tableaux.\
With this decomposition, it is possible to rewrite the coset representative $\mathcal{V}(\xi)$ in Eq. (\[corep\]) as $$\begin{aligned}
\mathcal{V}(\xi)= exp\bigg(\underbrace{\!\sum_{a\geq b}h_{b}^{\
a}(\xi)K^{b}_{\ a}}_{Level\
0}\!\bigg)exp\bigg(\underbrace{\!\sum A_{a_1\, a_2 \,a_3}(\xi)\, R^{a_1\, a_2 \,a_3}+ \ldots }_{Level\ \geq\,1}\!\bigg)\, ,\end{aligned}$$ where the first exponential contains only level zero operators and the second one the positive root generators of levels strictly greater than zero. Defining $$\begin{aligned}
\frac{dv(\xi)}{d\xi}=\frac{d\mathcal{V}}{d\xi}\mathcal{V}^{-1} \ , \qquad
\frac{d{\tilde{v}}(\xi)}{d\xi}=-\Omega _{1}\, \frac{dv(\xi)}{d\xi}=
\tilde{\mathcal{V}}^{-1}\frac{d\tilde{\mathcal{V}} }{d\xi}\ ,\qquad
\mathcal{P}=\frac{1}{2}\, \bigg (\frac{dv}{d\xi}+\frac{d{\tilde{v}}}{d\xi} \bigg ) \, ,\end{aligned}$$ where $\Omega_1$ is the temporal involution which allows identification of index $1$ to a time coordinate, one obtains in terms of the $\xi$-dependent fields, an action $\mathcal{S}_{\G}$ invariant under global $\G$ transformations, defined on the coset $\g/\mathcal{K}^{+++}$ : $$\begin{aligned}
\label{actionG}
\mathcal{S}_{\mathcal{G}^{+++}}=\int \, d\xi \,
\frac{1}{n(\xi)}\big<\mathcal{P} \mid \mathcal{P}
\big> \, ,\end{aligned}$$ where $n(\xi)$ is an arbitrary lapse function ensuring reparametrisation invariance on the world-line and ’$< , >$’ is the $\G$ invariant bilinear form ensuring invariance of the action (\[actionG\]) under the global transformations $\G$.\
The fundamental question is now : is there a link between the action invariant under the Kac-Moody algebras and the action of the space-time theory? In particular, considering our example, we would like to relate the action invariant under $E_8^{+++}$ and the $11$- dimensional supergravity. To do that, we must interpret the parameter $\xi$ which the fields of the $\sigma$-model depend on and find the significance of the infinity of fields $\phi_a(\xi)$. In this context, we are going to study the Kac-Moody algebra $\gdp$.
$\gdp$- invariant actions
=========================
To make connections between this new formalism and the covariant space-time theories, it is interesting to analyze the several actions invariant under overextended Kac-Moody algebra $\gdp$ (double extension of $\g$). The Dynkin diagram of $\gdp$ is obtained by deleting the root $\alpha_1$ in the Dynkin diagram of $\G$. The $\G$-invariant action leads to two distinct theories: $\gdp_{C}$ and $\gdp_{B}$ obtained both by a truncation of a infinity of $\G$-fields, putting to zero all the fields multiplying generators involving the deleted root $\alpha_1$. This truncation is realized consistently with all equations of motion, i.e. it implies that all the solutions of the equations of motions of $\mathcal{S}_{\gdp}$ are also solutions of the equations of motion of $\mathcal{S}_{\G}$ [@Englert2004ph].
$\gdp_{C}-$ invariant action
----------------------------
The recent study of the properties of cosmological solutions in the vicinity of a space-like singularity revealed an overextended symmetry $\gdp_C$. The action $\mathcal{S}_{\gdp_C}$ restricted to a defined number of lowest levels is equal to the corresponding space-time theory in which the fields depend only on the time coordinate (see Table 1) [@Damour2002cu]. The parameter $\xi$ is defined as a time coordinate and this $\gdp_C$-theory carries a Euclidean signature in $D-1$ dimensions.
$\gdp_{B}$-invariant action
---------------------------
The action $\mathcal{S}_{\G}$ contains another $\gdp$-invariant action $\mathcal{S}_{\gdp_{B}}$, obtained by performing the same consistent truncation as the one for $\gdp_{C}$, but now performed after a $\G$ Weyl transformation. As the Weyl transformations modify the identification of the indices $1$ and $2$ which become respectively space and time, the resulting action $\mathcal{S}_{\gdp_B}$ is different from $\mathcal{S}_{\gdp_C}$. This gives an action $\mathcal{S}_{\gdp_B}$ with a Lorentz signature for the metric and with the parameter $\xi$ identified to the missing space coordinate instead of $t$. This $\gdp_B$-theory admits exact solutions identical to the ones of covariant Einstein and field equations describing intersecting extremal brane solutions smeared in all direction but one (see Table 2) [@Englert2003py; @Englert2004it].
Moreover, intersections rules are neatly encoded in $\gdp_{B}$ by orthogonality conditions on the positive real roots characterizing each branes.
Weyl transformations and their consequences
===========================================
In this section, we review some consequences of Weyl transformations. First, a Weyl transformation on a generator $T$ of $\G$ can be expressed as a conjugaison by a group element $U_W$ of $\G$: $T \longrightarrow U_W T U_W^{-1}$. Because of the non-commutativity of Weyl transformation with the temporal involution $\Omega$: $$\begin{aligned}
U_W \Omega T U_W^{-1}= \Omega' U_W T U_W^{-1} \, ,\end{aligned}$$ different Lorentz signatures $(t,s)$ (where $ t \,(s)$ is the number of time (space) coordinates) can be obtained [@Englert2004ph]. More precisely, Weyl reflections of the gravity line do not change the global Lorentz signature $(t,s)$ but it change only the identification of time coordinates. In fact, only Weyl reflections with respect to roots not belonging to the gravity line can change the global signature of the theory.\
For instance the signatures found for $E_8^{+++}$ are $(1,10)$, $(2,9)$, $(5,6)$, $(6,5)$ and $(9,2)$ [@Keurentjes2004bv]. These signatures match perfectly with the signatures changing dualities. Indeed, we can interpret the Weyl transformation with respect to root $\alpha_{11}$ (not belonging to gravity line) as a double T-duality in the direction $9$ and $10$ with an exchange of these directions [@Obers1998rn; @Englert2003zs]. Moreover, these signatures match also with the exotic phases of M-theories (M’ and M\*) [@Hull1998ym].\
The previous construction has been generalized and the signatures for all $\G$ have been found [@deBuyl2005it; @Keurentjes2005jw].
Conclusions and perspectives
============================
Many properties of space time are neatly encoded in Kac-Moody algebras (branes, intersection rules, T-duality, …). So the study of $\gdp$ and $\G$ constitutes an interesting approach to understand the gravitational theories coupled to matter which is conceptually different from the Einstein approach. We have now to answer to the fundamental question of all this approach: are these symmetries only a consequence of the compactification process or are they effectively symmetries of the uncompactified theory? This question is related to the significance of the infinite number of fields. Progress is made in this direction [@prog].
Acknowledgments {#acknowledgments .unnumbered}
===============
I cordially thank Laurent Houart for his encouragement and his stimulation. I also thank Sophie and Pierre de Buyl, Nathan Goldman for their technical support and Vincent Wens for his reading of the paper.
[\[00\]]{}
E. Cremmer and B. Julia, Nucl. Phys. **B159**, 141 (1979). N. Marcus ans J. H. Schawrz, Nucl. Phys. **B228**, 145 (1983). H. Nicolai, lectures given at Cargese Summer School on Particle Physics, Cargese, France, Aug 3-21, 1987. B. Julia, invited talk given at AMS-SIAM Summer Seminar on Applications of Group Theory in Physics and Mathematical Physics, Chicago, Jul 6-16, 1982. P. West, Class. Quant. Grav. **18**, 4443 (2001). N. D. Lambert and P. C. West, Nucl. Phys. **B615**, 117 (2001). F. Englert, L. Houart, A. Taormina and P. West, JHEP **09**, 020 (2003). E. Cremmer, B. Julia, H. Lu and C. N. Pope, (1999). F. Englert and L. Houart, JHEP **01**, 002 (2004). H. Nicolai and T. Fischbacher, (2003). F. Englert, M. Henneaux and L. Houart, JHEP **02**, 070 (2005). T. Damour, M. Henneaux and H. Nicolai, Phys. Rev. Lett. **89**, 221601 (2002). F. Englert and L. Houart, JHEP **05**, 059 (2004). A. Keurentjes, Nucl. Phys. **B697**, 302 (2004). N. A. Obers and B. Pioline, (1998). C. M. Hull, JHEP **11**, 017 (1998). S. de Buyl, L. Houart and N. Tabti , JHEP **06**, 084 (2005). A. Keurentjes, (2005). F. Englert, L. Houart, A. Kleinschmidt, H. Nicolai and N. Tabti, Work in progress.
[^1]: Talk presented at the 2nd RTN Workshop of the RTN project ’Constituents, Fundamental Forces and Symmetries of the Universe’, Napoli, October 9-13, 2006.
[^2]: F.R.I.A Researcher.
[^3]: A necessary condition that $\mathcal{L}_{3}$ is invariant under $\g$ and not only under $GL(D-3, \mathbb{R})$ is that the $\bar{\alpha}$ in Eq.(\[lagred\]) are identified to positive roots of $\g$.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We demonstrate how galaxy morphologies can be represented by weighted sums of ‘eigengalaxies’ and how eigengalaxies can be used in a probabilistic framework to enable principled and simplified approaches in a variety of applications. Eigengalaxies can be derived from a Principal Component Analysis (PCA) of sets of single- or multi-band images. They encode the image space equivalent of basis vectors that can be combined to describe the structural properties of large samples of galaxies in a massively reduced manner. As an illustration, we show how a sample of 10,243 galaxies in the [*Hubble Space Telescope*]{} CANDELS survey, which have visual classifications from Galaxy Zoo, can be represented by just 12 eigengalaxies whilst retaining 96% explained variance. We show how the emphasis of certain eigengalaxy components correspond to visual features (e.g. disc, point source, etc.), and explore the correspondence between combinations of eigengalaxies and features defined in the Galaxy Zoo-CANDELS catalogue. We also describe a probabilistic extension to PCA (PPCA) which enables the eigengalaxy framework to assign probabilities to galaxies and characterise a whole population as a generative distribution. We present four practical applications of the probabilistic eigengalaxy framework that are particularly relevant for the next generation of large imaging surveys: we (i) show how low probability galaxies make for natural candidates for outlier detection (ii) demonstrate how missing data can be predicted (iii) show how a similarity search can be performed on exemplars (iv) demonstrate how unsupervised clustering of objects can be implemented.'
author:
- |
Emir Uzeirbegovic,[^1] James E. Geach and Sugata Kaviraj\
\
Centre for Astrophysics Research, School of Physics, Astronomy & Mathematics, University of Hertfordshire, Hatfield, AL10 9AB\
Centre of Data Innovation Research, School of Physics, Astronomy & Mathematics, University of Hertfordshire, Hatfield, AL10 9AB
bibliography:
- 'ref.bib'
title: 'Eigengalaxies: describing galaxy morphology using principal components in image space'
---
=1
\[firstpage\]
methods: data analysis, methods: statistical, galaxies: structure, techniques: image processing
Introduction
============
The distribution of light in galaxies, commonly referred to as galaxy ‘morphology’, is a fundamental observable property. Morphology strongly correlates with the physical properties of a galaxy, such as its stellar mass [e.g. @Bundy2005], star formation rate [e.g. @Bluck2014; @Smethurst2015; @Willett2015], surface brightness [e.g. @Martin2019], rest frame colour [e.g. @Strateva2001; @Skibba2009; @Bamford2009] and local environment [e.g. @Dressler1997; @Postman2005] and reveals key information about the processes that have shaped its evolution over cosmic time [e.g. @Martin2018b; @Jackson2020]. For example, the smooth light distributions of elliptical galaxies, which are a result of the largely random orbits of their stars [e.g. @Cappellari2011], are signposts of a merger-rich evolutionary history [e.g. @Conselice2006]. On the other hand, the presence of a disc indicates a relatively quiescent formation history, in which the galaxy has grown primarily through accretion of gas from the cosmic web [@Codis2012; @Martin2018c]. In a similar vein, morphological details such as extended tidal features suggest recent mergers and/or interactions [e.g. @Kaviraj2014b; @Kaviraj2019; @Jackson2019], with the surface brightness of these tidal features typically scaling with the mass ratios of the mergers in question [e.g. @Peirani2010; @Kaviraj2010].
Apart from being a fundamental component of galaxy evolution studies, morphological information has a wide range of applications across astrophysical science. For example, it can be a key prior in photometric redshift pipelines [e.g. @Soo2018; @Menou2018] which underpin much of observational cosmology and weak lensing studies, is used as contextual data in the classification of transient light curves [e.g. @Djorgovski2012; @Wollaeger2018] and is an essential ingredient in the study of the processes that drive active galactic nuclei [e.g. @Schawinski2014; @Kaviraj2015]. The morphological analysis of galaxy populations, especially in the large surveys that underpin our statistical understanding of galaxy evolution, is therefore of fundamental importance.
A vast literature exists on methods for measuring galaxy morphology. Popular techniques range from those that describe a galaxy’s light distribution using a small number of parameters [e.g. @deVaucouleurs1948; @Sersic1963; @Simard2002; @Odewahn2002; @Lackner2012] to non-parametric approaches such as ‘CAS’ [e.g. @Abraham1994; @Conselice2003; @Menanteau2006] or Gini-$M_{20}$ [e.g. @Lotz2004; @Scarlata2007; @Peth2016], where the light distribution is reduced to a single value. The convergence of large observational surveys and rapidly increasing computing power has recently brought machine learning to the fore in morphological studies. While the use of machine learning can be traced back at least as far back as the 1990s [e.g. @Lahav1995], there has been a recent explosion of studies that apply such techniques to the exploration of galaxy morphology, particularly in large survey datasets [e.g. @Huertas-Company2015; @Ostrovski2017; @Schawinski2017; @Hocking2018; @Goulding2018; @Cheng2019; @Martin2020].
Automated techniques lend themselves particularly well to the analysis of large surveys, but the most accurate method of morphological classification is arguably visual inspection [e.g. @Kaviraj2007]. Indeed the genesis of this subject can be traced back to the visual ‘tuning-fork’ classifications by @Hubble1926, where galaxies were classified into a so-called sequence of ellipticals and spirals. It is remarkable that this classification system still underpins the broad morphological classes into which galaxies are split in modern studies of galaxy evolution. Although visual inspection of large observational surveys is time-intensive, the advent of massively distributed systems like Galaxy Zoo has revolutionised its use on survey datasets [e.g. @Lintott2011; @Simmons2017; @Willett2017]. Galaxy Zoo has used more than a million citizen-science volunteers to classify large contemporary surveys, like the Sloan Digital Sky Survey (SDSS) and the [*Hubble Space Telescope*]{} ([*HST*]{}) Legacy Surveys and has provided the benchmark against which the accuracy of automated techniques have been routinely compared [e.g. @Huertas-Company2015; @Dieleman2015; @Beck2018; @Walmsley2019; @Ma2019].
Principal component analysis (PCA) is a multivariate statistical technique used to convert a set of variables to a set of uncorrelated variables termed principal components (PCs). The PCs are defined iteratively such that each component accounts for the most possible remaining variance in a dataset and is orthogonal to all prior components. We consider the line of approach inspired by the work of [@sirovich1987low], in which they coined the term ‘eigenfaces’ as a synonym for the eigenvectors (presented as images) that result from a PCA decomposition of a set of 2D images, considered together as the row vectors of a common matrix. Those authors showed that a set of $N\times N$ images described as vectors in $N^2$-space could be projected to a comparatively small $M$-space whilst preserving the majority of the variance. The $M$ basis vectors of the $M$-space became the template eigenfaces, where each face image in the original set could be faithfully reconstructed as a linear combination of the templates.
Similar methods have been applied in astronomy. For example, [@de2004machine] used PCA to project a set of 310 disparate images to a lower rank space to facilitate further classification steps. They referred to the basis eigenvectors as ‘eigengalaxies’ and used their weightings to test multiple machine learning methods against each other for the purpose of galaxy classification. [@li2005empirical] used PCA to decompose stellar spectra from the STELIB spectroscopic stellar library [@le2003stelib] and Two Micron All Sky Survey [2MASS, @skrutskie2006two] near-infrared photometry to derive ‘eigenspectra’ and then used them to fit the observed spectra of a selection of galaxies from the SDSS [@fukugita1996sloan] Data Release 1. @anderson2004fast generated many galaxies from a parametric model, calculated the eigenvectors of the generated set, and then used the reduced eigenspace to find the nearest synthetic model to any given real galaxy image as a way of discovering correspondence between galaxies and known structures. @wild2014new used PCA to concisely describe a large number of model spectral energy distributions (SEDs). They termed the retained eigenvectors as ‘super-colours’ and used them to impute SEDs from sparse samples in observed galaxies.
In this paper, we further consider the utility of PCs in image space, working with multi-band imaging and focusing on galaxy morphology. We start by finding a concise variance-preserving latent space representation using PCA, and as others have done before, refer to the resultant eigenvectors as ‘eigengalaxies’. Taking this concept further, we show how the result can be converted into a generative model which can assign likelihoods to galaxies, and demonstrate how the model enables the simple and principled implementation of a set of cross-cutting applications that may be of use to the community in a variety of scenarios, not necessarily limited to galaxy morphology studies.
The paper is structured as follows. In Section2 we describe both the sample and our methodology and discuss the interpretation of eigengalaxies and its relevance to the interpretation of galaxy morphology. We then show how this result is converted to a generating distribution using probabilistic PCA (PPCA). In Section3 we show how this probabilistic eigengalaxy framework leads to principled approaches in various applications. In particular, we (i) show how low probability galaxies make for natural candidates for outlier detection (ii) show how missing data can be predicted (iii) demonstrate how similarity searches for images can be implemented given an exemplar (iv) demonstrate how a natural unsupervised clustering of objects can be produced. We then outline how the methods may be used together in more advanced applications and how it is relevant to big datasets such as the upcoming Large Synoptic Survey Telescope (LSST) project [@Robertson2017]. Section4 concludes and summarises our findings, and provides a link to the codes developed for this work.
Eigengalaxies
=============
![Contour plot of explained variance corresponding to the required number of eigengalaxies versus sample size. Higher explained variances tend to require proportionally more eigengalaxies as sample size increases, however at roughly 96% explained variance is asymptotic to a constant.[]{data-label="fig:pca-over-samples"}](pca-over-samples)
![Histogram showing the frequency of the ratio of the sum of deltas (difference between original and reconstructed pixel flux over all bands) and the sum of the flux in the original images, for the PCA decomposition of the 10,243 galaxies in the GZ-CANDELS GOODS–S sample. 99% of objects lie between $\pm 2.5\%$. We account for 96% of variance by design and the distribution above shows that the reconstructed images bear a good likeness (in terms of reconstructed flux) to the originals.[]{data-label="fig:flux-diff-hist"}](flux-diff-hist)
Data from [*HST*]{} CANDELS
---------------------------
[*HST*]{} CANDELS [@grogin2011candels; @koekemoer2011candels] offers a high-resolution probe of galaxy evolution. The survey consists of optical and near-infrared (WFC3/UVIS/IR) images from the Wide Field Camera 3 (WFC3) and optical images from the Advanced Camera for Surveys (ACS) in five well-studied extragalactic survey fields. Here, we focus on GOODS-S, one of the deep tier (at least four-orbit effective depth) fields. We select a sample of 10,243 galaxies present in the ‘Galaxy Zoo: CANDELS’ (GZ-CANDELS) GOODS-S catalogue [@simmons2016galaxy], that fall within the region jointly covered by the F814W, F125W and F160W bands. The majority of objects are at $z<3$ [@simmons2016galaxy]. The motivation for using the GZ-CANDELS catalogue is that our results can be interpreted in the context of the morphological and structural classifications provided by the Galaxy Zoo citizen science project.
Object registration and preparation
-----------------------------------
The aim of registration is to remove the uninteresting sources of variation which would otherwise increase the number of eigengalaxies required. We identify pixel scale matching, image centering, cropping, range clipping and image rotation as the most important confounders to account for. For each object in the catalog we take a 1.8$''$ cut-out, using the catalogued sky coordinate as a centroid. ACS images are down-sampled by a factor of two to match the pixel scale of the WFC3 images (0.06$''$ pixel$^{-1}$), so that for each band we obtain a $30\times 30$-pixel image at the position of each galaxy. The cut-out dimensions are selected with prior experimentation because they result in the least superfluous background for most targets.
We perform some simple post-processing to the data. In particular, we truncate the pixel distribution at the 99$^\text{th}$ percentile to avoid anomalous spikes in flux in some galaxies from accounting for too much variance. Next, we produce a temporary composite image of the array by using, for each pixel, its maximum magnitude across all bands. We then use the locations of all pixels in our composite with values exceeding the 75$^\text{th}$ percentile to create a matrix of the coordinates of bright pixels. The first principal component of the two column matrix of pixel coordinates is a vector of weights $(x_0,y_0)$ which can be interpreted as the linear transformation of our 2D coordinates to the 1D space that preserves most variance (i.e. some line passing through the origin of the original 2D space). The angle between the original [*x*]{}-axis in our image and the new variance-preserving one is then given by $\theta =\operatorname{arctan2}(y_0,x_0)$. Finally, we transform the original array by rotating every band in turn by $\theta -\pi/2$ radians (in order to make vertical what would otherwise be a rotation to the horizontal plane) which results in vertically-aligned brightness.
![Cumulative explained variance illustrating how much variance each additional feature accounts for. In PCA successive components account for less variance than the previous component. In this instance, only [*two*]{} eigengalaxies are required to account for $\sim$85% of explained variance yet 12 eigengalaxies are required to account for 96% of explained variance.[]{data-label="fig:cum-var"}](cum-var)
Decomposition
-------------
The set of objects, each a registered $30\times 30\times 3$ array, are flattened into $1\times 2700$ row vectors and concatenated to form a feature matrix. We perform PCA[^2] on the resultant matrix, but prior to that we conduct some experiments to determine a reasonable level of explained variance to target. In theory, in the worst case scenario, as the ratio of explained variance approaches unity, the number of eigengalaxies required may be close to the total sample size. We perform PCA for varying sample sizes and create a contour plot to depict the relationship between sample size and the number of eigengalaxies required for any given level of explained variance (Figure \[fig:pca-over-samples\]). The $n^{\rm th}$ eigenvector is always the same regardless of what explained variance is targeted, because eigenvectors are ordered by explained variance and orthogonal to each other, so targeting a higher explained variance just means retaining more eigenvectors. We aim to target the highest level of explained variance asymptotic to a constant number of eigengalaxies as the sample size increases. This is important because we intend the methods illustrated here to be applicable to very large datasets in which the number of eigengalaxies targeted would have to be on an asymptotic curve. This is discussed further in Section 3. After choosing a target variance, each $1\times 2700$ eigenvector is reshaped into a $30\times 30\times 3$ array: these are the eigengalaxies. Reshaping the previously flattened array is done by filling a $30\times30\times3$ array one band at a time in column order. The matrix of coefficients defining how the eigengalaxies are combined to reconstruct any given galaxy is termed the ‘score matrix’ and is also calculated.
Results and interpretation
--------------------------
![The 12 eigengalaxies accounting for 96% variance in the GZ-CANDELS GOODS-S sample. Each row is as an eigengalaxy and each column is a band. All images are scaled identically, where white indicates relative emphasis and black indicates relative de-emphasis of the region. Each image is $30\times30$ pixels, or $1.8''\times 1.8''$.[]{data-label="fig:components"}](components)
Figure \[fig:pca-over-samples\] presents a contour plot of explained variance corresponding to various combinations of sample size and eigengalaxy number. We chose to retain 96% variance as it is the highest target variance with near zero or asymptotic growth as sample size increases. Only 12 eigengalaxies were required to achieve an explained variance of approximately 96% for 10,243 objects. Let $F_0$ be the collection of flux densities of the pixels in the original images and $F_\Delta$ be the corresponding differences in flux between the original pixels and those in images reconstructed using the eigengalaxies. Figure \[fig:flux-diff-hist\] presents a histogram of (i.e. the ratio of the sum of deltas and the sum of flux over all bands). It shows a symmetrical distribution centred at zero. 99% of objects lie between $< 2.5\%$.
Figure \[fig:cum-var\] shows the cumulative variance explained by successive eigengalaxies. PCA successively maximises the variance in each orthogonal eigengalaxy and each additional eigengalaxy therefore accounts for less variance. In this instance, only [*two*]{} eigengalaxies are required to account for $\sim$85% of explained variance, yet 12 eigengalaxies are required to account for 96% of explained variance. Figure \[fig:components\] presents the 12 eigengalaxies as grayscale images. The effect of image rotation to vertically align brightness can be seen in the relative emphasis of the central vertical strip. One may also wonder how structure with high spatial frequency – such as spirals – would be reconstructed given the features of the eigengalaxies. Finessed features are a function of coordination of adjacent pixel columns and although visually distinct typically account for a tiny amount of flux variance. Therefore, fine-structured spiral features (for example) will become washed out in reconstructions, and not appear as sharp as in the real image.
Examining the first two eigengalaxies we see that they are intensity dominated. It is unsurprising that most variance is to be found in accounting for relative brightness, given the GZ-CANDELS sample spans ([*H*]{}-band) magnitudes down to $\text{F160W} < 28.3$mag [@simmons2016galaxy]. Subsequent eigengalaxies become more complex and structured. How can we interpret them? It is sometimes difficult to ‘eyeball’ the feature implications of any given eigengalaxy but it can be made clearer by contrasting samples which extremely emphasise ($>$$99^\text{th}$ percentile) or de-emphasise ($<$1$^\text{st}$ percentile) the eigengalaxy as shown in Figure \[fig:sample-contrasts\]. In this example, it appears that extreme negative weights on eigengalaxy component 3 corresponds to the presence of extended resolved features such as (but not exclusively) discs; conversely, extreme positive eigengalaxy 3 weights result in point sources with no, or insignificant, resolved structure. Thus, selection on relative component weighting could lead to practical applications such as star/galaxy separation or the identification of quasars.
![The left panel shows a sample of objects where eigengalaxy component 3 has an extreme negative weight, illustrating how different eigengalaxy components can correspond to visual features. In this case, eigengalaxy 3 seems to correspond to resolved disc-like structures, which are completely absent in the images in the right panel, which are point sources. Each 1.8$''$ thumbnail image is an RGB composite of the F160W, F125W, and F814W bands.[]{data-label="fig:sample-contrasts"}](03-lo "fig:"){width="23.50000%"} ![The left panel shows a sample of objects where eigengalaxy component 3 has an extreme negative weight, illustrating how different eigengalaxy components can correspond to visual features. In this case, eigengalaxy 3 seems to correspond to resolved disc-like structures, which are completely absent in the images in the right panel, which are point sources. Each 1.8$''$ thumbnail image is an RGB composite of the F160W, F125W, and F814W bands.[]{data-label="fig:sample-contrasts"}](03-hi "fig:"){width="23.50000%"}
Where individual eigengalaxies do not coincide with a feature of interest, the correspondence can be sought with a combination of eigengalaxies. One way to establish this correspondence is to train a classifier to discover the relationship between the eigengalaxy scores and a set of labels, and then to interpret the classifier to better understand the correspondence. Note that the performance of classifiers for use cases such as these will be substantially dependent on the purity of labelled classes in the training set. We match the number of galaxies in each class to make the [*ex ante*]{} probability of any class roughly equal. We reserve half of the data at random for training, withhold the other half for testing, and fit a random forest [@liaw2002classification] classifier using the GZ-CANDELS classifications as a response and the score matrix as the explanatory variables. The random forest classifier is an ensemble learning technique which aggregates the classifications of a large number of decision trees. Decision trees themselves have low bias but a high variance on out of sample predictions because they tend to over-fit data. The random forest corrects for this by randomly partitioning the data that each tree can see, which has the effect of de-correlating tree classifications. The mean or mode of the resultant tree classifications is then used to produce the final result. Random forests can have a drastically lower variance than that of a single decision tree fitting, whilst retaining much of a decision tree’s interpretability.
Figure \[fig:confusion\] shows the error matrix (a cross tabulation of true class labels versus predicted class labels) for the withheld data. Correspondence to complex GZ-CANDELS rules is good despite some inherent class ambiguity. GZ-CANDELS classifications are based on combinations of cut-off points corresponding to fractions of votes by human classifiers. Classifications are presented as mutually exclusive whereas it is possible for a spiral galaxy to be simultaneously edge-on and/or clumpy for example, leading to classification ambiguity. Figure \[fig:feature-importance\] illustrates the importance weighting for each eigengalaxy in random forest classifications. It shows a particular emphasis on a subset of the eigengalaxies when classifying edge-on, spiral and clumpy galaxies against each other.
The eigengalaxies themselves encode something elementary about the (usually) complex multi-band light distribution in the transformed galaxy sample, and therefore can be used in a sense to ‘define’ morphology in a non-parametric, data-driven way through the score matrix, where the assumptions of PCA are reasonable; that is, that features are orthogonal, with each maximising the remainder of the residual variance, and the dataset is well described by a mean and covariance structure.
![Error matrix depicting the correspondence between true labels and those predicted by a random forest for edge-on (0), spiral (1) and clumpy (2) galaxies. It shows good ($\gtrsim$75%) correspondence to GZ-CANDELS classification rules which are based on Boolean combinations of cut-off conditions for fractions of user votes on different features.[]{data-label="fig:confusion"}](confusion)
![Importance weighting for each eigengalaxy in our random forest classifications. This shows a particular emphasis on eigengalaxies 1, 4 and 12 when classifying edge-on, spiral and clumpy galaxies against each other.[]{data-label="fig:feature-importance"}](feature-importance)
PCA as a generative model
-------------------------
Extending PCA to a probabilistic generative model greatly enhances its capabilities to applications like predicting missing data, generating new data and drawing inferences about the likelihood of data, which PCA alone is not capable of. [@tipping1999probabilistic] show that PCA is equivalent to the maximum likelihood (ML) solution of the factor model, which aims to map an observed vector space to an unobserved latent space by a linear transformation: $t=Wx+\mu+\epsilon$. Here $t\in \mathbb{R}^d$ is the observation vector, $x \in \mathbb{R}^q$ is the latent vector, $W$ is a $d\times q$ loading matrix (the coefficients mapping an observed vector to the latent space) relating $t$ to $x$ where $q<d$, $\mu$ is an offset and $\epsilon$ is a residual error. [@tipping1999probabilistic] show that given $x \sim N(0,I)$, and $\epsilon \sim N(0,\sigma^2)$ (i.e. isotropic error), then $t \sim N(\mu,WW^T+\sigma^2I)$, and the ML solution yields $W$ equivalent to the eigenvectors of the covariance matrix (the PCA decomposition).
Given a PCA decomposition and an ML estimate of $\sigma^2$, the established equivalence means that we can write down a generative factor model. This offers two key opportunities further discussed later: (i) the factor model is Gaussian and we can simultaneously estimate the loading matrix and missing data in an expectation maximisation setting, and thus predict missing values and compute PCA at the same time[^3]. This enables PCA even when data is missing, and it also enables the prediction of missing values; (ii) given a generative factor model fitted within our eigengalaxy framework, we can assign a probability to the generation of any given galaxy $x$ using the Gaussian density $p(x\mid \mu,WW^T+\sigma^2I)$. Sorting galaxies by their likelihood therefore provides a natural measure of ‘rarity’ which may be used to identify outliers in a quantitative manner. In the next section, we explore some practical applications of these opportunities.
Applications
============
In this section we explore some applications of the probabilistic eigengalaxy framework described above.
![25 of the ‘rarest’ galaxies from our sample. Rarity is quantified by the likelihood of the galaxy being generated by the PPCA generating function. The image shows anomalous detections and artefacts along with real, but intrinsically rare types of objects. Each 1.8$''$ thumbnail image is an RGB composite of the F160W, F125W, and F814W bands.[]{data-label="fig:5-most-rare"}](most-rare)
![The frequency of the ratio of the sum of deltas (difference between the original and predicted pixel flux) and the sum of the flux of the original images for galaxies with missing data. The body of the distribution extends to $\pm 20\%$ error but $\sim 72\%$ of objects are predicted within $\pm 10\%$ error, indicating that PPCA can predict missing data in a 2D sense with a high level of fidelity.[]{data-label="fig:recovery-rates"}](recovery-pca)
Outlier detection
-----------------
{width="29.00000%"} {width="29.00000%"} {width="29.00000%"}
A key utility of outlier detection is to make discoverable rare phenomena buried in enormous datasets. This may include searches for exotic galaxies and rare objects but also the identification of anomalous detections and pipeline errors. The two-part challenge is first to define an ‘outlier’ in a way useful to astronomy, and the second is to scale the detection algorithm to the size of the data.
@dutta2007distributed implement a distributed version of PCA using random projection and sampling to approximate eigenvectors for the purpose of outlier detection and apply it to the 2MASS [@skrutskie2006two] and SDSS [@fukugita1996sloan] datasets. In defining an outlier they search for galaxies which are over-represented by the last eigenvector. Other large scale PCA methods include incremental PCA [@ross2008incremental], which approximates PCA by processing data in batches commensurate with the available random access memory. [@baron2017weirdest] utilise a random forest and fit to discriminate between real and synthetic data[^4]. For every pair of objects, they count the number of trees in which each pair is labelled ‘real’ in the same leaf. The output is a $N\times N$ similarity matrix which is then searched for ‘outliers’, defined as objects with a large average distance from all other objects.
We focus on a simple and principled definition for an ‘outlier’, proceeding directly from our eigengalaxy framework. Given that PPCA allows us to cast our eigengalaxy result to a probabilistic model, a natural definition for an outlier is an object with a low probability of being generated. Outliers can thus be interpreted as the objects least representative of the object population described by the mean and covariance structure of the factor model. Note that the approach of calculating PCA for a set of features, converting it to PPCA and then using our outlier definition is general and could be applied not only to imaging, but also to spectroscopy, light curves, catalogues and other kinds of data. We define a formal outlier description within the eigengalaxy framework: given a generative factor model $N(Wx + \mu,\sigma^2I)$, an outlier $x$ is an object such that $p(x\mid Wx + \mu,\sigma^2I)<T$ where $T$ is a threshold probability density, and can be set according to the purpose at hand.
To illustrate the concept with the GZ-CANDELS dataset, we use our derived eigengalaxies to assign a log likelihood to every galaxy, using the [*score\_samples*]{} method on the [sklearn.decomposition.PCA]{} object which implements the [@tipping1999probabilistic] factor model representation to calculate likelihood. We sort the data by likelihood and present in Figure \[fig:5-most-rare\] the 25 objects least likely to be generated. The image shows not only anomalous detections and artefacts but also systems that are known to be rare, such as dust lanes which are signposts of recent minor mergers [see e.g. @Kaviraj2012b], ongoing mergers [see e.g. @Darg2010] and edge-on spirals which appear to be accreting a blue companion. Outlier detection therefore offers an efficient way to identify examples of rare objects in large surveys.
Predicting missing data
-----------------------
In many situations one might be missing a particular band, for example due to bad data, partial coverage with a certain bandpass, obliteration by [*Starlink*]{} satellite trails, etc. In these cases we can consider how well we can predict the missing data using eigengalaxies. [@tipping1999probabilistic] define an expectation maximisation (EM) algorithm for PPCA in the presence of missing data which works by treating missing values as jointly distributed with the latent variables and maximising the expectation of the joint likelihood function. As a demonstration, we randomly omit a band with equal probability from our dataset for 5% of rows chosen at random. We use the data as processed in Section 2.2 and an efficient variational EM version of the [@tipping1999probabilistic] algorithm [^5] set out in [@porta2005active] to fit our PPCA model. We use the delta sum over flux sum ratio as in Figure \[fig:flux-diff-hist\] for the reconstruction error to describe the prediction error. Figure \[fig:recovery-rates\] illustrates the distribution of prediction error. The distribution is roughly symmetrical and is centred on zero. The body of the distribution extends to $\pm 20\%$ error but $\sim 72\%$ of objects are accounted for within $\pm 10\%$ error showing that high fidelity prediction is possible for most objects. Figure \[fig:impute-error\] provides some examples of predicted data for different bands. It is noteworthy that by using information contained in the eigengalaxies, PPCA is successful in estimating total flux even when it bears no resemblance to that of the other bands. The ability to predict missing images offers a route to ‘filling in’ missing data, such as predicting photometric data points which are absent in the observations in order to reconstruct missing parts of a galaxy’s SED.
{width="24.50000%"} {width="24.50000%"} {width="24.50000%"} {width="24.50000%"}
Searching for galaxies similar to an exemplar
---------------------------------------------
Given a survey with a large number of objects with diverse variety, and an interest in galaxies of a specific kind of object, it is useful to be able to present an exemplar galaxy and use it to quickly search for all other galaxies with similar features. The utility and suitability of a similarity search for any particular use case will depend primarily on how the objects are being described, and how the similarity between their descriptions is being calculated. For example, [@protopapas2006finding] use cross-correlation as a proxy for similarity between light curves for the purpose of detecting outliers whereby light curves with the lowest average similarity are defined as outliers. [@sart2010accelerating] utilise dynamic time warps [@berndt1994using] to measure the similarity between light curves. [@Hocking2018] compare various measures, including Euclidean distance and Pearson’s correlation coefficient, and settle on using cosine distance to measure similarity from a description generated by growing neural gas prior to hierarchical clustering.
The eigengalaxy framework projects galaxies to real row vectors in an 12D space. The dimensions are not arbitrary but represent the eigengalaxies which together reconstruct the galaxies whilst preserving 96% of variance. Further, the value of each dimension is easy to interpret: it represents the weights required for each eigengalaxy to reconstruct the original galaxies as a linear combination of eigengalaxies. The closer row-vector weights are to each other, the more similar their constitution in terms of eigengalaxies and thus the more similar their reconstruction. In this case, Euclidean distances are preferable over, say, cosine distance or a correlation measure, because we want to be sensitive primarily to the magnitude of row vector differences. Formally, let $r$ be our exemplar galaxy expressed as a row vector and let $X$ be our dataset (an $N\times 11$ matrix). Let $Y$ represent the column vector of Euclidean distances of length $N$. $Y_n$ is given by:
$$Y_{n}=\left(\sum_{i=1}^{11}{\left(X_{n,i}-r_i\right)^2}\right)^{\sfrac{1}{2}}.$$
Let $I$ be the indices of $Y$. We sort $I$ with respect to the ascending order of $Y$, which gives us all objects in order of similarity to the exemplar. To demonstrate, we pick an exemplar galaxy and implement the procedure above. Figure \[fig:similar\] illustrates four examples of similarity searches, showing the 24 most similar galaxies to the exemplar. The image shows searches for spirals, edge-on spirals, ellipticals and recent mergers from left to right respectively. Most of the features are invariant to arbitrary brightness and the object registration has excluded aspects such as galaxy rotation, rendering this elementary algorithm remarkably effective.
Unsupervised clustering
-----------------------
Unsupervised clustering is particularly useful for massive astronomy datasets because it provides a method of investigating very large collections of objects efficiently. If the clustering method is effective at grouping objects with similar features together, then one can study a dataset by characterising its [*morphological clusters*]{} rather than examining each object separately. There is some precedence for this in astronomy. For example, [@Martin2020] follow [@Hocking2018] in using growing neural gas and hierarchical clustering directly on pixel data to identify structurally distinct clusters. [@almeida2010automatic] and [@almeida2013automated] utilise [*k*]{}-means to classify spectra from SDSS into fewer base types. [@valenzuela2018unsupervised] use [*k*]{}-medoids[^6] to cluster and map sequences of light curve segments to variational trees.
The interpretation of the eigengalaxy framework described in Section 2.4 naturally extends to a simple unsupervised clustering treatment. We can create a distance matrix (an $N \times N$ matrix in which each cell indicates a distance between object $j$ at row $j$ and object $i$ at row $i$) by calculating the Euclidean distance between every galaxy. The distance matrix provides an input for a broad range of unsupervised clustering algorithms. This provides us with an opportunity to define and discover groups of galaxies based on the similarity of their multi-band morphologies as encoded by the eigengalaxies.
To demonstrate this, we produce a distance matrix for the GZ-CANDELS dataset as outlined above. We use affinity propagation [@dueck2009affinity] which produces 536 clusters from the total sample of 10,243 objects, with clusters varying in size from 1 to 316 with a median size of 4. Affinity propagation has a tunable ‘preference’ parameter that allows the coarseness of clustering to be adjusted; here we use the default value (the median of the distance matrix). Figure \[fig:ap\] illustrates samples of galaxies from four morphological clusters. It is noteworthy that affinity propagation is an exemplar-based algorithm, so that each cluster is actually characterised by an exemplar galaxy. Thus, in this instance the whole dataset is summarised by 536 exemplars (5% of the sample) which could, in principle, be sorted by rarity (see Section 3.1) and then viewed as one large image. This massive reduction in the scale of the full dataset highlights the efficiency gains that can be made by using this method for exploring the extremely large imaging surveys of the future.
{width="24.50000%"} {width="24.50000%"} {width="24.50000%"} {width="24.50000%"}
Combining methods
-----------------
All the methods described above utilise the same framework and can be used together to create additional capabilities simply based on the eigengalaxy weights in the score matrix. There are many possible combinations but we highlight three:
- Missing value prediction $\to$ all methods. Predicting missing band values, dead pixels, etc. can be used to ‘complete’ data so that it may be considered on par with the rest of the data. It may then be used with any of the other methods.
- Outlier detection $\to$ similarity search / clustering / classification. We may discover an interesting outlier (e.g. a rare object such as a gravitationally lensed galaxy) of which we would like to find many more examples (similarity search). We may want to cluster the outliers to determine a self-similar morphological clusters. We may also want to label a training set of outliers, train a classifier and then select only the outliers of interest.
- Clustering $\to$ missing value prediction. To make the prediction of missing values more accurate, we may first generate eigengalaxies using the available bands and use it to create a clustering. This produces a set of self-similar classes to which missing values have more in common than the whole population. For each cluster we could then separately derive eigengalaxies and predict missing values for objects within that cluster.
Big data
--------
The GZ-CANDELS dataset is relatively small and can be easily processed. In this section we consider problems and adaptations which may be necessary to use these methods with extremely large datasets, such as LSST [@Robertson2017], [*Euclid*]{} [@Refregier2010] and the Square Kilometre Array [@Weltman2020] which may contain billions of observed objects. We assert that the eigengalaxy framework offers a novel solution to processing such big data by making it possible to find outliers, search for and cluster objects using only the reduced score matrix form of a dataset. For example, if LSST was to provide a corresponding score matrix, eigengalaxies and galaxy likelihoods, all the methods described here would be available based on a small fraction of the full dataset.
An important consideration in making the methods presented here feasible for big data is the number of eigengalaxies required to achieve a practically useful level of explained variance. @aggarwal2001surprising provides a deeper discussion on common issues for algorithms operating in high dimensional space. In our case we would expect the following problems given too many eigengalaxies:
1. The score matrix form of the data may itself be too large to be useful. In our case, as the number of eigengalaxies approaches 10,243, the total size of the data converges onto the total size of the cut-out collection.
2. Euclidean distance loses interpretability and the power to distinguish objects as the number of dimensions grows.
3. Some techniques like fully calculating a distance matrix or calculating the similarity of every object with reference to an exemplar would not be computationally scalable.
Taking LSST as an example, we would expect that the number of eigengalaxies required to explain, say, 96% variance to be more than in our GZ-CANDELS experiments for two primary reasons: (i) it is a deeper survey and it therefore observes more low mass galaxies, the morphological mix of which is yet unknown [e.g. @Martin2019], and (ii) we would expect bright galaxies to have more detail such as extended debris and tidal features [e.g. @Duc2011] and therefore exhibit more variance. However, the additional sources of variance are limited, and we would expect a reasonable asymptotic number of eigengalaxies to emerge for useful level of explained variance at relatively small sample sizes.
The kind of variance that needs to be captured and the sufficient ratio of it to retain ultimately depends on the intended application. If the number of eigengalaxies is too high we could explore various ways to reduce the variance not required for the intended purpose in order to stay within an eigengalaxy quota at a useful explained variance ratio. Potentially applicable techniques include using smaller cut-outs, down-sampling pixels, using fewer bands, combining bands, stretching the range of the flux densities to make differences less subtle, and partitioning the data by some other variable (e.g. brightness, location, etc.) and then conducting the analysis per partition. If the number of eigengalaxies was workable from a data size perspective but a problem for Euclidean distance then we could use other distance measures more robust in high dimensional space. Finally, in every case we would need to replace exhaustive similarity searches with more scalable methods such as the nearest neighbour algorithm [@beis1997shape], and affinity propagation would need to be replaced with a less expensive clustering method such as [*k*]{}-means (or [*k*]{}-medoids if Euclidean distance was unacceptable).
Conclusions and Summary
=======================
We have demonstrated how a large sample of galaxies can be described in terms of a small set of template ‘eigengalaxies’. Eigengalaxies can be considered to be the image-space equivalent of ‘basis vectors’, interpreted as orthogonal features. This offers a completely data-driven approach to characterising galaxy morphology. Using a sample of 10,243 galaxies from the Galaxy Zoo–CANDELS survey, chosen to lie in the deep GOODS–S field, we use three-band [*HST*]{} ACS and WFC3 imaging to derive just 12 eigengalaxies that are sufficient to explain 96% of the morphological variance in the sample. How eigengalaxies correspond to real features in galaxies can be explored by examining samples where particular eigengalaxies are emphasized or de-emphasized. For example, we have shown how the eigengalaxies can be mapped to semantically convenient labels using Galaxy Zoo classifications. We have further shown how eigengalaxies can be embedded into a probabilistic framework and so enable a variety of useful applications. We explored four potential applications of our framework:
1. we have shown how PPCA can be used to assign likelihoods and identify outliers as objects with a low probability of being generated.
2. we have shown how PPCA can be used to predict missing values in imaging, for example due to bad data or partial coverage in a given band.
3. we have shown how the projection of eigengalaxies on to a 12D space naturally facilitates similarity searches, where galaxies can be sorted relative to their Euclidean distance from an exemplar, thus quickly returning samples of galaxies that are morphologically similar to the exemplar as defined by their eigengalaxy components.
4. we have shown how the Euclidean distances between galaxies in the 12D space can be compiled into a distance matrix that can provide the input for unsupervised clustering algorithms to discover groups of similar objects. We have demonstrated this using affinity propagation to show how morphologically similar groupings can be identified in large samples.
Finally, we have described how the methods may be used together in more advanced use cases and how working in eigengalaxy space may present a novel solution to outlier detection, as well as search and clustering problems in forthcoming massive imaging datasets such as LSST. We argue that these results and illustrations underscore the suitability of our PCA based probabilistic eigengalaxy framework for the study of morphology, especially in the era of big data astronomy, where representational efficiency and relevance will pay dividends.
Code for all results and figures can be found at <https://emiruz.com/eigengalaxies>.
Acknowledgements {#acknowledgements .unnumbered}
================
J.E.G. is supported by the Royal Society. S.K acknowledges a Senior Research Fellowship from Worcester College Oxford. This work is based on observations taken by the CANDELS Multi-Cycle Treasury Program with the NASA/ESA HST, which is operated by the Association of Universities for Research in Astronomy, Inc., under NASA contract NAS5-26555. This research made use of Astropy,[^7] a community-developed core Python package for Astronomy [@astropy:2013; @astropy:2018].
[^1]: [email protected]
[^2]: We use the Python package [sklearn.decomposition.pca]{}.
[^3]: Although not explored in this paper, we also note that PPCA makes a ‘mixture of PCAs’ model [@tipping1999mixtures] possible. This enables the discovery of linear sub-spaces and the simultaneous PCA fitting of them, which may be more suited to datasets with which PCA struggles (e.g. data which cannot be well described by a mean and covariance structure). It may also reduce the overall dimensionality requirements for each linear sub-space compared to global PCA.
[^4]: These authors use the flux at each wavelength as a feature set, generating synthetic data by sampling from the marginal distribution of each feature.
[^5]: Available in Python package [pyppca]{}.
[^6]: In [*k*]{}-medoids, data points become cluster centres, unlike [*k*]{}-means where the cluster centre is not necessarily correspondent to a data point.
[^7]: http://www.astropy.org
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'With any integral lattice $\Lambda$ in $n$-dimensional euclidean space we associate an elementary abelian 2-group $I(\Lambda)$ whose elements represent parts of the dual lattice that are similar to $\Lambda$. There are corresponding involutions on modular forms for which the theta series of $\Lambda$ is an eigenform; previous work has focused on this connection. In the present paper $I(\Lambda)$ is considered as a quotient of some finite 2-subgroup of $\operatorname{O}_n({{\mathbb R}})$. We establish upper bounds, depending only on $n$, for the order of $I(\Lambda)$, and we study the occurrence of similarities of specific types.'
author:
- 'H.-G. Quebbemann, E. M. Rains'
title: On the involutions fixing the class of a lattice
---
Introduction {#introduction .unnumbered}
============
The arithmetic of lattices in euclidean spaces contains a counterpart to the action of Atkin-Lehner involutions on elliptic modular forms. An involution here takes the isometry class of a lattice to that of a rescaled partial dual. There are remarkable cases of (non-unimodular) lattices that are invariant under this action; such a lattice is called strongly modular. Some prominent examples have arisen as sections of the Leech lattice or in the study of finite rational matrix groups. See [@N], [@Q], [@R-S], [@S-S].
A basic question that has remained open is when in general strong modularity can occur. For instance, given $n=4r$ and $\ell =p_1\cdot \ldots \cdot p_s$ with distinct primes $p_i$ ($s>0$), there exist even lattices of dimension $n$ and level $\ell$ whose genera have the necessary property of being invariant under all $2^s$ involutions. But as we shall see, a class can have this property only if $s\le 2 v_2(n)$, where $v_2$ denotes the 2-adic exponent valuation. More generally, we find that at most $4^{[n/2]}$ involutions fix the class of a given $n$-dimensional integral lattice. In cases of squarefree level here $[n/2]$ can be replaced by $v_2(n)$. These bounds are attained for any $n$ by certain lattices built up from the case $n=2$ (which was studied in detail in [@R1]).
When one considers a single involution, again the question arises what additional conditions must an invariant genus satisfy in order to contain an invariant class. Some restrictions have been found by studying congruence properties mod 2 of the relevant Atkin-Lehner eigenforms ([@R-S], Appendix, and [@R2]). A more elementary discussion related to this question is given in the final section of this paper.
Involutions and groups
======================
We shall study the involutory invariance of an integral lattice by applying representation theory to its “isomorphism group”. This object, which is defined below, has occurred in [@N] as a subset of $\operatorname{GL}_n({{\mathbb R}})$, and as a group it had a short appearance in [@R-S Theorem 6].
We fix positive integers $\ell$ and $n$. Let ${{\mathcal L}}$ denote the set of lattices $\Lambda$ in euclidean ${{\mathbb R}}^n$ such that $\Lambda\subseteq \Lambda^*$ and $\ell\Lambda^*\subseteq
\Lambda$, where $\Lambda^*$ is the dual lattice. Then the orthogonal group $\operatorname{O}_n({{\mathbb R}})$ acts on ${{\mathcal L}}$. Furthermore we have the elementary abelian 2-group $W(\ell)$ of positive divisors $m\|\ell$ (i.e., $m|\ell$ and $gcd(m,\ell/m) =1$), with multiplication $m*m' = mm'/gcd(m,m')^2$. Defining (as in [@Q Section 1]) $$\Lambda_m = \sqrt{m}(\Lambda^* \cap m^{-1} \Lambda)$$ we also have a group action of $W(\ell)$ on ${{\mathcal L}}$, which clearly commutes with the action of $\operatorname{O}_n({{\mathbb R}})$. In particular, either group acts on the orbits of the other one.
Now fix $\Lambda$, and let $\ell$ be its level in the sense that $\ell$ is minimal for $\ell \Lambda^* \subseteq \Lambda$ (so $\ell$ divides $\det \Lambda = \# \Lambda^*/\Lambda$) . Then the following groups are canonically associated with $\Lambda$. First, $\operatorname{O}_n({{\mathbb R}})$ contains the usual stabilizer $\operatorname{Aut}(\Lambda)$ and the stabilizer of the $W(\ell)$-orbit $$\operatorname{Iso}(\Lambda) = \{\sigma\in \operatorname{O}_n({{\mathbb R}}) \mid \Lambda = \sigma\Lambda_m
{\quad}\text{for some}{\quad}m\in W(\ell)\}.$$ Next, in $W(\ell)$ there are again the stabilizer of $\Lambda$ and the stabilizer of its $\operatorname{O}_n({{\mathbb R}})$-orbit, denoted by $A(\Lambda)$ and $I(\Lambda)$, respectively. Note that if $m \in I(\Lambda)$, then $\det \Lambda = \det \Lambda_m$, which implies $m^{n/2} \| \det \Lambda$. Actually $A(\Lambda)$ is trivial. In fact, if $\Lambda = \Lambda_m$, then $m$ must be a square, say $m=k^2$, and $(1/\sqrt{k})\Lambda$ is an integral lattice of level $\ell/k$ and determinant prime to $m$; so $m=1$. As an immediate consequence we obtain a group isomorphism $$\operatorname{Iso}(\Lambda)/\operatorname{Aut}(\Lambda) \cong I(\Lambda)$$ induced by $\sigma\longmapsto m: \Lambda = \sigma\Lambda_m$. Note also that the squares in $I(\Lambda)$ form a subgroup $B(\Lambda)$ which corresponds to the quotient of $\operatorname{Iso}(\Lambda)\cap \operatorname{GL}(\Lambda\otimes{{\mathbb Q}})$ by $\operatorname{Aut}(\Lambda)$. Of course, $B(\Lambda)= I(\Lambda)$ for odd $n$.
If $I(\Lambda) = W(\ell)$, then $\Lambda$ is said to be strongly modular. Two “trivial” constructions of such lattices will be used. Namely, if $\Lambda$ and $\Lambda'$ are strongly modular with coprime levels $\ell,\ell'$, then so are, with level $\ell\ell'$,
- the orthogonal sum $\sqrt{k'}\Lambda\oplus \sqrt{k}\Lambda'$ in the case of squares $\ell = k^2$, $\ell'= k^{'2}$
- the tensor product $\Lambda\otimes \Lambda'$.
A general bound
===============
To measure $I(\Lambda)$ we first replace the linear group $\operatorname{Iso}(\Lambda)$ by a 2-Sylow subgroup $G$, so that $I(\Lambda) \cong G/P$ for $P = G\cap \operatorname{Aut}(\Lambda)$. If $n$ is odd, an even larger elementary abelian quotient of $G$ arises from $P \cap
\operatorname{SL}_n({{\mathbb R}})$ because $P$ contains $-\operatorname{id}$. Put $$d(G) = d(G/Q)$$ for any finite $p$-group $G$, where $Q$ denotes the minimal normal subgroup such that $G/Q$ is elementary abelian, in which case $d$ means dimension over ${{\mathbb F}}_p$. We would like to use that $d(G) \le n$ if $G$ is a linear group of degree $n$. Actually this inequality is true in general for odd $p$, while for 2-subgroups of $\operatorname{GL}_n({{\mathbb C}})$ one has $[3n/2]$ as the best possible general upper bound ([@I], [@W]).
\[thm1\] If $G$ is a finite $2$-subgroup of $\operatorname{GL}_n({{\mathbb R}})$, then $d(G) \le n$.
This result can be extracted from [@W Lemma 4], in the case where the natural representation of $G$ on ${{\mathbb R}}^n$ is monomial, but the method equally applies when this representation is induced from a 2-dimensional dihedral representation of a subgroup. (The inductive proof in any case requires to make a stronger statement, implying that if $H$ is a normal subgroup of $G$, then $d(H) \le n$ and $d(G/H) \le n$ where at least one inequality is strict. Of course, over ${{\mathbb C}}$ this already fails for $n=1$.) The remaining cases of reducible or not absolutely irreducible groups are easily dealt with, using induction on $n$ or an upper bound over ${{\mathbb C}}$ in dimension $\nu = n/2$ (the crude bound $2\nu-1$ suffices).
\[thm2\] For any integral lattice $\Lambda \subset {{\mathbb R}}^n$, $$d(I(\Lambda)) \le 2[n/2].$$ In particular, if $\Lambda$ is strongly modular, at most $2[n/2]$ distinct primes divide the level $\ell$. Lattices that attain this bound exist in any dimension $n$.
The bound is an immediate consequence of the preceding theorem and the remarks made at the beginning of this section. It is attained by strongly modular lattices that arise by the first construction mentioned at the end of Section 1. All we must know is that, for $n=2$ and primes $p\not= q$, there exists such a lattice with level $p^2q^2$; for this see [@R1 Corollary 10]. For general even $n$ we then take scaled orthogonal sums of such lattices, and for odd $n$ we add a 1-dimensional summand.
It should be noted at this point that also many pairs of primes $p\not= q$ admit a strongly modular 2-dimensional lattice with level $pq$; the condition is that $p$ and $q$ must be quadratic residues of each other ([@Q], [@R1]). Taking now tensor products of such lattices, the bound above is still attained for $n=4$, but otherwise this construction suggests a different bound which will be obtained in the next section.
It may be expected, however, that orthogonally indecomposable lattices attaining the bound of Theorem \[thm2\] exist in any dimension. To obtain an example for $n=3$, consider a solution of $a^2 + b^2 = c^2$ where $a,b,c$ are coprime positive integers with $a+c$ divisible by 4. Take the orthogonal sum $(a) \oplus \left(\begin{smallmatrix} c & b\\b &
c\end{smallmatrix}\right)$, and construct its Kneser 2-neighbour with respect to the sum of the first two base vectors. For this second lattice $\Lambda$, which like the first has level $\ell = a^2$, we find the Gram matrix $$C =
\begin{pmatrix}
\frac{1}{4}(a+c) & \frac{b}{2} & \frac{1}{2}(c-a)\\
\frac{b}{2} & c & b\\
\frac{1}{2}(c-a) & b & a+c
\end{pmatrix}
= S(\ell C^{-1}) S^t, \qquad S = \begin{pmatrix}
0 & 0 & 1\\
0 & -1 & 0\\
1 & 0 & 0
\end{pmatrix}\, .$$ This shows that again $\{1,\ell\} \subseteq I(\Lambda)$. For $(a,b,c) = (15,8,17)$, computation of $\Lambda_9$ moreover shows that $\Lambda$ is strongly modular; it is clearly indecomposable. ($C$ is reduced in the sense of Seeber-Eisenstein [@O] whenever $a = 4t^2-1$ and $b=4t$, $t>1$. For $t>2$, however, $I(\Lambda)$ does not contain $m=(2t-1)^2 \in W(\ell)$. In fact, $\Lambda$ has minimal norm $2t^2$, while $\Lambda_m$ contains the vector of norm 8 given by $(1/\sqrt{m})(2,-2,1)$ with respect to $C$.)
The squarefree case
===================
If we only consider lattices $\Lambda$ with squarefree level or, in the general case, factor out $B(\Lambda) = I(\Lambda) \cap \{k^2\mid k=1,2,\ldots\}$, then the result of Section 2 can be improved considerably. We now make use of the fact that any element of $\operatorname{Iso}(\Lambda)$ up to some scalar factor $1/\sqrt{m}$ is defined over ${{\mathbb Z}}$, and so $\operatorname{Iso}(\Lambda)$ is conjugate to a subgroup of $$\operatorname{GL}_n({{\mathbb Q}})^+ = \langle \operatorname{GL}_n({{\mathbb Q}}), \; \{\sqrt{p}\cdot\operatorname{id}\mid p
\text{ prime}\}\rangle \, .$$
\[thm3\] Let $G$ be a finite $2$-subgroup of $\operatorname{GL}_n({{\mathbb Q}})^+$, and define $N = G\cap \operatorname{GL}_n({{\mathbb Q}})$. Then $d(G/N) \le 2 v_2(n)$.
We can write $G = G_1 \cup G_2 \cup \ldots$, where $G_m = G\cap \sqrt{m} \operatorname{GL}_n({{\mathbb Q}})$ is a coset of $N=G_1$ or is empty. We may assume that $N$ contains $-\operatorname{id}$ (or adjoin it). Let $\varrho$ denote the natural representation of $G$ on $V = K^n$, where $K$ is the field obtained from ${{\mathbb Q}}$ by adjoining the relevant square roots $\sqrt{m}$ (i.e., those for which $G_m$ is nonempty) together with $\zeta = e^{2\pi i/r}$, where $r$ is the exponent of $G$. Let $\chi$ be the character of $\varrho$. Then the values of $\chi$ lie in the 2-power cyclotomic field ${{\mathbb Q}}(\zeta)$, while on the other hand $\chi(\sigma) \in
{{\mathbb Q}}\sqrt{m}$ for $\sigma \in G_m$. This shows that $\chi$ vanishes outside $N\cup G_2$. Set $M = N$ if $\chi$ also vanishes on $G_2$, and $M = N\cup G_2$ otherwise. Denoting the restriction of $\chi$ to the subgroup $M$ by $\chi_M$, we obtain $$\langle \chi,\chi\rangle
= \frac{1}{\# G} \sum_{\sigma\in G} |\chi(\sigma)|^2
= \frac{1}{\# G} \sum_{\sigma\in M} |\chi(\sigma)|^2
= \frac{1}{(G:M)} \langle \chi_M, \chi_M\rangle\, .$$ It follows that $(G:M) \le n^2$, since $\langle \chi,\chi\rangle \ge 1$ and $\langle \chi_M,\chi_M\rangle \le n^2$. However, the theorem moreover asserts that $(G:N)$ divides $n^2$. To prove this we first observe that the Galois group $\Gamma = \operatorname{Gal}(K/ {{\mathbb Q}})$ acts on the $G$-invariant subspaces of $V$. If such a subspace $U\not= 0$ is defined over ${{\mathbb Q}}$ (i.e., invariant also under $\Gamma$), then the kernel of the restriction of $\varrho$ to $U$ clearly is contained in $N$, and in the case $U\not= V$ it suffices to apply the theorem on $U$ and $V/U$ (using induction on $n$). Therefore we now assume that no such proper subspace exists. This implies that the isotypical summands of $\varrho_M$ are permuted transitively by the (semidirect) product $G\cdot \Gamma$. So we have $$\chi_M = a(\psi_1 + \cdots + \psi_b)$$ where $\psi_1,\ldots,\psi_b$ are pairwise inequivalent irreducible characters of $M$ and all are of the same degree $c$. Then $n= abc$ and $$n^2 = bc^2 \langle \chi_M, \chi_M\rangle = bc^2 \langle \chi,
\chi\rangle (G:M)\, .$$ Finally suppose that $(M:N) = 2$. Then it suffices to show that $bc^2$ is even. Since $M$ is a 2-group, $c$ is a power of 2, and we only have to show that $b$ is even if $c=1$. Let $\psi$ be one of the linear characters $\psi_i$. If $\psi^2 = 1$, set ${\widetilde}{\psi} = \lambda\psi$ where $\lambda$ is the linear character of $M$ with kernel $N$, and if $\psi^2 \not= 1$ set ${\widetilde}{\psi} = \psi^{-1}$. It follows that ${\widetilde}{\psi} \not= \psi$ and ${\widetilde}{\psi}$ also appears in $\chi_M$ (to see this in the first case, use a Galois automorphism mapping $\sqrt{2}$ to $-\sqrt{2}$). Therefore the linear characters $\psi_i$ appear in pairs.
\[thm4\] For any integral lattice $\Lambda \subset {{\mathbb R}}^n$, $$d(I(\Lambda)/B(\Lambda)) \le 2v_2(n)\, .$$ In particular, if $\Lambda$ is strongly modular with squarefree level $\ell$, then $\ell$ is the product of at most $2v_2(n)$ primes. Lattices that attain this bound exist in any dimension $n$.
Recall that $I(\Lambda)/B(\Lambda) \cong \operatorname{Iso}(\Lambda) / \operatorname{Iso}(\Lambda)\cap
\operatorname{GL}(\Lambda\otimes {{\mathbb Q}})$, where the last two groups again may be replaced by Sylow 2-subgroups. So the bound immediately follows from the preceding theorem. For $n=2^e f$, $f$ odd, this bound is attained by the $f$-fold orthogonal sum of the tensor product of $e$ strongly modular 2-dimensional lattices with coprime levels $p_i q_i$ ($p_i \not=
q_i$ primes). As was noted already, enough such lattices exist.
If more is known on the action of a Sylow 2-subgroup of $\operatorname{Aut}(\Lambda)$, then of course the proof of Theorem \[thm3\] gives more information on $I(\Lambda)$. Suppose, for example, that this action is multiplicity-free. Then in that proof the equation $n = abc$ holds with $a=1$ and $b = \langle
\chi_M,\chi_M\rangle = (G:M) \langle \chi,\chi\rangle$, so for $M=N$ (in particular, for odd $\ell$) we obtain $$d(I(\Lambda) / B(\Lambda)) \le v_2(n)\, .$$ Given $\ell = p_1\cdot \ldots \cdot p_s$ with $s$ distinct primes $p_i$, this bound is attained by the tensor product of 2-dimensional lattices having the Gram matrix $\left(\begin{smallmatrix} 1 & 0\\0 & p_i\end{smallmatrix}\right)$ or $\left(\begin{smallmatrix} 2 & 1\\1 & q_i\end{smallmatrix}\right)$, $q_i = (p_i+1)/2$. The condition of multiplicity one is satisfied because each of these lattices admits a reflection and the corresponding automorphisms of $\Lambda$ generate an elementary abelian 2-group acting by its regular representation.
Moreover, it is this last bound, $s \le v_2(n)$ for a strongly modular lattice, which some of the “extremal” lattices discovered during the past few years attain.
Single involutions
==================
In this section we fix a positive integer $m$ which is not of the form $k^2$ or $2k^2$ for an integer $k$. If $m$ belongs to $I(\Lambda)$ for an integral lattice $\Lambda$, then as before $\Lambda$ is the image of $\Lambda^* \cap m^{-1}\Lambda$ under a similarity $\tau=\sqrt{m}\sigma$ where $\sigma \in \operatorname{O}_n({{\mathbb R}})$ is of 2-power order; that is, we have $f(\tau) =0$ with $f= X^{2r}-m^r$ for some $r=2^k$, $k \geq 0$. Let such a polynomial $f$ be fixed, and define the algebra $$E = {{\mathbb Q}}[X]/(f) = {{\mathbb Q}}[\alpha], \quad \alpha = X+(f).$$ In $E$ we consider the root of unity $\zeta = \alpha^2/m$ and the order $R =
{{\mathbb Z}}[\alpha,\zeta]$ (over which our lattices will be defined). Now $f$ factors into the rationally irreducible polynomials $$f_0 = X^2-m, \quad f_j = X^{2^j}+m^{2^{j-1}} (1 \leq j \leq k).$$ So we have $E \cong E_0 \times ... \times E_k$ where $E_j
= {{\mathbb Q}}[X]/(f_j) ={{\mathbb Q}}(\alpha_j)$ is a quadratic extension of the $2^j$-th cyclotomic field ${{\mathbb Q}}(\zeta_j)$, with $\alpha_j^2 =m\zeta_j$. Defining $\overline{\alpha} = m \alpha^{-1}$ we obtain an involutory automorphism “bar” on $E$, and correspondingly on each $E_j$. For $j\geq 1$ this involution of $E_j$ is not the identity, so then $E_j$ also is a (totally complex) quadratic extension of the totally real field ${{\mathbb Q}}(\beta_j)$, $\beta_j=\alpha_j +\overline{\alpha_j}$, with $\beta_j^2
=m(1+\zeta_j)(1+\zeta_j^{-1})$. Note that $R = \overline{R}$.
By an inner product space over $E$, we mean a finitely generated $E$-module $V$ carrying an $E$-valued, totally positive definite hermitian form $h$ for the involution defined above. Clearly such a hermitian module splits as an orthogonal sum of hermitian spaces over the fields $E_j$ (for $j=0$ “hermitian” here means “symmetric bilinear”). By definition, $h$ is totally positive definite if each of these summands is. In this case $V$ also carries the rational inner product $$u\cdot v = (2r)^{-1} \operatorname{Tr}_{E/{{\mathbb Q}}} h(u,v)$$ where $\operatorname{Tr}_{E/{{\mathbb Q}}}$ denotes the sum of the field traces. Given an $R$-lattice $\Lambda \subset V$, we denote its $R$-dual with respect to $h$ by $\Lambda^h$ and its ${{\mathbb Z}}$-dual with respect to the dot product by $\Lambda^*$.
\[thm5\] If $(V,h)$ is an inner product space over $E$ and $\Lambda$ is an $R$-lattice on $V$, then the duals are related by $\Lambda^h=\tau(\Lambda^*)$, where $\tau(v)=\alpha v$. In particular, $\Lambda$ is contained in $\Lambda^h$ with index prime to $m$ precisely when $$\Lambda \subseteq \Lambda^*, \quad m \in I(\Lambda),
\quad \Lambda=\tau(\Lambda^*\cap m^{-1}\Lambda).$$ Conversely, if these conditions are satisfied by a lattice $\Lambda \subset {{\mathbb R}}^n$ and a similarity $\tau$ with $f(\tau)=0$, then $\Lambda$ and $\tau$ arise from $R$ in the way above.
For the ${{\mathbb Z}}$-basis $1,\zeta,...,\zeta^{r-1},\alpha,\alpha\zeta,...,\alpha\zeta^{r-1}$ of $R$ it is easily seen that all elements have trace 0, except for $\operatorname{Tr}_{E/{{\mathbb Q}}}(1)=2r$. Therefore, given $\varepsilon \in E$, the condition $(2r)^{-1}\operatorname{Tr}_{E/{{\mathbb Q}}}(\varepsilon R) \subseteq {{\mathbb Z}}$ is equivalent to $\varepsilon \in \alpha^{-1}R$. This proves the first part. As to the converse, the action of $\tau$ makes $V = \Lambda \otimes {{\mathbb Q}}$ into an $E$-module, and since $\Lambda$ is stable under both $\tau$ and $(1/m)\tau^2$, it is an $R$-lattice. We have to show that the inner product arises from one over $E$ in the way described. For fixed $u$ and $v$ in $V$ consider the linear map $t$ from $E$ to ${{\mathbb Q}}$ defined by $t(\varepsilon)= (\varepsilon u)\cdot v$. There is a unique $\varepsilon '$ in $E$ such that $$\operatorname{Tr}_{E/{{\mathbb Q}}} (\varepsilon \varepsilon') = 2r t(\varepsilon)$$ for all $\varepsilon \in E$, and we define $h(u,v)$ to be this element $\varepsilon'$. The verification of further details will be omitted. (Compare [@M]; the discussion restricted there to isometries carries over to similarities).
In the following $m$ is supposed to be odd. The theorem may be applied to find restrictions for the rational quadratic space $V = \Lambda \otimes {{\mathbb Q}}$ when $I(\Lambda)$ contains $m$. Recall that in this case $\det \Lambda = m^{n/2} d'$ where $m$ and $d'$ are relatively prime. The necessary condition that $V$ admits the similarity factor $m$ just means that (i) $m$ is a square modulo each prime $p>2$ dividing $d'$ with odd multiplicity and (ii) $d'$ is a square modulo each prime $p$ dividing $m$ with odd multiplicity. This can be seen, for instance, by [@S], Chapter 5, Corollary 3.6.
Let $\Lambda$ be an even lattice. We use the rational invariants defined at an arbitrary prime number $p$ by the Gauss sum $$\gamma_p(\Lambda) = (\operatorname{det}_p \Lambda)^{-1/2} \sum \exp(\pi i v\cdot v),$$ where the summation extends over the elements $v+\Lambda$ in the $p$-primary component of $\Lambda^*/\Lambda$; $\det_p \Lambda$ denotes the order of this group. The product $\prod \gamma_p(\Lambda)$ taken over all primes $p$ is $i^{n/2}$. (See [@S], Chapter 5, [§]{} 8.)
Recall that if $m \in I(\Lambda)$, then $V$ is defined over the direct product of the fields $E_0 ={{\mathbb Q}}(\sqrt{m})$, $E_1 = {{\mathbb Q}}(\sqrt{-m})$, $E_2, \ldots, E_k,$ where $E_j$ for $j >1$ has degree $2^{j-1}$ over ${{\mathbb Q}}(i)$. Assume for the moment that the contributions from $E_0$ and $E_1$ are zero. Then the dimension $n$ is a multiple of 4, and the rational quadratic form on $V$ arises from a hermitian form over ${{\mathbb Q}}(i)$, so it is an orthogonal sum $\phi \perp \phi$ and has square determinant. We compute that for $p^s \|m$, $$\gamma_p(\Lambda) =
\begin{cases}
1 & \text{if} {\quad}p^s \equiv 1 \pmod{4}\\
(-1)^{n/4} & \text{if} {\quad}p^s \equiv 3 \pmod{4},
\end{cases}$$ which gives $$\prod_{p|m} \gamma_p(\Lambda) =
(-1)^{\frac{n}{4} \cdot \frac{m-1}{2}}.$$ When $m \equiv 7 \pmod{8}$, the assumption used here is eliminated by the following theorem (first proved in [@R2] by less elementary arguments, a special case can be found in [@R-S]).
\[thm6\] Let $m$ be congruent to $7$ mod $8$, and let $m'$ be a positive integer relatively prime to $m$ such that $-m$ is a square mod $m'$. Then for any even lattice $\Lambda$ of level dividing $mm'$ with $m \in I(\Lambda)$, $$\prod_{p|m'} \gamma_p(\Lambda)
=
\prod_{p|m'}
\left( \frac{\sqrt{-m}}{\det_p\Lambda} \right),$$ where in the factor corresponding to $p$, $\sqrt{-m}$ represents either square root of $-m$ in the $p$-adic integers.
First, note that for odd $p|m'$ such that $m$ is not a square modulo $p$, $\det_p\Lambda$ must be a square (since otherwise $V$ does not admit the similarity factor $m$). In particular, since $-m$ is assumed to be a square modulo $p$, we find that either $\det_p\Lambda$ is a square or $-1$ is a square modulo $p$, and thus the given Legendre symbols are all well-defined.
Now consider the $R$-lattice structure for the order $R$ as before. Let $R_j$ be the image of $R$ in $E_j$ for $j=0, \dots, k$, and define the order $$R'= R_0 \times \ldots \times R_k.$$ Since the index of $R$ in $R'$ is a power of 2, there exists a positive integer $t$ such that $R'(2^t\Lambda)$ is an even lattice satisfying the hypotheses (possibly $m'$ has to be multiplied by a power of 2), with the same Gauss sums and $p$-determinants modulo squares as $\Lambda$. Since $R'$ is a product of orders preserved by the involution, $R'(2^t\Lambda)$ splits as a product of lattices, one over each order. Since the desired identity is preserved under taking products, we may reduce to the case that $\Lambda$ is a lattice over one of the factors $R_j$ of $R'$.
Case $j>1$. As noted already, here $\Lambda$ has square determinant and dimension divisible by 4. The computations preceding the theorem and the product formula give $$\prod_{p|m'} \gamma_p(\Lambda) = (-1)^{n/4} \prod_{p|m} \gamma_p(\Lambda)=1$$ as required.
Case $j=1$, ${{\mathbb Q}}(\sqrt{-m})$. The rational quadratic form on $V$ now splits as $\phi\perp m\phi$; since $-m$ is a square in ${{\mathbb Z}}_p$, $V\otimes {{\mathbb Q}}_p$ is hyperbolic for all $p|m'$. In particular, $\det_p \Lambda$ is a square and $\gamma_p(\Lambda)=1$ for all $p|m'$.
Case $j=0$, ${{\mathbb Q}}(\sqrt{m})$. Then $\Lambda$ contains a sublattice of index prime to $m$ that splits as an orthogonal sum of 2-dimensional lattices preserved by ${{\mathbb Z}}[\sqrt{m}]$. (Indeed, if $v\in\Lambda$ has norm prime to $m$, then the sum of $\Lambda \cap {{\mathbb Q}}(\sqrt{m})v$ and $\Lambda \cap ({{\mathbb Q}}(\sqrt{m})v)^\perp$ has index prime to $m$; proceed by induction). It thus suffices to consider the case of a primitive 2-dimensional even lattice of level exactly $mm'$, in which case the claim follows from Theorem 11 of [@R1].
For example, let $4|n$ and let $\ell > 1$ be squarefree. Given signs $\gamma_p = \pm 1$ for the primes $p|\ell$ such that $\prod \gamma_p = (-1)^{n/4}$, there exist even lattices $\Lambda$ of dimension $n$, level $\ell$, determinant $\ell^{n/2}$ and rational invariant $\gamma_p$ at each $p$. In this situation all divisors of $\ell$ are similarity factors of $\Lambda \otimes {{\mathbb Q}}$. However, when $\ell = mm'$ where $m \equiv 7 \pmod{8}$ and $-m$ is a square modulo $m'$, we obtain $$\prod_{p|m'} \gamma_p =1$$ as a necessary condition for having $m \in I(\Lambda)$. For $m$ incongruent to 7 mod 8 there is no similar restriction in general.
[X]{} , [*The number of generators of a linear $p$-group*]{}, Can. J. Math. [**24**]{} (1972) 851–858. , [*On isometries of inner product spaces*]{}, Invent. math. [**8**]{} (1969) 83–97. , [*The normalizer action and strongly modular lattices*]{}, L’Enseignement Math. [**43**]{} (1997) 67–76. , [*On indecomposable quadratic forms*]{}, J. Reine Angew. Math. [**317**]{} (1980) 120–156. , [*Atkin-Lehner eigenforms and strongly modular lattices*]{}, L’Enseignement Math. [**43**]{} (1997) 55–65. , [*Class groups and modular lattices*]{}, J. Number Theory [**88**]{} (2001) 211–224. , [*Non-existence results for modular lattices*]{}, Preprint 1999. , [*The shadow theory of modular and unimodular lattices*]{}, J. Number Theory [**73**]{} (1998) 359–389. , [*Extremal lattices*]{}, in [*Algorithmic Algebra and Number Theory*]{}, eds. B.H. Matzat, G.-M. Greuel, G. Hiss, Springer-Verlag 1998, pp. 139–170. , [*Quadratic and Hermitian Forms*]{}, Springer-Verlag 1985. , [*The rank of a linear $p$-group, an apology*]{}, J. London Math. Soc. (2)[**21**]{} (1980) 237–243.
-------------------------- -----------------------------------
Fachbereich 6 Mathematik Center for Communication Research
Universität Oldenburg Institute for Defense Analyses
26111 Oldenburg, Germany Princeton, NJ 08540, U.S.A.
quebbemann@mathematik. [email protected]
uni-oldenburg.de
-------------------------- -----------------------------------
| {
"pile_set_name": "ArXiv"
} |
`arXiv:0901.0411`\
`DAMTP-2009-1`
**Two-loop Integrability of Planar\
$\mathcal{N}=6$ Superconformal Chern-Simons Theory**
<span style="font-variant:small-caps;">Benjamin I. Zwiebel</span>
*DAMTP, Centre for Mathematical Sciences\
University of Cambridge\
Wilberforce Road, Cambridge CB3 0WA, UK*
`[email protected]`
**Abstract**
Bethe ansatz equations have been proposed for the asymptotic spectral problem of $AdS_4/CFT_3$. This proposal assumes integrability, but the previous verification of weak-coupling integrability covered only the ${\mathfrak{su}}(4)$ sector of the ABJM gauge theory. Here we derive the complete planar two-loop dilatation generator of $\mathcal{N}=6$ superconformal Chern-Simons theory from ${\mathfrak{osp}}(6|4)$ superconformal symmetry. For the ${\mathfrak{osp}}(4|2)$ sector, we prove integrability through a Yangian construction. We argue that integrability extends to the full planar two-loop dilatation generator, confirming the applicability of the Bethe equations at weak coupling. Further confirmation follows from an analytic computation of the two-loop twist-one spectrum.
Introduction
============
The $\mathcal{N}=6$ superconformal $U(N) \times U(N)$ Chern-Simons theory of Aharony, Bergman, Jafferis, and Maldacena has a ’t Hooft limit in which the dual description reduces to type IIA string theory on $AdS_4 \times CP^3$ [@Aharony:2008ug]. This limit is $N \to \infty$ and Chern-Simons level $k \to \infty$ with coupling $\lambda=N/k$ held fixed. Excitingly, the asymptotic spectral problem of this $AdS_4/CFT_3$ duality may be solvable by an ${\mathfrak{osp}}(6|4)$ Bethe ansatz, as first argued by Minahan and Zarembo based on their two-loop ($\mathcal{O}(\lambda^2)$) gauge theory calculation[@Minahan:2008hf] (also see [@Bak:2008cp]). Paralleling the case of $AdS_5/CFT_4$ [@Minahan:2002ve; @Beisert:2003tq; @Bena:2003wd; @Arutyunov:2004vx; @Staudacher:2004tk; @Beisert:2005tm; @Janik:2006dc; @Arutyunov:2006ak; @Beisert:2006ez], this apparent integrability is likely to lead to a better understanding of $AdS_4/CFT_3$ and the relationship between gauge/string duality and integrability.
There has been much related activity. Integrability of classical string theory on $AdS_4 \times CP^3$ has been proved for a subsector described by an $OSP(6|4)/(U(3) \times SO(1,3))$ coset sigma model [@Arutyunov:2008if; @Stefanski:2008ik]. The conjectured weak-coupling Bethe equations [@Minahan:2008hf] have been extended to an all-loop proposal [@Gromov:2008qe][^1], with the corresponding two-magnon S-matrix obtained in [@Ahn:2008aa]. This all-loop Bethe ansatz has passed strong-coupling tests in a $SU(2) \times SU(2)$ sector[@Astolfi:2008ji; @Sundin:2008vt], and seeming disagreement with semi-classical string computations [@McLoughlin:2008ms; @Alday:2008ut; @Krishnan:2008zs] is eliminated by a one-loop correction [@McLoughlin:2008he] to the magnon dispersion relation. Nonetheless, discrepancies between one-loop strong-coupling string calculations and the Bethe ansatz remain [@McLoughlin:2008he], and the classical integrability of the *complete* $AdS_4 \times CP^3$ Green-Schwarz action [@Gomis:2008jt] is unverified.
This work, however, will focus on the weak-coupling planar gauge theory. As is the case for $\mathcal{N}=4$ SYM, the spectral problem of computing anomalous dimensions is usefully formulated in terms of a spin chain. Local gauge-invariant operators are mapped to spin-chain states, and the anomalous part of the dilatation generator is mapped to a spin-chain Hamiltonian whose eigenvalues are the spectrum of anomalous dimensions. Then, if this Hamiltonian is integrable, one can replace the eigenvalue problem with a dramatically simpler problem of solving a system of Bethe equations, see [@Beisert:2004ry] for a review.
For the ABJM gauge theory, [@Minahan:2008hf] calculated the leading-order two-loop planar dilatation generator or Hamiltonian in the ${\mathfrak{su}}(4)$ sector, confirming two-loop integrability within that sector. The ${\mathfrak{su}}(4)$ sector corresponds to local operators composed of alternating scalars $\phi_i$ and $\bar{\phi}^i$, $i=1,\ldots4$, or a spin chain with alternating sites occupied by a spin transforming in the $\mathbf{4}$ or $\mathbf{\bar{4}}$ of ${\mathfrak{su}}(4)$. The spin chain is alternating because the $\phi_i$ transform as $(\mathbf{N}, \mathbf{\bar{N}})$ with respect to the $U(N) \times U(N)$ gauge group, while the $\bar{\phi}^i$ transform as $(\mathbf{\bar{N}},\mathbf{N})$.
To describe the full set of local operators, we must extend these conjugate finite-dimensional representations of the ${\mathfrak{su}}(4) = {\mathfrak{so}}(6)$ $R$ symmetry to two highest-weight irreducible representations of the $\mathcal{N}=6$ superconformal Lie algebra, ${\mathfrak{osp}}(6|4)$. In addition to the ${\mathfrak{su}}(4)$ scalars and their supersymmetric partners, $\bar{\psi}^i$ or $\psi_i$, these representations or modules includes arbitrarily many (symmetrized traceless) covariant derivatives acting on each scalar or fermion. The validity of the weak-coupling ${\mathfrak{osp}}(6|4)$ Bethe equation conjecture requires the significant assumption that two-loop integrability extends to the alternating spin chain with sites hosting these infinite-dimensional ${\mathfrak{osp}}(6|4)$ modules.
Before working on extending the two-loop ${\mathfrak{su}}(4)$ sector dilatation generator to the full ${\mathfrak{osp}}(6|4)$ spin chain, it is useful to review key steps used to compute the leading-order one-loop planar dilatation generator of $\mathcal{N}=4$ SYM. For that theory, the one-loop dilatation generator acts on two (identical) ${\mathfrak{psu}}(2,2|4)$ modules. An important observation is that the irreducible multiplets in the tensor product of two modules match up one-to-one with the corresponding multiplets in the ${\mathfrak{sl}}(2)$ sector[^2]. Combined with superconformal invariance, this enabled the lift of the ${\mathfrak{sl}}(2)$ sector dilatation generator[^3] to the complete one-loop dilatation generator [@Beisert:2003jj]. In fact, superconformal symmetry completely fixes the leading-order Hamiltonian (up to normalization); the ${\mathfrak{sl}}(2)$ sector one-loop dilatation generator was later computed [@Beisert:2004ry] just by requiring closure of the residual superconformal symmetry algebra of the sector, which includes an extra ${\mathfrak{su}}(1|1)$ symmetry generated by “hidden” supercharges.
We follow a similar path for the planar ABJM theory, using superconformal symmetry directly[^4] to compute the leading-order dilatation generator up to normalization[^5]. Again the first step involves restricting to a sector; we find the leading-order ${\mathfrak{osp}}(4|2)$ sector dilatation generator using residual superconformal symmetry. This sector corresponds to the set of local operators that are $1/12$ BPS at $\lambda=0$ with respect to a fixed pair of supercharges, which we label $\hat{{\mathfrak{Q}}}$ and $\hat{{\mathfrak{S}}}$. Similar to the case of the ${\mathfrak{sl}}(2)$ sector of $\mathcal{N}=4$ SYM, these supercharges generate an important additional ${\mathfrak{su}}(1|1)$, which commutes with ${\mathfrak{osp}}(4|2)$. By definition, all states in this sector are annihilated by these supercharges at $\lambda=0$, so we then have $$\delta {\mathfrak{D}}_2 = {\{\hat{{\mathfrak{Q}}}_1,\hat{{\mathfrak{S}}}_1\}}, \label{eq:introd2}$$ where subscripts refer to the order in $\lambda$. In other words, the two-loop dilatation generator within this sector is fixed once we know the *one-loop* corrections to these supercharges. Furthermore, requiring vanishing commutators with ${\mathfrak{osp}}(4|2)$ generators at leading order fixes $\hat{{\mathfrak{Q}}}_1$ and $\hat{{\mathfrak{S}}}_1$ up to overall normalization, giving the two-loop dilatation generator (the normalization can be identified from [@Minahan:2008hf]).
Superconformal invariance enables us to lift the leading-order dilatation generator within the ${\mathfrak{osp}}(4|2)$ sector to the full theory. However, in comparison to the analogous $\mathcal{N}=4$ SYM calculation, additional intermediate steps are needed. This is largely because here the leading-order Hamiltonian acts on three adjacent sites. The expression that immediately follows from the right side of (\[eq:introd2\]) is not in a convenient form for the lift. Instead we first prove integrability within this sector by constructing an ${\mathfrak{osp}}(4|2)$ Yangian that commutes with the Hamiltonian (The construction of the Yangian symmetry of the one-loop dilatation generator of $\mathcal{N}=4$ SYM was completed in [@Dolan:2003uh; @Dolan:2004ps]). This identifies the Hamiltonian as one that follows from an ${\mathfrak{osp}}(4|2)$ R-matrix construction. Now the Hamiltonian is written in terms of projectors onto irreducible modules appearing in the tensor products of two one-site modules, and now the lift becomes straightforward. Superconformal invariance, combined with some analysis of tensor products of one-site modules, allows us to lift this second expression for the ${\mathfrak{osp}}(4|2)$ Hamiltonian to the full two-loop ${\mathfrak{osp}}(6|4)$ spin-chain Hamiltonian.
In this work we do not explicitly verify that the complete Hamiltonian has an ${\mathfrak{osp}}(6|4)$ Yangian symmetry. However, it takes the exact form of an integrable alternating ${\mathfrak{osp}}(6|4)$ spin-chain Hamiltonian, assuming the existence of an ${\mathfrak{osp}}(6|4)$ R-matrix (This is parallel to a result of [@Beisert:2003yb] for $\mathcal{N}=4$ SYM). Combined with the explicit proof within the ${\mathfrak{osp}}(4|2)$ sector, this is convincing evidence that the complete two-loop planar model is integrable and that the two-loop spectral problem is solved by the Bethe ansatz.
In Section \[sec:alg\] we introduce the ABJM ${\mathfrak{osp}}(6|4)$ spin-chain model and the restriction to the ${\mathfrak{osp}}(4|2)$ sector, and Section \[sec:oneloop\] derives the two-loop dilatation generator for this sector. Here we also compute analytically the twist-one spectrum from the dilatation generator, which matches the Bethe ansatz prediction. The Yangian proof of integrability appears in Section \[sec:yangian\], and the following section computes the corresponding R-matrix expression for the Hamiltonian. Section \[sec:lift\] gives the unique lift to the complete two-loop planar dilatation generator, and we conclude with the following section.
The ${\mathfrak{osp}}(6|4)$ spin chain and its ${\mathfrak{osp}}(4|2)$ subsector \[sec:alg\]
============================================================================================
The ABJM gauge theory’s $\mathcal{N}=6$ superconformal symmetry, which has been verified in [@Benna:2008zy; @Bandres:2008ry], corresponds to the ${\mathfrak{osp}}(6|4)$ Lie algebra. After reviewing this algebra we introduce the corresponding module that is used for building the spin-chain description of gauge-invariant local operators. We then explain the restriction to the ${\mathfrak{osp}}(4|2)$ sector, and this sector’s algebra and spin module. Finally, we introduce a light-cone superspace basis for the ${\mathfrak{osp}}(4|2)$ module. For a recent general analysis of representations of the three dimensional superconformal groups, see [@Dolan:2008vc].
The complete algebra and spin module
------------------------------------
For the ${\mathfrak{osp}}(6|4)$ algebra generators we use one ${\mathfrak{su}}(2)$ spinor index, $\alpha, \beta = 1, 2$ and an ${\mathfrak{su}}(4)$ index $i, j=1, 2, 3, 4$. The ${\mathfrak{sp}}(4)$ elements are ${\mathfrak{su}}(2)$ Lorentz generators ${\mathfrak{L}}^\alpha{}_\beta$, translation and special conformal generators ${\mathfrak{P}}_{\alpha\beta}={\mathfrak{P}}_{\beta\alpha}$ and ${\mathfrak{K}}^{\alpha\beta}={\mathfrak{K}}^{\beta\alpha}$, and the dilatation generator ${\mathfrak{D}}$. The remaining bosonic elements are ${\mathfrak{su}}(4)$ ${\mathfrak{R}}$-symmetry generators ${\mathfrak{R}}^i{}_j$. Additionally, there are 24 supercharges with $${\mathfrak{Q}}_{ij,\alpha} = - {\mathfrak{Q}}_{ji,\alpha}, \quad {\mathfrak{S}}^{kl,\beta} = - {\mathfrak{S}}^{lk,\beta}.$$ The nonvanishing commutators with the ${\mathfrak{sp}}(4)$ generators are $$\begin{aligned}
{[{\mathfrak{L}}^\alpha{}_\beta,{\mathfrak{L}}^\gamma{}_\delta]} &= \delta^\alpha_\delta {\mathfrak{L}}^\gamma{}_\beta - \delta^\gamma_\beta {\mathfrak{L}}^\alpha{}_\delta, & {[{\mathfrak{L}}^\alpha{}_\beta,{\mathfrak{P}}_{\gamma\delta}]}& = 2 \delta^\alpha_{\{\gamma} {\mathfrak{P}}_{\delta\}\beta} - \delta^\alpha_\beta {\mathfrak{P}}_{\gamma\delta},
\notag \\
{[{\mathfrak{L}}^\alpha{}_\beta,{\mathfrak{K}}^{\gamma\delta}]}& = -2 \delta_\beta^{\{\gamma} {\mathfrak{K}}^{\delta\}\alpha} + \delta^\alpha_\beta {\mathfrak{K}}^{\gamma\delta}, &
{[{\mathfrak{L}}^\alpha{}_\beta,{\mathfrak{Q}}_\gamma]} &= \delta^\alpha_\gamma {\mathfrak{Q}}_\beta -{{{\textstyle\frac{1}{2}}}}\delta^\alpha_\beta {\mathfrak{Q}}_\gamma,
\notag \\
{[{\mathfrak{L}}^\alpha{}_\beta,{\mathfrak{S}}^\gamma]} & = -\delta^\gamma_\beta {\mathfrak{S}}^\alpha +{{{\textstyle\frac{1}{2}}}}\delta^\alpha_\beta {\mathfrak{S}}^\gamma, & {[{\mathfrak{K}}^{\alpha \beta},{\mathfrak{P}}_{\gamma\delta}]} & = 4\delta^{\{\alpha}_{\{\gamma} {\mathfrak{L}}^{\beta\}}{}_{\delta\}} + 4\delta^{\{\alpha}_{\{\gamma} \delta^{\beta\}}_{\delta\}} {\mathfrak{D}},
\notag \\
{[{\mathfrak{K}}^{\alpha \beta},{\mathfrak{Q}}_{kl, \gamma}]} & = {{{\textstyle\frac{1}{2}}}}\varepsilon_{klij}\big(\delta^\alpha_\gamma {\mathfrak{S}}^{ij, \beta} + \delta^\beta_\gamma {\mathfrak{S}}^{ij, \alpha} \big), & &
\notag \\
{[{\mathfrak{P}}_{\alpha \beta},{\mathfrak{S}}^{kl,\gamma}]} & = -{{{\textstyle\frac{1}{2}}}}\varepsilon^{klij}\big(\delta_\alpha^\gamma {\mathfrak{Q}}_{ij,\beta} + \delta_\beta^\gamma {\mathfrak{Q}}_{ij,\alpha} \big), &&\end{aligned}$$ and the dimensions of $\{{\mathfrak{P}}, {\mathfrak{K}}, {\mathfrak{Q}}, {\mathfrak{S}}\}$ are $\{1,-1,{{{\textstyle\frac{1}{2}}}},-{{{\textstyle\frac{1}{2}}}}\}$.
The commutators with the ${\mathfrak{so}}(6) = {\mathfrak{su}}(4)$ ${\mathfrak{R}}$ generators are $$\begin{aligned}
{[{\mathfrak{R}}^i{}_j,{\mathfrak{R}}^k{}_l]} & = \delta^i_l{\mathfrak{R}}^k{}_j - \delta^k_j{\mathfrak{R}}^i{}_l,& {[{\mathfrak{R}}^i{}_j,{\mathfrak{Q}}_{kl}]} &=2 \delta^i_{[l} {\mathfrak{Q}}_{k]j}-{{{\textstyle\frac{1}{2}}}}\delta^i_j {\mathfrak{Q}}_{kl}, \notag \\
{[{\mathfrak{R}}^i{}_j,{\mathfrak{S}}^{kl}]} & = -2 \delta_j^{[l} {\mathfrak{S}}^{k]i}+{{{\textstyle\frac{1}{2}}}}\delta^i_j {\mathfrak{S}}^{kl}. & &\end{aligned}$$ Finally, the nonvanishing anticommutators are $$\begin{aligned}
{\{{\mathfrak{Q}}_{ij,\alpha},{\mathfrak{Q}}_{kl,\beta}\}} & = -\varepsilon_{ijkl} {\mathfrak{P}}_{\alpha\beta}, &
{\{{\mathfrak{S}}^{ij,\alpha},{\mathfrak{S}}^{kl,\beta}\}} & = -\varepsilon^{ijkl} {\mathfrak{K}}^{\alpha\beta},
\notag \\
{\{{\mathfrak{Q}}_{ij,\beta},{\mathfrak{S}}^{kl,\gamma}\}} & = 4 \delta^\gamma_\beta \delta^{[k}_{[i}{\mathfrak{R}}^{l]}{}_{j]} + 2 \delta^k_{[j}\delta^l_{i]}{\mathfrak{L}} ^\gamma{}_\beta + 2 \delta^k_{[j}\delta^l_{i]}\delta^\gamma_\beta {\mathfrak{D}}. && \label{eq:anticommutators}\end{aligned}$$ There are multiple choices of positive roots of the Lie algebra. For instance, we can choose the raising generators as $${\mathfrak{L}}^2{}_1, \quad {\mathfrak{K}}, \quad {\mathfrak{R}}^i{}_j, \, i > j, \quad {\mathfrak{S}}.$$ and the Hermitian conjugate lowering generators $${\mathfrak{L}}^1{}_2, \quad {\mathfrak{P}}, \quad {\mathfrak{R}}^i{}_j, \, i < j, \quad {\mathfrak{Q}}.$$ The Cartan generators are then the diagonal generators of ${\mathfrak{L}}$ and ${\mathfrak{R}}$ (the traceless conditions mean that only $1+3$ of these are independent) and the dilatation generator.
The spin chain has alternating highest-weight modules. As stated in the introduction, this is because the matter fields are in bifundamental representations of the $U(N) \times U(N)$ gauge group. We also note that ABJM gauge theory has an additional ${\mathfrak{u}}(1)$ symmetry, under which the modules have alternating charge $\pm 1$. The first representation $\mathcal{V}_\phi$ has highest-weight element ${\mathopen{\big|}\phi_1^{(0,0)}\mathclose{\bigr \rangle}}$ and consists of $${\mathopen{\big|}\phi_i^{(n_1, n_2)}\mathclose{\bigr \rangle}}, \quad {\mathopen{\big|}(\bar{\psi}^i)^{(n_1, n_2)}\mathclose{\bigr \rangle}}, \quad n_i =0, 1, 2, \ldots$$ where the superscripts correspond to the number of ${\mathfrak{su}}(2)$ indices carried by covariant derivatives acting on the fields of the ABJM model, $\phi$ and $\bar{\psi}$. All Lorentz indices are symmetrized on individual module elements, as needed for an irreducible representation. Alternate sites of the chain, instead host a representation $\mathcal{V}_{\bar \phi}$ with highest-weight element ${\mathopen{\big|}(\bar{\phi}^4)^{(0,0)}\mathclose{\bigr \rangle}}$ spanned by $${\mathopen{\big|}(\bar{\phi}^i)^{(n_1, n_2)}\mathclose{\bigr \rangle}}, \quad {\mathopen{\big|}\psi_i^{(n_1, n_2)}\mathclose{\bigr \rangle}}, \quad n_i =0, 1, 2, \ldots$$ The ${\mathfrak{su}}(4)$ sector corresponds to the $\phi$ and $\bar{\phi}$ states with all $n_i=0$. Also, of course, $n_1+n_2$ is even (odd) for bosons (fermions).
The ${\mathfrak{osp}}(6|4)$ variations were given in [@Gaiotto:2008cg; @Hosomichi:2008jb]. We will not need the explicit action of the entire ${\mathfrak{osp}}(6|4)$ generators at leading order, which can be identified (up to a physically irrelevant choice of basis) straightforwardly by requiring closure of the algebra. However, a few details are useful. The diagonal Lorentz generators and the classical dilatation generator act (independently of ${\mathfrak{R}}$ symmetry indices) as \^1\_1[X\^[(n\_1, n\_2)]{}]{} [[&=&]{}]{}-\^2\_2[X\^[(n\_1, n\_2)]{}]{} = (n\_1-n\_2) [X\^[(n\_1, n\_2)]{}]{}, [\
]{}\_0 [X\^[(n\_1, n\_2)]{}]{} [[&=&]{}]{}(n\_1+n\_2+1)[X\^[(n\_1, n\_2)]{}]{}. \[eq:lorentanddilatationaction\] All unbarred elements transform in the $\mathbf{4}$ of ${\mathfrak{su}}(4)$ and barred elements transform (naturally) in the $\mathbf{\bar{4}}$, $$\label{eq:rsymaction}
{\mathfrak{R}}^i{}_j {\mathopen{\big|}X_k\mathclose{\bigr \rangle}} = \delta^i_k {\mathopen{\big|}X_j\mathclose{\bigr \rangle}} - {{{\textstyle\frac{1}{4}}}}\delta^i_j {\mathopen{\big|}X_k\mathclose{\bigr \rangle}}, \quad {\mathfrak{R}}^i{}_j {\mathopen{\big|}\bar{X}^k\mathclose{\bigr \rangle}} = -\delta^k_j {\mathopen{\big|}\bar{X}^i\mathclose{\bigr \rangle}} + {{{\textstyle\frac{1}{4}}}}\delta^i_j {\mathopen{\big|}\bar{X}^k\mathclose{\bigr \rangle}}.$$ Also, schematically we have the ${\mathfrak{R}}$-symmetry index dependence of the leading-order supercharges, $$\begin{aligned}
{\mathfrak{Q}}_{ij}{\mathopen{\big|}X_k\mathclose{\bigr \rangle}} & \sim \varepsilon_{ijkm}{\mathopen{\big|}\bar{Y}^m\mathclose{\bigr \rangle}}, & {\mathfrak{Q}}_{ij}{\mathopen{\big|}\bar{X}^k\mathclose{\bigr \rangle}} & \sim \delta^k_{[i}{\mathopen{\big|}Y_{j]}\mathclose{\bigr \rangle}},
\notag \\
{\mathfrak{S}}^{ij}{\mathopen{\big|}\bar{X}^k\mathclose{\bigr \rangle}} & \sim \varepsilon^{ijkm}{\mathopen{\big|}Y_m\mathclose{\bigr \rangle}}, & {\mathfrak{S}}^{ij}{\mathopen{\big|}X_k\mathclose{\bigr \rangle}} & \sim \delta_k^{[i}{\mathopen{\big|}\bar{Y}^{j]}\mathclose{\bigr \rangle}}. \label{eq:qaction}\end{aligned}$$ Of course, there is also dependence on the Lorentz indices, but we will not need these precise factors for the full algebra. Simply we note that all interactions allowed by quantum numbers appear with nonzero coefficient. So ${\mathfrak{S}}^\alpha$ has nonvanishing action only on module elements with $n_{\alpha}>0$.
For $\lambda=0$ the spin-chain states transform in the tensor product of the single-site ${\mathfrak{osp}}(6|4)$ representations, but beyond leading order they transform in a deformed representation; the ${\mathfrak{osp}}(6|4)$ generators act on the spin chain via interactions that couple multiple sites. Beside the manifest ${\mathfrak{R}}$ and ${\mathfrak{L}}$ symmetry generators, all generators receive corrections for $\lambda \neq 0$. Perturbatively, these interactions act on an increasing number of sites with each order in $\lambda$. Counting powers of the coupling constant implies that at $\mathcal{O}(\lambda)$, interactions can act on up to a total of four sites (e.g. two initial and two final sites), and up to six sites at $\mathcal{O}(\lambda^2)$. In the planar limit, which is our focus, these interactions act on adjacent modules.
Finally, ${\mathfrak{osp}}(6|4)$ has a quadratic Casimir $J^2$, \[eq:osp64jsquared\] J\^2 [[&=&]{}]{}([\[\_[ij,]{},\^[ij, ]{}\]]{} - 2 \^i\_j \^j\_i + 2 \^\_\^\_+ 4 \^2 - [{\_,\^}]{}). On highest-weight states, $J^2$ simplifies to J\^2 [[&=&]{}]{}(D(D+3) + s(s+2) + 3 R\^1\_1+2 R\^2\_2 + R\^3\_3 - \_[i=1]{}\^4 (R\^i\_i)\^2 ) [\
]{} [[&=&]{}]{}(D(D+3) + s(s+2) [\
&&]{} - q\_1(q\_1+2) - q\_2(q\_2+2) - (2 p + q\_1 + q\_2)\^2 - (2 p + q\_1 + q\_2)). Here $D$ is the dimension and $s$ is the Lorentz spin. The first expression uses eigenvalues of all diagonal entries of the traceless matrix of ${\mathfrak{R}}$-symmetry generators, while the second uses the standard ${\mathfrak{su}}(4)$ Dynkin labels, $$q_1 = R^2_2-R^1_1, \quad p=R^3_3 - R^2_2, \quad q_2 = R^4_4-R^3_3.$$ The ${\mathfrak{osp}}(6|4)$ spin $j$ satisfies $j(j+1) = J^2$, the eigenvalue of ${\mathfrak{J}}^2$. The tensor product of a conjugate pair of modules $\mathcal{V}_\phi$ and $\mathcal{V}_{\bar \phi}$ has one highest-weight state for each nonnegative integer spin $j$ (and no other highest-weight states), $$\label{eq:conjugateproduct}
\mathcal{V}_\phi \otimes \mathcal{V}_{\bar \phi} = \sum_{j=0}^\infty \mathcal{V}_j.$$ Similarly, a like pair of modules has has one highest-weight state with spin $(j-1/2)$ for each nonnegative integer $j$, $$\label{eq:likeproduct}
\mathcal{V}_\phi \otimes \mathcal{V}_\phi = \sum_{j=0}^\infty \mathcal{V}_{j-1/2}, \quad \mathcal{V}_{\bar \phi} \otimes \mathcal{V}_{\bar \phi} = \sum_{j=0}^\infty \mathcal{V}_{j-1/2}.$$
Restriction to the ${\mathfrak{osp}}(4|2)$ sector \[sec:restrictosp42\]
-----------------------------------------------------------------------
Consider the set of states that at leading order are annihilated by ${\mathfrak{Q}}_{12,2}$ and ${\mathfrak{S}}^{12,2}$, which is a $1/12$ BPS condition[^6]. According to the last algebra relation of (\[eq:anticommutators\]), they satisfy $$R^1{}_1 + R^2{}_2 - L^2{}_2 - D_0 = 0, \label{eq:1/12bps}$$ where ordinary font denotes the eigenvalue of the corresponding generator in Gothic font, and $D_0$ is the classical dimension of a state. Since states only mix with other states with the same Lorentz and ${\mathfrak{R}}$ symmetry quantum numbers, and we choose a renormalization scheme where only states with the same classical dimension mix, it follows that this set of states is closed to all orders in perturbation theory.
Under this restriction, ${\mathfrak{osp}}(6|4)$ reduces to the set of generators that commute with the left side of (\[eq:1/12bps\]), and from the above algebra relations we find a residual ${\mathfrak{u}}(1) \ltimes {\mathfrak{su}}(1|1) \times {\mathfrak{osp}}(4|2)$ algebra. Since at leading order the ${\mathfrak{su}}(1|1)$ algebra (generated by ${\mathfrak{Q}}_{12,2}$ and ${\mathfrak{S}}^{12,2}$) acts trivially, we call this sector the ${\mathfrak{osp}}(4|2)$ sector. $ {\mathfrak{osp}}(4|2)$ includes a ${\mathfrak{sl}}(2)$ subalgebra of ${\mathfrak{sp}}(4)$, two ${\mathfrak{su}}(2)$ algebras from the original ${\mathfrak{su}}(4)$ ${\mathfrak{R}}$ symmetry, and eight supercharges. More precisely, and introducing a convenient notation, the ${\mathfrak{osp}}(4|2)$ generators are related to those for the full ${\mathfrak{osp}}(6|4)$ theory as (all index variables run from 1 to 2) $$\begin{aligned}
{\mathfrak{J}}^{11} & = {{{\textstyle\frac{1}{2}}}}{\mathfrak{P}}_{11},
&
{\mathfrak{J}}^{22} & = {{{\textstyle\frac{1}{2}}}}{\mathfrak{K}}^{11},
&
{\mathfrak{J}}^{12} & = -{{\textstyle\frac{1}{2}}} {\mathfrak{L}} + {\mathfrak{D}} + {{{\textstyle\frac{1}{2}}}}\delta{\mathfrak{D}},
\notag \\
{\mathfrak{R}}^{ab}& = {{{\textstyle\frac{1}{2}}}}\varepsilon^{ac} {\mathfrak{R}}^b_c +{{{\textstyle\frac{1}{2}}}}\varepsilon^{bc} {\mathfrak{R}}^a_c,
&
\tilde{{\mathfrak{R}}}^{\mathfrak{ab}} & = {{{\textstyle\frac{1}{2}}}}\varepsilon^{\mathfrak{ac}} {\mathfrak{R}}^{\mathfrak{b}+2}_{\mathfrak{c}+2} +{{{\textstyle\frac{1}{2}}}}\varepsilon^{\mathfrak{bc}} {\mathfrak{R}}^{\mathfrak{a}+2}_{\mathfrak{c}+2}, & &
\notag \\
{\mathfrak{Q}}^{a1\mathfrak{b}} & = \varepsilon^{ac}\varepsilon^{\mathfrak{bd}} {\mathfrak{Q}}_{c(\mathfrak{d}+2),1},
&
{\mathfrak{Q}}^{a2\mathfrak{b}} & = - {\mathfrak{S}}^{a(\mathfrak{b}+2),1}. & & \label{eq:osp42generators}\end{aligned}$$ Here we have also introduced ${\mathfrak{L}}$, which generates an additional ${\mathfrak{u}}(1)$ and is related to ${\mathfrak{osp}}(6|4)$ generators as $${\mathfrak{L}} = {\mathfrak{R}}^1_1 + {\mathfrak{R}}^2_2 = -{\mathfrak{L}}^1_1 + {\mathfrak{D}}_0, \label{eq:definel}$$ The second equality is satisfied within this sector due to (\[eq:1/12bps\]). ${\mathfrak{L}}$ commutes with the ${\mathfrak{osp}}(4|2)$ generators. As we will see below, due to the restricted field content in this sector ${\mathfrak{L}}$ simply gives half of the number of spin-chain sites (the number of pairs of conjugate representations).
The rank-one subalgebras take the standard form $${[J^{AB},J^{CD}]} = \varepsilon^{CB}J^{AD} - \varepsilon^{AD}J^{CB}, \quad {[J^{AB},X^{C}]} = {{{\textstyle\frac{1}{2}}}}\varepsilon^{CB}X^{A} + {{{\textstyle\frac{1}{2}}}}\varepsilon^{CA}X^{B}, \label{eq:rankonealgebras}$$ and the anticommutators are $${\{{\mathfrak{Q}}^{a\beta\mathfrak{c}},{\mathfrak{Q}}^{d\epsilon\mathfrak{f}}\}} = -\varepsilon^{\beta\epsilon}\varepsilon^{\mathfrak{cf}} {\mathfrak{R}}^{ad} -\varepsilon^{ad}\varepsilon^{\beta\epsilon} \tilde{{\mathfrak{R}}}^{\mathfrak{cf}} + 2 \varepsilon^{ad}\varepsilon^{\mathfrak{cf}} {\mathfrak{J}}^{\beta\epsilon}. \label{eq:osp42anticom}$$ We use hatted notation for the ${\mathfrak{su}}(1|1)$ algebra supercharges (note the reversed order of ${\mathfrak{R}}$ indices for the second one), $$\lambda \hat{{\mathfrak{Q}}} = {\mathfrak{Q}}_{12,2}, \quad \lambda \hat{{\mathfrak{S}}} = {\mathfrak{S}}^{21,2}.$$ In contrast to the introduction, we have now included a factor of $\lambda$ in the definitions since these generators act nontrivially first at $\mathcal{O}(\lambda)$ in this sector. $\hat{{\mathfrak{Q}}}$ and $\hat{{\mathfrak{S}}}$ are nilpotent and the only nonvanishing (anti)commutator for ${\mathfrak{su}}(1|1)$ is $$\label{eq:qhatshatacomm}
{\{\hat{{\mathfrak{Q}}},\hat{{\mathfrak{S}}}\}} = \frac{1}{\lambda^2} \delta {\mathfrak{D}} = \mathcal{H}.$$ Here we have introduced $\mathcal{H}$, the anomalous part of the dilatation generator divided by $\lambda^2$. We will also normalize the (one-loop) supercharges so that $$\hat{{\mathfrak{S}}} = (\hat{{\mathfrak{Q}}})^\dagger.$$ We can do this since only the product of their overall normalizations appears in the Hamiltonian.
Again, ${\mathfrak{su}}(1|1)$ and ${\mathfrak{osp}}(4|2)$ commute, and therefore $\delta {\mathfrak{D}}$ is a shared central charge. The ${\mathfrak{u}}(1)$ length generator ${\mathfrak{L}}$ satisfies $$\label{eq:lcomm}
{[{\mathfrak{L}},\hat{{\mathfrak{Q}}}]} = \hat{{\mathfrak{Q}}}, \quad {[{\mathfrak{L}},\hat{{\mathfrak{S}}}]} = -\hat{{\mathfrak{S}}}.$$ In this work we will only consider the leading (nonvanishing) contributions to the generators. So after this section, ${\mathfrak{Q}}$, ${\mathfrak{J}}$, $\hat{{\mathfrak{Q}}}$, $\hat{{\mathfrak{S}}}$ and $\mathcal{H}$ will refer only to $\mathcal{O}(\lambda^0)$ terms. We will still explicitly refer to the order of the leading anomalous piece of the dilatation generator, $\delta {\mathfrak{D}}_2$.
The ${\mathfrak{osp}}(4|2)$ module and leading-order representation
-------------------------------------------------------------------
The two modules of the ${\mathfrak{osp}}(4|2)$ sector $\mathcal{V}_\phi^{(4|2)}$ and $\mathcal{V}_{\bar \phi}^{(4|2)}$ can be obtained by acting with the generators (\[eq:osp42generators\]) on the highest weight states ${\mathopen{\big|}\phi_1^{(0,0)}\mathclose{\bigr \rangle}}$ and ${\mathopen{\big|}(\bar{\phi}^4)^{(0,0)}\mathclose{\bigr \rangle}}$. From (\[eq:rsymaction\] - \[eq:qaction\]), we conclude states in this sector have lower ${\mathfrak{R}}$ indices $1$ or $2$ or upper indices $3$ or $4$. Then, using (\[eq:lorentanddilatationaction\]), we see that (\[eq:1/12bps\]) implies that all states in this sector have only nonzero values for the first Lorentz excitation number, so we can use a single argument for the Lorentz indices of states as $${\mathopen{\big|}\phi_a^{(n)}\mathclose{\bigr \rangle}} = {\mathopen{\big|}\phi_a^{(2n, 0)}\mathclose{\bigr \rangle}} \sim \mathcal{D}_{11}^{n} \phi_a, \quad {\mathopen{\big|}\psi_a^{(n)}\mathclose{\bigr \rangle}} = {\mathopen{\big|}\psi_a^{(2n+1, 0)}\mathclose{\bigr \rangle}} \sim \mathcal{D}_{11}^{n} \psi_{1a}, \quad a=1,2.$$ Again, the conjugate fields have upper ${\mathfrak{R}}$ indices $3$, $4$ in the notation for the full theory; it is convenient to replace them with lower indices $\mathfrak{a,b}=1,2$, resulting in [|\_\^[(n)]{}]{} [[&=&]{}]{}\_ [(|\^[+2]{})\^[(2n, 0)]{}]{} \~\_ \_[11]{}\^[n]{}|\^[+2]{}, [\
]{} [|\_\^[(n)]{}]{} [[&=&]{}]{}\_ [(|\^[+2]{})\^[(2n+1, 0)]{}]{} \~\_ \_[11]{}\^[n]{}|\^[+2]{}\_1. (\[eq:definel\]) and (\[eq:lorentanddilatationaction\]) now imply in this sector that any state ${\mathopen{\big|}X\mathclose{\bigr \rangle}}$ satisfies $${\mathfrak{L}} {\mathopen{\big|}X\mathclose{\bigr \rangle}} = {{{\textstyle\frac{1}{2}}}}{\mathopen{\big|}X\mathclose{\bigr \rangle}}.$$ In other words, ${\mathfrak{L}}$ gives $L$ on an ${\mathfrak{osp}}(4|2)$ spin chain state of $2 L$ sites, as mentioned previously.
In the just-introduced notation, (\[eq:rsymaction\]) implies that the ${\mathfrak{su}}(2)$ subalgebras’ generators ${\mathfrak{R}}$ act canonically on the the unbarred states, and the $\tilde{{\mathfrak{R}}}$ act the same way on barred states, $${\mathfrak{R}}^{ab}{\mathopen{\big|}X_c\mathclose{\bigr \rangle}} = {{{\textstyle\frac{1}{2}}}}\delta^a_c \varepsilon^{bd} {\mathopen{\big|}X_d\mathclose{\bigr \rangle}} + {{{\textstyle\frac{1}{2}}}}\delta^b_c \varepsilon^{ad} {\mathopen{\big|}X_d\mathclose{\bigr \rangle}}, \quad \tilde{{\mathfrak{R}}}^{\mathfrak{ab}}{\mathopen{\big|}\bar{X}_{\mathfrak{c}}\mathclose{\bigr \rangle}} = {{{\textstyle\frac{1}{2}}}}\delta^{\mathfrak{a}}_{\mathfrak{c}} \varepsilon^{\mathfrak{bd}} {\mathopen{\big|}\bar{X}_{\mathfrak{d}}\mathclose{\bigr \rangle}} + {{{\textstyle\frac{1}{2}}}}\delta^{\mathfrak{b}}_{\mathfrak{c}} \varepsilon^{\mathfrak{ad}} {\mathopen{\big|}\bar{X}_{\mathfrak{d}}\mathclose{\bigr \rangle}}.$$ Closure of the algebra (\[eq:rankonealgebras\] - \[eq:osp42anticom\]) fixes the supercharges’ action (up to physically irrelevant possible changes of basis) as $$\begin{aligned}
{\mathfrak{Q}}^{a1\mathfrak{b}} {\mathopen{\big|}\phi_c^{(n)}\mathclose{\bigr \rangle}} & = \delta^a_c \varepsilon^{\mathfrak{bd}} \sqrt{2n + 1} {\mathopen{\big|}\bar{\psi}^{(n)}_{\mathfrak{d}}\mathclose{\bigr \rangle}}, & {\mathfrak{Q}}^{a1\mathfrak{b}} {\mathopen{\big|}\psi_c^{(n)}\mathclose{\bigr \rangle}} & = \delta^a_c \varepsilon^{\mathfrak{bd}} \sqrt{2n + 2} {\mathopen{\big|}\bar{\phi}^{(n+1)}_{\mathfrak{d}}\mathclose{\bigr \rangle}},
\notag \\
{\mathfrak{Q}}^{a2\mathfrak{b}} {\mathopen{\big|}\phi_c^{(n)}\mathclose{\bigr \rangle}} & = \delta^a_c \varepsilon^{\mathfrak{bd}} \sqrt{2n } {\mathopen{\big|}\bar{\psi}^{(n-1)}_{\mathfrak{d}}\mathclose{\bigr \rangle}}, & {\mathfrak{Q}}^{a2\mathfrak{b}} {\mathopen{\big|}\psi_c^{(n)}\mathclose{\bigr \rangle}} & = \delta^a_c \varepsilon^{\mathfrak{bd}} \sqrt{2n + 1} {\mathopen{\big|}\bar{\phi}^{(n)}_{\mathfrak{d}}\mathclose{\bigr \rangle}}
\notag \\
{\mathfrak{Q}}^{a1\mathfrak{b}} {\mathopen{\big|}\bar{\phi}_{\mathfrak{c}}^{(n)}\mathclose{\bigr \rangle}} & = - \delta^\mathfrak{b}_\mathfrak{c} \varepsilon^{ad} \sqrt{2n + 1} {\mathopen{\big|}\psi^{(n)}_{d}\mathclose{\bigr \rangle}}, & {\mathfrak{Q}}^{a1\mathfrak{b}} {\mathopen{\big|}\bar{\psi}_{\mathfrak{c}}^{(n)}\mathclose{\bigr \rangle}} & = - \delta^\mathfrak{b}_\mathfrak{c} \varepsilon^{ad} \sqrt{2n + 2} {\mathopen{\big|}\phi^{(n+1)}_{d}\mathclose{\bigr \rangle}},
\notag \\
{\mathfrak{Q}}^{a2\mathfrak{b}} {\mathopen{\big|}\bar{\phi}_{\mathfrak{c}}^{(n)}\mathclose{\bigr \rangle}} & = - \delta^\mathfrak{b}_\mathfrak{c} \varepsilon^{ad} \sqrt{2n } {\mathopen{\big|}\psi^{(n-1)}_{d}\mathclose{\bigr \rangle}}, & {\mathfrak{Q}}^{a2\mathfrak{b}} {\mathopen{\big|}\bar{\psi}_{\mathfrak{c}}^{(n)}\mathclose{\bigr \rangle}} & = - \delta^\mathfrak{b}_\mathfrak{c} \varepsilon^{ad} \sqrt{2n + 1} {\mathopen{\big|}\phi^{(n)}_{d}\mathclose{\bigr \rangle}}.\end{aligned}$$ Note the symmetry between the supercharge actions on barred and unbarred states, only differing by a minus sign and appropriate interchanges of Gothic and Latin indices. Finally, the action of the ${\mathfrak{J}}$ (independent of ${\mathfrak{R}}$ and $\tilde{{\mathfrak{R}}}$ indices) is the same for barred and unbarred states. Representing both $\phi_a$ and $\bar{\phi}_{\mathfrak{a}}$ with $\phi$, and $\psi_a$ and $\bar{\psi}_{\mathfrak{a}}$ with $\psi$, we have $$\begin{aligned}
{\mathfrak{J}}^{11} {\mathopen{\big|}\phi^{(n)}\mathclose{\bigr \rangle}} & = \sqrt{(n+{{{\textstyle\frac{1}{2}}}})(n+1)} {\mathopen{\big|}\phi^{(n+1)}\mathclose{\bigr \rangle}}, & {\mathfrak{J}}^{11} {\mathopen{\big|}\psi^{(n)}\mathclose{\bigr \rangle}} & = \sqrt{(n+1)(n+{{\textstyle\frac{3}{2}}})} {\mathopen{\big|}\psi^{(n+1)}\mathclose{\bigr \rangle}},
\notag \\
{\mathfrak{J}}^{22} {\mathopen{\big|}\phi^{(n)}\mathclose{\bigr \rangle}} & = \sqrt{(n-{{{\textstyle\frac{1}{2}}}})n} {\mathopen{\big|}\phi^{(n-1)}\mathclose{\bigr \rangle}}, & {\mathfrak{J}}^{22} {\mathopen{\big|}\psi^{(n)}\mathclose{\bigr \rangle}} & = \sqrt{n(n+{{{\textstyle\frac{1}{2}}}})} {\mathopen{\big|}\psi^{(n-1)}\mathclose{\bigr \rangle}},
\notag \\
{\mathfrak{J}}^{12} {\mathopen{\big|}\phi^{(n)}\mathclose{\bigr \rangle}} & = (n +{{\textstyle\frac{1}{4}}}) {\mathopen{\big|}\phi^{(n)}\mathclose{\bigr \rangle}}, & {\mathfrak{J}}^{12} {\mathopen{\big|}\psi^{(n)}\mathclose{\bigr \rangle}} & = (n +{{\textstyle\frac{3}{4}}}) {\mathopen{\big|}\psi^{(n)}\mathclose{\bigr \rangle}}.\end{aligned}$$ In the basis we have chosen, Hermiticity is manifest, $$({\mathfrak{Q}}^{a1 \mathfrak{b}})^\dagger = \varepsilon_{ac}\varepsilon_{\mathfrak{bd}}{\mathfrak{Q}}^{c2\mathfrak{d}}, \quad ({\mathfrak{J}}^{11})^\dagger = {\mathfrak{J}}^{22}, \quad ({\mathfrak{J}}^{12})^\dagger = {\mathfrak{J}}^{12}, \quad ({\mathfrak{X}}^{AB})^\dagger = \varepsilon_{AD}\varepsilon_{CB} {\mathfrak{X}}^{CD},$$ where ${\mathfrak{X}}$ denotes ${\mathfrak{R}}$ or $\tilde{{\mathfrak{R}}}$.
The quadratic Casimir for this sector is $${\mathfrak{J}}^2=-{{{\textstyle\frac{1}{4}}}}\varepsilon_{ad}\varepsilon_{cb} {\mathfrak{R}}^{ab}{\mathfrak{R}}^{cd} -{{{\textstyle\frac{1}{4}}}}\varepsilon_{\mathfrak{ad}}\varepsilon_{\mathfrak{cb}} \tilde{{\mathfrak{R}}}^{\mathfrak{ab}}{{\mathfrak{R}}}^{\mathfrak{cd}} + {{{\textstyle\frac{1}{2}}}}\varepsilon_{\alpha \delta}\varepsilon_{\gamma \beta} {\mathfrak{J}}^{\alpha \beta} {\mathfrak{J}}^{\gamma \delta} -{{{\textstyle\frac{1}{4}}}}\varepsilon_{ad}\varepsilon_{\beta \epsilon } \varepsilon_{\mathfrak{cd}} {\mathfrak{Q}}^{a\beta\mathfrak{c}}{\mathfrak{Q}}^{d\epsilon\mathfrak{d}}. \label{eq:defineosp42casimir}$$ It follows from the algebra commutation relations that highest-weight states, which are annihilated by ${\mathfrak{R}}^{22}$, $\tilde{{\mathfrak{R}}}^{22}$, ${\mathfrak{J}}^{22}$ and ${\mathfrak{Q}}^{a2\mathfrak{b}}$, have quadratic Casimir eigenvalue $$\label{eq:twositeCartan}
J^2 = - {{{\textstyle\frac{1}{2}}}}R^{12}(R^{12}-1)- {{{\textstyle\frac{1}{2}}}}\tilde{R}^{12}(\tilde{R}^{12} - 1) + J^{12}(J^{12} + 1).$$ As for ${\mathfrak{osp}}(6|4)$, a eigenstate of the quadratic Casimir has ${\mathfrak{osp}}(4|2)$ spin $j$ given by $J^2 = j(j+1)$.
It is a straightforward exercise to find the highest-weight states for two-site states. For a conjugate pair of modules, there is one highest-weight state (and irreducible highest-weight ${\mathfrak{osp}}(4|2)$ module) for each nonnegative integer spin $j$, matching precisely the result previously given for the full ${\mathfrak{osp}}(6|4)$ modules (\[eq:conjugateproduct\]), $$\label{eq:42conjugateproduct}
\mathcal{V}_\phi^{(4|2)} \otimes \mathcal{V}_{\bar \phi}^{(4|2)} = \sum_{j=0}^\infty \mathcal{V}^{(4|2)}_j.$$ The corresponding Cartan charges are $$\label{eq:twositedifferenthigestweight}
[ R^{12}, J^{12}, \tilde{R}^{12} ] = [-{{{\textstyle\frac{1}{2}}}}, {{{\textstyle\frac{1}{2}}}}, -{{{\textstyle\frac{1}{2}}}}] \quad \text{and} \quad [ R^{12}, J^{12}, \tilde{R}^{12} ] = [0, j, 0] \quad j =1, 2, \ldots$$ We will also need the case of two identical modules (which would appear as two next-nearest neighbor sites on the alternating chain). Again there is precise agreement with the ${\mathfrak{osp}}(6|4)$ result (\[eq:likeproduct\]), $$\label{eq:42likeproduct}
\mathcal{V}_\phi^{(4|2)} \otimes \mathcal{V}^{(4|2)}_\phi = \sum_{j=0}^\infty \mathcal{V}^{(4|2)}_{j-1/2}, \quad \mathcal{V}^{(4|2)}_{\bar \phi} \otimes \mathcal{V}^{(4|2)}_{\bar \phi} = \sum_{j=0}^\infty \mathcal{V}^{(4|2)}_{j-1/2}.$$ and the ${\mathfrak{osp}}(4|2)$ Cartan charges are $$\label{eq:twositeidenticalhigestweight}
[-1, {{{\textstyle\frac{1}{2}}}}, 0] \quad \text{and} \quad [0, j + {{{\textstyle\frac{1}{2}}}}, 0], \quad j = 0, 1, \ldots$$ Finally, the ${\mathfrak{osp}}(4|2)$ subsector has ${\mathfrak{sl}}(2)$ subsector(s) in which only the ${\mathfrak{J}}$ act nontrivially[^7]. The two modules are spanned by $${\mathopen{\big|}\phi_1^{(n)}\mathclose{\bigr \rangle}} \quad \text{and} \quad {\mathopen{\big|}\psi_1^{(n)}\mathclose{\bigr \rangle}}.$$ The ${\mathfrak{sl}}(2)$ sector two-site highest-weight states (which are descendants in the ${\mathfrak{osp}}(4|2)$ sector) have ${\mathfrak{sl}}(2)$ spins that take the same value as for the larger ${\mathfrak{osp}}(4|2)$ sector, and therefore as in ${\mathfrak{osp}}(6|4)$ as well.
Light-cone superspace basis
---------------------------
We will find it very useful to use a light-cone superspace [@Kogut:1969xa; @Brink:1982pd; @Mandelstam:1982cb; @Belitsky:2004yg] basis for the modules[^8]. This basis parameterizes $\mathcal{V}_{\phi}^{(4|2)}$ ($\mathcal{V}_{\bar \phi}^{(4|2)}$) with continuous variables: $x$, a ${\mathfrak{su}}(2)$ doublet $\theta^a$ ($\bar{\theta}^{\mathfrak{a}}$), and an anticommuting ${\mathfrak{su}}(2)$ doublet $\bar{\eta}^{\mathfrak{a}}$ ($\eta^a$). The spin-chain states are labeled ${\mathopen{\big|}x, \theta, \bar{\eta}\mathclose{\bigr \rangle}}$, or ${\mathopen{\big|}x, \bar{\theta}, \eta\mathclose{\bigr \rangle}}$, which are defined via sums over the entire modules,
\[eq:definelightcone\] [x, , |]{} [[&=&]{}]{}\_[n=0]{}\^ [\_a\^[(n)]{}]{} + [|\_\^[(n)]{}]{}, [\
]{}[x, |, ]{} [[&=&]{}]{}\_[n=0]{}\^ [|\_\^[(n)]{}]{} + [\_a\^[(n)]{}]{}. In this basis the leading-order generators are represented by differential operators. As an example, we derive the representation for ${\mathfrak{Q}}^{a1\mathfrak{c}}$ acting on $\mathcal{V}_{\phi}^{(4|2)}$, \^[a1]{} [x, , |]{} [[&=&]{}]{}\_[n=0]{}\^ ( \^b \^[a1]{} [\_b\^[(n)]{}]{} - |\^ \^[a1]{} [\_\^[(n)]{}]{} ) [\
]{}[[&=&]{}]{} \_[n=0]{}\^ ( \^a \^ [|\_\^[(n)]{}]{} + \^[ab]{} |\^ [\_b\^[(n+1)]{}]{} ) [\
]{}[[&=&]{}]{}(\^a |\^ + \_x \^a |\^ ) [x, , |]{} . Our conventions for raising and lowering indices lead to, for $\mathcal{V}_{\phi}^{(4|2)}$, $$\begin{aligned}
\partial^{a} & = \varepsilon^{ab} \partial_{\theta^{b}} = \varepsilon^{ab} \partial_b, & \bar{\partial}^{\mathfrak{c}}& = \varepsilon^{\mathfrak{cd}} \partial_{\bar{\eta}^{\mathfrak{d}}} = \varepsilon^{\mathfrak{cd}} \bar{\partial}_{\mathfrak{d}}, \notag \\
\partial_{a} & = \partial^b \varepsilon_{ba}, & \partial_{\mathfrak{c}}& = \bar{\partial}^{\mathfrak{d}} \varepsilon_{\mathfrak{dc}} .\end{aligned}$$ For $\mathcal{V}_{\bar \phi}^{(4|2)}$, the same equations apply for $\theta$ replaced by $\eta$ and $\bar{\eta}$ replaced by $\bar{\theta}$. So, for example, $\partial^a$ is bosonic acting on $\mathcal{V}_{\phi}^{(4|2)}$, and fermionic on $\mathcal{V}_{\bar \phi}^{(4|2)}$. We repeat the above result for ${\mathfrak{Q}}^{a1\mathfrak{c}}$ and add the parallel expressions for the other supercharges and for $\mathcal{V}_{\bar \phi}^{(4|2)}$ (abbreviating the states by suppressing an $\eta$ or $\bar{\eta}$), $$\begin{aligned}
{\mathfrak{Q}}^{a1\mathfrak{c}} {\mathopen{\big|}x, \theta\mathclose{\bigr \rangle}} &=\Big(\theta^a \bar{\partial}^{\mathfrak{c}} + \partial_x \partial^a \bar{\eta}^{\mathfrak{c}} \Big) {\mathopen{\big|}x, \theta\mathclose{\bigr \rangle}} , & {\mathfrak{Q}}^{a2\mathfrak{c}} {\mathopen{\big|}x, \theta\mathclose{\bigr \rangle}} &=\Big(2 x \theta^a \bar{\partial}^{\mathfrak{c}} + (2 x \partial_x +1)\partial^a \bar{\eta}^{\mathfrak{c}} \Big) {\mathopen{\big|}x, \theta\mathclose{\bigr \rangle}} ,
\notag \\
{\mathfrak{Q}}^{a1\mathfrak{c}} {\mathopen{\big|}x, \bar{\theta}\mathclose{\bigr \rangle}} &=-\Big(\bar{\theta}^{\mathfrak{c}} \partial^{a} + \partial_x \bar{\partial}^{\mathfrak{c}} \eta^a \Big) {\mathopen{\big|}x, \bar{\theta}\mathclose{\bigr \rangle}} , & {\mathfrak{Q}}^{a2\mathfrak{c}} {\mathopen{\big|}x, \bar{\theta}\mathclose{\bigr \rangle}} &=-\Big(2 x \bar{\theta}^{\mathfrak{c}} \partial^{a} + (2 x \partial_x +1) \bar{\partial}^{\mathfrak{c}} \eta^a \Big) {\mathopen{\big|}x, \bar{\theta}\mathclose{\bigr \rangle}}.\end{aligned}$$ The remaining generators’ actions can be written the same way[^9] for both types of modules, $$\begin{aligned}
\label{eq:osp42lightcone}
{\mathfrak{R}}^{ab} &= \theta^{\{a} \partial^{b\}} + \eta^{\{a} \partial^{b\}}, & \bar{{\mathfrak{R}}}^{\mathfrak{cd}} &= \bar{\theta}^{\{\mathfrak{c}} \partial^{\mathfrak{d}\}} + \bar{\eta}^{\{\mathfrak{c}} \partial^{\mathfrak{d}\}},
\notag \\
{\mathfrak{J}}^{11} & = \partial_x, & {\mathfrak{J}}^{22} & = 2 x^2 \partial_x + x - 2 x \Big(\varepsilon_{ab} \eta^a \partial^b + \varepsilon_{\mathfrak{cd}} \bar{\eta}^c \bar{\partial}^{\mathfrak{d}}\Big),
\notag \\
{\mathfrak{J}}^{12} & = x \partial_x + \frac{1}{4} -\frac{1}{2} \Big(\varepsilon_{ab} \eta^a \partial^b + \varepsilon_{\mathfrak{cd}} \bar{\eta}^c \bar{\partial}^{\mathfrak{d}}\Big). & &\end{aligned}$$
Leading corrections in the ${\mathfrak{osp}}(4|2)$ sector \[sec:oneloop\]
=========================================================================
We will now compute the leading actions of $\hat{{\mathfrak{Q}}}$ and $\hat{{\mathfrak{S}}}$, which immediately give the two-loop dilatation generator.
Structure of $\hat{{\mathfrak{Q}}}$ and $\hat{{\mathfrak{S}}}$ interactions
---------------------------------------------------------------------------
(\[eq:lcomm\]) implies that $\hat{{\mathfrak{Q}}}$ and $\hat{{\mathfrak{S}}}$ are dynamic; they change the length of the spin chain as do spin-chain generators in $\mathcal{N}=4$ SYM [@Beisert:2003ys]. $\hat{{\mathfrak{Q}}}$ inserts two sites, and $\hat{{\mathfrak{S}}}$ removes two sites. Note that this means these supercharges have well-defined actions only on cyclic states, which is all that is required since these spin-chain states represent single-trace local operators. At one-loop, consistency with the coupling constant dependence of the interactions of the Lagrangian requires $\hat{{\mathfrak{Q}}}$ to replace one site with three, and $\hat{{\mathfrak{S}}}$ to replace three sites with one. Let $\mathcal{U}$ be the generator that shifts all sites by two to the right (with the last site $2 L$ going to site $2$ for example). By definition a cyclic alternating state ${\mathopen{\big|}Y\mathclose{\bigr \rangle}}$ satisfies $\mathcal{U}{\mathopen{\big|}Y\mathclose{\bigr \rangle}}={\mathopen{\big|}Y\mathclose{\bigr \rangle}}$. $\hat{{\mathfrak{Q}}}$ then acts as $$\hat{{\mathfrak{Q}}} {\mathopen{\big|}Y\mathclose{\bigr \rangle}} = \frac{L}{L +1} \sum_{i=0}^{L} \mathcal{U}^{-i} \Big(\hat{{\mathfrak{Q}}}(1) + \hat{{\mathfrak{Q}}}(2)\Big){\mathopen{\big|}Y\mathclose{\bigr \rangle}} .$$ and $\hat{{\mathfrak{Q}}}(i)$ gives the action of $\hat{{\mathfrak{Q}}}$ the $i$th site, which we determine below. Similarly $$\hat{{\mathfrak{S}}} {\mathopen{\big|}Y\mathclose{\bigr \rangle}} = \frac{L}{L -1} \sum_{i=0}^{L-2} \mathcal{U}^{-i} \Big(\hat{{\mathfrak{S}}}(1,2,3) + \hat{{\mathfrak{S}}}(2, 3, 4)\Big){\mathopen{\big|}Y\mathclose{\bigr \rangle}} .$$ In both these expressions a minus signs must be included for each crossing of two fermions (or for a supercharge crossing a fermion).
Constraints from manifest ${\mathfrak{R}}$ and $\tilde{{\mathfrak{R}}}$ symmetries and consistency with the classical scaling dimension assignments further severely restrict the supercharge actions. For instance, acting on a scalar initial state, we (almost) immediately can restrict to an ansatz for the action of $\hat{{\mathfrak{Q}}}$ on one site, \[eq:qhatansatz\] [\_a\^[(m)]{}]{} [[&=&]{}]{}\_[n+p < m]{} ( c\_1(m, n, p) \^[bc]{} [\_a\^[(n)]{} \_b\^[(p)]{} \_c\^[(m-n-p-1)]{}]{} [\
&&]{}- c\_1’(m, n, p) \^[bc]{} [\_c\^[(m-n-p-1)]{} \_b\^[(p)]{} \_a\^[(n)]{}]{} ) [\
&&]{}+ \^ (c\_2(m, n, p) [\_a\^[(n)]{} |\_\^[(p)]{} |\_\^[(m-n-p-1)]{}]{} - c\_2’(m, n, p) [|\_\^[(m-n-p-1)]{} |\_\^[(p)]{} \_a\^[(n)]{}]{}). [\
&&]{} Here and below summations will be over all nonnegative integers satisfying an inequality. In this case we have $n=0, 1, \ldots m-1$ and $p = 0, 1, \ldots m-n-1$. This ansatz (\[eq:qhatansatz\]) is the most general one consistent with the constraints mentioned above and two further observations. First, note the ${\mathfrak{su}}(2)$ identity, $$\varepsilon^{bc} {\mathopen{\big|}X_bY_cZ_a\mathclose{\bigr \rangle}} + \varepsilon^{bc} {\mathopen{\big|}X_aY_bZ_c\mathclose{\bigr \rangle}}= \varepsilon^{bc} {\mathopen{\big|}X_bY_aZ_c\mathclose{\bigr \rangle}},$$ which eliminates the need for another term. Second, a priori a scalar to three fermions interaction would be allowed, but this turns out to be inconsistent with the requirement that $\hat{{\mathfrak{Q}}}$ commutes with all of the ${\mathfrak{osp}}(4|2)$ generators (one can check this using the same methods we will now use to determine the $c_i$).
Constraints from commutator with ${\mathfrak{J}}^{11}$
------------------------------------------------------
Rather than considering the commutators with all ${\mathfrak{osp}}(4|2)$ generators, let us first just consider $${[\hat{{\mathfrak{Q}}},{\mathfrak{J}}^{11}]}=0 \text{ on cyclic states.}$$ As initially noted for $\mathcal{N}=4$ SYM [@Beisert:2003ys], the fact that the algebra only must be satisfied on cyclic spin chain states allows for “gauge transformations”. Here, consistency with the basic constraints used above and with alternating modules restricts such a gauge transformation to \[eq:gauge\] [\[,\^[11]{}\]]{}[X]{} [[&=&]{}]{}(-1)\^X g\_1 \^ [X|\_\^[(0)]{}|\_\^[(0)]{}]{} - g\_2 \^ [|\_\^[(0)]{}|\_\^[(0)]{}X]{} [\
&&]{}+ (-1)\^X g\_3 \^[ab]{}[X\_a\^[(0)]{}\_b\^[(0)]{}]{} - g\_4 \^[ab]{}[\_a\^[(0)]{}\_b\^[(0)]{}X]{}, [\
]{}[\[,\^[11]{}\]]{}[|[X]{}]{} [[&=&]{}]{}(-1)\^[|[X]{}]{} g\_2 \^ [|[X]{}|\_\^[(0)]{}|\_\^[(0)]{}]{} - g\_1 \^ [|\_\^[(0)]{}|\_\^[(0)]{}|[X]{}]{} [\
&&]{}+ (-1)\^[|[X]{}]{} g\_4 \^[ab]{}[|[X]{}\_a\^[(0)]{}\_b\^[(0)]{}]{} - g\_3 \^[ab]{}[\_a\^[(0)]{}\_b\^[(0)]{}|[X]{}]{}. $X$ represents any element of $\mathcal{V}_{\phi}^{(4|2)}$, $\bar{X}$ represents any element in $\mathcal{V}_{\bar \phi}^{(4|2)}$, and $(-1)^X$ gives $(-1)$ for fermionic $X$ and $1$ otherwise. To see that this interaction gives zero on cyclic states, consider the two terms with coefficient $g_1$. The first $g_1$ term inserts an $\tilde{{\mathfrak{R}}}$ singlet to the right of $\mathcal{V}_{\phi}^{(4|2)}$ sites. But on alternating cyclic (or infinite) chains, this is canceled by the second $g_1$ term, which inserts the same $\tilde{{\mathfrak{R}}}$ singlet to the left of $\mathcal{V}_{\bar \phi}^{(4|2)}$ sites with a relative minus sign. The cancellations for the other $g_i$ terms work in the same way.
Since ${\mathfrak{J}}^{11}$ does not affect ${\mathfrak{R}}$ charges, we obtain independent equations for the four coefficient functions appearing in (\[eq:qhatansatz\]). For instance the unprimed $c_1$ terms contribute \[eq:j11constraint\] [\[,\^[11]{}\]]{}\_[c\_1]{} [\_a\^[(m)]{}]{} [[&=&]{}]{}\_[n+p m]{} ( c\_1(m+1, n, p) [\
&&]{}- c\_1(m, n-1, p) - c\_1(m, n, p-1) [\
&&]{} -c\_1(m, n, p))\^[bc]{} [\_a\^[(n)]{}\_b\^[(p)]{}\_c\^[(m-n-p)]{}]{} [\
]{} [[&=&]{}]{}g\_3 \^[bc]{} [\_a\^[(0)]{} \_b\^[(0)]{} \_c\^[(m)]{}]{}. From acting on a scalar with $m=0$, we find the only coefficient with first argument $1$, $c_1(1, 0,0) = \sqrt{2} g_3$. But now it is straightforward to see that all coefficients are determined inductively. Assume all coefficients with first argument less than or equal to $m_0$ are known. Then (\[eq:j11constraint\]) determines all coefficients with first argument $m_0+1$ in terms of these known coefficients and $g_3$. The solution for all arguments is[^10] $$c_1(m,n, p) = \frac{\sqrt{2} g_3r_-(m)}{(m-n)r_-(n)r_+(p)r_-(m-n-p-1)}, \quad r_{\pm}(x) = \frac{\sqrt{x!}}{\sqrt{(2 x \pm 1)!!}} .$$ Repeating for the other three coefficient functions entering (\[eq:qhatansatz\]), again yields solutions determined inductively from a single coefficient. These solutions depend in total on the four $g_i$ coefficients appearing in (\[eq:gauge\]). Furthermore, (\[eq:gauge\]) implies that these are the only free parameters for $\hat{{\mathfrak{Q}}}$ for acting on scalars in $\mathcal{V}_{\bar \phi}^{(4|2)}$. For $\hat{{\mathfrak{Q}}}$ acting on fermions, in addition to structures paralleling the four terms of (\[eq:qhatansatz\]), there is an additional fermion-to-three-scalar interaction that is possible, which commutes exactly with ${\mathfrak{J}}^{11}$. Therefore, the ${\mathfrak{J}}^{11}$ constraint allows for two more free parameter (fermions and conjugate fermions), in addition to the four $g_i$. However, there are more constraints.
Anticommutator with ${\mathfrak{Q}}^{a1\mathfrak{b}}$ and solution for $\hat{{\mathfrak{Q}}}$ and $\hat{{\mathfrak{S}}}$
------------------------------------------------------------------------------------------------------------------------
Of course the ${\mathfrak{osp}}(4|2)$ supercharges relate coefficients unrelated by the ${\mathfrak{J}}^{11}$. Then requiring the anticommutator with ${\mathfrak{Q}}^{a1\mathfrak{b}}$ to vanish on cyclic states fixes all of the above independent coefficients to be proportional to a single free parameter; $\hat{{\mathfrak{Q}}}$ is determined by symmetry up to overall normalization. This free parameter is fixed by any single nonvanishing two-loop anomalous dimension, since $\hat{{\mathfrak{S}}} = (\hat{{\mathfrak{Q}}})^\dagger$, and (\[eq:qhatshatacomm\]) ${\{\hat{{\mathfrak{Q}}},\hat{{\mathfrak{S}}}\}} = \delta {\mathfrak{D}}_2$. The correct choice turns out to be $g_3=1/2$, as we will see below. With this normalization, the gauge transformation for the anticommutator between supercharges is \[eq:qqhatgauge\] [{,\^[a1]{}}]{}[X]{} [[&=&]{}]{}\^[ac]{} \^([\_c\^[(0)]{}|\^[(0)]{}\_X]{}- [X|\^[(0)]{}\_\_c\^[(0)]{}]{} ) , [\
]{}[{,\^[a1]{}}]{}[|[X]{}]{} [[&=&]{}]{}\^[ac]{} \^( [|\^[(0)]{}\_\_c\^[(0)]{}|[X]{}]{} -[|[X]{}\_c\^[(0)]{}|\^[(0)]{}\_]{} ). The corresponding complete solution for $\hat{{\mathfrak{Q}}}$ acting on $\mathcal{V}_{\phi}^{(4|2)}$ depends on three coefficient functions $c_i$ (note that $c_1$ is as previously, but $c_2$ is different from the $c_2$ appearing in (\[eq:qhatansatz\])), \[eq:qhatsolution\] [\_a\^[(m)]{}]{} [[&=&]{}]{}\_[n+p < m]{} c\_1(m, n, p) \^[bc]{} ( [\_a\^[(n)]{} \_b\^[(p)]{} \_c\^[(m-n-p-1)]{}]{} - [\_c\^[(m-n-p-1)]{} \_b\^[(p)]{} \_a\^[(n)]{}]{} ) [\
&&]{}+ c\_1(m, n, m-n-p-1) \^ ( [\_a\^[(n)]{} |\_\^[(p)]{} |\_\^[(m-n-p-1)]{}]{} - [|\_\^[(m-n-p-1)]{} |\_\^[(p)]{} \_a\^[(n)]{}]{}), [\
]{}[|\_\^[(m)]{}]{} [[&=&]{}]{}\_[n+p < m]{} c\_2(m, n, p) \^[bc]{} ( [|\_\^[(n)]{} \_b\^[(p)]{} \_c\^[(m-n-p-1)]{}]{} + [\_c\^[(m-n-p-1)]{} \_b\^[(p)]{} |\_\^[(n)]{}]{} ) [\
&&]{}+ c\_2(m, n, m-n-p-1) \^ ( [|\_\^[(n)]{} |\_\^[(p)]{} |\_\^[(m-n-p-1)]{}]{} + [|\_\^[(m-n-p-1)]{} |\_\^[(p)]{} |\_\^[(n)]{}]{}) [\
&&]{}+ \_[n+p m]{} c\_3(m, n, p) \^[bc]{} [\_b\^[(n)]{}|\_\^[(p)]{}\_c\^[(m-n-p)]{}]{}. The action on $\mathcal{V}_{\bar \phi}^{(4|2)}$ simply involves switching barred and unbarred module elements and an overall minus sign, \[eq:qhatsolutionconjugate\] [|\_\^[(m)]{}]{} [[&=&]{}]{}-\_[n+p < m]{} c\_1(m, n, p)\^( [|\_\^[(n)]{} |\_\^[(p)]{} |\_\^[(m-n-p-1)]{}]{} - [|\_\^[(m-n-p-1)]{} |\_\^[(p)]{} |\_\^[(n)]{}]{} ) [\
&&]{}+ c\_1(m, n, m-n-p-1) \^[bc]{} ( [|\_\^[(n)]{} \_b\^[(p)]{} \_c\^[(m-n-p-1)]{}]{} - [\_c\^[(m-n-p-1)]{} \_b\^[(p)]{} |\_\^[(n)]{}]{}), [\
]{}[\_a\^[(m)]{}]{} [[&=&]{}]{}-\_[n+p < m]{} (c\_2(m, n, p)\^( [\_a\^[(n)]{} |\_\^[(p)]{} |\_\^[(m-n-p-1)]{}]{} + [|\_\^[(m-n-p-1)]{} |\_\^[(p)]{} \_a\^[(n)]{}]{} ) [\
&&]{}+ c\_2(m, n, m-n-p-1) \^[bc]{} ( [\_a\^[(n)]{} \_b\^[(p)]{} \_c\^[(m-n-p-1)]{}]{} + [\_c\^[(m-n-p-1)]{} \_b\^[(p)]{} \_a\^[(n)]{}]{}) ) [\
&&]{}- \_[n+p m]{} c\_3(m, n, p) \^ [|\_\^[(n)]{}\_a\^[(p)]{}|\_\^[(m-n-p)]{}]{}. With the normalization $g_3=1/2$, $c_1$ becomes $$c_1(m,n, p) = \frac{ r_-(m)}{\sqrt{2}(m-n)r_-(n)r_+(p)r_-(m-n-p-1)}.$$ The next coefficient function takes a very similar form, just switching some $r_+$ and $r_-$ and an overall minus sign, $$c_2(m,n, p) = -\frac{ r_+(m)}{\sqrt{2}(m-n)r_+(n)r_+(p)r_-(m-n-p-1)}.$$ Finally, the last coefficient function takes a similar form, without the $\sqrt{2}(m-n)$ factor in the denominator, $$c_3(m,n, p) = -\frac{ r_+(m)}{r_-(n)r_-(p)r_-(m-n-p)}.$$ Initial (final) fermions lead to $r_+$ factors in the numerator (denominator) and initial (final) scalars to $r_-$ factors in the numerator (denominator).
Recall that the ${\mathfrak{su}}(1|1)$ supercharges are nilpotent. One could check that $\hat{{\mathfrak{Q}}}^2=0$ by working out all of its one-to-five site interactions in terms of the $c_i$. However, this is redundant; a few basic facts about the spin modules plus the vanishing commutators (up to gauge transformations) of $\hat{{\mathfrak{Q}}}$ with the ${\mathfrak{osp}}(4|2)$ generators already ensure that $\hat{{\mathfrak{Q}}}^2=0$. We leave it as an exercise for the reader to work out this argument.
Manifest Hermiticity of the leading ${\mathfrak{osp}}(4|2)$ generators implies that above we could have considered $\hat{{\mathfrak{S}}}$ instead of $\hat{{\mathfrak{Q}}}$, and we would have then found a unique solution to $\hat{{\mathfrak{S}}}$ (up to normalization). This was anticipated by the previously stated equality $\hat{{\mathfrak{S}}} = (\hat{{\mathfrak{Q}}})^\dagger$. Using this equality it is straightforward to work out the interactions of $\hat{{\mathfrak{S}}}$ by switching initial and final states of (\[eq:qhatsolution\]-\[eq:qhatsolutionconjugate\]). For instance, the first line of (\[eq:qhatsolution\]) implies [\_a\^[(m)]{}\_b\^[(n)]{}\_c\^[(p)]{}]{} [[&=&]{}]{}c\_1(m+n+p+1, m, n)\_[bc]{}[\_a\^[(m+n+p+1)]{}]{} [\
&&]{}+c\_1(m+n+p+1, p, n)\_[ab]{}[\_c\^[(m+n+p+1)]{}]{}. The complete solution for $\hat{{\mathfrak{S}}}$ is given in the light-cone basis in Appendix \[sec:lightcone\].
From the expression for $\hat{{\mathfrak{Q}}}$, we see that the (two-loop) ${\mathfrak{osp}}(4|2)$ sector has an additional discrete symmetry under *spin-chain* parity $\mathbf{p}$, which reverses the order of the spin chain sites with an extra signs for each crossing of fermions. After application of parity each site will have the opposite type of representation. Note that $\mathbf{p}$ is distinct from space-time parity, and, unlike the case of $\mathcal{N}=4$ SYM, this operation is distinct also from charge conjugation symmetry. While the leading order ${\mathfrak{osp}}(4|2)$ generators are (trivially) parity even, $\hat{{\mathfrak{Q}}}$ and $\hat{{\mathfrak{S}}}$ are parity odd[^11].
Solution for $\hat{{\mathfrak{Q}}}$ in light-cone superspace basis
------------------------------------------------------------------
In the light-cone basis the $r_{\pm}$ factors are absorbed into the normalization of the states in the expansion of ${\mathopen{\big|}x, \theta, \bar{\eta}\mathclose{\bigr \rangle}}$, while the $(m-n)$ factors in the denominators can be accounted for with an integral as \[eq:lightconeqhat\] [x, , |]{} [[&=&]{}]{}\_0\^x y ( \_[ab]{} \^a\_2 \^b\_3 + \_ \^\_2 \^\_3 ) [x, , |; y, |\_2, \_2;y, \_3, |\_3]{} [\
&&]{} + \_0\^x y ( \_[ab]{} \^a\_1 \^b\_2 + \_ \_1\^ \_2\^ ) [ y, \_1, |\_1; y, |\_2, \_2; x, , |]{} [\
&&]{} -\_ |\^ \_2\^ \_[cd]{} \_1\^c\_3\^d [ x, \_1, |\_1;x, |\_2, \_2; x, \_3, |\_3 ]{} . Here the subscripts label on which site the partial derivatives act. For example, in the first term $\partial_2^a=\varepsilon^{ac} \partial/\partial \eta_2^c$. The expression for acting on $\mathcal{V}_{\bar \phi}^{(4|2)}$ just follows from switching ${\mathfrak{su}}(2)$ indices in epsilon tensors and in the derivatives ($\partial^a \leftrightarrow \partial^{\mathfrak{a}}$), and switching all states with their conjugates (${\mathopen{\big|}x, \theta, \bar{\eta}\mathclose{\bigr \rangle}} \leftrightarrow {\mathopen{\big|}x, \bar{\theta}, \eta\mathclose{\bigr \rangle}}$). Also, there is an extra overall minus sign for acting on $\mathcal{V}_{\bar \phi}^{(4|2)}$. As an example, we check one term. Expanding the $\theta$ component of the first term of the first line of (\[eq:lightconeqhat\]) we find \[eq:checklightconeqhat\] [x, , |]{}\_ [[&=&]{}]{}\_0\^x y \_[ab]{} \^a\_2 \^b\_3 [x, , |; y, |\_2, \_2;y, \_3, |\_3]{} \_ [\
]{}[[&=&]{}]{}\_0\^x y \_[m\_1, m\_2,m\_3]{} x\^[m\_1]{}y\^[m\_2+m\_3]{} \^a \^[bc]{} [\_a\^[(m\_1)]{}\_b\^[(m\_2)]{}\_c\^[(m\_3)]{}]{} [\
]{}[[&=&]{}]{}\_[n+p < m]{} \^[bc]{} [\_a\^[(n)]{}\_b\^[(p)]{}\_c\^[(m-n-p-1)]{}]{} [\
]{}[[&=&]{}]{}\_m \^a \_[n+p < m]{} c\_1(m, n, p) \^[bc]{} [\_a\^[(n)]{}\_b\^[(p)]{}\_c\^[(m-n-p-1)]{}]{}. To reach the third line we did the integral, substituted for $m_3$ using $m=m_1+m_2+m_3+1$, and then replaced $m_1,m_2$ with $n, p$. The combinatoric factor simplifies to $\sqrt{(2m)!}c_1(m, n, p)/m!$, yielding the last line after reordering factors. On the other hand, acting with the first term of the first line of (\[eq:qhatsolution\]) on the left side of (\[eq:checklightconeqhat\]) clearly leads to the same result, as needed. One can the check remaining terms, involving also those with initial fermions, in a similar fashion.
Because the light-cone basis is not manifestly Hermitian, $\hat{{\mathfrak{S}}}$ takes a more involved form (requiring integration over two auxiliary variables, rather than just one), again see Appendix \[sec:lightcone\].
Hamiltonian, wrapping interactions, and twist-one spectrum
----------------------------------------------------------
The two loop dilatation generator for the ${\mathfrak{osp}}(4|2)$ sector now follows from the anticommutator (\[eq:qhatshatacomm\]) $${\{\hat{{\mathfrak{Q}}},\hat{{\mathfrak{S}}}\}} = \delta {\mathfrak{D}}_2. \label{eq:ham}$$ Since $\hat{{\mathfrak{Q}}}$ and $\hat{{\mathfrak{S}}}$ are spin-chain parity odd, the Hamiltonian is parity even. We will not explicitly compute the expansion of the Hamiltonian in terms of interactions, since there is a simpler way to prove integrability. Still, we have used (\[eq:ham\]), (\[eq:qhatsolution\]-\[eq:qhatsolutionconjugate\]), and `Mathematica` to check the spectrum for many (cyclic) spin chain states of relatively low dimension and length, finding complete agreement with the Bethe ansatz predictions of [@Minahan:2008hf]. This is empirical confirmation of the integrability that we will prove in the next section.
It is important that there are physically equivalent expressions for the Hamiltonian which have different expansions in terms of local interactions. For periodic spin chains, including the cyclic spin chains of this work, there is always freedom to add chain derivatives to spin-chain generators including the Hamiltonian. Chain derivatives are nonzero interactions that vanish on periodic states. An example three-site chain derivative acts on a periodic spin chain of length $2L$ as $$\sum_{i=1}^{L} \Big({\mathfrak{L}}(2i-1){\mathfrak{X}}(2i,2i +1) -{\mathfrak{X}}(2i,2i +1){\mathfrak{L}}(2i+2)\Big),$$ where again ${\mathfrak{L}}$ is the length generator which just gives $1/2$ when acting on an individual site. ${\mathfrak{X}}$ here can be an arbitrary (length-preserving) two-site generator . If ${\mathfrak{X}}$ is fermionic, though, there would be extra signs. Of course, there is another chain derivative where ${\mathfrak{X}}$ acts on sites $(2i-1, 2i)$ instead. Similarly, there are physically equivalent expressions that act on a (superficially) different number of sites, due to interactions including spectator sites. For example a two-site generator ${\mathfrak{Z}}$ can be written equivalently as a three-site generator, $$\sum_{I=1}^{2L} {\mathfrak{Z}}(i, i+1) = 2 \sum_{i=1}^{2L} \Big({\mathfrak{Z}}(i, i+1) {\mathfrak{L}}(i+2)\Big).$$ Importantly (\[eq:qhatshatacomm\]) implies that there is no wrapping contribution for $\mathcal{N}=6$ Chern-Simons until four loops, as is also the case for $\mathcal{N}=4$ SYM [@Beisert:2005fw; @Kotikov:2007cy]. Each ${\mathfrak{osp}}(6|4)$ highest weight two-site state is in one-to-one correspondence with a descendant which is an ${\mathfrak{osp}}(4|2)$ sector highest-weight. Furthermore, in this sector ${\hat{{\mathfrak{S}}}}$ annihilates two-site states and $\hat{{\mathfrak{Q}}}$ combines non-BPS states with four-site states in long multiplets for $\lambda \neq 0$. Wrapping interactions for $\hat{{\mathfrak{Q}}}$ first appears when there are 3-to-5 site interactions, i.e. at $\mathcal{O}(\lambda^3)$, which leads to a wrapping contribution to the four-loop dilatation generator.
In fact, it is straightforward to compute the anomalous dimensions of two-site states at two loops. These states have twist one, and according to (\[eq:twositeCartan\]) there is one highest-weight twist-one state for each nonnegative integer ${\mathfrak{osp}}(4|2)$ spin $s$. A convenient representative of the $s$th multiplet is in the ${\mathfrak{sl}}(2)$ sector and has Lorentz spin $(s+1/2)$, $${\mathopen{\big|}{\mathnormal{\oldPsi}}_s\mathclose{\bigr \rangle}}= \sum_{m=0}^s (-1)^m \sqrt{\binom{2 s +1}{2m}} {\mathopen{\big|}\phi_1^{(m)} \psi_1^{(s-m)}\mathclose{\bigr \rangle}}.$$ ${\mathopen{\big|}{\mathnormal{\oldPsi}}_s\mathclose{\bigr \rangle}}$ is annihilated by ${\mathfrak{J}}^{22}$, and is therefore an eigenstate of the dilatation generator since there are no other such states with the same quantum numbers. Then the two-loop contribution to the anomalous dimension, $\Delta_{s,2}$, simply equals the coefficient of ${\mathopen{\big|}\phi_1^{(s)}\psi_1^{(0)}\mathclose{\bigr \rangle}}$ for $\mathcal{H}{\mathopen{\big|}{\mathnormal{\oldPsi}}_s\mathclose{\bigr \rangle}}$ divided by $(-1)^s \sqrt{2s+1}$. Since ${\hat{{\mathfrak{S}}}}$ annihilates two-site states, the Hamiltonian reduces to $ \hat{{\mathfrak{S}}}\hat{{\mathfrak{Q}}}$.
We can organize the contributions of $\hat{{\mathfrak{S}}}\hat{{\mathfrak{Q}}}$ as follows. There are (diagonal) terms from the product acting only on the first site, or only on the second site. As explained in Appendix \[sec:lightcone\], the one-site part of the Hamiltonian is $2 S_1(2m)$ for ${\mathopen{\big|}\phi^{(m)}\mathclose{\bigr \rangle}}$ and $2 S_1(2m +1)$ for ${\mathopen{\big|}\psi^{(m+1)}\mathclose{\bigr \rangle}}$. Also, there is a contribution from the fermion-to-three-boson interaction of $\hat{{\mathfrak{Q}}}$ combined with the conjugate interaction of $\hat{{\mathfrak{S}}}$, $${\mathopen{\big|}\phi_1^{(0)} \psi_1^{(s)}\mathclose{\bigr \rangle}} \rightarrow \varepsilon^{\mathfrak{bc}} {\mathopen{\big|}\phi_1^{(0)} \bar{\phi}_{\mathfrak{b}}^{(0)}\phi_1^{(s)}\bar{\phi}_{\mathfrak{c}}^{(0)}\mathclose{\bigr \rangle}} \rightarrow {\mathopen{\big|}\phi_1^{(s)} \psi_1^{(0)}\mathclose{\bigr \rangle}}.$$ Note that the second arrow refer to $\hat{{\mathfrak{S}}}$ acting on the last and first two sites, $\hat{{\mathfrak{S}}}(4,1,2)$. Finally, for the generic terms, $\hat{{\mathfrak{Q}}}$ inserts a ${\mathopen{\big|}\psi_1^{(0)}\mathclose{\bigr \rangle}}$ (and two additional module elements) and $\hat{{\mathfrak{S}}}$ replaces the other three sites with ${\mathopen{\big|}\phi_1^{(s)}\mathclose{\bigr \rangle}}$, yielding a multiple of ${\mathopen{\big|}\phi_1^{(s)}\psi_1^{(0)}\mathclose{\bigr \rangle}}$. Combining all of these contributions yields the $s$th two-loop anomalous dimension, \[eq:delta2sum\] \_[s,2]{} [[&=&]{}]{}\_[m=0]{}\^[s]{} A(m), [\
]{}A(m) [[&=&]{}]{}2 \_[ms]{} (S\_1(2s) + S\_1(1)) - + \_[n=0]{}\^[m-1]{} 2 c\_1(m,n, 0)c\_1(s, n, s-m) [\
&&]{}+ \_[n=0]{}\^[s-m-1]{} (- 2 c\_2(s-m, 0, n) c\_1(s, s-m-n-1, n) [\
&&]{}+ 8 c\_2(s-m, 0, n)c\_1(s, m, n) + 2 c\_2(s-m, n, 0)c\_1(s, m, n) ). It is simple to find the pattern numerically by evaluating the sum for low values of $s$, but in fact we have also done the sum analytically for arbitrary (nonnegative integer) $s$. For this it is significantly easier to use the light-cone superspace expressions for the supercharges, and we give more details about this in Appendix \[sec:lightcone\]. We find the spectrum in terms of the harmonic numbers and a generalized harmonic sum, $$\label{eq:twistonespectrum}
\Delta_s = 4 \lambda^2 \big(S_1(s) - S_{-1}(s)\big) + \mathcal{O}(\lambda^3),$$ which is similar to the twist-two spectrum of $\mathcal{N}=4$ SYM, $8 \lambda_{\mathcal{N}=4} S_1(s)$.
As we will see, the R-matrix construction requires at least four sites. Still the Bethe ansatz correctly gives this twist-one spectrum because it naturally accounts for the action of $\hat{{\mathfrak{Q}}}$ on two-site states described earlier. For zero-momentum solutions of the Bethe ansatz equations of [@Gromov:2008qe][^12], $\hat{{\mathfrak{Q}}}$ acts via $L \rightarrow (L+1)$ simultaneous with the removal of a single $u_3$ root at $0$, with all other roots unchanged. It is straightforward to check that this transformation does not change the energy or momentum (zero), and that it carries the same Cartan charges as $\hat{{\mathfrak{Q}}}$ does. It follows that the spectrum for $L=1$ states is the same as the spectrum of $L=2$ states without a $u_3$ root at zero (and only $u_4$, $\bar{u}_4$ and $u_3$ roots excited). A similar phenomenon occurs in the ${\mathfrak{psu}}(1,1|2)$ sector of $\mathcal{N}=4$ SYM [@Beisert:2005fw].
In Appendix \[sec:bethe\] we check analytically that the Bethe ansatz prediction gives precisely (\[eq:twistonespectrum\]). This is already very strong evidence in favor of the leading-order Bethe equations of [@Minahan:2008hf]. As found by [@Gromov:2008qe], the large $s$ behavior of $\Delta_s$ gives a cusp anomalous dimension of $f(\lambda) = 4 \lambda^2$, though this disagrees by a factor of four with the value give in [@Aharony:2008ug], based on [@Gaiotto:2007qi]. The author does not know the origin of this discrepancy, which was already noted in [@Gromov:2008qe].
Proof of ${\mathfrak{osp}}(4|2)$ sector integrability \[sec:yangian\]
=====================================================================
In this section we prove that the two-loop dilatation generator for the ${\mathfrak{osp}}(4|2)$ sector is integrable by constructing an ${\mathfrak{osp}}(4|2)$ Yangian that commutes with the leading-order ${\mathfrak{su}}(1|1)$ generators, and therefore with the two-loop dilatation generator.
Leading order Yangian
---------------------
A Yangian was used in a similar context for $\mathcal{N}=4$ SYM in [@Dolan:2003uh]. This is an infinite-dimensional symmetry algebra generated by the ordinary Lie algebra generators ${\mathfrak{J}}^A$ and bilocal products of Lie algebra generators $$\label{eq:definey}
{\mathfrak{Y}}^A = f^{A}{}_{CB} \sum_{i<j} {\mathfrak{J}}^B(i) {\mathfrak{J}}^C(j),$$ where $f^{A}{}_{CB}$ are the structure constants, with indices lowered (raised) using the (inverse) Cartan-Killing form. Note that the ${\mathfrak{Y}}^A $ are incompatible with periodic boundary conditions. As a result, generically the Yangian symmetry is only unbroken for infinite-length chains, as will be the case here. The $ {\mathfrak{Y}}^A$ manifestly transform in the adjoint of the Lie algebra. To consistently generate the entire algebra, the only additional requirement is that these ${\mathfrak{Y}}^A$ satisfy Serre relations $$\label{eq:serre}
3 {[{\mathfrak{Y}}^{[A },{[{\mathfrak{J}}^B,{\mathfrak{Y}}^{C\}\}\}} = - (-1)^{(EM)} f^{AK}{}_{D} f^B{}_{E}{}^{L} f^C{}_{F}{}^{M} f_{KLM}\{ {\mathfrak{J}}^D, {\mathfrak{J}}^E, {\mathfrak{J}}^F \},$$ where the curly brackets on the right side refer to the totally symmetric triple product, including a factor of $1/6$. Note that the indices on the left side are anti-symmetrized, with a factor of $1/6$. The signs and ordering of indices in the structure constants properly account for fermionic statistics for super Yangians [@Zwiebel:2006cb]. For ${\mathfrak{osp}}(4|2)$, a possible basis is $\{{\mathfrak{J}}^A \} = \{{\mathfrak{Q}}^{a\beta\mathfrak{c}},\, {\mathfrak{R}}^{ab}, \tilde{{\mathfrak{R}}}^{\mathfrak{ab}}, {\mathfrak{J}}^{\alpha\beta} \}$.
As reviewed in [@Dolan:2004ps], it is sufficient to check that the Serre relations are satisfied for a one-site chain. This is because the Yangian, a Hopf algebra, has a coproduct which gives the Yangian’s action on tensor products ((\[eq:definey\]) actually follows from the coproduct). If the Yangian’s Serre relations are satisfied on a single module, because of the coproduct they will be satisfied on chains of arbitrary length. For a one-site chain, the left-side of the Serre relation (\[eq:serre\]) vanishes, so we simply need to confirm that the right side vanishes. Programming the generators in $\texttt{Mathematica}$, we have confirmed that the Serre relations are satisfied when acting on any element of a single ${\mathfrak{osp}}(4|2)$ module (of either type)[^13], as required.
Vanishing commutator with $\hat{{\mathfrak{Q}}}$ and $\hat{{\mathfrak{S}}}$
---------------------------------------------------------------------------
We will now explicitly show that the Yangian generators ${\mathfrak{Y}}^{aa}$, which have the same ${\mathfrak{osp}}(4|2)$ indices as ${\mathfrak{R}}^{11}$ or ${\mathfrak{R}}^{22}$, commute with $\hat{{\mathfrak{Q}}}$ on infinite-length chains. At the end of this section we will infer from this that the ${\mathfrak{osp}}(4|2)$ Hamiltonian is integrable.
First, consider the commutators between $\hat{{\mathfrak{Q}}}$ and the ${\mathfrak{osp}}(4|2)$ Lie algebra generators. Because ${\mathfrak{R}}$ symmetry is manifest, the commutator of ${\mathfrak{R}}$ with $\hat{{\mathfrak{Q}}}$ vanishes locally, not just up to a gauge transformation. Also, ${\mathfrak{Q}}^{a2\mathfrak{b}}$ commutes with $\hat{{\mathfrak{Q}}}$ locally as well (this commutator has classical dimension $0$, and a gauge transformation that inserts two sites has a minimum dimension $1$, twice the dimension of scalars without derivatives). On the other hand, the commutator with ${\mathfrak{Q}}^{a1\mathfrak{b}}$ gives the gauge transformation (\[eq:qqhatgauge\]) that acts as [{\^[a1]{},}]{} [[&=&]{}]{}\_i \^[a]{}\_i -\^[a]{}\_i , [\
]{}\^[a]{}[x, , |]{} [[&=&]{}]{}\^a\_1|\^\_2 [0, \_1, |\_1;0, |\_2, \_2;x, , |]{}, [\
]{}\^[a]{}[x, , |]{} [[&=&]{}]{}|\^\_2\^a\_3 [x, , |; 0, |\_2, \_2; 0, \_3, |\_3;]{}, [\
]{}\^[a]{}[x, |, ]{} [[&=&]{}]{}|\^\_1\^a\_2 [0, |\_1, \_1;0, \_2, |\_2;x, |, ]{}, [\
]{}\^[a]{}[x, |, ]{} [[&=&]{}]{}\^a\_2 |\^\_3[x, |,; 0, \_2, |\_2; 0, |\_3, \_3;]{}. In terms of components, the $\acute{Z}$ and $\grave{Z}$ insert two scalars without derivatives on adjacent sites. Our convention for site indices is that $\hat{{\mathfrak{Q}}}_i$, $\acute{Z}_i$ or $\grave{Z}_i$ or act on site $i$ and inserts sites $i+1$ and $i+2$ . Therefore, the initial (and final) sites $1$ through $i-1$ are unaffected by $\hat{{\mathfrak{Q}}}_i$, while for $j >i$, an initial site $j$ becomes site $j+2$, with these sites otherwise unchanged.
Before focusing on the ${\mathfrak{Y}}^{aa}$, we consider general features of the commutator between $\hat{{\mathfrak{Q}}}$ and bilocal generators. Consider the commutator involving one-site (bosonic) generators ${\mathfrak{J}}^A$ and ${\mathfrak{J}}^B$, $$\label{eq:commbilocalqhat}
{[\sum_{i < j} {\mathfrak{J}}_i^A {\mathfrak{J}}_j^B,\hat{{\mathfrak{Q}}}]} = \sum_{i < j} {\mathfrak{J}}_i^A {[{\mathfrak{J}}^B,\hat{{\mathfrak{Q}}}]}_j + \sum_{i < j} {[{\mathfrak{J}}^A,\hat{{\mathfrak{Q}}}]}_i {\mathfrak{J}}_j^B + \text{local}.$$ Since $\hat{{\mathfrak{Q}}}$ has one-to-three site interactions, the commutator with an individual ${\mathfrak{J}}$ also gives a (spin-chain-local) one-to-three site generator, for which the subscript refers to the single site on which this new local generator acts. The terms of (\[eq:commbilocalqhat\]) emerge as follows. The commutator vanishes when the ${\mathfrak{J}}$ act on sites that $\hat{{\mathfrak{Q}}}$ does not act on or insert. The terms where ${\mathfrak{J}}^B$ acts on a site inserted or acted on by $\hat{{\mathfrak{Q}}}$ but ${\mathfrak{J}}^A$ acts completely to the left of $\hat{{\mathfrak{Q}}}$ simplifies to the first term of the right side of (\[eq:commbilocalqhat\]). The reflected terms, with ${\mathfrak{J}}^A$ and ${\mathfrak{J}}^B$ and right and left switched, yield the second term. Finally, there are terms on which both ${\mathfrak{J}}$ act on sites inserted by $\hat{{\mathfrak{Q}}}$, which we call local because these combine into a homogeneous one-to-three site interaction $$\label{eq:local}
\sum_{i} \Big( {\mathfrak{J}}^A_i ({\mathfrak{J}}^B_{i+1} +{\mathfrak{J}}^B_{i+2} ) +{\mathfrak{J}}^A_{i+1}{\mathfrak{J}}^B_{i+2} \Big) {\mathfrak{\hat{Q}}}_i.$$ Finally we turn to the ${\mathfrak{Y}}^{aa}$. From (\[eq:definey\]) we find 4 \^[aa]{} [[&=&]{}]{} \_[i < j]{} 2 \_[bc]{} \^[bc]{}\_i \^[ca]{}\_j - \_ \^[a1 ]{}\_i \^[a2 ]{}\_j + \_ \^[a2 ]{}\_i \^[a1 ]{}\_j [\
]{}[[&=&]{}]{}\^[aa]{}\_ + \^[aa]{}\_[\^1]{} + \^[aa]{}\_[\^2]{}. \[eq:Ycc\] where the factor of $4$ on the left side is for convenience. Using (\[eq:commbilocalqhat\]) and (\[eq:Ycc\]), the commutators described above, and taking into account statistics, we find 4[\[\^[aa]{},\]]{} [[&=&]{}]{}\_[i < j]{} \_ [{\^[a1]{},}]{}\_i \_j\^[a2]{} - \_[i < j]{} \_ \_i\^[a2]{} [{\^[a1]{},}]{}\_j + [\
]{}[[&=&]{}]{}\_[i < j]{} \_ (\_i\^[a]{} -\_i\^[a]{}) \_j\^[a2]{} - \_[i < j]{} \_ \_i\^[a2]{}(\_j\^[a]{} -\_j\^[a]{}) + [\
]{}[[&=&]{}]{}- \_[i]{} \_ \_i\^[a]{} \_[i+1]{}\^[a2]{} - \_i \_ \_i\^[a2]{}\_[i+1]{}\^[a]{} + . \[eq:bilocalYqhatcomm\] To reach the last line, we used the cancellations between $\acute{Z}$ and $\grave{Z}$ acting upon adjacent sites, similar to the cancellation explained after (\[eq:gauge\]). In Appendix \[sec:localycomm\] we evaluate the remaining local piece, which involves computing (\[eq:local\]) with the ${\mathfrak{J}}$ there replaced with the generators that appear in (\[eq:Ycc\]). The result leads to the precise cancellation on infinite-length chains, $${[{\mathfrak{Y}}^{aa},\hat{{\mathfrak{Q}}}]}=0.$$ Since $\hat{{\mathfrak{Q}}}$ commutes with all ${\mathfrak{osp}}(4|2)$ Lie algebra generators and because the Yangian generators transform in the adjoint of ${\mathfrak{osp}}(4|2)$, this is sufficient to imply that all generators commute with $\hat{{\mathfrak{Q}}}$. For instance, any of the fermionic Yangian generator can be written as $${\mathfrak{Y}}^{a\beta\mathfrak{c}} = \pm {[{\mathfrak{Q}}^{d\beta\mathfrak{c}},{\mathfrak{Y}}^{aa}]}, \quad d \neq a,$$ which then necessarily commute with $\hat{{\mathfrak{Q}}}$ since the ${\mathfrak{osp}}(4|2)$ supercharges do (on infinite-length chains, as needed). Similarly, one can extend this to the remaining Yangian generators. Hermiticity then implies that $\hat{{\mathfrak{S}}}$ also commutes with the Yangian. Therefore, the two-loop ${\mathfrak{osp}}(4|2)$ sector dilatation generator $$\delta{\mathfrak{D}}_2 = {\{\hat{{\mathfrak{Q}}},\hat{{\mathfrak{S}}}\}}$$ has an ${\mathfrak{osp}}(4|2)$ Yangian symmetry and is integrable.
R-matrix construction of the ${\mathfrak{osp}}(4|2)$ sector Hamiltonian \[sec:rmatrix\]
=======================================================================================
The ${\mathfrak{osp}}(4|2)$ sector spin chain and its Yangian symmetry can be restricted consistently to the ${\mathfrak{sl}}(2)$ sector. For such a ${\mathfrak{sl}}(2)$ alternating spin chain we can use the known universal ${\mathfrak{sl}}(2)$ R-matrix [@Kulish:1981gi] to construct a transfer matrix as a function of two spectral parameters, $u$ and $\alpha$. We will see below that in our case $\alpha=0$. As is well-known, the expansion about $u=\infty$ gives the Yangian symmetry, while the expansion about $u=0$ gives the complete set of local conserved charges, $\mathcal{Q}_I$, that commute with the Yangian on infinite-length chains. Then the Hamiltonian for the ${\mathfrak{sl}}(2)$ sector of the ABJM spin chain must be a linear combination of the $\mathcal{Q}_I$, up to physically irrelevant chain derivatives. The restriction to nearest- and next-nearest-neighbor interactions identifies this Hamiltonian uniquely, up to coefficients that can be fixed by acting on a few states.
Since the ${\mathfrak{sl}}(2)$ sector Hamiltonian originates from an R-matrix construction, we can conclude that the ${\mathfrak{osp}}(4|2)$ sector Hamiltonian also follows from an R-matrix construction. The argument is as follows. In Appendix \[sec:maptosl2\] we show that there is a unique lift from the ${\mathfrak{sl}}(2)$ sector Hamiltonian to the ${\mathfrak{osp}}(4|2)$ sector. The (Lie algebra invariant) R-matrix only acts on two sites at a time. Also, recall that there is a one-to-one map between irreducible modules in the tensor product of two-sites in the ${\mathfrak{sl}}(2)$ subsector and those in the ${\mathfrak{osp}}(4|2)$ sector. Then the ${\mathfrak{osp}}(4|2)$ sector Hamiltonian must take the form that would follow from an ${\mathfrak{osp}}(4|2)$ R-matrix construction (assuming the existence of such R-matrices). The Yangian construction of the previous section confirms the existence of R-matrices for the ${\mathfrak{osp}}(4|2)$ spin-chain modules.
In this section we review the general R-matrix construction of the transfer matrix and conserved charges for an alternating spin chain. Based on the above argument, we then apply this construction to the ${\mathfrak{osp}}(4|2)$ sector. We deduce the action of the ${\mathfrak{osp}}(4|2)$ R-matrix on the spin-chain modules from the universal ${\mathfrak{sl}}(2)$ R-matrix, obtaining another expression for the Hamiltonian. This new expression for the Hamiltonian will enable us to obtain the full ${\mathfrak{osp}}(6|4)$ two-loop dilatation generator in the next section.
The transfer matrix and local conserved charges
-----------------------------------------------
The following discussion parallels the recent construction for the alternating ${\mathfrak{su}}(4)$ spin chain [@Minahan:2008hf], and the original general construction of [@deVega:1991rc].
We start with a R-matrix, which satisfies the Yang-Baxter equation $$R_{12}(u-v)R_{13}(u)R_{23}(v)= R_{23}(v)R_{13}(u)R_{12}(u-v), \label{eq:ybe}$$ where $R_{ij}$ is the R-matrix acting on sites $i$ and $j$. For alternating chains, it is sufficient for the R-matrix to satisfy the Yang-Baxter for each of the $2^3=8$ possible ways to assign one of the two representations to sites $1,2,3$.
Now we consider an alternating chain with the two representations distinguished by the presence or absence of a bar, $1, \bar{2} \ldots (2L-1), \overline{2L}$. We build two monodromy matrices from the R-matrix, $$\mathcal{T}_a(u, \alpha) = \prod_{i=1}^L R_{a,2i-1}(u)R_{a, \bar{2i}}(u+\alpha), \quad \mathcal{T}_{\bar{b}}(u, \beta) = \prod_{i=1}^L R_{\bar{b},2i-1}(u+\beta)R_{\bar{b}, \bar{2i}}(u).$$ Since the R-matrices satisfy the Yang-Baxter equation, these monodromy matrices also satisfy the Yang-Baxter equations R\_[ab]{}(u-v) \_a(u) \_b(v)[[&=&]{}]{}\_b(v) \_a(u)R\_[ab]{}(u-v), [\
]{}R\_[|[a]{}|[b]{}]{}(u-v) \_[|[a]{}]{}(u) \_[|[b]{}]{}(v) [[&=&]{}]{}\_[|[b]{}]{}(v) \_[|[a]{}]{}(u)R\_[|[a]{}|[b]{}]{}(u-v). Moreover, if $\beta=-\alpha$, which we will choose from now on, the mixed Yang-Baxter equation is also satisfied $$R_{a\bar{b}}(u+\alpha-v) \mathcal{T}_{a}(u, \alpha)\mathcal{T}_{\bar{b}}(v, -\alpha)= \mathcal{T}_{\bar{b}}(v, -\alpha) \mathcal{T}_{a}(u,\alpha)R_{a\bar{b}}(u+\alpha-v).$$ Taking the trace, and using the invertibility of the R-matrix, we infer that the transfer matrices, $$T(u, \alpha) = \mathrm{Tr}_a \mathcal{T}_a(u, \alpha), \quad \bar{T}(u, -\alpha) = \mathrm{Tr}_{\bar{b}} \mathcal{T}_{\bar{b}}(u, -\alpha),$$ satisfy $${[T(u, \alpha),T(v, \alpha)]}=0, \quad {[\bar{T}(u, -\alpha),\bar{T}(v, -\alpha)]}=0, \quad {[T(u, \alpha),\bar{T}(v, -\alpha)]}=0.$$ In particular, this implies the existence of (up to) $2L$ commuting generators. The expansion of $T(u, \alpha)$ about $u=0$ gives $L$ commuting generators, which also commute with the $L$ commuting generators coming from the expansion of $\bar{T}(u, -\alpha)$ about $u=0$. However, we will only consider the first two terms in the expansions. The leading terms, the transfer matrices evaluated at zero spectral parameter, yield the generators \_1 [[&=&]{}]{}\_[i=1]{}\^[L-1]{}R\_[2i+3, 2i+1]{}(0) \_[i=1]{}\^L R\_[2i-1,2i]{}() , [\
]{} |\_1 [[&=&]{}]{}\_[i=1]{}\^[L]{}R\_[2i-1, 2i]{}(-) \_[i=1]{}\^[L-1]{} R\_[2i,2i+2]{}(0). Here we have stopped including bars to distinguish representations, which are of one type for odd-numbered sites, and the other for even-numbered sites. Also, these expressions require that, when acting on two identical representations, at $u=0$ the $R$-matrix is proportional to the permutation generator, which will be the case for the R-matrices we consider. Simplifying the product using $R^{-1}(\alpha) = R(-\alpha)$, another property of our R-matrices, we obtain the two-site shift generator $$\mathcal{Q}_1\bar{\mathcal{Q}}_1 = \prod_{i=1}^{L-1}R_{2i+3, 2i+1}(0)\prod_{i=1}^{L-1} R_{2i,2i+2}(0).$$ Expanding to $\mathcal{O}(u)$, the next charges are defined through T(u)[[&=&]{}]{}\_1 + u \_1 \_2 + …, [\
]{}|[T]{}(u) [[&=&]{}]{}|\_1 + u |\_1 |\_2 + … The charges have nearest-neighbor and next-nearest-neighbor contributions $$\mathcal{Q}_2 = (\mathcal{Q}_2)_{NN} + (\mathcal{Q}_2)_{NNN}, \quad \bar{\mathcal{Q}}_2 = (\bar{\mathcal{Q}}_2)_{NN} + (\bar{\mathcal{Q}}_2)_{NNN}.$$ The nearest-neighbor contribution can be chosen symmetrically as $$(\mathcal{Q}_2)_{NN} = \sum_{i=1}^{2L} R_{i,i+1}(-\alpha)R'_{i,i+1}(\alpha), \quad (\bar{\mathcal{Q}}_2)_{NN} = \sum_{i=1}^{2L} R_{i,i+1}(\alpha)R'_{i,i+1}(-\alpha)$$ provided the next nearest-neighbor contributions are (\_2)\_[NNN]{} [[&=&]{}]{}\_[i=1]{}\^L ( R\_[2i-1, 2i]{}()R\_[2i-1, 2i+1]{}’(0)R\_[2i-1, 2i+1]{}(0)R\_[2i-1, 2i]{}(-) [\
&&]{}+ R\_[2i, 2i+1]{}(-)R\_[2i-1, 2i+1]{}’(0)R\_[2i-1, 2i+1]{}(0)R\_[2i, 2i+1]{}()), [\
]{}(|\_2)\_[NNN]{} [[&=&]{}]{}\_[i=1]{}\^L (R\_[2i, 2i+1]{}(-)R\_[2i, 2i+2]{}’(0)R\_[2i, 2i+2]{}(0)R\_[2i, 2i+1]{}() [\
&&]{}+ R\_[2i+1, 2i+2]{}()R\_[2i, 2i+2]{}’(0)R\_[2i, 2i+2]{}(0)R\_[2i+1, 2i+2]{}(-)). To compute the next-nearest-neighbor terms we inserted a factor of $R(0)R^{-1}(0)=1$, and to obtain symmetric expressions we used the vanishing commutator between the $\mathcal{Q}_i$. There are additional possibilities that differ by chain derivatives.
The ${\mathfrak{osp}}(4|2)$ case
--------------------------------
As explained above, the relevant ${\mathfrak{osp}}(4|2)$ R-matrices are determined by the universal R-matrix of ${\mathfrak{sl}}(2)$ [@Kulish:1981gi]. This also occurred for the one-loop $\mathcal{N}=4$ SYM spin [@Beisert:2003yb]. The result is a sum over the irreducible representations of the tensor product of two sites, labeled by ${\mathfrak{osp}}(4|2)$ spin $j$, weighted by a certain ratio of Gamma functions, $$\label{eq:sl2universal}
R_{12}(u) = \sum_{j} (-1)^j f(cu) \frac{\Gamma(j+1+c u)\Gamma(1- c u)}{\Gamma(j+1-c u)\Gamma(1+ c u)}\mathcal{P}^{(j)}_{12}.$$ Here $\mathcal{P}^{(j)}$ is the projector that acts as the identity on ${\mathfrak{osp}}(4|2)$ states with spin $j$, and gives zero on all other states. Using (\[eq:42conjugateproduct\]) and (\[eq:42likeproduct\]), for the case of sites $1,2$ in alternate representations, the sum is over all nonnegative $j$, while for identical representations, $j$ takes values $n-{{{\textstyle\frac{1}{2}}}}$ for all nonnegative $n$. For our purposes, the function of the spectral parameter $f$ and the constant $c$ can be replaced simply with freedom in the normalization of the local charges. We set $c$ to 1 and choose $f$ to cancel any overall factors of $\pm i$ from the $(-1)^j$ factors. The Yang-Baxter equation (\[eq:ybe\]) is still satisfied even if we choose different (constant) values of $f$ for different pairs of representations. As in the previous section, with these conventions $R^{-1}(u)= R(-u)$, and $R(0)$ acts as the permutation operator on identical representations.
We are almost ready to simply insert the expression (\[eq:sl2universal\]) for the R-matrices into the expressions for $\mathcal{Q}_2$ and $\bar{\mathcal{Q}}_2$ given in the last subsection, but there are four coefficients to fix. These are $\alpha$, the coefficients of $\mathcal{Q}_2$ and $\bar{\mathcal{Q}}_2$, and the coefficient of the identity operator, which we are also free to add without spoiling integrability. In principle one could compute four eigenvalues to fix these coefficients. However, since the Hamiltonian is even under spin-chain parity, $\alpha$ must be zero. Also, symmetry under charge conjugation implies that $\mathcal{Q}_2$ and $\bar{\mathcal{Q}}_2$ have equal coefficients. We have found the final two coefficients by comparison with eigenvalues of the (previous expression for the) Hamiltonian. $\mathcal{Q}_2$ and $\bar{\mathcal{Q}}_2$ have coefficient $\lambda^2/2$ and the identity has coefficient $2 \lambda^2 \log 2$, \_2 [[&=&]{}]{}(\_2 + |\_2)\_[|=0]{} + 4 L 2 [\
]{}[[&=&]{}]{}\_[i=1]{}\^[2L]{} ( R\_[i, i+1]{}\^[(0)]{}R’\_[i, i+1]{}(0) + 2 2 + [\
&&]{} (R\_[i, i+1]{}\^[(0)]{} R\_[i, i+2]{}\^[(0)]{}R\_[i, i+2]{}’(0)R\_[i, i+1]{}\^[(0)]{} + R\_[i+1, i+2]{}\^[(0)]{} R\_[i, i+2]{}\^[(0)]{}R\_[i, i+2]{}’(0)R\_[i+1, i+2]{}\^[(0)]{}) ). $\mathcal{Q}_2$ and $\bar{\mathcal{Q}}_2$ combine nicely here, and we used the more compact notation $R^{(0)}$ for the R-matrix evaluated at zero spectral parameter, $R(0)$. Next, evaluating (\[eq:sl2universal\]) and its derivative at $u=0$, the dependence on $j$ reduces to factors of $(-1)^j$ and the harmonic numbers, $S_1(j)$ [^14], \[eq:osp42projectorham\] \_2 [[&=&]{}]{}\_[i=1]{}\^[2L]{} (2 2 + \_[j=0]{}\^S\_1(j)\^[(j)]{}\_[i, i+1]{} [\
&&]{} + \_[j\_1,j\_2,j\_3=0]{}\^(-1)\^[j\_1+j\_3]{} S\_1(j\_2-) (\^[(j\_1)]{}\_[i, i+1]{} \^[(j\_2-1/2)]{}\_[i, i+2]{}\^[(j\_3)]{}\_[i, i+1]{} + \^[(j\_1)]{}\_[i+1, i+2]{} \^[(j\_2-1/2)]{}\_[i, i+2]{}\^[(j\_3)]{}\_[i+1, i+2]{} ) ). [\
&&]{} Using the explicit form for the projectors given in Appendix \[sec:projectors\] and `Mathematica`, we have checked that this spin-chain Hamiltonian for the two-loop (${\mathfrak{osp}}(4|2)$ sector) dilatation generator leads to the correct two-magnon S-matrix, and that its spectrum for low numbers of excitations and short states agrees with Bethe ansatz predictions and the alternative expression for the Hamiltonian as the anticommutator of $\hat{{\mathfrak{Q}}}$ and $\hat{{\mathfrak{S}}}$. Still, the anticommutator of $\hat{{\mathfrak{Q}}}$ and $\hat{{\mathfrak{S}}}$ gives a slightly more general form that applies to two-site states also, while this R-matrix expression requires chains of length four.
The lift to the complete ${\mathfrak{osp}}(6|4)$ chain \[sec:lift\]
===================================================================
Here we derive the two-loop planar ${\mathfrak{osp}}(6|4)$ dilatation generator using superconformal invariance. We then observe that this spin-chain Hamiltonian is integrable, assuming the existence of an ${\mathfrak{osp}}(6|4)$ R-matrix for like or conjugate spin-chain modules. We argue that there is no reason to doubt this assumption.
Unique lift from ${\mathfrak{osp}}(4|2)$ to ${\mathfrak{osp}}(6|4)$ \[sec:uniquelift\]
--------------------------------------------------------------------------------------
By adding chain derivatives to replace one-site or two-site interactions with three-site interactions, one can write the ${\mathfrak{osp}}(6|4)$ two-loop Hamiltonian completely in terms of a Hamiltonian density $\mathcal{H}_{i,i+1,i+2}$ of three-site to three-site interactions[^15], $$\mathcal{H}{\mathopen{\big|}X_1\ldots X_{2L}\mathclose{\bigr \rangle}} = \sum_{i=1}^{2L} \mathcal{H}_{i,i+1,i+2} {\mathopen{\big|}X_1\ldots X_{2L}\mathclose{\bigr \rangle}}.$$ ${\mathfrak{osp}}(6|4)$ invariance allows us to use the freedom to add chain derivatives so the Hamiltonian *density* commutes with the leading-order ${\mathfrak{osp}}(6|4)$ generators. Therefore the Hamiltonian is completely specified by the Hamiltonian density’s action on all three-site (without cyclicity condition) highest-weight states. Furthermore, the Hamiltonian density mixes highest-weight states only with other highest-weight states with the same ${\mathfrak{su}}(4)$ Cartan charges, classical dimension, and Lorentz spin. For fixed values of these five Cartan charges there are finitely many linearly-independent three-site highest-weight states, ${\mathopen{\big|}\Omega_I\mathclose{\bigr \rangle}}$. In the sector with basis ${\mathopen{\big|}\Omega_I\mathclose{\bigr \rangle}}$, the Hamiltonian density acts as $$\mathcal{H}_{123} {\mathopen{\big|}\Omega_I\mathclose{\bigr \rangle}} = C^J_I {\mathopen{\big|}\Omega_J\mathclose{\bigr \rangle}},$$ for some coefficients $C^J_I$. The full set of such $C^J_I$ gives the Hamiltonian density and the Hamiltonian. If the Cartan charges of the$ {\mathopen{\big|}\Omega_I\mathclose{\bigr \rangle}}$ satisfy a BPS condition, then these states are in an ${\mathfrak{osp}}(4|2)$ sector[^16], and we have already determined the action of the Hamiltonian (density). If not, as we show below, one can act with a combination of supercharges $\prod {\mathfrak{Q}}$ (determined by the Cartan charges only) so that $$\big(\prod {\mathfrak{Q}}\big) {\mathopen{\big|}\Omega_I\mathclose{\bigr \rangle}} = M^{I'}_I {\mathopen{\big|}\Omega'_{I'}\mathclose{\bigr \rangle}}, \label{eq:maptoosp42}$$ where the ${\mathopen{\big|}\Omega_{I'}\mathclose{\bigr \rangle}}$ are contained within an ${\mathfrak{osp}}(4|2)$ sector and the matrix $M$ is invertible. We have $$\mathcal{H}_{123} {\mathopen{\big|}\Omega'_{I'}\mathclose{\bigr \rangle}} = C'^{J'}_{I'} {\mathopen{\big|}\Omega'_{J'}\mathclose{\bigr \rangle}},$$ for coefficients $C'^{J'}_{I'}$ determined by the known ${\mathfrak{osp}}(4|2)$ sector Hamiltonian. Then, the needed coefficients of the full Hamiltonian are given by $$C = M C' M^{-1}, \quad \text{or} \quad C^J_I = M^{I'}_{I} C'^{J'}_{I'} ((M)^{-1})^{J}_{J'}.$$ Therefore, as claimed, there is a unique lift of the ${\mathfrak{osp}}(4|2)$ sector Hamiltonian to ${\mathfrak{osp}}(6|4)$. We will give this Hamiltonian below, after first proving the existence of the invertible map (\[eq:maptoosp42\]). In Appendix \[sec:maptosl2\] we go one step further and use the same type of argument to show that there is a unique lift from the ${\mathfrak{sl}}(2)$ sector to the ${\mathfrak{osp}}(4|2)$ sector (and therefore to ${\mathfrak{osp}}(6|4)$).
Invertible map between ${\mathfrak{osp}}(6|4)$ and ${\mathfrak{osp}}(4|2)$ states \[sec:invertiblemap\]
-------------------------------------------------------------------------------------------------------
We will first show that for any basis ${\mathopen{\big|}\Omega_I\mathclose{\bigr \rangle}}$ of three-site highest-weight states with identical Cartan charges there exists a product of supercharges that maps the ${\mathopen{\big|}\Omega_I\mathclose{\bigr \rangle}}$ to a ${\mathfrak{osp}}(4|2)$ sector. It will be straightforward afterward to show that this map is invertible.
Again we choose the roots of ${\mathfrak{osp}}(6|4)$ so that highest-weight states are annihilated by the following raising generators, $${\mathfrak{L}}^2_1, \quad {\mathfrak{K}}, \quad {\mathfrak{R}}^i_j \, \, i >j, \quad {\mathfrak{S}}. \label{eq:rraise}$$ Act on the ${\mathopen{\big|}\Omega_I\mathclose{\bigr \rangle}}$ with ${\mathfrak{Q}}_{12,1}$. If the Cartan charges of the ${\mathopen{\big|}\Omega_I\mathclose{\bigr \rangle}}$ satisfy the BPS condition for ${\mathfrak{Q}}_{12,1}$ and ${\mathfrak{S}}^{12,1}$, this will vanish, the ${\mathopen{\big|}\Omega\mathclose{\bigr \rangle}}_I$ are in an ${\mathfrak{osp}}(4|2)$ sector, and the needed map is trivial. So we can assume $ {\mathfrak{Q}}_{12,1} {\mathopen{\big|}\Omega_I\mathclose{\bigr \rangle}}\neq 0$ .
Since the highest-weight states ${\mathopen{\big|}\Omega_I\mathclose{\bigr \rangle}}$ are annihilated by the ${\mathfrak{R}}$ of (\[eq:rraise\]) it is a simple exercise to check that now the ${\mathopen{\big|}\Omega_I\mathclose{\bigr \rangle}}$ must transform as $\mathbf{4}$ or $\mathbf{\bar{4}}$ under ${\mathfrak{R}}$. This is because ${\mathfrak{Q}}_{12,1}$ has nonvanishing action only on lower ${\mathfrak{R}}$ indices $3$ or $4$ or upper indices $1$ or $2$, and because the ${\mathopen{\big|}\Omega_I\mathclose{\bigr \rangle}}$ have spin-chain length three. For simplicity, assume the ${\mathopen{\big|}\Omega_I\mathclose{\bigr \rangle}}$ transform as $\mathbf{4}$. With appropriate interchange of indices this argument can be repeated for the $\mathbf{\bar{4}}$ case.
Next consider $${\mathfrak{Q}}_{13,1} {\mathfrak{Q}}_{12,1} {\mathopen{\big|}\Omega_I\mathclose{\bigr \rangle}}.$$ If these vanish then the ${\mathfrak{Q}}_{12,1}{\mathopen{\big|}\Omega_I\mathclose{\bigr \rangle}}$ satisfy the BPS condition for ${\mathfrak{Q}}_{13,1}$ and ${\mathfrak{S}}^{13,1}$, and therefore ${\mathfrak{Q}}_{12,1}$ gives the required map to a ${\mathfrak{osp}}(4|2)$ sector[^17]. To see this use the last commutation relation of (\[eq:anticommutators\]) and the annihilation of the ${\mathopen{\big|}\Omega_I\mathclose{\bigr \rangle}}$ by the ${\mathfrak{R}}$ raising generator ${\mathfrak{R}}^3_2$. Similarly, if the $${\mathfrak{Q}}_{14,1}{\mathfrak{Q}}_{13,1} {\mathfrak{Q}}_{12,1} {\mathopen{\big|}\Omega_I\mathclose{\bigr \rangle}}$$ vanish, ${\mathfrak{Q}}_{13,1} {\mathfrak{Q}}_{12,1}$ gives the map to an ${\mathfrak{osp}}(4|2)$ subsector.
Finally, assume that ${\mathfrak{Q}}_{14,1}{\mathfrak{Q}}_{13,1} {\mathfrak{Q}}_{12,1} {\mathopen{\big|}\Omega_I\mathclose{\bigr \rangle}} \neq 0$. With respect to ${\mathfrak{su}}(4)$ the scalars of the modules could transform as (fundamental, anti-fundamental, fundamental) or the conjugate. Since the argument would be essentially the same in either case, we assume the first possibility. For Lorentz spin $s$ and classical dimension $N + 3/2$ , the most general possibility is $${\mathfrak{Q}}_{14,1}{\mathfrak{Q}}_{13,1} {\mathfrak{Q}}_{12,1} {\mathopen{\big|}\Omega_I\mathclose{\bigr \rangle}} = \sum_{n_{i,j}} a_I(n_{i, j}) {\mathopen{\big|}\phi_1^{(n_{1,1}, n_{2,1})} \psi_1^{(n_{1,2}, n_{2,2})}\phi_1^{(N + s -n_{1,1} - n_{1,2},N-s -n_{2,1} - n_{2,2})}\mathclose{\bigr \rangle}} \label{eq:genformdescendant}$$ where the $n_{i,j}$ are nonnegative integers such that all superscript arguments are also nonnegative integers consistent with spin statistics, and $a_I(n_{i,j})$ are some coefficients. Now ${\mathfrak{S}}^{12,2}$ must still annihilate (\[eq:genformdescendant\]), since it anticommutes with ${\mathfrak{Q}}_{13, 1}$ and ${\mathfrak{Q}}_{14, 1}$ and gives ${\mathfrak{R}}^2_1$ when anticommuted with ${\mathfrak{Q}}_{12, 1}$. ${\mathfrak{S}}^{12,2}$ acts on single-sites with lower $1$ indices as \^[12,2]{} [\_1\^[(n\_1, n\_2)]{}]{} [[&=&]{}]{}b(n\_2) [(|\^2)\^[(n\_1, n\_2-1)]{}]{}, b(n\_2) = 0 n\_2=0. [\
]{}\^[12,2]{} [\_1\^[(n\_1, n\_2)]{}]{} [[&=&]{}]{}c(n\_2) [(|\^2)\^[(n\_1, n\_2-1)]{}]{}, c(n\_2) = 0 n\_2=0, where all that matters here are the quantum number of the states and whether the coefficients $b$ and $c$ are nonzero. From this it follows that ${\mathfrak{S}}^{12,2}$ annihilates (\[eq:genformdescendant\]) only if all of the second Lorentz index excitation numbers $n_{2,1}$, $n_{2,2}$ and $N-s-n_{2,1}-n_{2,2}$ are zero. But then all of the states of (\[eq:genformdescendant\]) are clearly in an ${\mathfrak{osp}}(4|2)$ (${\mathfrak{sl}}(2)$) sector, and in this case ${\mathfrak{Q}}_{14,1}{\mathfrak{Q}}_{13,1} {\mathfrak{Q}}_{12,1}$ gives the required map.
We have shown that applying a product of supercharges (zero, one, two, or three depending on the Cartan charges) gives a map to an ${\mathfrak{osp}}(4|2)$ sector. This is as abbreviated in (\[eq:maptoosp42\]). To show that this map is invertible, we simply need to show the linear independence of the $$\big(\prod {\mathfrak{Q}}\big){\mathopen{\big|}\Omega_I\mathclose{\bigr \rangle}},$$ where as usual we are focusing on a linearly-independent basis given by three-site highest-weight states ${\mathopen{\big|}\Omega_I\mathclose{\bigr \rangle}}$ with identical Cartan charges. If there were some linear combination of the ${\mathopen{\big|}\Omega_I\mathclose{\bigr \rangle}}$ that were annihilated by the relevant $ \prod {\mathfrak{Q}}$, the above construction implies this linear combination would satisfy another BPS condition. This is a contradiction because the BPS conditions depend only on the Cartan charges (and because the ${\mathopen{\big|}\Omega_I\mathclose{\bigr \rangle}}$ are assumed linearly independent). Therefore, the maps of (\[eq:maptoosp42\]) to the ${\mathfrak{osp}}(4|2)$ sector are invertible, completing the proof.
The two-loop ${\mathfrak{osp}}(6|4)$ spin-chain Hamiltonian and integrability
-----------------------------------------------------------------------------
Recall that there is a one-to-one mapping between highest-weight two-site states in the ${\mathfrak{osp}}(4|2)$ subsector and the full theory, and that these highest-weight states have the same value of ${\mathfrak{osp}}(4|2)$ or ${\mathfrak{osp}}(6|4)$ spin. Therefore, the unique lift is given by replacing projections onto ${\mathfrak{osp}}(4|2)$ spin in (\[eq:osp42projectorham\]) with projections onto the corresponding ${\mathfrak{osp}}(6|4)$ spin. Now the complete planar Hamiltonian is given by the same formal expression as for the ${\mathfrak{osp}}(4|2)$ sector, \[eq:completeham\] \_2 [[&=&]{}]{}\_[i=1]{}\^[2L]{} ( 2 2 + \_[j=0]{}\^S\_1(j)\^[(j)]{}\_[i, i+1]{} [\
&&]{} + \_[j\_1,j\_2,j\_3=0]{}\^(-1)\^[j\_1+j\_3]{} S\_1(j\_2-) (\^[(j\_1)]{}\_[i, i+1]{} \^[(j\_2-1/2)]{}\_[i, i+2]{}\^[(j\_3)]{}\_[i, i+1]{} + \^[(j\_1)]{}\_[i+1, i+2]{} \^[(j\_2-1/2)]{}\_[i, i+2]{}\^[(j\_3)]{}\_[i+1, i+2]{} ) ). [\
&&]{} Of course, the differences from the ${\mathfrak{osp}}(4|2)$ sector Hamiltonian are that the projectors act on the full ${\mathfrak{osp}}(6|4)$ modules, and the $j_i$ correspond to ${\mathfrak{osp}}(6|4)$ spin. Again, there is a sum over the spin-chain sites labeled by $i$, with $\mathcal{P}^{(j)}_{i,k}$ acting on sites $i$ and $k$. It would be nice to have the expressions for these projectors in components, extending those given for the ${\mathfrak{osp}}(4|2)$ sector in Appendix \[sec:projectors\].
Assuming the existence of an ${\mathfrak{osp}}(6|4)$ R-matrix acting on like or conjugate pairs of the two types of ${\mathfrak{osp}}(6|4)$ modules, a parallel derivation to the one given for ${\mathfrak{osp}}(4|2)$ in Section \[sec:rmatrix\] would apply, and would lead to the complete Hamiltonian given in (\[eq:completeham\]). Therefore, up to this assumption, we have shown that planar $\mathcal{N}=6$ superconformal Chern-Simons theory is integrable at two-loops. A similar assumption was used in the $\mathcal{N}=4$ SYM case [@Beisert:2003yb].
As noted earlier, the existence of Yangian symmetry for the ${\mathfrak{osp}}(4|2)$ sector implies the existence of the corresponding ${\mathfrak{osp}}(4|2)$ R-matrices for the modules appearing in the spin chain. This is indication that there is no problem constructing R-matrices for ${\mathfrak{osp}}$ algebras. It seems that it would be sufficient to confirm that the Serre relations (\[eq:serre\]) are satisfied for the complete ${\mathfrak{osp}}(6|4)$ modules, since that would imply the existence of a ${\mathfrak{osp}}(6|4)$ Yangian and the corresponding R-matrices. However, such a check is beyond the scope of this work.
In fact, it reasonable to assume even a *universal* ${\mathfrak{osp}}(6|4)$ R-matrix[^18], which would give the R-matrix for arbitrary ${\mathfrak{osp}}(6|4)$ representations. There is a general construction of universal R-matrices for Yangians of bosonic simple Lie algebras [@Khoroshkin:1994uk] later extended to ${\mathfrak{sl}}(m|n), m\neq n$ [@Stukopin:2005st]. The latter construction was further modified for a recent derivation of the leading-order spin-chain S-matrices of $\mathcal{N}=4$ SYM and ABJM [@Spill:2008yr]. It should be possible to apply a similar construction to ${\mathfrak{osp}}$ algebras as well[^19].
Finally, we emphasize that the invariance of the Hamiltonian density with respect to ${\mathfrak{osp}}(6|4)$ only applies at leading order. At the next order in $\lambda$, the Hamiltonian will only commute exactly with at most a proper subset of the ${\mathfrak{osp}}(6|4)$ supercharges, while the commutators with the other supercharges will vanish only when applied to cyclic alternating chains. This can be see already in the ${\mathfrak{osp}}(4|2)$ sector. While the expression (\[eq:osp42projectorham\]) for the Hamiltonian commutes with $\mathcal{O}(\lambda^0)$ ${\mathfrak{osp}}(4|2)$ generators manifestly, it only commutes with the $\mathcal{O}(\lambda^1)$ supercharges $\hat{{\mathfrak{Q}}}$ and $\hat{{\mathfrak{S}}}$ up to gauge transformations. A similar phenomenon occurs in the ${\mathfrak{psu}}(1,1|2)$ sector of $\mathcal{N}=4$ SYM [@Beisert:2007sk]. It can be traced back to the algebra of supersymmetry variations only closing up to gauge transformations, which appear for the spin-chain as (\[eq:qqhatgauge\]). Importantly, these gauge transformations appear at subleading order and do not obstruct the leading-order ${\mathfrak{osp}}(6|4)$ invariance used to lift the Hamiltonian to the full theory.
Conclusions \[sec:conc\]
========================
We have shown that the two-loop planar dilatation generator of $\mathcal{N}=6$ superconformal Chern-Simons theory is fixed by superconformal invariance up to overall normalization, and can be written compactly as in (\[eq:completeham\]). Through a Yangian construction we have proved integrability for the ${\mathfrak{osp}}(4|2)$ sector, and for the full model assuming the existence of an ${\mathfrak{osp}}(6|4)$ R-matrix. This confirms the conjectured two-loop ${\mathfrak{osp}}(6|4)$ Bethe equations of Minahan and Zarembo. We also analytically computed the twist-one spectrum of the model, both from the Bethe equations and from the dilatation generator, finding $\Delta_s= 4 \lambda^2(S_1(s)-S_{-1}(s)) + \mathcal{O}(\lambda^3)$.
It seems unreasonable to doubt complete two-loop integrability. Still, an explicit proof would be better. Constructing the ${\mathfrak{osp}}(6|4)$ Yangian that commutes with the Hamiltonian (\[eq:completeham\]) may be the simplest approach.
Further algebraic constructions of spin-chain generators for the ABJM gauge theory are possible. It would be good to calculate the complete $\mathcal{O}(\lambda)$ generators, extending the calculation here of two $\mathcal{O}(\lambda)$ supercharges acting within the ${\mathfrak{osp}}(4|2)$ sector. It would also be wonderful to obtain higher-loop corrections. For $\mathcal{N}=4$ SYM there is evidence of recursive structure to these corrections, at least in sectors of the theory [@Zwiebel:2008gr; @Bargheer:2008jt]. [@Bargheer:2008jt] also argues that such recursive structure appears in a compact sector for the ABJM spin chain (assuming higher-loop integrability). Perhaps this recursive structure extends to the ${\mathfrak{osp}}(4|2)$ sector, which could make an algebraic computation of the four-loop dilatation generator tractable. This would still require a direct field theory calculation of $h(\lambda)$, which appears in the one-magnon dispersion relation.
We found it useful to work with a representation of the spin module in terms of continuous variables, a light-cone superspace basis. We expect this and similar representations to be helpful for gaining new insights about the $\mathcal{N}=4$ SYM and the ABJM spin chains.
It is straightforward to lift the ${\mathfrak{osp}}(4|2)$ dilatation generator to the complete ${\mathfrak{osp}}(4|2)$ two-loop dilatation generator, including nonplanar corrections. $\hat{{\mathfrak{Q}}}$ ($\hat{{\mathfrak{S}}}$) has a unique nonplanar lift since it acts on one initial (final) module. The nonplanar two-loop dilatation generator than follows from the anticommutator (\[eq:qhatshatacomm\]), and this should match the ${\mathfrak{su}}(2) \times {\mathfrak{su}}(2)$ subsector result of [@Kristjansen:2008ib]. There is a very similar observation for the ${\mathfrak{psu}}(1,1|2)$ sector of $\mathcal{N}=4$ SYM [@Zwiebel:2005er]. Also for $\mathcal{N}=4$ SYM, an analysis of the gauge group structure of the Feynman diagrams that contribute to one-loop anomalous dimensions led to a unique lift from the planar limit to the full nonplanar theory [@Beisert:2003jj]. The product gauge group and bifundamental fields of ABJM probably make the corresponding two-loop lift here more difficult. Still, a lift from the nonplanar ${\mathfrak{osp}}(4|2)$ sector to the full theory should be possible, especially given the one-to-one maps of Section \[sec:lift\] between highest-weight three-site states of the ${\mathfrak{osp}}(4|2)$ and ${\mathfrak{osp}}(6|4)$ sectors.
The ${\mathfrak{osp}}(4|2)$ sector dilatation generator, including nonplanar corrections, should be useful for seeking the gauge duals of 1/12 BPS black holes in $AdS_4$. For recent related work and comments see [@Bhattacharya:2008bja]. It would also be interesting to calculate (nonplanar) anomalous dimensions of near ($1/12$) BPS states, as done previously [@Berkooz:2008gc] for near ($1/16$) BPS operators in $\mathcal{N}=4$ SYM.
Finally, this work’s confirmation of weak-coupling integrability further motivates study of the many topics related to the integrability of the ABJM gauge theory and its string theory dual. As for $AdS_5/CFT_4$, these topics range far beyond the weak-coupling spin chain.
Acknowledgements {#acknowledgements .unnumbered}
----------------
I am very grateful to Niklas Beisert, Andrei Belitsky, Diego Hofman, and Eric Ragoucy for discussions. I also thank Georgios Papathanasiou and Marcus Spradlin for pointing out important typos in Appendix D of the first version of this work.
Additional light-cone superspace computations \[sec:lightcone\]
================================================================
Working in light-cone superspace given by (\[eq:definelightcone\]) often simplifies calculations. We gave the actions of $\hat{{\mathfrak{Q}}}$ in light-cone superspace in (\[eq:lightconeqhat\]), and in this appendix we give the corresponding expression for $\hat{{\mathfrak{S}}}$. We then use this expression to compute the one-site interactions of $\mathcal{H}$ and and the two-loop twist-one spectrum.
$\hat{{\mathfrak{S}}}$ in light-cone superspace
-----------------------------------------------
We abbreviate $${\mathopen{\big|}x_1, \theta_1, \bar{\eta}_1;x_2, \bar{\theta}_2, \eta_2 ;
x_3, \theta_3, \bar{\eta}_3\mathclose{\bigr \rangle}} = {\mathopen{\big|}1;\bar{2},3\mathclose{\bigr \rangle}}.$$ Then, $\hat{{\mathfrak{S}}}$ acts as \[eq:lightconeshat\] [1;|[2]{},3]{} [[&=&]{}]{}- \_0\^1 \_0\^1 [\
&&]{}(( \_[bc]{} \_2\^b \_3\^c \_[x\_2]{} + \_|\_2\^ |\_3\^ \_[x\_3]{} ) [(1-t\_2) x\_1 + t\_1 t\_2 x\_2 + (1-t\_1) t\_2x\_3, \_1, (1-t\_2) |\_1]{}\
&+& ( \_[bc]{} \_1\^b \_2\^c \_[x\_2]{} + \_|\_1\^ |\_2\^ \_[x\_1]{} ) [ (1-t\_1)t\_2 x\_1 + t\_1 t\_2 x\_2 + (1-t\_2) x\_3, \_3, (1-t\_2) |\_3]{}\
&+& |\^\_2 |\_ \_[bc]{} \_1\^b \_3\^c [ t\_1 t\_2 x\_1 + (1-t\_2) x\_2 + (1-t\_1) t\_2 x\_3,, t\_2 |]{} ). [\
&&]{} To check this, first expand both sides according to (\[eq:definelightcone\]). Using the Beta integral one can then show that this is equivalent to the action of $\hat{{\mathfrak{S}}}$ given previously (the Hermitian conjugate of the $\hat{{\mathfrak{Q}}}$ action (\[eq:qhatsolution\])). Since the light-cone superspace basis is not manifestly Hermitian, here $\hat{{\mathfrak{S}}}$ takes a more involved form than $\hat{{\mathfrak{Q}}}$ does. To obtain the action on the conjugate state, ${\mathopen{\big|}\bar{1};2\;\bar{3}\mathclose{\bigr \rangle}}$, remove the overall minus sign and replace all $\theta,\eta$ with $\bar{\theta},\bar{\eta}$ and vice-versa (interchanging all Latin and Gothic indices).
One-site interactions of $\mathcal{H}$
--------------------------------------
Because physical spin-chain states are cyclic and have at least two sites, in general one can not uniquely classify interactions as one-site interactions rather than two-site (or three-site) interactions. However, we can isolate the contribution to $\mathcal{H}$ from $\hat{{\mathfrak{Q}}}$ replacing one site with three new sites followed by $\hat{{\mathfrak{S}}}$ replacing the same three sites with a single site. This is what we mean here by $\mathcal{H}_{\text{one site}}$. Using the light-cone superspace expressions (\[eq:lightconeqhat\]) and (\[eq:lightconeshat\]), we find the following expression for $ \mathcal{H}_{\text{one site}}{\mathopen{\big|}x, \theta, \bar{\eta}\mathclose{\bigr \rangle}}$, $$\begin{gathered}
\frac{1}{ 2 \pi} \int_0^1 \frac{\mathrm{d} t_1}{\sqrt{t_1}\sqrt{1-t_1}} \int_0^1 \frac{\mathrm{d}t_2}{t_2\sqrt{1-t_2}} \int_0^x \mathrm{d}\!y \, \times \notag \\
\bigg( 2 (\partial_{y_2} + \partial_{y_3}) {\mathopen{\big|}(1-t_2)x + t_1t_2 y_2 + (1-t_1)t_2 y_3, \theta, (1-t_2) \bar{\eta}\mathclose{\bigr \rangle}}_{|^{y_2,y_3=y}}
\notag \\
+ 2 \Big(\bar{\eta}^{\mathfrak{c}}\bar{\partial}_{\mathfrak{c}} \partial_x -\theta^c \partial_{c} \partial_{y_2} \Big){\mathopen{\big|}(1-t_1)t_2 x + t_1t_2y_2 + (1-t_2)y,\theta, (1-t_2) \bar{\eta}\mathclose{\bigr \rangle}}_{|^{y_2=y}}
\notag \\
+ 2 (\partial_{y_1} + \partial_{y_2}) {\mathopen{\big|}(1-t_1)t_2 y_1 + t_1t_2 y_2 + (1-t_2) x, \theta, (1-t_2) \bar{\eta}\mathclose{\bigr \rangle}}_{|^{y_1,y_2=y}} \bigg)
\notag \\
+ \frac{1}{2 \pi} \int_0^1 \frac{\mathrm{d}t_1}{\sqrt{t_1}\sqrt{1-t_1}} \int_0^1 \frac{\mathrm{d}t_2}{t_2\sqrt{1-t_2}} 2 \bar{\eta}^{\mathfrak{a}} \bar{\partial}_{\mathfrak{a}} {\mathopen{\big|}x, \theta, t_2\bar{\eta}\mathclose{\bigr \rangle}}.
\end{gathered}$$ Simplifying and doing integrals of derivatives and Beta integrals yields \_[1]{} [[&=&]{}]{}\_0\^1 ( 2 [x, , (1-t\_2) |]{} -2 [(1-t\_2)x , , (1-t\_2) |]{}) [\
&&]{} + \_0\^1 \_0\^1 [\
&&]{} ( |\^ |\_ -t\_1 ) ([x, , (1-t\_2) |]{} -[(1-t\_1)t\_2 x, , (1-t\_2)|]{} ) [\
&&]{} + 2 |\^ |\_ [x, , |]{}. We now expand the light-cone basis in components, with the $x^m \theta$ terms giving the action on ${\mathopen{\big|}\phi^{(m)}\mathclose{\bigr \rangle}}$ and the $x^m \bar{\eta}$ terms giving the action on ${\mathopen{\big|}\bar{\psi}^{(m)}\mathclose{\bigr \rangle}}$. The integrand of the middle two lines then simplifies to a finite sum of products of powers of the $t_i$ and $(1-t_i)$, so that the integrals again reduce to Beta integrals. Doing these integrals and the sum, and combining with contributions from the first and last line finally yields the result $$\mathcal{H}_{\text{one site}}{\mathopen{\big|}\phi_a^{(m)}\mathclose{\bigr \rangle}} = 2 S_1(2m) {\mathopen{\big|}\phi_a^{(m)}\mathclose{\bigr \rangle}}, \quad \mathcal{H}_{\text{one site}}{\mathopen{\big|}\bar{\psi}_{\mathfrak{a}}^{(m)}\mathclose{\bigr \rangle}} = 2 S_1(2m+1) {\mathopen{\big|}\bar{\psi}_{\mathfrak{a}}^{(m)}\mathclose{\bigr \rangle}}.$$ Of course, the same coefficients appear for the conjugate scalars and fermions. We also used the identity: $$S_1(m-{{{\textstyle\frac{1}{2}}}}) + S_1(m) + 2 \log 2 = 2 S_1(2m),$$ where the definition of the harmonic numbers that applies for nonintegers is the difference involving the digamma function $$S(x) = \psi(x+1)-\psi(1).$$
Two-loop twist-one spectrum
----------------------------
Working in light-cone superspace using (\[eq:lightconeqhat\]) and (\[eq:lightconeshat\]), we find the parallel expression to (\[eq:delta2sum\]), \[eq:delta2integral\] \_[s,2]{} [[&=&]{}]{}\_[m=0]{}\^[s]{} , [\
]{}(m) [[&=&]{}]{}\_[ms]{} (2 S(2s) + 2 S(1)) - [\
&&]{}+ y ((1-t\_1)t\_2y +(1-t\_2))\^[m-1]{} [\
&&]{}+ y (1-t\_1t\_2)\^[s-m-1]{}y\^[s-m-1]{} [\
&&]{}- y (t\_1+(1-t\_1)y)\^[s-m-1]{} [\
&&]{}- 4 y y\^[s-m-1]{}. The ranges of integration variables are all zero to one. The first integral should only be included in the sum for $m>0$, and the last three only for $m < s$. To obtain (\[eq:delta2integral\]) we expanded the light-cone basis in components and absorbed (most of) the wavefunction and light-cone basis normalizations into the integrands. Next, the $y$ integrals are elementary, and, using binomial expansions, all of the $t_i$ integrals can be done using the Beta integral. The sums can also be done[^20], provided we use the identity $$\frac{\sqrt{\pi}\Gamma(s+{{{\textstyle\frac{1}{2}}}})}{2}\sum_{m=1}^s\frac{(-1)^{m}}{m\Gamma(m+3/2)\Gamma(s-m+{{{\textstyle\frac{1}{2}}}})} = 2 S_{-1}(s) - S_1(s) + 2 -\frac{(-1)^s}{s+{{{\textstyle\frac{1}{2}}}}}.$$ An equivalent version for integer $s$ that is also valid for more general $s$ is $$\frac{{}_3F_2(1,{{\textstyle\frac{3}{2}}},s+1; s+2, s+{{\textstyle\frac{5}{2}}};1)}{(s+1)(2s+1)(2s+3)}=S_1(\frac{s-1}{2}) - S_1(\frac{s}{2}) + \frac{1}{s+{{{\textstyle\frac{1}{2}}}}}.$$ We have proved this identity through a somewhat involved calculation. Key steps include considering the difference of the identity at $s=s'$ and at $s=s'+1$, using the series expansion of hypergeometric functions, reintroducing Beta integrals (as well as another elementary integral), and switching orders of summation and integration. It would be nice to find an elegant proof of this identity, or better, a more elegant way to evaluate $\Delta_{s,2}$ directly from the light-cone superspace expressions for the supercharges. In any case, the final result is as given in (\[eq:twistonespectrum\]), $$\Delta_{s,2} = 4 \Big(S_1(s)-S_{-1}(s)\Big).$$
Bethe ansatz solution for two-loop twist-one spectrum \[sec:bethe\]
===================================================================
Here we work with the $\eta=-1$ (leading-order) Bethe equations of [@Gromov:2008qe]. The ${\mathfrak{sl}}(2)$ sector highest-weight state with Lorentz spin $(s+1/2)$ has $s$ pairs of roots $u_{4, k}=\bar{u}_{4, k}$. Labeling both types of roots $u_k$, the Bethe equations reduce to $$\frac{u_k+i/2}{u_k-i/2}=\prod_{j}\frac{u_k-u_j-i}{u_k-u_j+i}.$$ As done for the parallel calculation in $\mathcal{N}=4$ SYM [@Eden:2006rx] based on [@Korchemsky:1995be], we introduce the Baxter polynomial $$Q_s(u) = C_s\prod_k(u-u_k)$$ that satisfies $$T_s(u)Q_s(u) = (u+i/2) Q_s(u+i)-(u-i/2)Q_s(u-i)$$ for some auxiliary polynomial $T_s$. $C_s$ is a $u$-independent normalization factor. Matching powers of $u$ on each side requires $T_s$ to be independent of $u$, and for this equation to have a solution for all $s\geq0$ $T_s=(2 s+1)i$. The solution to a difference equation of this form is a Meixner polynomial (closely related to Jacobi and Krawtchouk polynomials) [@Koekoek:1996xx], $$Q_s(u)={}_2F_1(-s, iu + {{{\textstyle\frac{1}{2}}}}; 1; 2).$$ The two-loop contribution to the anomalous dimension is given by $$\sum_k \frac{2 \lambda^2}{u_k^2 + {{{\textstyle\frac{1}{4}}}}},$$ (the factor of $2$ is for the $u_4$ and $\bar{u}_4$ roots), which is proportional to the ratio of the coefficient of $u$ to the constant term in $Q_s(u+i/2)$. In terms of the expansion, $$\frac{Q_s(u+i/2)}{C_s} = \prod_k(u +i/2-u_k) = c^{(s)}_s u^s + c^{(s)}_{s-1} u^{s-1} + \ldots c^{(s)}_1 u + c^{(s)}_0,$$ this ratio is $4 i \, \lambda^2 c^{(s)}_1 / c^{(s)}_0$. The series expansion of the hypergeometric function gives $$Q_s(u +i/2)={}_2F_1(-s, iu ; 1; 2) = \sum_{k=0}^s \frac{(-s)_k (iu)_k}{(1)_k} \frac{2^k}{k!}, \quad (a)_k=\Gamma(a+k)/\Gamma(a),$$ which implies $$c^{(s)}_0 = 1, \quad c^{(s)}_1 = i \sum_{k=1}^s \frac{(-s)_k (k-1)! }{(1)_k} \frac{2^k}{ k!} = i \sum_{k=1}^s \frac{s! (-2)^k}{k(s-k)! k!}.$$ To find $c_1^{(s)}$ we evaluate c\^[(s)]{}\_1 - c\^[(s-1)]{}\_1 [[&=&]{}]{}i \_[k=1]{}\^s - i \_[k=1]{}\^[s-1]{} [\
]{}[[&=&]{}]{} (-2)\^s+ \_[k=1]{}\^[s-1]{} [\
]{}[[&=&]{}]{} ((-1)\^s -1), where we completed the binomial expansion of $(1- 2)^s$ to reach the last line. Since $c_1^{(0)} =0 $, $c_1^{(s)} = i\Big(S_{-1}(s) - S_1(s) \Big)$ and $$\Delta_{s,2} = 4 i \frac{c^{(s)}_1}{c^{(s)}_0} = 4 \Big(S_1(s) - S_{-1}(s)\Big).$$ As $s \rightarrow \infty$, $S_1(s) \rightarrow \log s$ while $S_{-1}(s) \rightarrow - \log 2$, so the cusp anomalous dimension has weak coupling expansion $$f(\lambda) = 4 \lambda^2 + \mathcal{O}(\lambda^3) .$$ We expect the methods of [@Kotikov:2008pv] could be used to obtain the four-loop correction to the twist-one spectrum, up to a currently unknown coefficient from $h(\lambda)$.
Local contribution to ${[{\mathfrak{Y}}^{aa},\hat{{\mathfrak{Q}}}]}$ \[sec:localycomm\]
=======================================================================================
In this appendix we complete the proof of ${\mathfrak{osp}}(4|2)$ Yangian symmetry of Section \[sec:yangian\] by computing what we called the local term of the commutator between $\hat{{\mathfrak{Q}}}$ and the ${\mathfrak{Y}}^{aa}$. Recall 4 \^[aa]{} [[&=&]{}]{} \_[i < j]{} 2 \_[bc]{} \^[bc]{}\_i \^[ca]{}\_j - \_ \^[a1 ]{}\_i \^[a2 ]{}\_j + \_ \^[a2 ]{}\_i \^[a1 ]{}\_j [\
]{}[[&=&]{}]{}\^[aa]{}\_ + \^[aa]{}\_[\^1]{} + \^[aa]{}\_[\^2]{}. So the local contributions of the form (\[eq:local\]) from ${\mathfrak{Y}}^{aa}_{{\mathfrak{R}}}$ are $$2 \varepsilon_{bc} \sum_{i} \Big( {\mathfrak{R}}^{ab}_i ({\mathfrak{R}}^{ac}_{i+1} +{\mathfrak{R}}^{ac}_{i+2} ) +{\mathfrak{R}}^{ab}_{i+1}{\mathfrak{R}}^{ac}_{i+2} \Big) {\mathfrak{\hat{Q}}}_i.$$ Using the expressions for the ${\mathfrak{R}}$ generators given in (\[eq:osp42lightcone\]), and the expression for $\hat{{\mathfrak{Q}}}$, \[eq:appqhat\] [x, , |]{} [[&=&]{}]{}\_0\^x y ( \_[ab]{} \^a\_2 \^b\_3 + \_ \^\_2 \^\_3 ) [x, , |; y, |\_2, \_2;y, \_3, |\_3]{} [\
&&]{} + \_0\^x y ( \_[ab]{} \^a\_1 \^b\_2 + \_ \_1\^ \_2\^ ) [ y, \_1, |\_1; y, |\_2, \_2; x, , |]{} [\
&&]{} -\_ |\^ \_2\^ \_[cd]{} \_1\^c\_3\^d [ x, \_1, |\_1;x, |\_2, \_2; x, \_3, |\_3 ]{} , we obtain 4 [\[\^[aa]{}\_,\]]{}\_ [x, , |]{} [[&=&]{}]{}\_0\^x y 2 \_2\^a \_3\^a [x, , |; y, |\_2, \_2;y, \_3, |\_3]{} [\
&&]{} + \_0\^x y 2 \_1\^a \_2\^a [ y, \_1, |\_1; y, |\_2, \_2; x, , |]{} [\
&&]{} - 2 \_|\^ |\_2\^ \_1\^a \_3\^a [ x, \_1, |\_1;x, |\_2, \_2; x, \_3, |\_3 ]{} . \[eq:rycom\] This above calculation is simplified using the vanishing of a single ${\mathfrak{R}}$ applied to ${\mathfrak{R}}$-singlet combinations of derivatives.
Next the local contributions from ${\mathfrak{Y}}^{aa}_{{\mathfrak{Q}}^1}$ are, after regrouping, $$\label{eq:twoparts}
- \varepsilon_{\mathfrak{bc}} \sum_{i} \Big(({\mathfrak{Q}}^{a1 \mathfrak{b}}_i + {\mathfrak{Q}}^{a1 \mathfrak{b}}_{i+1}){\mathfrak{Q}}^{a2 \mathfrak{c}}_{i+2} + {\mathfrak{Q}}^{a1 \mathfrak{b}}_i {\mathfrak{Q}}^{a2 \mathfrak{c}}_{i+1} \Big) {\mathfrak{\hat{Q}}}_i.$$ First we compute the action of the first term (including the two supercharge terms in parenthesis) on ${\mathopen{\big|}x, \theta, \bar{\eta}\mathclose{\bigr \rangle}}$. The three lines of (\[eq:appqhat\]) yield respectively, $$\begin{aligned}
& & \int_0^x \mathrm{d}y \, \bigg(-\varepsilon_{\mathfrak{bc}} {\mathfrak{Q}}_1^{a1\mathfrak{b}} \Big(2 y \partial_2^a\bar{\partial}_3^{\mathfrak{c}} +(2 y \partial_{y_3}+1)\bar{\partial}_2^{\mathfrak{c}} \partial_3^a \Big) \label{eq:11}
\notag \\
& & -4 \partial_2^a\partial_3^a y \partial_{y_3} -2 \partial_2^a\partial_3^a \bigg) {\mathopen{\big|}x, \theta, \bar{\eta}; y, \bar{\theta}_2, \eta_2;y, \theta_3, \bar{\eta}_3\mathclose{\bigr \rangle}}
\\
&&-\varepsilon_{\mathfrak{bc}} \partial_1^a \bar{\partial}_2^{\mathfrak{b}}{\mathfrak{Q}}_3^{a2\mathfrak{c}} \Big({\mathopen{\big|}x,\theta_1, \bar{\eta}_1;x, \bar{\theta}_2, \eta_2;x, \theta, \bar{\eta}\mathclose{\bigr \rangle}}-{\mathopen{\big|}0,\theta_1, \bar{\eta}_1;0, \bar{\theta}_2, \eta_2;x, \theta, \bar{\eta}\mathclose{\bigr \rangle}}\Big) \label{eq:12}
\\
&& +2x \partial_1^a\partial_2^a \varepsilon_{\mathfrak{bc}}\bar{\eta}^{\mathfrak{b}} \bar{\partial}_3^{\mathfrak{c}} {\mathopen{\big|} x, \theta_1, \bar{\eta}_1;x, \bar{\theta}_2, \eta_2; x, \theta_3, \bar{\eta}_3 \mathclose{\bigr \rangle}}. \label{eq:13}
\end{aligned}$$ In the first line we introduced the notation $\partial_{y_i}$ ($i=3$ in this case), which should be understood as $\partial_y$ acting only on the $i$th site. Similarly, we compute the action of the last term of (\[eq:twoparts\]) on ${\mathopen{\big|}x, \theta, \bar{\eta}\mathclose{\bigr \rangle}}$, again splitting according to the three lines of $\hat{{\mathfrak{Q}}}$ in (\[eq:appqhat\]), yielding $$\begin{aligned}
& & \int_0^x \mathrm{d}y \, \varepsilon_{\mathfrak{bc}} {\mathfrak{Q}}_1^{a1\mathfrak{b}}\Big(-(2 y\partial_{y_2} +1)\bar{\partial}_2^{\mathfrak{c}}\partial_3^a + 2 y \partial_2^a \bar{\partial}_3^{\mathfrak{c}}\Big){\mathopen{\big|}x, \theta, \bar{\eta}; y, \bar{\theta}_2, \eta_2;y, \theta_3, \bar{\eta}_3\mathclose{\bigr \rangle}} \label{eq:21}
\\
& & + \int_0^x \mathrm{d}y \, 4 y \partial_{y_1} \partial_1^a \partial_2^a {\mathopen{\big|} y, \theta_1, \bar{\eta}_1; y, \bar{\theta}_2, \eta_2; x, \theta, \bar{\eta}\mathclose{\bigr \rangle}} \label{eq:22}
\\
& & + 2 x \varepsilon_{\mathfrak{bc}} \bar{\partial}_1^{\mathfrak{b}} \partial_2^a \partial_3^a \bar{\eta}^{\mathfrak{c}} {\mathopen{\big|} x, \theta_1, \bar{\eta}_1;x, \bar{\theta}_2, \eta_2; x, \theta_3, \bar{\eta}_3 \mathclose{\bigr \rangle}}. \label{eq:23}\end{aligned}$$ It remains to compute the local terms from the last part of ${\mathfrak{Y}}^{aa}$ (\[eq:Ycc\]), labeled ${\mathfrak{Y}}^{aa}_{{\mathfrak{Q}}^2}$. However, as we noted previously, $\hat{{\mathfrak{Q}}}$ is odd under spin-chain parity $\mathbf{p}$, and it is straightforward to check that the Yangian generators are as well. Therefore, we find $${[{\mathfrak{Y}}^{aa}_{{\mathfrak{Q}}^2},\hat{{\mathfrak{Q}}}]}_{\text{local}}= \mathbf{p} {[{\mathfrak{Y}}^{aa}_{{\mathfrak{Q}}^1},\hat{{\mathfrak{Q}}}]}_{\text{local}}\mathbf{p}^{-1},$$ where $4 {[{\mathfrak{Y}}^{aa}_{{\mathfrak{Q}}^1},\hat{{\mathfrak{Q}}}]}_{\text{local}}$ is given by the sum of (\[eq:11\]-\[eq:13\]) and (\[eq:21\]-\[eq:23\]). So now we add (\[eq:11\]-\[eq:13\]) and (\[eq:21\]-\[eq:23\]), their images under $\mathbf{p}$, and (\[eq:rycom\]) to give the full local contribution to the commutator. Many terms cancel[^21]. Also, the second term of (\[eq:11\]) combines with the first term of (\[eq:21\]) to give \[eq:last\] -2 \_0\^x y \_y(y \_ \_1\^[a1]{} |\_2\^ \_3\^a [x, , |;y, |\_2, \_2;y, \_3,|\_3]{} ) [[&=&]{}]{}\
-2 x \_ \_1\^[a1]{} |\_2\^ \_3\^a [x, , |;x, |\_2, \_2;x, \_3,|\_3]{}. & & All that remains is the last term of (\[eq:rycom\]), (\[eq:12\]) and its $\mathbf{p}$ image, and this last expression (\[eq:last\]) and its $\mathbf{p}$ image. We write out all these terms, expanding the remaining ${\mathfrak{Q}}$ factors, yielding $$\begin{aligned}
& & - 2 \varepsilon_{\mathfrak{bc}}\bar{\eta}^{\mathfrak{b}} \bar{\partial}_2^{\mathfrak{c}} \partial_1^a \partial_3^a {\mathopen{\big|} x, \theta_1, \bar{\eta}_1;x, \bar{\theta}_2, \eta_2; x, \theta_3, \bar{\eta}_3 \mathclose{\bigr \rangle}}
\notag \\
& & -\varepsilon_{\mathfrak{bc}} \partial_1^a \bar{\partial}_2^{\mathfrak{b}} (2 x \theta^a \bar{\partial}_3^{\mathfrak{c}} + 2 x \partial_{x_3} \partial_3^a \bar{\eta}^{\mathfrak{c}} + \partial_3^a \bar{\eta}^{\mathfrak{c}} ) {\mathopen{\big|}x,\theta_1, \bar{\eta}_1;x, \bar{\theta}_2, \eta_2;x, \theta, \bar{\eta}\mathclose{\bigr \rangle}} - \text{parity}
\notag \\
& & -2 x \varepsilon_{\mathfrak{bc}} (\theta^a \bar{\partial}_1^{\mathfrak{b}} + \partial_{x_1} \partial_1^a \bar{\eta}^\mathfrak{b}) \bar{\partial}_2^{\mathfrak{c}} \partial_3^a {\mathopen{\big|}x, \theta, \bar{\eta};x, \bar{\theta}_2, \eta_2;x, \theta_3,\bar{\eta}_3\mathclose{\bigr \rangle}} - \text{parity}
\notag \\
& & + \varepsilon_{\mathfrak{bc}} \partial_1^a \bar{\partial}_2^{\mathfrak{b}}{\mathfrak{Q}}_3^{a2\mathfrak{c}} {\mathopen{\big|}0,\theta_1, \bar{\eta}_1;0, \bar{\theta}_2, \eta_2;x, \theta, \bar{\eta}\mathclose{\bigr \rangle}} + \text{parity}
{\nonumber\\}{{\mathrel{}&=&\mathrel{}}}\varepsilon_{\mathfrak{bc}} \partial_1^a \bar{\partial}_2^{\mathfrak{b}}{\mathfrak{Q}}_3^{a2\mathfrak{c}} {\mathopen{\big|}0,\theta_1, \bar{\eta}_1;0, \bar{\theta}_2, \eta_2;x, \theta, \bar{\eta}\mathclose{\bigr \rangle}} - \varepsilon_{\mathfrak{bc}} {\mathfrak{Q}}_1^{a2\mathfrak{b}} \bar{\partial}_2^{\mathfrak{c}} \partial_3^a {\mathopen{\big|}x, \theta, \bar{\eta};0, \bar{\theta}_2, \eta_2;0,\theta_3, \bar{\eta}_3\mathclose{\bigr \rangle}}
\notag \\
& \mapsto & \sum_i \varepsilon_{\mathfrak{bc}} \acute{Z}^{a\mathfrak{b}}_i{\mathfrak{Q}}_{i+1}^{a2\mathfrak{c}} +\sum_i \varepsilon_{\mathfrak{bc}} {\mathfrak{Q}}_i^{a2\mathfrak{c}} \grave{Z}^{a\mathfrak{b}}_{i+1}.\end{aligned}$$ The initial expression simplified as follows. The first line canceled against the terms without $x$ coefficients in the second line, and the remaining terms proportional to $x$ canceled in four pairs, leaving only the fourth line. The second-to-last line is simply expanding the fourth line (the parity term). We used the symbol $\mapsto$ in the last line for two reasons. We have lifted the one-to-three local site interaction to its homogeneous action on an infinite chain. Also, this expression includes a second set of local actions, on initial sites ${\mathopen{\big|}x, \bar{\theta}, \eta\mathclose{\bigr \rangle}}.$ For these the calculation is completely analogous. Starting from the action of $\hat{{\mathfrak{Q}}}$ on such sites, [x, |, ]{} [[&=&]{}]{}-\_0\^x y ( \_ |\^\_2 |\^\_3 + \_[ab]{} \^a\_2 \^b\_3 ) [x, |, ; y, \_2, |\_2;y, |\_3, \_3]{} [\
&&]{} - \_0\^x y ( \_ |\^\_1 |\^\_2 + \_[ab]{} \^a\_1 \^b\_2 ) [ y, |\_1, \_1; y, \_2, |\_2; x, |, ]{} [\
&&]{} + \_[ab]{} \^[a]{} \_2\^b \_ |\^\_1 |\^\_3 [ x, |\_1, \_1;x, \_2, |\_2; x, |\_3, \_3 ]{} . We again can compute the contributions to the local terms of ${\mathfrak{Y}}_{{\mathfrak{R}}}$, ${\mathfrak{Y}}_{{\mathfrak{Q}}^1}$ and (using $\mathbf{p}$) ${\mathfrak{Y}}_{{\mathfrak{Q}}^2}$. After simplification, we find & & \_ |\_1\^ \_2\^a \^[a 2 ]{}\_3 [ 0, |\_1, \_1; 0, \_2, |\_2; x, |, ]{} + \_ \^[a 2 ]{}\_1 \_2\^a |\_3\^ [ x, |, ; 0, \_2, |\_2; 0, |\_3, \_3]{}\
& & \_i \_ \^[a]{}\_i\_[i+1]{}\^[a2]{} +\_i \_ \_i\^[a2]{} \^[a]{}\_[i+1]{}. As claimed, this gives a precise cancellation with the remaining contribution from bilocal terms of (\[eq:bilocalYqhatcomm\]).
Projectors in the spin-module representation \[sec:projectors\]
===============================================================
The results of this section parallel those given for the ${\mathfrak{psu}}(1,1|2)$ sector of $\mathcal{N}=4$ SYM spin chain in Appendix E of [@Zwiebel:2008gr]. The expression for the dilatation generator (\[eq:osp42projectorham\]) depends on certain ${\mathfrak{osp}}(4|2)$ invariant generators. These generators act on pairs of different types of modules, and on pairs of identical modules. First, for the case of different types of modules we introduce the weighted sum of projectors, $$\mathcal{P}(c) = \sum_{i=0}^{\infty} c_i \mathcal{P}^{(i)},$$ where again $\mathcal{P}^{(i)}$ is the projector for ${\mathfrak{osp}}(4|2)$ spin $i$. So for $c_{i'} = \delta_{ii'}$, $\mathcal{P}(c)$ reduces to $\mathcal{P}^{(i)}$. More generally, $\mathcal{P}(c)$ can be written in terms of six component functions that depend on the $c_i$ (we suppress the argument $c$ of $\mathcal{P}$). [\_a\^[(j)]{} \_b\^[(n-j)]{}]{} [[&=&]{}]{}\_[k=0]{}\^ n ( p\_1(n, j, k) [\_a\^[(k)]{} \_b\^[(n-k)]{}]{} + p\_2(n, j, k) [\_b\^[(k)]{} \_a\^[(n-k)]{}]{} ), [\
&&]{}+ \_[k=0]{}\^n p\_3(n, j, k) \_[ab]{}\^ [|\_\^[(k)]{} |\_\^[(n-k)]{}]{}, [\
]{}[\_a\^[(j)]{} |\_\^[(n-j)]{}]{} [[&=&]{}]{}\_[k=0]{}\^ n p\_4(n, j, k) [\_a\^[(k)]{} |\_\^[(n-k)]{}]{} + \_[k=0]{}\^ [n-1]{} p\_5(n, j, k) [|\_\^[(k)]{} \_a\^[(n-k-1)]{}]{}, [\
]{}[|\_\^[(j)]{} \_b\^[(n-j)]{}]{} [[&=&]{}]{}\_[k=0]{}\^n p\_6(n, j, k) [|\_\^[(k)]{} \_b\^[(n-k)]{}]{} + \_[k=0]{}\^ [n+1]{} p\_5(n+1, k, j) [\_b\^[(k)]{} |\_\^[(n-k+1)]{}]{}, [\
]{}[|\_\^[(j)]{} |\_\^[(n-j)]{}]{} [[&=&]{}]{}\_[k=0]{}\^n (p\_1(n, n-j, n-k) [|\_\^[(k)]{} |\_\^[(n-k)]{}]{} + p\_2(n, n-j, n-k) [|\_\^[(k)]{} |\_\^[(n-k)]{}]{}) [\
&&]{}+ \_[k=0]{}\^n p\_3(n, k, j) \_ \^[cd]{} [\_c\^[(k)]{} \_d\^[(n-k)]{}]{}. The $p_l$ for $l=1,2$ only differ by a few minus signs, $$\begin{gathered}
p_l(n, j, k) = \sqrt{\frac{\pi \Gamma(j+{{{\textstyle\frac{1}{2}}}})\Gamma(k+{{{\textstyle\frac{1}{2}}}})(2(n-j)+1)!(2(n-k)+1)!}{j!k!}} \sum_{i=0}^n \bigg( \frac{2^{j+k-2n}(n-i)!}{4(n+i+1)!}
\notag \\
{}_3F_2^{\text{reg}}(-{{{\textstyle\frac{1}{2}}}}-i, -i, -j;{{{\textstyle\frac{1}{2}}}}, 1-i-j+n;1)C_l(i) {}_3F_2^{\text{reg}}(-{{{\textstyle\frac{1}{2}}}}-i, -i, -k;{{{\textstyle\frac{1}{2}}}}, 1-i-k+n;1) \bigg), \notag\end{gathered}$$ C\_1(i) [[&=&]{}]{}(i-1) c\_[i-1]{} + (2 + \_[i,0]{})c\_i + c\_[i+1]{}, [\
]{}C\_2(i) [[&=&]{}]{}-(i-1) c\_[i-1]{} + (2 -\_[i,0]{})c\_i - c\_[i+1]{}. \[eq:definepl\] The regularized hypergeometric functions are defined as ordinary hypergeometric functions divided by gamma functions, as $${}_3F_2^{\text{reg}}(a_1, a_2, a_3; b_1, b_2; z) = \frac{{}_3F_2(a_1, a_2, a_3; b_1, b_2; z)}{\Gamma(b_1)\Gamma(b_2)}.$$ The remaining four component functions can then be written relatively compactly in terms of these two. p\_4(n, j, k) [[&=&]{}]{} p\_1(n, n-j, n-k) + p\_2(n, n-j, n-k) [\
&&]{}+ (n-j-1) p\_2(n, j+1, k), [\
]{}p\_3(n, j, k) [[&=&]{}]{}- p\_1(n, n-j, n-k) + p\_4(n, j, k), [\
]{}p\_5(n, j, k) [[&=&]{}]{}(n-k-1)( p\_2(n-1, j, k) - p\_3(n-1, k, j-1)), [\
]{}p\_6(n, j, k) [[&=&]{}]{} p\_1(n, j, k) - p\_3(n, j, k). Note that the first (second) term in the expression for $p_5$ should be set to zero when $n=j$ ($j=0$) (as written, they are $0 \times \infty$). For use in (\[eq:osp42projectorham\]), one only needs to set $c_i = (-1)^i$, or $c_i = S_1(i)$.
For identical representations we instead define $\mathcal{P}(c)$ as $$\mathcal{P}(c) = \sum_{i=0}^{\infty} c_{i} \mathcal{P}^{(i-1/2)}.$$ Recall that in this case the ${\mathfrak{osp}}(4|2)$ spin is half-integer valued. Now the expansion of $\mathcal{P}(c)$ is [\_a\^[(j)]{} \_b\^[(n-j)]{}]{} [[&=&]{}]{}\_[k=0]{}\^ n ( p\_7(n, j, k) [\_a\^[(k)]{} \_b\^[(n-k)]{}]{} + p\_8(n, j, k) [\_b\^[(k)]{} \_a\^[(n-k)]{}]{} ) [\
&&]{}+ \_[k=0]{}\^ [n-1]{} p\_9(n, j, k) \_[ab]{}\^ [|\_\^[(k)]{} |\_\^[(n-k-1)]{}]{} [\
]{}[\_a\^[(j)]{} |\_\^[(n-j)]{}]{} [[&=&]{}]{}\_[k=0]{}\^ n (p\_[10]{}(n, j, k) [\_a\^[(k)]{} |\_\^[(n-k)]{}]{} + p\_[11]{}(n, j, k) [|\_\^[(k)]{} \_a\^[(n-k)]{}]{}), [\
]{}[|\_\^[(j)]{}\_b\^[(n-j)]{} ]{} [[&=&]{}]{}\_[k=0]{}\^ n (p\_[10]{}(n, n-j, n-k) [|\_\^[(k)]{}\_b\^[(n-k)]{} ]{} + p\_[11]{}(n, n-j, n-k) [\_b\^[k]{} |\_\^[(n-k)]{}]{}), [\
]{}[|\_\^[(j)]{}|\_\^[(n-j)]{} ]{} [[&=&]{}]{}\_[k=0]{}\^ n (p\_[12]{}(n, j, k)[|\_\^[(k)]{}|\_\^[(n-k)]{} ]{} + p\_[13]{}(n, j, k)[|\_\^[(k)]{}|\_\^[(n-k)]{} ]{} ) [\
&&]{}+ \_[k=0]{}\^[n+1]{} p\_9(n+1, k, j) \_ \^[cd]{}[\_c\^[(k)]{} \_d\^[(n-k+1)]{}]{}. $p_{10}$ actually takes the same form as given in (\[eq:definepl\]), with $$C_{10}(i) = 2 c_i + 2 c_{i+1}.$$ For $p_m$, $m=7, 8$ there are some shifts of arguments, $$\begin{gathered}
p_m(n, j, k) = \sqrt{\frac{\pi \Gamma(j+{{{\textstyle\frac{1}{2}}}})\Gamma(k+{{{\textstyle\frac{1}{2}}}})(2(n-j))!(2(n-k))!}{j!k!}} \sum_{i=0}^n \bigg( \frac{2^{1+j+k-2n}(n-i)!}{4(n+i)!}
\notag \\
{}_3F_2^{\text{reg}}({{{\textstyle\frac{1}{2}}}}-i, -i, -j;{{{\textstyle\frac{1}{2}}}}, 1-i-j+n;1)D_m(i) {}_3F_2^{\text{reg}}({{{\textstyle\frac{1}{2}}}}-i, -i, -k;{{{\textstyle\frac{1}{2}}}}, 1-i-k+n;1) \bigg), \notag\end{gathered}$$ D\_7(i) [[&=&]{}]{}(i-1) c\_[i-1]{} + (2- \_[i, 0]{}) c\_i + c\_[i+1]{}, [\
]{}D\_8(i) [[&=&]{}]{}-(i-1) c\_[i-1]{} + (2- \_[i, 0]{})c\_i - c\_[i+1]{}. The remaining four component functions again can be written in terms of ones defined earlier. p\_9(n, j, k) [[&=&]{}]{}(n-k-1)( p\_7(n, k+1, j) [\
&&]{}- p\_[10]{}(n-1, n-k-1, n-j) ), [\
]{}p\_[11]{}(n, j, k) [[&=&]{}]{} p\_8(n+1, n-j+1, n-k) [\
&&]{}+ p\_9(n+1, n-k, n-j), [\
]{}p\_[12]{}(n, j, k) [[&=&]{}]{}p\_[10]{}(n, n-j, n-k) + p\_9(n+1, j+1, k) [\
]{}p\_[13]{}(n, j, k) [[&=&]{}]{}- p\_[11]{}(n,n-j, n-k) - p\_9(n+1, j+1, k). [\
&&]{} Note that the second term in the expression for $p_9$ should be set to zero for $j=0$. For the conjugate case of $\mathcal{V}_{\bar \phi}^{(4|2)}$, simply remove (add) a bar from (to) all (un)barred module elements, keeping the same component functions. For (\[eq:osp42projectorham\]), in these cases we need to evaluate $\mathcal{P}(c)$ only for $c_i = S_1(i-1/2)$.
Unique lift from ${\mathfrak{sl}}(2)$ to ${\mathfrak{osp}}(4|2)$ \[sec:maptosl2\]
=================================================================================
The general argument of Section \[sec:uniquelift\] can be straightforwardly applied to lift the Hamiltonian uniquely from the ${\mathfrak{sl}}(2)$ sector to the ${\mathfrak{osp}}(4|2)$ sector. However, we still need to construct invertible maps between the ${\mathfrak{osp}}(4|2)$ sector and the ${\mathfrak{sl}}(2)$ sector, as done between the ${\mathfrak{osp}}(6|4)$ and ${\mathfrak{osp}}(4|2)$ sectors in Section (\[sec:invertiblemap\]). We complete this construction here. Again, simply constructing the map from the larger to smaller sector will be sufficient. Invertibility will follow, as in the last paragraph of Section \[sec:invertiblemap\].
The ${\mathfrak{osp}}(4|2)$ raising generators are ${\mathfrak{R}}^{22}$, $\tilde{{\mathfrak{R}}}^{22}$, ${\mathfrak{J}}^{22}$, and ${\mathfrak{Q}}^{a2\mathfrak{b}}$. Let ${\mathopen{\big|}\Omega_I\mathclose{\bigr \rangle}}$ be a linearly-independent basis of three-site highest-weight ${\mathfrak{osp}}(4|2)$ sector states with identical Cartan charges. Of course, we need not consider ${\mathopen{\big|}\Omega_I\mathclose{\bigr \rangle}}$ already in the ${\mathfrak{sl}}(2)$ sector (actually there are no such three-site ${\mathfrak{osp}}(4|2)$ highest-weight states). So, first assume the ${\mathopen{\big|}\Omega_I\mathclose{\bigr \rangle}}$ transform with respect to ${\mathfrak{su}}(2)_{{\mathfrak{R}}} \otimes {\mathfrak{su}}(2)_{\tilde{{\mathfrak{R}}}}$ as $(\mathbf{3} ,\mathbf{2}),$ or equivalently that the ${\mathopen{\big|}\Omega_I\mathclose{\bigr \rangle}}$ have ${\mathfrak{su}}(2)$ Cartan eigenvalues $[R^{12}, \tilde{R}^{12}]$ = $[-1, -1/2]$ (The third Cartan charge is unimportant for this section). Then consider $${\mathfrak{Q}}^{211}{\mathopen{\big|}\Omega_I\mathclose{\bigr \rangle}}.$$ These states have (${\mathfrak{su}}(2)$) Cartan charges $[-3/2, 0]$ and, equivalently, are in the ${\mathfrak{sl}}(2)$ subsector. However, a priori the ${\mathfrak{Q}}^{211}{\mathopen{\big|}\Omega_I\mathclose{\bigr \rangle}}$ could be zero. But if they were zero, then the highest-weight states ${\mathopen{\big|}\Omega_I\mathclose{\bigr \rangle}}$ would satisfy a second $1/12$ BPS condition with respect to ${\mathfrak{Q}}^{211}\sim {\mathfrak{Q}}_{14,1}$ and ${\mathfrak{Q}}^{122}\sim {\mathfrak{S}}^{14,1}$. This BPS condition(s) is inconsistent (for any possible value of $J^{12}$) with the ${\mathfrak{su}}(2)$ Cartan charges of the ${\mathopen{\big|}\Omega_I\mathclose{\bigr \rangle}}$. Therefore, ${\mathfrak{Q}}^{211}{\mathopen{\big|}\Omega_I\mathclose{\bigr \rangle}}$ is nonvanishing and in this case ${\mathfrak{Q}}^{211}$ gives the invertible map to the ${\mathfrak{sl}}(2)$ sector.
Up to interchanging ${\mathfrak{su}}(2)_{{\mathfrak{R}}}$ and ${\mathfrak{su}}(2)_{\tilde{{\mathfrak{R}}}}$ charges, the only other possibility for the Cartan charges of the ${\mathopen{\big|}\Omega_I\mathclose{\bigr \rangle}}$ is $[R^{12}, \tilde{R}^{12}]$ = $[-1/2, 0]$. Now descendants in the ${\mathfrak{sl}}(2)$ sector are given by $${\mathfrak{Q}}^{212}{\mathfrak{Q}}^{211} {\mathopen{\big|}\Omega_I\mathclose{\bigr \rangle}} = -{\mathfrak{Q}}^{211} {\mathfrak{Q}}^{212}{\mathopen{\big|}\Omega_I\mathclose{\bigr \rangle}},$$ where the equality follows from the vanishing anticommutator between these supercharges. The same BPS condition as in the previous case would apply if the ${\mathfrak{Q}}^{211} {\mathopen{\big|}\Omega_I\mathclose{\bigr \rangle}}$ were zero, and this is inconsistent with Cartan charges $[-1/2, 0]$. Therefore, we only need to show that it is impossible to simultaneously satisfy $${\mathfrak{Q}}^{212}{\mathfrak{Q}}^{211} {\mathopen{\big|}\Omega_I\mathclose{\bigr \rangle}} =0 \quad \text{and} \quad {\mathfrak{Q}}^{211} {\mathopen{\big|}\Omega_I\mathclose{\bigr \rangle}} \neq 0.$$ ${\mathfrak{Q}}^{211} {\mathopen{\big|}\Omega_I\mathclose{\bigr \rangle}} $ have Cartan charges $[-1, 1/2]$. Also, they are annihilated by ${\mathfrak{Q}}^{121} =({\mathfrak{Q}}^{212})^\dagger $ because of the anticommutation relation (\[eq:osp42anticom\]) and because $\tilde{{\mathfrak{R}}}^{11}$ annihilates the ${\mathopen{\big|}\Omega_I\mathclose{\bigr \rangle}}$ (they carries no $\tilde{{\mathfrak{R}}}$ charge). Therefore, if ${\mathfrak{Q}}^{212}$ annihilated the ${\mathfrak{Q}}^{211} {\mathopen{\big|}\Omega_I\mathclose{\bigr \rangle}} $, ${\mathfrak{Q}}^{211} {\mathopen{\big|}\Omega_I\mathclose{\bigr \rangle}} $ would satisfy another BPS condition, which again turns out to be incompatible with their Cartan charges. We conclude that in this case ${\mathfrak{Q}}^{212}{\mathfrak{Q}}^{211} {\mathopen{\big|}\Omega_I\mathclose{\bigr \rangle}}$ gives nonvanishing descendants in the ${\mathfrak{sl}}(2)$ sector, and therefore we can always construct the required invertible maps between the the ${\mathfrak{osp}}(4|2)$ sector and a ${\mathfrak{sl}}(2)$ sector.
[10]{}
O. Aharony, O. Bergman, D. L. Jafferis and J. Maldacena, *“[$\mathcal{N}=6$ superconformal Chern-Simons-matter theories, M2-branes and their gravity duals]{}”*, , [`arXiv:0806.1218`](http://arXiv.org/abs/arXiv:0806.1218) \[hep-th\]. J. A. Minahan and K. Zarembo, *“[The Bethe ansatz for superconformal Chern-Simons]{}”*, , [`arXiv:0806.3951`](http://arXiv.org/abs/arXiv:0806.3951) \[hep-th\]. D. Bak and S. J. Rey, *“[Integrable Spin Chain in Superconformal Chern-Simons Theory]{}”*, , [`arXiv:0807.2063`](http://arXiv.org/abs/arXiv:0807.2063) \[hep-th\]. J. A. Minahan and K. Zarembo, *“The Bethe-ansatz for [$\mathcal{N}=\mathord{}$4]{} super Yang-Mills”*, , [`arXiv:hep-th/0212208`](http://arXiv.org/abs/arXiv:hep-th/0212208). N. Beisert, C. Kristjansen and M. Staudacher, *“The Dilatation Operator of [$\mathcal{N}=\mathord{}$4]{} Conformal Super Yang-Mills Theory”*, , [`arXiv:hep-th/0303060`](http://arXiv.org/abs/arXiv:hep-th/0303060). I. Bena, J. Polchinski and R. Roiban, *“Hidden symmetries of the [$AdS_5\times S^5$]{} superstring”*, , [`arXiv:hep-th/0305116`](http://arXiv.org/abs/arXiv:hep-th/0305116). G. Arutyunov, S. Frolov and M. Staudacher, *“Bethe ansatz for quantum strings”*, , [`arXiv:hep-th/0406256`](http://arXiv.org/abs/arXiv:hep-th/0406256). M. Staudacher, *“The factorized S-matrix of CFT/AdS”*, , [`arXiv:hep-th/0412188`](http://arXiv.org/abs/arXiv:hep-th/0412188). N. Beisert, *“The SU(2$|$2) Dynamic S-Matrix”*, [`arXiv:hep-th/0511082`](http://arXiv.org/abs/arXiv:hep-th/0511082). R. A. Janik, *“The $AdS_5\times S^5$ superstring worldsheet S-matrix and crossing symmetry”*, , [`arXiv:hep-th/0603038`](http://arXiv.org/abs/arXiv:hep-th/0603038). G. Arutyunov, S. Frolov, J. Plefka and M. Zamaklar, *“The Off-shell Symmetry Algebra of the Light-cone $AdS_5\times S^5$ Superstring”*, , [`arXiv:hep-th/0609157`](http://arXiv.org/abs/arXiv:hep-th/0609157). N. Beisert, B. Eden and M. Staudacher, *“Transcendentality and Crossing”*, , [`arXiv:hep-th/0610251`](http://arXiv.org/abs/arXiv:hep-th/0610251). G. Arutyunov and S. Frolov, *“[Superstrings on $AdS_4 \times CP^3$ as a Coset Sigma-model]{}”*, , [`arXiv:0806.4940`](http://arXiv.org/abs/arXiv:0806.4940) \[hep-th\]. B. Stefa[ń]{}ski, Jr., *“[Green-Schwarz action for Type IIA strings on $AdS_4\times CP^3$]{}”*, , [`arXiv:0806.4948`](http://arXiv.org/abs/arXiv:0806.4948) \[hep-th\]. N. Gromov and P. Vieira, *“[The all loop $AdS_4/CFT_3$ Bethe ansatz]{}”*, , [`arXiv:0807.0777`](http://arXiv.org/abs/arXiv:0807.0777) \[hep-th\]. T. Nishioka and T. Takayanagi, *“[On Type IIA Penrose Limit and $\mathcal{N}=6$ Chern-Simons Theories]{}”*, , [`arXiv:0806.3391`](http://arXiv.org/abs/arXiv:0806.3391) \[hep-th\]. D. Gaiotto, S. Giombi and X. Yin, *“[Spin Chains in $\mathcal{N}=6$ Superconformal Chern-Simons-Matter Theory]{}”*, , [`arXiv:0806.4589`](http://arXiv.org/abs/arXiv:0806.4589) \[hep-th\]. G. Grignani, T. Harmark and M. Orselli, *“[The $SU(2) \times SU(2)$ sector in the string dual of $\mathcal{N}=6$ superconformal Chern-Simons theory]{}”*, , [`arXiv:0806.4959`](http://arXiv.org/abs/arXiv:0806.4959) \[hep-th\]. D. Berenstein and D. Trancanelli, *“[Three-dimensional $\mathcal{N}=6$ SCFT’s and their membrane dynamics]{}”*, , [`arXiv:0808.2503`](http://arXiv.org/abs/arXiv:0808.2503) \[hep-th\]. C. Ahn and R. I. Nepomechie, *“[$\mathcal{N}=6$ super Chern-Simons theory S-matrix and all-loop Bethe ansatz equations]{}”*, , [`arXiv:0807.1924`](http://arXiv.org/abs/arXiv:0807.1924) \[hep-th\]. D. Astolfi, V. G. M. Puletti, G. Grignani, T. Harmark and M. Orselli, *“[Finite-size corrections in the $SU(2) \times SU(2)$ sector of type IIA string theory on $AdS_4 \times CP^3$]{}”*, , [`arXiv:0807.1527`](http://arXiv.org/abs/arXiv:0807.1527) \[hep-th\]. P. Sundin, *“[The $AdS_4 \times CP^3$ string and its Bethe equations in the near plane wave limit]{}”*, , [`arXiv:0811.2775`](http://arXiv.org/abs/arXiv:0811.2775) \[hep-th\]. T. McLoughlin and R. Roiban, *“[Spinning strings at one-loop in $AdS_4 \times P^3$]{}”*, , [`arXiv:0807.3965`](http://arXiv.org/abs/arXiv:0807.3965) \[hep-th\]. L. F. Alday, G. Arutyunov and D. Bykov, *“[Semiclassical Quantization of Spinning Strings in $AdS_4 \times
CP^3$]{}”*, , [`arXiv:0807.4400`](http://arXiv.org/abs/arXiv:0807.4400) \[hep-th\]. C. Krishnan, *“[$AdS4/CFT_3$ at One Loop]{}”*, , [`arXiv0807.4561`](http://arXiv.org/abs/arXiv0807.4561) \[hep-th\]. T. McLoughlin, R. Roiban and A. A. Tseytlin, *“[Quantum spinning strings in $AdS_4 \times CP^3$: testing the Bethe Ansatz proposal]{}”*, , [`arXiv:0809.4038`](http://arXiv.org/abs/arXiv:0809.4038) \[hep-th\]. J. Gomis, D. Sorokin and L. Wulff, *“[The complete $AdS_4 \times CP^3$ superspace for the type IIA superstring and D-branes]{}”*, , [`arXiv:0811.1566`](http://arXiv.org/abs/arXiv:0811.1566) \[hep-th\]. N. Beisert, *“The Dilatation Operator of [$\mathcal{N}=\mathord{}$4]{} Super Yang-Mills Theory and Integrability”*, , [`arXiv:hep-th/0407277`](http://arXiv.org/abs/arXiv:hep-th/0407277). L. N. Lipatov, *“Evolution equations in QCD”*, in: *“Perspectives in hadronic physics”*, ed.: S. Boffi, C. Ciofi Degli Atti and M. Giannini, World Scientific (1998), Singapore. F. A. Dolan and H. Osborn, *“Superconformal symmetry, correlation functions and the operator product expansion”*, , [`arXiv:hep-th/0112251`](http://arXiv.org/abs/arXiv:hep-th/0112251). N. Beisert, *“The Complete One-Loop Dilatation Operator of [$\mathcal{N}=\mathord{}$4]{} Super Yang-Mills Theory”*, , [`arXiv:hep-th/0307015`](http://arXiv.org/abs/arXiv:hep-th/0307015). C. Sochichiu, *“[Dilatation operator in 3d]{}”*, , [`arXiv:0811.2669`](http://arXiv.org/abs/arXiv:0811.2669) \[hep-th\]. L. Dolan, C. R. Nappi and E. Witten, *“A Relation Between Approaches to Integrability in Superconformal Yang-Mills Theory”*, , [`arXiv:hep-th/0308089`](http://arXiv.org/abs/arXiv:hep-th/0308089). L. Dolan, C. R. Nappi and E. Witten, *“Yangian symmetry in $D=$4 superconformal Yang-Mills theory”*, [`arXiv:hep-th/0401243`](http://arXiv.org/abs/arXiv:hep-th/0401243), in: *“Quantum theory and symmetries”*, ed.: P. C. Argyres et al., World Scientific (2004), Singapore. N. Beisert and M. Staudacher, *“The [$\mathcal{N}=\mathord{}$4]{} SYM Integrable Super Spin Chain”*, , [`arXiv:hep-th/0307042`](http://arXiv.org/abs/arXiv:hep-th/0307042). M. Benna, I. Klebanov, T. Klose and M. Smedback, *“[Superconformal Chern-Simons Theories and $AdS_4/CFT_3$ Correspondence]{}”*, , [`arXiv:0806.1519`](http://arXiv.org/abs/arXiv:0806.1519) \[hep-th\]. M. A. Bandres, A. E. Lipstein and J. H. Schwarz, *“[Studies of the ABJM Theory in a Formulation with Manifest $SU(4)$ R-Symmetry]{}”*, , [`arXiv:0807.0880`](http://arXiv.org/abs/arXiv:0807.0880) \[hep-th\]. F. A. Dolan, *“[On Superconformal Characters and Partition Functions in Three Dimensions]{}”*, [`arXiv:0811.2740`](http://arXiv.org/abs/arXiv:0811.2740) \[hep-th\]. K. Hosomichi, K. M. Lee, S. Lee, S. Lee and J. Park, *“[$\mathcal{N}=5,6$ Superconformal Chern-Simons Theories and M2-branes on Orbifolds]{}”*, , [`arXiv:0806.4977`](http://arXiv.org/abs/arXiv:0806.4977) \[hep-th\]. J. B. Kogut and D. E. Soper, *“[Quantum Electrodynamics in the Infinite Momentum Frame]{}”*, . L. Brink, O. Lindgren and B. E. W. Nilsson, *“[$\mathcal{N}=4$ Yang-Mills Theory on the Light Cone]{}”*, . S. Mandelstam, *“[Light Cone Superspace and the Ultraviolet Finiteness of the $\mathcal{N}=4$ Model]{}”*, . A. V. Belitsky, S. E. Derkachov, G. P. Korchemsky and A. N. Manashov, *“Quantum integrability in (super) Yang-Mills theory on the light-cone”*, , [`arXiv:hep-th/0403085`](http://arXiv.org/abs/arXiv:hep-th/0403085). N. Beisert, *“The SU(2$|$3) Dynamic Spin Chain”*, , [`arXiv:hep-th/0310252`](http://arXiv.org/abs/arXiv:hep-th/0310252). N. Beisert and M. Staudacher, *“Long-Range PSU(2,2$|$4) Bethe Ans[ä]{}tze for Gauge Theory and Strings”*, , [`arXiv:hep-th/0504190`](http://arXiv.org/abs/arXiv:hep-th/0504190). A. V. Kotikov, L. N. Lipatov, A. Rej, M. Staudacher and V. N. Velizhanin, *“[Dressing and Wrapping]{}”*, , [`arXiv:0704.3586`](http://arXiv.org/abs/arXiv:0704.3586) \[hep-th\]. D. Gaiotto and X. Yin, *“[Notes on superconformal Chern-Simons-matter theories]{}”*, , [`arXiv:0704.3740`](http://arXiv.org/abs/arXiv:0704.3740) \[hep-th\]. B. I. Zwiebel, *“Yangian symmetry at two-loops for the ${\mathfrak{su}}(2|1)$ sector of [$\mathcal{N}=\mathord{}$4]{} SYM”*, , [`arXiv:hep-th/0610283`](http://arXiv.org/abs/arXiv:hep-th/0610283). P. P. Kulish, N. Y. Reshetikhin and E. K. Sklyanin, *“Yang-Baxter equation and representation theory: I”*, . H. J. de Vega and F. Woynarovich, *“[New integrable quantum chains combining different kinds of spins]{}”*, . S. M. Khoroshkin and V. N. Tolstoi, *“[Yangian double and rational R matrix]{}”*, [`arXiv:hep-th/9406194`](http://arXiv.org/abs/arXiv:hep-th/9406194). V. Stukopin, *“Quantum Double of Yangian of Lie Superalgebra $A(m,n)$ and computation of Universal $R$-matrix”*, [`arXiv:math/0504302`](http://arXiv.org/abs/arXiv:math/0504302). F. Spill, *“[Weakly coupled $\mathcal{N}=4$ Super Yang-Mills and $\mathcal{N}=6$ Chern-Simons theories from ${\mathfrak{u}}(2|2)$ Yangian symmetry]{}”*, , [`arXiv:0810.3897`](http://arXiv.org/abs/arXiv:0810.3897) \[hep-th\]. N. Beisert and B. I. Zwiebel, *“[On Symmetry Enhancement in the ${\mathfrak{psu}}(1,1|2)$ Sector of $\mathcal{N}=4$ SYM]{}”*, , [`arXiv:0707.1031`](http://arXiv.org/abs/arXiv:0707.1031) \[hep-th\]. B. I. Zwiebel, *“[Iterative Structure of the $\mathcal{N}=4$ SYM Spin Chain]{}”*, , [`arXiv:0806.1786`](http://arXiv.org/abs/arXiv:0806.1786) \[hep-th\]. T. Bargheer, N. Beisert and F. Loebbert, *“[Boosting Nearest-Neighbour to Long-Range Integrable Spin Chains]{}”*, , [`arXiv:0807.5081`](http://arXiv.org/abs/arXiv:0807.5081) \[hep-th\]. C. Kristjansen, M. Orselli and K. Zoubos, *“[Non-planar ABJM Theory and Integrability]{}”*, , [`arXiv:0811.2150`](http://arXiv.org/abs/arXiv:0811.2150) \[hep-th\]. B. I. Zwiebel, *“[$\mathcal{N}=\mathord{}$4]{} SYM to two loops: Compact expressions for the non-compact symmetry algebra of the ${\mathfrak{su}}(1,1|2)$ sector”*, , [`arXiv:hep-th/0511109`](http://arXiv.org/abs/arXiv:hep-th/0511109). J. Bhattacharya and S. Minwalla, *“[Superconformal Indices for $\mathcal{N}=6$ Chern Simons Theories]{}”*, , [`arXiv:0806.3251`](http://arXiv.org/abs/arXiv:0806.3251) \[hep-th\]. M. Berkooz and D. Reichmann, *“[Weakly Renormalized Near $1/16$ SUSY Fermi Liquid Operators in $\mathcal{N} = 4$ SYM]{}”*, , [`arXiv:0807.0559`](http://arXiv.org/abs/arXiv:0807.0559) \[hep-th\]. B. Eden and M. Staudacher, *“Integrability and transcendentality”*, , [`arXiv:hep-th/0603157`](http://arXiv.org/abs/arXiv:hep-th/0603157). G. P. Korchemsky, *“[Quasiclassical QCD pomeron]{}”*, , [`arXiv:hep-th/9508025`](http://arXiv.org/abs/arXiv:hep-th/9508025). R. Koekoek and R. F. Swarttouw, *“[The Askey-scheme of hypergeometric orthogonal polynomials and its q-analogue]{}”*, [`arXiv:math/9602214`](http://arXiv.org/abs/arXiv:math/9602214). A. V. Kotikov, A. Rej and S. Zieme, *“[Analytic three-loop Solutions for $\mathcal{N}=4$ SYM Twist Operators]{}”*, , [`arXiv:0810.0691`](http://arXiv.org/abs/arXiv:0810.0691) \[hep-th\].
[^1]: The all-loop proposal depends on an unknown interpolating function that appears in the one-magnon dispersion relation [@Nishioka:2008gz; @Gaiotto:2008cg],[@Grignani:2008is; @Berenstein:2008dc].
[^2]: $\mathcal{N}=4$ SYM actually has two types of ${\mathfrak{sl}}(2)$ sectors, with bosonic and fermionic module elements respectively. For our purposes it is not important which sector was used for each of the following steps, so we will not again refer to this distinction.
[^3]: The one-loop ${\mathfrak{sl}}(2)$ sector dilatation generator is fixed by the twist-two spectrum, which was also computed earlier in [@Lipatov:1997vu; @Dolan:2001tt].
[^4]: As in the $\mathcal{N}=4$ SYM calculations, we also use basic structural properties of the gauge theory, such as the range of planar interactions.
[^5]: An alternative approach would be to use the construction of [@Sochichiu:2008tw] of the two-loop dilatation generator for arbitrary three-dimensional Lorentz-invariant renormalizable theories. That is likely to be a much more difficult approach since it would not take full advantage of superconformal symmetry.
[^6]: Beyond leading order, as we will see, generic such states acquire anomalous dimension so that they do not satisfy the exact $1/12$ BPS condition $R^1{}_1 + R^2{}_2 - L^2{}_2 - D = 0$. Here the full scaling dimension is $D=D_0 + \delta D$, which includes an anomalous contribution.
[^7]: Of course, ${\mathfrak{L}}$ also still acts within this sector, and is proportional to the action of ${\mathfrak{R}}^{12}$.
[^8]: I thank A. Belitsky for suggesting such a basis, actually in the context of $\mathcal{N}=4$ SYM.
[^9]: Provided we project the right sides to their bosonic components. For example, the first term for ${\mathfrak{R}}^{ab}$ only includes $\theta^{\{a}\varepsilon^{b\}c} \partial_{\theta^c}$ and *not* $\theta^{\{a}\varepsilon^{b\}c} \partial_{\eta^c}$.
[^10]: We could also write $r_{\pm}(x) = \sqrt{\frac{\sqrt{\pi/2}\Gamma(x+1)}{2^{x\pm {{{\textstyle\frac{1}{2}}}}}\Gamma(x+1\pm{{{\textstyle\frac{1}{2}}}})}} \rightarrow \sqrt{\frac{\sqrt{\pi}\Gamma(x+1)}{\Gamma(x+1\pm{{{\textstyle\frac{1}{2}}}})}}$. Up to a factor of $\sqrt{2}$ in one case, we can use the second expression since the powers of $2$ will cancel in the expressions for the $c_i$ given here and later.
[^11]: Here we have chosen not to define $\mathbf{p}$ with a factor of $(-1)^L$ for chains with $2L$ sites, which would make all generators parity even.
[^12]: With grading $\eta=1$. For the opposite grading, one adds rather than removes a $u_3$ root at zero.
[^13]: We have obtained extra confirmation by also performed a number of checks of the Serre relation on two-site alternating or homogeneous chains. Of course, as stated above, the Serre relations are guaranteed to be satisfied because of the one-site result.
[^14]: It is also possible to absorb the identity component into the next-nearest neighbor terms since a permutation squared equals the identity, as does the sum (with unit weight) over the projectors for next-nearest neighbors. The coefficient would then be $(-1)^{j_1+j_3}({{{\textstyle\frac{1}{2}}}}S_1(j_2-1/2) + \log 2)$, which is rational.
[^15]: The absence of one- or two-site interactions slightly simplifies our argument, but is not essential.
[^16]: There are twelve isomorphic ${\mathfrak{osp}}(4|2)$ sectors from different choices of the $1/12$ BPS condition.
[^17]: Note that either all or none of the ${\mathfrak{Q}}_{12,1}{\mathopen{\big|}\Omega_I\mathclose{\bigr \rangle}}$ satisfy the BPS condition since they all have the same Cartan charges.
[^18]: I thank E. Ragoucy for helpful related discussions.
[^19]: A degenerate Cartan matrix is an obstruction to constructing a R-matrix, but that is not a problem for ${\mathfrak{osp}}(6|4)$ since its Cartan matrix is invertible.
[^20]: We have used to evaluate these sums, symbolically as a function of $s$.
[^21]: Specifically, the cancellations are: last term of (\[eq:11\]) with first term of (\[eq:rycom\]) (and $\mathbf{p}$ image cancellation), first term of (\[eq:11\]) with second term of (\[eq:21\]), (\[eq:13\]) with $\mathbf{p}$ of (\[eq:23\]) (and $\mathbf{p}$ image), second-to-last term of (\[eq:11\]) with $\mathbf{p}$ of (\[eq:22\]) (and $\mathbf{p}$ image).
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We discuss the physical consequences of general phase space deformations on the minisuperspace of phantom cosmology. Based on the principle of physically equivalent descriptions in the deformed theory, we investigate for what values of the deformation parameters the arising descriptions are physically equivalent. We also construct and solve the quantum model and derive the semiclassical dynamics.'
address: |
$^a$ Departamento de Matemáticas, Facultad de Ciencias\
Universidad Autónoma de Baja California, Ensenada, Baja California, México.\
$^b$Department of Theoretical Physics, University of the Basque Country UPV/EHU, P.O. Box 644, 48080 Bilbao, Spain.\
$^c$ Departamento de Física de la Universidad de Guanajuato,\
A.P. E-143, C.P. 37150, León, Guanajuato, México.\
author:
- 'J. L. López$^a$'
- 'M. Sabido$^{b,c}$'
- 'C. Yee-Romero$^a$'
title: 'Phase space deformations in phantom cosmology.'
---
Noncommutative cosmology, deformed phase space models, late time acceleration.
Introduction
============
One of the most intriguing aspects in physics, is the current acceleration of the Universe. Is it a consequence of modifications to general relativity (GR) or is it a new kind of matter that drives this acceleration?. Although having some theoretical problems, the best answer to this question is the cosmological constant $\Lambda$. An alternative candidate that has been successful for the description of dark energy is the scalar field [@SF1; @ratra; @SF2; @SF3; @Nojiri; @Pad]. In particular, scalar fields with a negative kinetic term have been considered in the literature. This type of fields are known as Phantom fields. They have some interesting properties that allows them to be considered as a strange but viable matter which could be relevant in the evolution of the Universe [@Rybakov]. In particular, a phantom field provides an effective negative pressure and a repulsive effect on the matter content of the Universe which in the long term could be responsible for the current accelerated expansion [@Perlmutter; @Riess1; @Riess2; @Riess3; @Kiefer]. Therefore it has been considered as the matter source of the late time accelerated expansion of the Universe [@Cald1; @Cline; @Hu].
When considering the spacetime structure of the Universe it is usually done in reference to GR. But when regarding the micro structure of spacetime we do not have a universally accepted quantum theory of gravity. Although there are several candidates, noncommutativity has been considered as an alternative to understand the small scale structure of the Universe and help in the construction of quantum theory of gravity. For this reason, noncommutative versions of gravity have been constructed [@Obregon1; @Calmet; @Wess; @NC2]. Noncommutativity is usually believed to be present near Planck’s scale and is consistent with a discrete nature of spacetime. Motivated by this idea, it is [justified to consider an inherently noncommutative spacetime at the early ages of the universe. Directly using noncommutative gravity is quite difficult, this is a consequence of the highly nonlinear character of these theories[@Obregon1; @Calmet; @Wess; @NC2]. Fortunately there is an alternative, in [@Obregon2] the authors introduce]{} the effects of noncommutativity using the Moyal product of functions on the Wheeler-DeWitt (WDW) equation.\
The effects of the noncommutative deformation at the classical level was studied by a WKB approximation of noncommutative quantum model [@eri1] and also by modifying the Poisson algebra [@Obregon3; @bastos]. More general minisuperspace deformations have been done in connection with $\Lambda$ [@NC3; @vakili; @Miguel1; @yee; @eri2; @Miguel2]. Phase space deformations give rise to two physically nonequivalent descriptions, the [*“C-frame"*]{} based on the original variables but with a modified interaction and the [*“NC-frame"*]{} frame based on the deformed variables but with the original interaction. Given this ambiguity, in [@Miguel2] a principle was proposed to restrict the value of the deformation parameters in order to make both descriptions physically equivalent.
In this work we consider a Friedmann-Robertson-Walker (FRW) cosmological model coupled to a phantom scalar field and study the physical consequences of introducing general phase space deformations on the minisuperspace of the theory. We will study both, the classical and quantum models and we also find that the semiclassical approximation of the deformed quantum model agrees with the classical model. The deformation parameter space is determined by considering the principle of physically equivalent frames [@Miguel2].
The paper is organized as follows. In section 2, the commutative model is presented. In section 3, the minisuperspace phase space deformation is implemented and the dynamics in the two frames is obtained. Also the deformation parameters are constrained by imposing physical equivalence between the [*“C-frame"*]{} and the [*“NC-frame"*]{}. The quantum analysis is done in section 4, we find the solution to the deformed WDW equation and fix the parameters in the deformation in order to make the quantization possible. We also show that the classical paths arise from the semiclassical approximation of the WDW equation obtained from the deformed Hamiltonian. Finally, section 5 is devoted for concluding remarks.
Phantom field model
===================
We start with the flat FRW metric
$$ds^2 = -N^2(t)dt^2 + a^2(t)[ dr^2 + r^2d\Omega]$$
$a(t)$ corresponds to the scale factor and $N(t)$ is lapse function. In this background the action of a minimally coupled phantom scalar field $\varphi (t)$ with constant potential is $$\label{accion}
S = \int dt \left \{\ -\frac{3a \dot{a}^2}{N} - a^3\left( \frac{\dot{\varphi}^2}{2N} + N\Lambda \right) \right \}\ ,$$ where we have set the units so $8\pi G = 1$. The minus sign in the kinetic term of the scalar action is the difference between the usual scalar field and the phantom scalar field [@Kiefer]. The canonical Hamiltonian derived from Eq.(\[accion\]) is
$$\label{ham}
-N \left[ \frac{P_a^2}{12a} +\frac{P_{\varphi}^2}{2a^3} - a^3 \Lambda \right]. \label{Ham1}$$
With the the following change of variables $$\begin{aligned}
x &=& \mu^{-1}a^{3/2}\sin{(\mu \varphi)}, \label{Trans1} \\ \nonumber
y &=& \mu^{-1}a^{3/2}\cos{(\mu \varphi)},\end{aligned}$$ and $\mu = \sqrt{3/8}$. The Hamiltonian Eq.(\[ham\]) can be rewritten as a sum of two harmonic oscillators $$H = N\left( \frac{1}{2}P_x^2 + \frac{\omega^2}{2}x^2 \right) + N\left( \frac{1}{2}P_y^2 + \frac{\omega^2}{2}y^2 \right), \label{Ham1}$$ where $\omega^2 = -\frac{3}{4}\Lambda$. When one considers the usual scalar field, the Hamiltonian is transformed to a “ghost oscillator” which is simply a difference of two harmonic oscillators [@Miguel1; @Miguel2].
Deformed Phase Space Model
==========================
There exist different approaches to incorporate noncommutativity into physical theories. Particularly, in cosmology there is a broadly explored path to study noncommutativity [@Obregon2], where noncommutativity is realized in the so called minisuperspace variables. We will follow a deformed phase space approach. The deformation is introduced by the Moyal bracket $\{f,g\}_{\alpha}=f\star_{\alpha}g-g\star_{\alpha}f$. By substituting the usual product with the Moyal product $(f\star{g})(x)=\exp{\left[\frac{1}{2}\alpha^{ab}\partial_{a}^{(1)}\partial_{b}^{(2)}\right]}f(x_{1})g(x_2)\vert_{x_1=x_2=x}$ such that $$\begin{aligned}
\alpha =
\left( {\begin{array}{cc}
\theta_{ij} & \delta_{ij}+\sigma_{ij} \\
- \delta_{ij}-\sigma_{ij} & \beta_{ij} \\
\end{array} } \right) .\end{aligned}$$ The $2\times 2$ antisymmetric matrices $\theta_{ij}$ and $\beta_{ij}$ represent the noncommutativity in the coordinates and momenta respectively. The $\alpha$ deformed algebra becomes
$$\{x_i,x_j\}_{\alpha}=\theta_{ij},\{x_i,p_j\}_{\alpha}=\delta_{ij}+\sigma_{ij},\; \{p_i,p_j\}_{\alpha}=\beta_{ij}.\label{alg}$$
In this work we use the particular deformations, $\theta_{ij}=-\theta\epsilon_{ij}$ and $\beta_{ij}=\beta\epsilon_{ij}$.
There is an alternative to derive a similar algebra to Eq.(\[alg\]). For this we follow the procedure given in [@NC3; @Miguel1]. Start with the transformation $$\begin{aligned}
\widehat{x}=x+\frac{\theta}{2}P_{y}, \qquad \widehat{y}=y-\frac{\theta}{2}P_{x},\nonumber \\
\widehat{P}_{x}=P_{x}-\frac{\beta}{2}y, \qquad \widehat{P}_{y}=P_{y}+\frac{\beta}{2}x, \label{nctrans}\end{aligned}$$ on the classical phase space variables $\{x,y,P_x,P_y\}$, these are the variables that satisfy the usual Poisson algebra. The new variables satisfy a deformed algebra $$\{\widehat{y},\widehat{x}\}=\theta,\; \{\widehat{x},\widehat{P}_{x}\}=\{\widehat{y},\widehat{P}_{y}\}=1+\sigma,\; \{\widehat{P}_y,\widehat{P}_x\}=\beta,\label{dpa}$$ where $\sigma=\theta\beta/4$. Furthermore, as in [@NC3; @Miguel1], we assume that the deformed variables satisfy the same relations as their commutative counterparts. The resulting algebra is the same, but the Poisson bracket is different in the two algebras. In Eq.(\[alg\]), the brackets are the $\alpha$ deformed ones related to the Moyal product, for the other algebra the brackets are the usual Poisson brackets.
To construct the deformed theory, we start with a Hamiltonian which is formally analogous to Eq.(\[Ham1\]) but constructed with the variables that obey the modified algebra Eq.(\[dpa\]), this gives the deformed Hamiltonian $$H_{nc} = \frac{1}{2} \left[ (P_x^2 + P_y^2) + \ell^2(xP_y-yP_x) + \widetilde{\omega}^2(x^2+y^2) \right], \label{Ham2}$$ where $\ell^2$ and $\widetilde{\omega}^2$ are given by
$$\ell^2 = \frac{\beta + \omega^2 \theta}{1 + \frac{\omega^2 \theta^2}{4}}, ~~~~~\tilde{\omega}^2 = \frac{\omega^2 + \frac{\beta^2}{4}}{1 + \frac{\omega^2 \theta^2}{4}}.$$
There is a significant difference between the transformed Hamiltonian of the scalar field cosmology model [@Miguel1] and Eq. (\[Ham2\]), the deformed Hamiltonian of the phantom cosmology model. Unlike the scalar field case, the crossed term involving position and momentum in the phantom model, corresponds to an angular momentum term. To understand the physics of the deformation, it is necessary to remember that the deformation Eq.(\[nctrans\]) defines two physical nonequivalent descriptions, the [*“C-frame"*]{} where the effects of the deformation are interpreted as a commutative space $(x,y)$ but with modification of the original interaction and the [*“NC-frame"*]{} where we work with the deformed variables $(\widehat{x},\widehat{y})$ and the original interaction. In general, the dynamics in the two frames is different but the physical interpretation in the [*“C-frame"*]{} is easier. In this frame we can interpret the deformed model as a bidimensional harmonic oscillator and the minisuperspace deformation comes into play as an angular momentum term in the Hamiltonian. For this reason we will do the calculations in the [*“C-frame"*]{}.
We obtain the equations of motion from the Hamiltonian Eq.(\[Ham2\]), which in the $(x,y)$ variables are given by
$$\begin{aligned}
\label{nceqm}
\dot{x}&=& P_{x}-\frac{1}{2}\ell^2{y},\qquad
\dot{y}= P_{y}+\frac{1}{2}\ell^2{x},\\
{\dot{P}_{x}}&=&-\frac{1}{2}\ell^2 P_{y}-\widetilde{\omega}^2{x}, \qquad
{\dot{P}_{y}}=\frac{1}{2}\ell^2 P_{x}-\widetilde{\omega}^2{y}, \nonumber\end{aligned}$$
and get $$\begin{aligned}
\ddot{x} + \ell^2 \dot{y} + \left( \widetilde{\omega}^2 -\frac{\ell^4}{4} \right)x &=&0, \label{EqMot} \\ \nonumber
\ddot{y} - \ell^2 \dot{x} + \left( \widetilde{\omega}^2 -\frac{\ell^4}{4} \right)y &=& 0.\end{aligned}$$ With the transformation $z=x+iy$ we can easily solve Eq.(\[EqMot\]). We have three different solutions depending on the sign of $\widetilde{\omega}^2$. For $\widetilde{\omega}^2 > 0$, we get
$$\begin{aligned}
\label{Sol1}
x(t) &=& A_1 \cos{[( \ell^2/2 + |\widetilde{\omega}|)t ]}
+ B_1 \cos{[(\ell^2/2 - |\widetilde{\omega}|)t]}, ~~~~~~\\
y(t) &=& A_1 \sin{[(\ell^2/2 +|\widetilde{\omega}|)t]}
+ B_1 \sin{[(\ell^2/2 - |\widetilde{\omega}|)t]},~~~~~~\nonumber\end{aligned}$$
$A_1,B_1$ are arbitrary constants that, from the Hamiltonian constraint, satisfy $(\ell^2+2|\widetilde{\omega}|)A_1-(\ell^2-2|\widetilde{\omega}|)B_1=0$. When $\widetilde{\omega}^2 < 0$ we have the solutions $$\begin{aligned}
\label{Sol2}
x(t) &=& \left[A_3 \cosh{(|\widetilde{\omega}| t)}+B_3 \sinh{(|\widetilde{\omega}| t)}\right]\cos{\left( \frac{\ell^2}{2} t \right)}, \\ \nonumber
y(t) &=& \left[A_3 \cosh{(|\widetilde{\omega}| t)}+B_3 \sinh{(|\widetilde{\omega}| t)}\right]\sin{\left( \frac{\ell^2}{2} t \right)},\end{aligned}$$ for this case the integrating constants satisfy the condition $A_3^2=B_3^2$. The last solution is for the case $\widetilde{\omega}= 0$, $$\begin{aligned}
\label{Sol3}
x(t) &=& \left(A_2 +B_2 t\right) \cos{ \left(\frac{\ell^2}{2} t \right)}, \\
y(t) &=& \left(A_2 +B_2 t\right) \sin{ \left(\frac{\ell^2}{2} t \right)}, \nonumber\end{aligned}$$ the constraint imposes $B_2=0$. The condition arises when the cosmological constant satisfies $2\beta = -3\Lambda$.
In analyzing the evolution of the Universe, we will focus our attention on its volume and calculate $a^3(t)$ in the two frames. Let us start in the [*“C-frame"*]{}, using the solutions for the model, Eq.(\[Sol1\]), Eq.(\[Sol2\]) and Eq.(\[Sol3\]), together with the change of variables in Eq.(\[Trans1\]) we get $$a^3(t) = \left \{
\begin{array}{ccc} \label{CVolume}
& V^{(1)}_0 + V^{(1)}_1 \cos^2{(\widetilde{\omega}t)} , ~~~~~~~~~~~ \widetilde{\omega}^2 > 0, \\
& ~ \\
& V^{(2)}_0 , ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ \widetilde{\omega}^2 = 0, \\
& ~ \\
& V^{(3)}_0 e^{2\tilde{\omega} t}, ~~~~~~~~~~~~~~~~~~~~~~~~~ \widetilde{\omega}^2 < 0, \\
\end{array} \right.$$ where $V^{(i)}_0,V^{(i)}_1$ are constructed from the integrating constants. For the first case, $V^{(1)}_0=(A_1-B_1)^2$ and $V^{(1)}_1=4A_1B_1$, we must remember that the constants are related. For the second case the volume is the square of a constant. Finally for the last case we have $V^{(3)}_0=A_3^2$, but if we take $A_3=-B_3$ we have an exponentially decaying volume.
To study the dynamics in the [*“NC-frame"*]{} we use the [*“C-frame”*]{} solutions. The volume is constructed with the [*“NC-frame"*]{} variables, $\widehat{a}^3(t) = m^2(\widehat{x}^2 + \widehat{y}^2)$ $$\widehat{a}^3(t) = \left\{
\begin{array}{ccc} \label{NCVolume}
& \widehat{V}^{(1)}_0 + \widehat{V}^{(1)}_1 \cos^2{(\widetilde{\omega}t)} , ~~~~~~~~~~~ \widetilde{\omega}^2 > 0, \\
& ~ \\
& \widehat{V}^{(2)}_0, ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ \widetilde{\omega}^2 = 0, \\
& ~ \\
& \widehat{V}^{(3)}_1e^{2\tilde{\omega} t}, ~~~~~~~~~~~~~~~~~~~~~~~~~~ \widetilde{\omega}^2 < 0, \\
\end{array} \right.$$ where $\widehat{V}^{(i)}_0,\widehat{V}^{(i)}_1$ are constants related to $V^{(i)}_0,V^{(i)}_1$. [It is important to mention that even if we have a positive volume in the [*“C-frame"*]{} we might need to impose some restriction in order to have a valid description in the [*“NC-frame"*]{}.]{} In order to determine the valid range of the deformation parameters we use the principle used in [@Miguel2] which states the following, [*“Deformed phase space models are only valid when the [*ÒNC-frameÓ*]{} and [-frameÓ]{} descriptions are physically equivalent”*]{}. To impose that the description in the two frames should be physically equivalent, we start by analyzing the expression for the [*“C-frame"*]{} volume. It is easy to show that the volume is always positive. In the first case one obtains that $V^{(1)}_0+V^{(1)}_1=(A_1+B_1)^2\ge0$ so is positive for any $t$. For the second case the volume is also positive, and finally for the third case the constant $V^{(3)}_1>0$. Then we conclude that for the three cases, the [*“C-frame"*]{} volume is always positive for any choice of the integration constants. [The constants in the [*“NC-frame"*]{} solutions are related to the [*“C-frame"*]{} constants as follows. For the first case, $\widehat{A}_1=(1+|\widetilde{\omega}|\theta/2)A_1$ and $\widehat{B}_1=(1-|\widetilde{\omega}|\theta/2)B_1$, for the second case the volume is the same and for the last case the constant in the [*“NC-frame"*]{} is related by $\widehat{A}^2_3=(1+\tilde{\omega}^2\theta^2/4)A_3^2$.]{}
We see from Eq. (\[CVolume\]) and Eq.(\[NCVolume\]) that for $\theta = 0$, the two frames are physically equivalent. We can conclude that $\theta=0$ and $\beta\ne0$ are valid values for the deformation parameters. Now we consider the general case $\theta \neq 0$ and $\beta\ne0$. It is clear that in the first two cases, that the volume has the same physical behavior in the two frames then there are no restriction for the deformation parameters. For the third case, we have the same behavior as long as we satisfy the condition $1+\frac{\theta^2 \widetilde{\omega}^2}{4} >0$. Furthermore, the evolution is that of an exponential scale factor for an accelerating Universe. Then comparing with a de-Sitter Universe we find (only for the third case), an effective positive cosmological constant
$$\Lambda_{Eff} = \frac{1}{3}\left( \frac{3\Lambda - \beta^2}{1-\frac{3\theta^2\Lambda}{16}} \right). \label{LambdaEf}$$
This effective cosmological constant is determined by the deformation parameters as well as the value of $\Lambda$. Since current observations point to a positive cosmological constant. This gives and extra condition that restricts the values of the deformation parameters by imposing that $\Lambda_{eff}>0$. After fixing the value of $\Lambda$, we have three definite restrictions:
1. $\Lambda_{Eff} > 0$. \[Conditions\]
2. $f(\beta,\Lambda) = 3\Lambda -\beta^2$ and $g(\theta,\Lambda) = 1-\frac{3\theta^2 \Lambda}{16}$ must have the same sign.\
3. $ 1+\frac{\theta^2 \widetilde{\omega}^{ 2}}{4} > 0$.
From ii), we can see in Fig.(1) and Fig.(2) the regions where these inequalities are satisfied. We did not consider $\Lambda \leq 0$, since condition ii) would not be satisfied, for this reason the plots only show the region where $\Lambda > 0$.
{width="7cm"} \[Figure1\]
{width="7cm"} \[Figure2\]
We want to emphasize that for the phantom scalar field we cannot define an effective cosmological constant when $\Lambda = 0$, as it was done for the usual scalar field in [@Miguel1]. When setting $\Lambda=0$ we will have $\widetilde{\omega}^2>0$ meaning that we are in the first case and therefore we will not have an exponential behavior. For this model the deformation maps an initial positive cosmological constant model to another positive cosmological constant model, the two cosmological constants have different values but they are both positive. This fact also leads to a more restrictive space of parameters than the one analyzed in [@Miguel2].
Noncommutative quantum model
============================
Now we proceed with the quantization of the deformed Hamiltonian. The usual canonical quantization approach $H_{nc} \psi = 0$, gives the corresponding WDW equation. The Hamiltonian $H_{nc}$, explicitly contains the angular momentum operator $(xP_y - yP_x) = L_z$ and can be written in terms of the Hamiltonian of a 2-dimensional harmonic oscillator $H_{xy}$, with frequency $\widetilde{\omega}^2$ plus an angular momentum term, that is $H_{nc} = H_{xy} + L_z$. Using the similarity of the WDW equation, with the 2-dimensional harmonic oscillator, we will follow the approach in [@quantum; @quantum2]. First we write the eigenvalue equation for the harmonic oscillator, then we will find the energy eigenvalues and finally impose the zero energy condition. After writing the quantum equation in polar coordinates, we obtain the equation for the energy eigenstates, and is given by
$$-\frac{\hbar^2}{2}\left[ \frac{\partial^2 \psi}{\partial r^2} + \frac{1}{r}\frac{\partial \psi}{\partial r} + \frac{1}{r^2}\frac{\partial^2 \psi}{\partial \phi^2} \right]
+ \frac{\widetilde{\omega}^2}{2}r^2\psi -\frac{i\hbar \ell^2}{2} \frac{\partial \psi}{\partial \phi} = E \psi. \label{WDW}$$
It is well known that the Hamiltonian of the 2d harmonic oscillator commutes with the angular momentum. This means that we have $[H_{xy}, L_z] = 0$, allowing us to write the wave function with eigenstates common to both operators. This breaks the degeneracy in energy eigenstates and every state is uniquely specified by the quantum numbers $n$ and $m$. Solving Eq.(\[WDW\]) we find the wave function of the Universe $\psi_{n,m} (r,\phi)$
$$\psi_{n,m}(r,\phi) = N \left( \frac{\widetilde{\omega}r^2}{\hbar} \right)^{\frac{|m|}{2}} L^{|m|}_n\left( \frac{\widetilde{\omega}r^2}{\hbar} \right)e^{im\phi}e^{-\frac{\tilde{\omega} r^2}{2 \hbar}}. \label{Wave}$$
where $n=0,1,2,...\:; ~m = 0, \pm 1, \pm 2,...,\pm n,$ and $L_n^{\alpha} (z)$ are the generalized Laguerre polynomials. The zero energy condition $E = 0$ from the WDW equation is trivially satisfied by $\ell = \widetilde{\omega} = 0$. For $\ell \neq 0, ~\widetilde{\omega} \neq 0$ (for instance for a single state of quantum numbers $n$ and $m$) we get the general equation
$$\frac{\ell^2}{\widetilde{\omega}} = -\frac{2(2n + |m| + 1)}{m}. \label{Condition2}$$
By using the definitions for $\ell^2$ and $\widetilde{\omega}^2 > 0$ we can write this equation as
$$\frac{16\beta -12\theta \Lambda}{\sqrt{(16-3\theta^2 \Lambda)(4\beta^2-12\Lambda)}} = -\frac{2(2n + |m| + 1)}{m}. \label{Condition3}$$
With respect to the condition in Eq. (\[Condition3\]), we need to find the values of $\theta$, $\beta$ and $\Lambda$ to satisfy the condition. The first thing to notice is that the right hand side of this equation takes only discrete values and to determine if there are possible values of the parameters to satisfy the equation we may proceed as follows. Analyze the behavior of the function $h(\theta,\beta,\Lambda) = \frac{16\beta -12\theta \Lambda}{\sqrt{16-3\theta^2 \Lambda}\sqrt{4\beta^2-12\Lambda}}$ and then see whether or not, there are permitted values of the parameters for which the function takes the value $h(\theta,\beta,\Lambda) = -\frac{2(2n + |m| + 1)}{m}$. The quantum solution is possible when $\widetilde{\omega}^2> 0$, which corresponds to the first solution in Eq. (\[CVolume\]) and Eq.(\[NCVolume\]). Therefore, the space of parameters is already depicted in Fig.(1) and Fig.(2), corresponding to the condition that the functions $f(\beta,\Lambda)$ and $g(\theta,\Lambda)$ have opposite sign. The valid values of parameters $\theta$, $\beta$ and $\Lambda$, corresponds to the gray region of Fig. 1 and white region of Fig. 2, or vice versa. We use the inverse transformation of Eq.(\[Trans1\]) with Eq. (\[Wave\]) $$r^2 = \frac{a^3}{\mu^2}, ~~~~~ \phi = \frac{\pi}{2}-\mu \varphi,$$ to find the wave function $\psi_{n,m} (a,\varphi)$ in terms of the scale factor $a$ and the phantom field $\varphi$.
We can construct a general wave packet
$$\psi(a,\varphi) = \sum_n^N c_{n,m}\psi_{(n,m)} (a,\varphi), \label{WaveP}$$
with $m$ fixed and $c_{n,m}$ arbitrary constants as long as the condition $E = \sum_n^N E_{(n,m)} = 0$ also holds. When this is the case, it is always possible to find values of $\theta$, $\beta$ and $\Lambda$ to satisfy the constriction Eq. (\[Condition2\]), under which these values are fixed unambiguously. For a given value of $m$, the sum in Eq.(\[WaveP\]) runs over odd/even values of $n$ depending on the odd/even value of $m$ respectively. The wave packet is normalizable by the use of the orthogonality relation of generalized Laguerre polynomials. The total energy of the wave packet will satisfy the generalized zero energy condition Eq.(\[Condition2\]), but replacing $n$ by $N$. From the arguments given above, we conclude that it is always possible to find, for a given $\Lambda > 0$ values for $\theta$ and $\beta$ for which the condition holds.
Classical paths from a WKB approximation
----------------------------------------
One can apply a WKB type approximation on the WDW equation $H_{nc} \Psi = 0$, to see if we get the classical solutions. First we notice that the effective classical cosmological constant given in Eq. (\[LambdaEf\]) is the same quantum cosmological constant in the Hamiltonian $H_{nc}$. Remembering that in the commutative Hamiltonian $4\omega^2 = -3\Lambda$, for the deformed Hamiltonian $4\widetilde{\omega}^2 = -3\Lambda_{Eff}$. This is the generalized deformed frequency and using the definition of $\widetilde{\omega}^2$ we get the same effective cosmological constant.
We now apply the standard procedure for the semiclassical approximation. We propose a wave function of the form
$$\psi \propto e^{\frac{i}{\hbar}S_1(x) + \frac{i}{\hbar}S_2(y)},$$
which upon substitution in $H_{nc} \Psi = 0$, in the limit $\hbar \rightarrow 0$ and using the approximation
$$\left( \frac{\partial S_1(x)}{\partial x} \right)^2 > > \frac{\partial^2 S_1(x)}{\partial x^2},~~~ \left( \frac{\partial S_2(y)}{\partial y} \right)^2 > > \frac{\partial^2 S_2(y)}{\partial y^2},$$
we get the Einstein-Hamilton-Jacobi (EHJ) equation
$$\begin{aligned}
&& \left( \frac{\partial S_1(x)}{\partial x} \right)^2 + \left( \frac{\partial S_2(y)}{\partial y} \right)^2 + \widetilde{\omega}^2(x^2 + y^2) \\ \nonumber
& +& \ell^2x\left( \frac{\partial S_2(y)}{\partial y} \right) -\ell^2 y \left( \frac{\partial S_1(x)}{\partial x} \right) = 0.\end{aligned}$$
With the identification $\frac{\partial S_1(x)}{\partial x} = P_x$, $\frac{\partial S_2(y)}{\partial y} = P_y$, the EHJ equation takes the form
$$\dot{x}^2 +\dot{y}^2 +\left( \widetilde{\omega}^2-\frac{\ell^4}{4} \right) \left(x^2+y^2\right) = 0.$$
Now we take the time derivative of the EHJ equation and divide the result by $\dot{x}\dot{y}$ to get $$\frac{\ddot{x}}{\dot{y}}+\left( \widetilde{\omega}^2-\frac{\ell^4}{4}\right)\frac{x}{\dot{y}}+\frac{\ddot{y}}{\dot{x}}+\left( \widetilde{\omega}^2-\frac{\ell^4}{4}\right)\frac{y}{\dot{x}}=0,$$ we can separate the expression by equating the first two terms to $-\ell^2$ and the last two to $\ell^2$. The two resulting equations are the equations of motion in Eq.(\[EqMot\]). We can be confident that the same classical paths arise in the classical limit of the WDW equation obtained from the deformed Hamiltonian.
Conclusions
===========
In this paper we have studied the effects of the minisuperspace phase space deformations in the dynamics of phantom scalar cosmology. The deformation was applied on the minisuperspace variables. From the deformed Hamiltonian we derived the equations of motion. It is known that from the deformation two generally nonequivalent descriptions arise. Using the solutions to Eq.(\[EqMot\]), we calculate the volume of the Universe in the [*“C-frame"*]{} and in the [*“NC-frame"*]{}. [To constrain the deformation parameters $\beta$ and $\theta$, we impose that for phase space deformations to be valid, the physical description in the two frames must be equivalent. From the value of $\widetilde{\omega}^2$ three different cases appear. For $\widetilde{\omega}^2>0$, we get a bouncing Universe and the solutions in the two frames are equivalent. For the last case, when the Hamiltonian constraint is applied, we have a de Sitter Universe and the parameter $\widetilde{\omega}^2$ can be interpreted as an effective positive cosmological constant $\Lambda_{Eff}$. The value of $\Lambda_{Eff}$ depends on the deformation parameters $\theta$ and $\beta$ as well as the original $\Lambda$.]{} If we set the original cosmological constant $\Lambda = 0$, then $\widetilde{\omega}^2>0$ and is not possible to have $\Lambda_{Eff} \geq 0$. This imposes a stronger restriction on the values of $\beta$ and $\theta$ and the parameter space is even more restricted, but even when this is the case we have a region of values where the two descriptions are physically equivalent. We quantized the model, obtaining the corresponding WDW equation and found the wave function of the Universe. We derived a relation that fixes the values of the parameters $\theta$, $\beta$ and $\Lambda$ in order to satisfy the zero energy condition of the WDW equation. We showed that the deformed phantom cosmological model is physically viable. Also, when $\widetilde{\omega}^2 > 0$ the quantization can be achieved and the introduction of the deformation breaks the degeneracy of the quantum states. A general condition on the parameters $\theta$ and $\beta$ arises from the zero energy condition that can be satisfied by the general wave packets. [It is worth to mention that there is a nice physical interpretation in the [*“C-frame"*]{}. It is related to the introduction to an effective perpendicular and constant “magnetic field" $\vec{B}$. This can be seen by taking the bidimensional harmonic oscillator coupled to the vector potential $\vec A=(\frac{B}{2}y,-\frac{B}{2} x)$ where the magnitude of the field will be $B=\ell^2$. Then we can see that the zero energy condition is related to the manipulation of this field and corresponds to a modification in the deformation parameters $(\theta,\beta)$.]{} Finally, we performed the WKB approximation and show that the same classical solutions arise in the classical limit of the WDW equation of the corresponding deformed Hamiltonian.
Acknowledgments {#acknowledgments .unnumbered}
===============
This work is supported by CONACyT research grants 167335, 257919, 258982.
J. L. López is supported by PRODEP postdoctoral grant 17-1947. This work is part of the PRODEP research network “Gravitación y Física Matemática".
M. S. is supported by DAIP1107/2016 and by the CONACyT program “Estancias sabáticas en el extranjero”, grant 31065.
References {#references .unnumbered}
==========
[cc]{}
B. Ratra and P. J. E. Peebles, Phys. Rev. [**D 37**]{} 3406 (1988).
P. J. E. Peebles and B. Ratra, Rev. Mod. Phys. [**75**]{}, 559 (2003).
L. P. Chimento and A. S. Jakubi, Int. J. Mod. Phys. [**D 5**]{} 71 (1996).
E. J. Copeland, M. Sami, and S. Tsujikawa, Int. J. Mod. Phys. [**D 15**]{} 1753 (2006).
K. Bamba, S. Capozziello, S. Nojiri, and S. D. Odintsov, Astrophys. Space. Sci. [**342**]{} 155 (2012).
T. Padmanabhan and T. Roy Choudhury, Phys. Rev. [**D 66**]{} 081301 (2002).
V. A. Rubakov, Theor. Math. Phys. [**149**]{} 1651 (2006).
S. Perlmutter [*et al.*]{} \[Supernova Cosmology Project Collaboration\], Astrophys. J. [**517**]{}, 565 (1999). A. G. Riess et al., Astron. J. [**116**]{} 1009 (1998).
A. G. Riess et al, Astrophys. J. [**560**]{} 49 (2001).
Narciso Benitez et al., Astrophys. J. [**577**]{} L1 (2002).
Mariusz Dabrowski, Phys. Rev. [**D 74**]{} 044022 (2006).
Robert R. Caldwell and Marc Kamionkowski, Ann. Rev. Nucl. Part. Sci. [**59**]{} 397 (2009).
J. M. Cline, S. Jeon, and G. D. Moore, Phys. Rev. [**D 70**]{} 043543 (2004).
W. Hu, Phys. Rev. [**D 71**]{} 047301 (2005).
H. Garcia-Compean, O. Obregon, C. Ramirez, and M. Sabido, Phys. Rev. [**D 68**]{} 045010 (2003).
X. Calmet and A. Kobakhidze, Phys. Rev. [**D 72**]{} 045010 (2005).
P. Aschieri, M.Dimitrijevic, F.Meyer and J.Wess, Class. Quant. Grav. [**23**]{} 1883 (2006).
T. Ohl and A. Schenkel, JHEP [**0910**]{} 052 (2009).
H. Garcia-Compean, O. Obregon, and C. Ramirez, Phys. Rev. Lett. [**88**]{} 161301 (2002).
E. Mena, O. Obregon and M. Sabido, Int. J. Mod. Phys. D [**18**]{} (2009) 95.
L. O. Pimentel and O.Obregon, Gen. Rel. Grav. [**38**]{} 553 (2006).
C. Bastos, O. Bertolami, N. Costa Dias and J. Nuno Prata, Phys. Rev. D [**78**]{} (2008) 023516.
B. Vakili, C. I. Azad, P. Pedram, and S. Jalalzadeh, Phys. Lett. [**B 687**]{} 119 (2010). B. Malekolkalami, K. Atazadeh and B. Vakili, Phys. Lett. B [**739**]{}, 400 (2014).
S. Pérez-Payán, M. Sabido and C. Yee-Romero, Phys. Rev. [**D 88**]{} 027503 (2013). M. Sabido and C. Yee-Romero, Phys. Lett. B [**757**]{} (2016) 57.
A. Crespo-Hernandez, E. A. Mena-Barboza and M. Sabido, Entropy [**19**]{}, 91 (2017).
Sinuhe Pérez-Payán, M. Sabido, E. Mena and C. Yee-Romero, Adv. High. Energy Phys. [**2014**]{} 958137 (2014)
S. S. Gousheh and H. R. Sepangi, Phys. Lett. A [**272**]{}, 304 (2000). S. Jalalzadeh, F. Ahmadi and H. R. Sepangi, JHEP [**0308**]{}, 012 (2003).
| {
"pile_set_name": "ArXiv"
} |
---
abstract: |
Kasraoui, Stanton and Zeng, and Kim, Stanton and Zeng introduced certain $q$-analogues of Laguerre and Charlier polynomials. The moments of these orthogonal polynomials have combinatorial models in terms of crossings in permutations and set partitions. The aim of this article is to prove simple formulas for the moments of the $q$-Laguerre and the $q$-Charlier polynomials, in the style of the Touchard-Riordan formula (which gives the moments of some $q$-Hermite polynomials, and also the distribution of crossings in matchings).
Our method mainly consists in the enumeration of weighted Motzkin paths, which are naturally associated with the moments. Some steps are bijective, in particular we describe a decomposition of paths which generalises a previous construction of Penaud for the case of the Touchard-Riordan formula. There are also some non-bijective steps using basic hypergeometric series, and continued fractions or, alternatively, functional equations.
address:
- 'LRI, CNRS and Université Paris-Sud, Bâtiment 490, 91405 Orsay, France'
- 'Institut für Algebra, Zahlentheorie und Diskrete Mathematik, Leibniz Universität Hannover, Welfengarten 1, 30167 Hannover, Deutschland'
author:
- 'Matthieu Josuat-Vergès'
- Martin Rubey
title: 'Crossings, Motzkin paths and Moments'
---
Introduction {#sec:introduction}
============
Our motivation is to derive in a uniform way generating functions for matchings, set partitions and permutations refined by the number of crossings. We achieve this by enumerating certain weighted Motzkin paths, which in turn prompt us to consider these counts as moments of certain families of orthogonal polynomials. In some cases, formulas for these moments are already known. However, the method of proof we present in this algorithm is quite general, and leads to very simple formulas.
Let us first define the notion of crossings in matchings, set partitions and permutations. To do so, it is best to draw the objects we are interested in in a certain standard way. We begin with the set of matchings (or fixed-point free involutions) ${\ensuremath{\mathcal{M}}}_{2n}$ of $\{1,\dots,2n\}$: these are drawn by putting the numbers from $1$ to $2n$ in this order on a straight line, and then connecting paired numbers by an arc. Of course, arcs are always drawn in a way such that any two arcs cross at most once, and no more than two arcs intersect at any point, see the first picture in Figure \[fig:matching\] for an example. Then, a [*crossing in a matching*]{} is, as one would expect, a pair of matched points $\{i,j\}$ and $\{k,l\}$ with $i<k<j<l$, pictorially:
(28,10)(-30,-3) (-30,0)[(1,0)[28]{}]{} (-28,0)(-28,0)[(0,-5)\[c\][$i$]{}]{} (-20,0)(-20,0)[(0,-5)\[c\][$k$]{}]{} (-12,0)(-12,0)[(0,-5)\[c\][$j$]{}]{} (-4,0)(-4,0)[(0,-5)\[c\][$l$]{}]{} (-28,0)(-20,12)(-12,0) (-20,0)(-12,12)(-4,0)
(10,7.5)(0,-4.5) (0,0.5)[(1,0)[9]{}]{} (0,0)(0,0.5) (1,0)(1,0.5) (2,0)(2,0.5) (3,0)(3,0.5) (4,0)(4,0.5) (5,0)(5,0.5) (6,0)(6,0.5) (7,0)(7,0.5) (8,0)(8,0.5) (9,0)(9,0.5) (0,0.5)(2.0,2.5)(4,0.5) (1,0.5)(2.0,1.5)(3,0.5) (2,0.5)(5.0,3.5)(8,0.5) (5,0.5)(5.5,1.0)(6,0.5) (7,0.5)(8.0,1.5)(9,0.5) (-0.5,-4)[(1,0)[10]{}]{} (-0.5,-4) (-0.5,-4)[(1,1)[1]{}]{} (-0.2,-3.4)[$1$]{} (0.5,-3) (0.5,-3)[(1,1)[1]{}]{} (0.8,-2.4)[$1$]{} (1.5,-2) (1.5,-2)[(1,1)[1]{}]{} (1.8,-1.4)[$1$]{} (2.5,-1) (2.5,-1)[(1,-1)[1]{}]{} (3.0,-1.4)[$q$]{} (3.5,-2) (3.5,-2)[(1,-1)[1]{}]{} (4.0,-2.4)[$q$]{} (4.5,-3) (4.5,-3)[(1,1)[1]{}]{} (4.8,-2.4)[$1$]{} (5.5,-2) (5.5,-2)[(1,-1)[1]{}]{} (6.0,-2.4)[$1$]{} (6.5,-3) (6.5,-3)[(1,1)[1]{}]{} (6.8,-2.4)[$1$]{} (7.5,-2) (7.5,-2)[(1,-1)[1]{}]{} (8.0,-2.4)[$q$]{} (8.5,-3) (8.5,-3)[(1,-1)[1]{}]{} (9.0,-3.4)[$1$]{} (9.5,-4)
Indeed, the motivating example for this article is the Touchard-Riordan formula, which gives, for each $n$, the generating polynomial according to crossings for perfect matchings of the set $\{1,\dots,2n\}$. Denoting by $\operatorname{cro}(M)$ the number of crossings of the matching $M$, we have:
\[th:Touchard\] $$\label{eq:Touchard}
\sum_{M\in{\ensuremath{\mathcal{M}}}_{2n}} q^{\operatorname{cro}(M)}
= \frac{1}{(1-q)^n}
\sum_{k\geq0}(-1)^k\left(\binom{2n}{n-k}-\binom{2n}{n-k-1}\right)
q^{\binom{k+1}{2}}.$$
This has been proved in the 1950’s by Touchard, although, curiously, it seems that the formula was not given explicitly. This was later rectified by Riordan.
Quite similar to matchings, a set partition can be depicted by connecting the numbers on the line which are in one block $B=\{b_1 <
b_2 < \dots < b_l\}$ by arcs $(b_1,b_2)$, $(b_2,b_3)$,…, $(b_{l-1},b_l)$, see Figure \[fig:setpartition\] for an example. Again, a [*crossing in a set partition*]{} is what one would expect: a pair of arcs $\{i,j\}$ and $\{k,l\}$ with $i<k<j<l$. Denoting the set of set partitions of $\{1,\dots, n\}$ by $\Pi_n$, the number of crossings in a set partition $\pi$ by $\operatorname{cro}(\pi)$ and the number of its blocks by $|\pi|$, we will obtain the following $q$-analogue of the Stirling numbers of the second kind:
\[th:Charlier\] $$\begin{aligned}
\label{eq:Charlier}
\sum_{ \substack{{ \pi\in\Pi_n }\\{|\pi|=k}} } q^{\operatorname{cro}(\pi)}
= \frac1{(1-q)^{n-k}} \sum\limits_{j=0}^{k}
\sum\limits_{i=j}^{n-k} (-1)^i \left(\tbinom n{k+i}\tbinom
n{k-j}-\tbinom n{k+i+1}\tbinom n{k-j-1}\right) {\genfrac{[}{]}{0pt}{}{i}{j}_q}
q^{\tbinom{j+1}2},
\end{aligned}$$ where ${\genfrac{[}{]}{0pt}{}{n}{k}_q} = \prod_{i=1}^k\frac{[n-k+i]_q}{[i]_q}$ is the $q$-binomial coefficient, and $[n]_q=1+q+\dots+q^{n-1}$.
(7,6.5)(0,-4.5) (0,0.5)[(1,0)[7]{}]{} (0,0)(0,0.5) (1,0)(1,0.5) (2,0)(2,0.5) (3,0)(3,0.5) (4,0)(4,0.5) (5,0)(5,0.5) (6,0)(6,0.5) (7,0)(7,0.5) (0,0.5)(2.0,2.5)(4,0.5) (1,0.5)(3.0,2.5)(5,0.5) (2,0.5)(2.5,1.0)(3,0.5) (4,0.5)(5.5,2.5)(7,0.5) (-0.5,-4)[(1,0)[8]{}]{} (-0.5,-4) (-0.5,-4)[(1,1)[1]{}]{} (-0.2,-3.4)[$y$]{} (0.5,-3) (0.5,-3)[(1,1)[1]{}]{} (0.8,-2.4)[$y$]{} (1.5,-2) (1.5,-2)[(1,1)[1]{}]{} (1.8,-1.4)[$y$]{} (2.5,-1) (2.5,-1)[(1,-1)[1]{}]{} (3.0,-1.4)[$1$]{} (3.5,-2) (3.5,-2)[(1,0)[1]{}]{} (3.8,-1.8)[$q$]{} (4.5,-2) (4.5,-2)[(1,-1)[1]{}]{} (5.0,-2.4)[$q$]{} (5.5,-3) (5.5,-3)[(1,0)[1]{}]{} (5.8,-2.8)[$y$]{} (6.5,-3) (6.5,-3)[(1,-1)[1]{}]{} (7.0,-3.4)[$1$]{} (7.5,-4)
There is an alternative notion of crossings for set partitions, due to Ehrenborg and Readdy [@ER96], coming from juggling patterns. Namely, we introduce an additional infinite arc extending to the right from every maximal element of each block, including singletons, see Figure \[fig:setpartitionER\] for an example. Denoting the number of crossings in such a drawing of a set partition $\pi$ by $\operatorname{cro}^*(\pi)$, we have:
\[th:Charlier\*\] $$\label{eq:Charlier*}
\sum_{\substack{{ \pi\in\Pi_n }\\{|\pi|=k}}} q^{\operatorname{cro}^*(\pi)}
=\frac{1}{(1-q)^{n-k}} \sum\limits_{j=0}^{n-k}
(-1)^j \tbinom n{k+j}{\genfrac{[}{]}{0pt}{}{k+j}{j}_q}.$$
This is not a new result: essentially, this formula was already known to Gould from another definition (the link with crossings is more recent as will appear below).
(7,7.5)(0,-4.5) (0,0.5)[(1,0)[7]{}]{} (0,0)(0,0.5) (1,0)(1,0.5) (2,0)(2,0.5) (3,0)(3,0.5) (4,0)(4,0.5) (5,0)(5,0.5) (6,0)(6,0.5) (7,0)(7,0.5) (0,0.5)(2.0,2.5)(4,0.5) (1,0.5)(3.0,2.5)(5,0.5) (2,0.5)(2.5,1.0)(3,0.5) (4,0.5)(5.5,2.5)(7,0.5) (3,0.5)(5.5,2.5)(8,2.5) (5,0.5)(6.5,2.0)(8,2.0) (6,0.5)(7.0,1.5)(8,1.5) (7,0.5)(7.5,1.0)(8,1.0) (-0.5,-4)[(1,0)[8]{}]{} (-0.5,-4) (-0.5,-4)[(1,1)[1]{}]{} (-0.2,-3.4)[$y$]{} (0.5,-3) (0.5,-3)[(1,1)[1]{}]{} (0.8,-2.4)[$y$]{} (1.5,-2) (1.5,-2)[(1,1)[1]{}]{} (1.8,-1.4)[$y$]{} (2.5,-1) (2.5,-1)[(1,-1)[1]{}]{} (3.0,-1.4)[$q^2$]{} (3.5,-2) (3.5,-2)[(1,0)[1]{}]{} (3.8,-1.8)[$q$]{} (4.5,-2) (4.5,-2)[(1,-1)[1]{}]{} (5.0,-2.4)[$q^2$]{} (5.5,-3) (5.5,-3)[(1,0)[1]{}]{} (5.8,-2.8)[$yq$]{} (6.5,-3) (6.5,-3)[(1,-1)[1]{}]{} (7.0,-3.4)[$1$]{} (7.5,-4)
Finally, to depict a permutation $\sigma$, we connect the number $i$ with $\sigma(i)$ with an arc *above* the line, if $i\leq\sigma(i)$, otherwise with an arc *below* the line, as done in Figure \[fig:permutation\]. The notion of crossing in a permutation was introduced by Corteel [@Cor07], and is slightly less straightforward: a pair of numbers $(i,k)$ constitutes a [*crossing in a permutation*]{}, if $i<k\leq\sigma(i)<\sigma(k)$ or $\sigma(i)<\sigma(k)<i<k$:
(29,14)(-30,-7) (-30,0)[(1,0)[28]{}]{} (-28,0)(-28,0)[(0,-5)\[c\][$i$]{}]{} (-20,0)(-20,0)[(0,-5)\[c\][$k$]{}]{} (-12,0)(-12,0)[(0,-5)\[c\][$\pi(i)$]{}]{} (-4,0)(-4,0)[(0,-5)\[c\][$\pi(k)$]{}]{} (-28,0)(-20,12)(-12,0) (-20,0)(-12,12)(-4,0)
(34,14)(-35,-7) (-35,-1)[or]{} (-30,0)[(1,0)[28]{}]{} (-28,0)(-28,0)[(0,-5)\[c\][$i$]{}]{} (-16,0)(-16,0)[(0,-5)\[c\][$k=\sigma(i)$]{}]{} (-4,0)(-4,0)[(0,-5)\[c\][$\sigma(k)$]{}]{} (-28,0)(-22,12)(-16,0) (-16,0)(-10,12)(-4,0)
(34,14)(-35,-7) (-35,-1)[or]{} (-30,0)[(1,0)[28]{}]{} (-28,0)(-28,5)[(0,-5)\[c\][$i$]{}]{} (-20,0)(-20,5)[(0,-5)\[c\][$k$]{}]{} (-12,0)(-12,5)[(0,-5)\[c\][$\sigma(i)$]{}]{} (-4,0)(-4,5)[(0,-5)\[c\][$\sigma(k)$]{}]{} (-28,0)(-20,-12)(-12,0) (-20,0)(-12,-12)(-4,0) (0,-1)[.]{}
Denoting the set of permutations of $\{1,\dots, n\}$ by $\mathfrak{S}_n$, and the number of weak exceedances, [*i.e.*]{} numbers $i$ with $\sigma(i)\geq i$, of a permutation $\sigma$ by $\operatorname{wex}(\sigma)$, we have:
\[th:Laguerre\] $$\label{eq:Laguerre}
\begin{split}
&\sum_{\sigma\in\mathfrak{S}_n}
y^{\operatorname{wex}(\sigma)} q^{\operatorname{cro}(\sigma)} \\
&= \frac 1{(1-q)^n} \sum\limits_{k=0}^n (-1)^k
\left(\sum\limits_{j=0}^{n-k} y^j
\Big( \tbinom{n}{j}\tbinom{n}{j+k} -
\tbinom{n}{j-1}\tbinom{n}{j+k+1}\Big) \right)
\left(\sum\limits_{i=0}^k y^iq^{i(k+1-i)} \right).
\end{split}$$
This theorem recently found a rather different proof by the first author [@MJV]. In the present article we provide an alternative, using a bijective decomposition of weighted Motzkin paths that gives a natural interpretation for the two inner sums.
The rest of this article is organised as follows. In Section \[sec:orth-poly\], we present some background material concerning the combinatorial theory of orthogonal polynomials. In Section \[sec:penaud-decomp\], we describe the decomposition of weighted Motzkin paths mentioned above, in full generality. Each Motzkin path will be decomposed into a Motzkin prefix and another Motzkin path satisfying certain additional conditions. In Section \[sec:prefixes\], we enumerate Motzkin prefixes, and in Section \[sec:paths2\] the other set of paths appearing in the decomposition are enumerated.
There are three appendices. In the first appendix we give an alternative point of view of the decomposition presented in Section \[sec:penaud-decomp\], using inverse relations. In the second appendix, we give a bijective proof of the formula for the generating function of the paths appearing in the decomposition in the case of set-partitions, using a sign-reversing involution. It is thus possible to give a fully bijective proof of Theorem \[th:Charlier\], analogous to Penaud’s proof of the Touchard-Riordan formula. Finally, in the last appendix we sketch a proof showing that one cannot expect closed forms for Motzkin prefixes with weights different from those considered in Section \[sec:prefixes\].
(7,6)(0,-4.5) (0,0.5)[(1,0)[7]{}]{} (0,0)(0,0.5) (1,0)(1,0.5) (2,0)(2,0.5) (3,0)(3,0.5) (4,0)(4,0.5) (5,0)(5,0.5) (6,0)(6,0.5) (7,0)(7,0.5) (0,0.5)(1.0,1.5)(2,0.5) (1,0.5)(2.0,1.5)(3,0.5) (2,0.5)(4.0,2.5)(6,0.5) (4,0.75) (6,0.5)(6.5,1.0)(7,0.5) (7,0.5)(6.0,-1.5)(5,0.5) (5,0.5)(3.0,-2.0)(1,0.5) (3,0.5)(1.5,-1.5)(0,0.5) (-0.5,-4)[(1,0)[8]{}]{} (-0.5,-4) (-0.5,-4)[(1,1)[1]{}]{} (-0.2,-3.4)[$y$]{} (0.5,-3) (0.5,-3)[(1,1)[1]{}]{} (0.6,-2.4)[$yq$]{} (1.5,-2) (1.5,-2)[(1,0)[1]{}]{} (1.8,-1.8)[$yq^2$]{} (2.5,-2) (2.5,-2)[(1,-1)[1]{}]{} (3.0,-2.4)[$q$]{} (3.5,-3) (3.5,-3)[(1,0)[1]{}]{} (3.8,-2.8)[$y$]{} (4.5,-3) (4.5,-3)[(1,0)[1]{}]{} (4.8,-2.8)[$1$]{} (5.5,-3) (5.5,-3)[(1,0)[1]{}]{} (5.8,-2.8)[$yq$]{} (6.5,-3) (6.5,-3)[(1,-1)[1]{}]{} (7.0,-3.4)[$1$]{} (7.5,-4)
Orthogonal Polynomials, moments and histoires {#sec:orth-poly}
=============================================
Motzkin paths are at the heart of the *combinatorial theory of orthogonal polynomials*, as developed by Flajolet [@Fla82] and Viennot [@Vie84]. This theory tells us, that the moments of any family of orthogonal polynomials are given by a certain weighted count of Motzkin paths.
More precisely, by Favard’s theorem, any monic sequence of orthogonal polynomials $(P_n)_{n\geq0}$ satisfies a three term recurrence of the form $$xP_n(x) = P_{n+1}(x)+ b_n P_n(x) + \lambda_n P_{n-1}(x),$$ where $b_n$ and $\lambda_n$ do not depend on $x$. Given this recurrence, the $n$^th^ moment $\mu^P_n$ of $P$ can be expressed as the weighted sum of Motzkin paths of length $n$, that is, paths taking up ($\nearrow$), down ($\searrow$) and level ($\rightarrow$) steps, starting and ending at height $0$, and not going below this height, where a horizontal step at height $h$ has weight $b_h$ and a down step starting at height $h$ has weight $\lambda_h$.
Histoires {#sec:histoires}
---------
Three basic examples of families of orthogonal polynomials are given by (rescalings of) the Hermite, Charlier and Laguerre polynomials, where the moments count matchings ($b_n=0$, $\lambda_n=n$), set partitions ($b_n=1+n$, $\lambda_n=n$) and permutations ($b_n=2n+1$, $\lambda_n=n^2$) respectively. It turns out that the Hermite, Charlier and Laguerre polynomials indeed have beautiful $q$-analogues such that the moments count the corresponding objects, and $q$ marks the number of crossings. We want to establish this correspondence via “histoires”:
Consider a family of orthogonal polynomials with coefficients $b_n$ and $\lambda_n$, and fix $a_n$ and $c_n$ such that $\lambda_n=a_{n-1} c_n$ for all $n$. Suppose that for every fixed $n$, the coefficients $a_n$, $b_n$ and $c_n$ are polynomials such that each monomial has coefficient 1 (as will appear shortly, this is general enough in our context).
We then call a weighted Motzkin path [*histoire*]{}, when the weight of an up step $\nearrow$ (respectively a level step $\rightarrow$ or a a down step $\searrow$) starting at level $h$ is one of the monomials appearing in $a_h$ (respectively $b_h$ or $c_h$).
We want to consider four different families of “histoires”, corresponding to $q$-analogues of the Hermite, Charlier and Laguerre polynomials.
There are weight-preserving bijections between
- matchings $M$ with weight $q^{\operatorname{cro}(M)}$, and “histoires de Hermite” defined by $b_n=0$, $a_n=1$ and $c_n=[n]_q$,
- set partitions $\pi$ with weight $y^{|\pi|}q^{\operatorname{cro}(\pi)}$, and “histoires de Charlier” defined by $b_n=y+[n]_q$, $a_n=y$ and $c_n=[n]_q$,
- set partitions $\pi$ with weight $y^{|\pi|}q^{\operatorname{cro}^*(\pi)}$, and “histoires de Charlier-$*$” defined by $b_n=yq^n+[n]_q$, $a_n=yq^n$ and $c_n=[n]_q$, and
- permutations $\sigma$ with weight $y^{\operatorname{wex}(\sigma)}q^{\operatorname{cro}(\sigma)}$, and “histoires de Laguerre” defined by $b_n=yq^n+[n]_q$, $a_n=yq^n$ and $c_n=[n]_q$.
These bijections are straightforward modifications of classical bijections used by Viennot [@Vie84]. We detail them here for convenience, but also because of their beauty… Examples can be found in Figures \[fig:matching\]–\[fig:permutation\].
The bijection connecting matchings and “histoires de Hermite”, such that crossings are recorded in the exponent of $q$, goes as follows: we traverse the matching, depicted in the standard way, from left to right, while we build up the Motzkin path step by step, also from left to right. For every arc connecting $i$ and $j$ with $i<j$, we call $i$ an [*opener*]{} and $j$ a [*closer*]{}. When we have traversed the matching up to and including number $\ell$, we call the openers $i\leq\ell$ with corresponding closers $j<\ell$ [*active*]{}. Openers are translated into up steps with weight $1$. Accordingly, when we encounter a closers $\ell$ it becomes a down step with weight $q^k$, where $k$ is the number of active openers between $\ell$ and the opener corresponding to $\ell$. It is a enjoyable exercise to see that this is indeed a bijection, and that a matching with $k$ crossings corresponds to a Dyck path of weight $q^k$.
The bijection between set partitions and “histoires de Charlier”, due to Anisse Kasraoui and Jiang Zeng [@KaZe06], is very similar: in addition to openers and closers, which are the non-maximal and non-minimal elements of the blocks of the set partition, we now also have [*singletons*]{}, which are neither openers nor closers. Elements that are openers and closers at the same time are called [*transients*]{}. Non-transient openers are translated into up steps with weight $y$, and singletons are translated into level steps with weight $y$. Non-transient closers $\ell$ are translated into down steps both with weight $q^k$, where $k$ is the number of active openers between $\ell$ and the opener corresponding to $\ell$. Finally, transient closers $\ell$ become level-steps with weight $q^k$, with $k$ as before.
To obtain a “histoire de Charlier-$*$” of a set partition, using the modified definition of crossings, we only have to multiply the weights of steps corresponding to closers and singletons by $q^k$, where $k$ is the number of crossings of the infinite arc with other arcs.
It remains to describe the bijection between permutations and “histoires de Laguerre”, due to Dominique Foata and Doron Zeilberger, which is usually done in a different way than in what follows, however. To obtain the Motzkin path itself, we ignore all the arcs below the line and also the loops corresponding to fixed points. What remains can be interpreted as a set partition, and thus determines a Motzkin path. Moreover, the weights of the down steps are computed as in the case of set partitions, except that the weight of each of those steps needs to be multiplied by $y$. The weights of the level steps that correspond to transients of the set partition are also computed as before, but are then multiplied by $yq$. Level steps that correspond to fixed points of the permutation get weight $y$. The weights of the remaining steps are computed by deleting all arcs above the line, and again interpreting what remains as a set partition. However, this set partition has to be traversed from right to left, and weights are accordingly put onto the up steps of the Motzkin path. Later, it will be more convenient to move the factor $y$ that appears in the weight of all the down steps onto the weight of the corresponding up steps, see Figure \[fig:permutation\] for an example.
Particular classes of orthogonal polynomials {#sec:ortho}
--------------------------------------------
In this section we relate the families of orthogonal polynomials introduced via their parameters $b_n$ and $\lambda_n$ in Section \[sec:histoires\] to classical families. We follow the Askey-Wilson scheme [@KoSw98] for their definition.
The [*continuous $q$-Hermite polynomials*]{} $H_n=H_n(x|q)$ can be defined [@KoSw98 Section 3.26] by the recurrence relation $$2 x H_n = H_{n+1} + (1-q^n) H_{n-1},$$ with $H_0=1$.
Define rescaled continuous $q$-Hermite polynomials $\tilde
H_n=\tilde H_n(x|q)$ as $$\tilde H_n(x|q)=(1-q)^{-n/2}H_n(x\tfrac{\sqrt{1-q}}{2}|q).$$ They satisfy the recurrence relation $$x \tilde H_n = \tilde H_{n+1} + [n]_q \tilde H_{n-1},$$ and their even moments are given by $$\label{eq:muTouchard}
\mu^{\tilde H}_{2n} = \sum_{M\in{\ensuremath{\mathcal{M}}}_{2n}} q^{\operatorname{cro}(M)}.$$ The odd moments are all zero.
The [*Al-Salam-Chihara*]{} polynomials $Q_n=Q_n(x;a,b|q)$ can be defined [@KoSw98 Section 3.8] by the recurrence relation $$2 x Q_n = Q_{n+1} + (a + b)q^n Q_n + (1-q^n)(1-abq^{n-1})Q_{n-1},$$ with $Q_0=1$. We consider two different specialisations of these polynomials. The first was introduced by Kim, Stanton and Zeng [@KSZ06], and in their Proposition 5 they also gave a formula for the moments. However, the formula that follows from our Theorem \[th:Charlier\] appears to be much simpler.
Define $q$-Charlier polynomials $\tilde C_n=\tilde C_n(x;y|q)$ as $$\tilde C_n(x;y|q)=\left(\tfrac{y}{1-q}\right)^{n/2}
Q_n\left(\sqrt{\tfrac{1-q}{4y}}\left(x-y-\tfrac{1}{1-q}\right);
\tfrac{-1}{\sqrt{y(1-q)}}, 0 \;\big\vert\; q\right).$$ They satisfy the recurrence relation $$x \tilde C_n = \tilde C_{n+1} + ( y + [n]_q) \tilde C_n + y [n]_q
\tilde C_{n-1}$$ and their moments are given by $$\mu^{\tilde C}_{n} = \sum_{\pi\in\Pi_n} y^{|\pi|}q^{\operatorname{cro}(\pi)}.$$
The other specialisation was introduced by Kasraoui, Stanton and Zeng [@KSZ08], however, without providing a formula for the moments (these are actually a particular case of octabasic $q$-Laguerre polynomials from [@SS]).
Define $q$-Laguerre polynomials $\tilde L_n=\tilde L_n(x;y|q)$ as $$\tilde L_n(x; y|q) = \left(\tfrac{\sqrt{y}}{q-1}\right)^n
Q_n\left(\tfrac{(q-1)x+y+1}{2\sqrt{y}}; \tfrac{1}{\sqrt{y}}, \sqrt{y} q \;\big\vert\; q\right).$$ They satisfy the recurrence relation: $$x \tilde L_n = \tilde L_{n+1} + ([n]_q + y[n+1]_q) \tilde L_n + y
[n]_q^2 \tilde L_{n-1}.$$ and their moments are given by $$\begin{aligned}
\label{eq:muLaguerre}
\mu^{\tilde L}_{n} = \sum_{\sigma\in\mathfrak{S}_n}
y^{\operatorname{wex}(\sigma)} q^{\operatorname{cro}(\sigma)}.
\end{aligned}$$
The [*Al-Salam-Carlitz I*]{} polynomials $U^{(a)}_n(x|q)$ can be defined [@KoSw98 Section 3.24] by the recurrence relation $$xU_{n}^{(a)}(x|q) = U_{n+1}^{(a)}(x|q) + (a+1)q^nU_{n}^{(a)}(x|q)
- q^{n-1} a(1-q^n)U_{n-1}^{(a)}(x|q),$$ with $U_{0}^{(a)}(x|q)=1$.
Define modified $q$-Charlier polynomials $\tilde C_n^*=\tilde
C_n^*(x;y|q) $ as: $$C^*_n(x;y|q) = y^n U_n^{\left(\frac{-1}{y(1-q)}\right)}
\left( \tfrac xy - \tfrac 1{y(1-q)} | q \right).$$ They satisfy the recurrence relation $$x\tilde C^*_n = \tilde C^*_{n+1}
+ ( yq^n + \left[n\right]_q ) \tilde C^*_n
+ y\left[n\right]_q q^{n-1} \tilde C^*_{n-1},$$ and their moments are given by $$\begin{aligned}
\label{eq:muCharlier*}
\mu^{\tilde C^*}_{n} = \sum_{\pi\in\Pi_n }y^{|\pi|}
q^{\operatorname{cro}^*(\pi)}.
\end{aligned}$$
The result from [@MSW] was actually stated with another statistic, but both correspond to Carlitz’ $q$-analogue of the Stirling numbers of the second kind $S[n,k]$, which are such that $S[n,k]=S[n-1,k-1]+[k]_qS[n-1,k]$, and $$\mu^{\tilde C^*}_{n} = \sum_{k=1}^{n} S[n,k]y^k.$$
Penaud’s decomposition {#sec:penaud-decomp}
======================
Let us first briefly recall Penaud’s strategy to prove the Touchard-Riordan formula for the moments of the rescaled continuous $q$-Hermite polynomials $\tilde H_n$. As already indicated in the introduction, his starting point was their combinatorial interpretation in terms of weighted Dyck paths, down steps starting at level $h\geq1$ having weight $[h]_q$, up steps having weight $1$.
As the total number of down steps in these paths is $n$, we may take out a factor $(1-q)^{-n}$, and instead consider paths with down steps having weight $1-q^h$, or, equivalently, consider paths with down steps having weight $1$ *or* $-q^h$.
The next step is to (bijectively) decompose each path into two objects: the first is a left factor of an unweighted Dyck path of length $n$ and final height $n-2k\geq0$, for some $k$. The second object, in some sense the remainder, is a weighted Dyck path of length $k$ with the same possibilities for the weights as in the original path, except that peaks (consisting of an up step immediately followed by a down step) of weight $1$ are not allowed. This decomposition will be generalised in Lemma \[lem:decomposition\] below.
The left factors are straightforward to count, the result being the ballot numbers $\binom{2n}{n-k}-\binom{2n}{n-k-1}$. For the remainders, Penaud presented a bijective proof that the sum of their weights is given by $(-1)^k q^{\binom{k+1}{2}}$. Summing over all $k$ we obtain the Touchard-Riordan formula .
The general setting {#sec:general-setting}
-------------------
\[defsMP\] Let $\mathcal{M}_n(a,b,c,d;q)$ be the set of weighted Motzkin paths of length $n$, such that the weight of
- an up step $\nearrow$ starting at level $h$ is either $1$ or $-q^{h+1}$,
- of a level step $\rightarrow$ starting at level $h$ is either $d$ or $(a+b)q^h$,
- a down step $\searrow$ starting at level $h$ is either $c$ or $-abq^{h-1}$.
Furthermore, let $\mathcal{M}^*_n(a,b,c;q)\subset\mathcal{M}_n(a,b,c,d;q)$ be the subset of paths that do not contain any
- level step $\rightarrow$ of weight $d$,
- peak $\nearrow\searrow$ such that the up step has weight $1$ and the down step has weight $c$.
Finally, let $\mathcal{P}_{n,k}(c,d)$ be the set of left factors of Motzkin paths of length $n$ and final height $k$, such that the weight of
- an up step $\nearrow$ is $1$,
- a level step $\rightarrow$ is $d$,
- a down step $\searrow$ is $c$.
With these definitions, the decomposition used by Penaud can be generalised in a natural way as follows:
\[lem:decomposition\] There is a bijection $\Delta$ between $\mathcal{M}_n(a,b,c,d;q)$ and the disjoint union of the sets $\mathcal{P}_{n,k}(c,d)\times\mathcal{M}^*_k(a,b,c;q)$ for $k\in\{0,\dots,n\}$.
Let $H$ be a path in $\mathcal{M}_n(a,b,c,d;q)$. Consider the maximal factors $f_1,\dots,f_j$ of $H$ that are Motzkin paths and have up steps of weight $1$, level steps of weight $d$ and down steps of weight $c$. We can thus factorise $H$ as $h_0f_1h_1f_2\dots f_jh_j$.
Since this factorisation is uniquely determined, we can define $\Delta(H)=(H_1,H_2)$ as follows: $$H_1 = (\nearrow)^{|h_0|} f_1 (\nearrow)^{|h_1|} f_2 \dots f_j (\nearrow)^{|h_j|}
\qquad\hbox{ and } \qquad
H_2 = h_0\dots h_j.$$
Thus, $H_1$ is obtained from $H$ by replacing each step in the $h_i$ by an up step $\nearrow$, and $H_2$ is obtained from $H$ by deleting the factors $f_i$. Since the $f_i$ are Motzkin paths, the weight of $H$ is just the product of the weights of $H_1$ and $H_2$. Furthermore, it is clear that $H_1$ is a path in $\mathcal{P}_{n,k}(c,d)$ with final height $k=|h_0|+|h_1|+\dots+|h_j|$. We observe that the $h_i$ cannot contain a level step $\rightarrow$ of weight $d$ or a peak $\nearrow\searrow$ such that the up step has weight $1$ and the down step has weight $c$, because then the factorisation of $H$ would not have been complete. Thus $H_2$ is a path in $\mathcal{M}^*_k(a,b,c;q)$.
It remains to verify that $\Delta$ is indeed a bijection. To do so, we describe the inverse map: let $(H_1,H_2)\in\mathcal{P}_{n,k}(c,d)\times \mathcal{M}^*_k(a,b,c;q)$ for some $k\in\{0,\dots,n\}$. Thus, there exists a unique factorisation $$H_1=(\nearrow)^{u_0} f_1 (\nearrow)^{u_1} f_2 \dots f_j
(\nearrow)^{u_j}$$ such that the $f_i$ are Motzkin paths and $k=\sum_{\ell=0}^j
u_\ell$. Write $H_2$ as $h_0\dots h_j$, where the factor $h_\ell$ has length $u_\ell$. Then $\Delta^{-1}(H_1,H_2) =
h_0f_1h_1f_2\dots f_jh_j$ is the preimage of $(H_1,H_2)$.
Specialising to matchings, set partitions and permutations {#sec:specialising}
----------------------------------------------------------
As remarked in the introduction of this section, we begin by multiplying the weighted sum of all Motzkin paths by an appropriate power of $1-q$. In the case of *matchings* of $\{1,\dots,2n\}$, we are in fact considering Dyck paths of length $2n$ where a down step starting at height $h$ has weight $[h]_q$. Multiplying the weighted sum with $(1-q)^n$, or, equivalently, multiplying the weight of each down step by $1-q$, we thus obtain Dyck paths having down steps starting at height $h$ weighted by $1-q^h$, which fits well into the model introduced in Definition \[defsMP\]: namely, the set $\mathcal{M}_n(a,b,c,d;q)$ with $a=0$, $b=0$, $c=1$ and $d=0$ consists precisely of these paths – except that they are all reversed.
In the case of *set partitions* of $\{1,\dots,n\}$, multiplying the weighted sum by $(1-q)^n$ and reversing all paths we see that we need to enumerate the set $\mathcal{M}_n(a,b,c,d;q)$ with $a=0$, $b=-1$, $c=y(1-q)$ and $d=1+y(1-q)$. When using the *modified* definition of crossings in *set partitions*, we obtain surprisingly different parameters, namely $a=-1$, $b=y(1-q)$, $c=0$ and $d=1$. Finally, the case of *permutations* of $\{1,\dots,n\}$ is covered by enumerating the set $\mathcal{M}_n(a,b,1,d;q)$ with $a=-1$, $b=-yq$, $c=y$ and $d=1+y$.
Counting $\mathcal{P}_{n,k}(c,d)$ {#sec:prefixes}
=================================
In general, formulas for the cardinality of $\mathcal{P}_{n,k}(c,d)$ can be found easily using Lagrange inversion [@Sta]. Consider the generating function $P_k = \sum_n \size{\mathcal{P}_{n,k}(c,d)} t^n$, we want to determine the coefficient of $t^{n+1}$ in $tP_k=(t
P_0)^{k+1}$. Observing the relationship $$t P_0 = t\left( 1 + d (t P_0) + c (t P_0)^2\right)$$ we find that $[t^n](t P_0)^k = \frac{k}{n}[z^{n-k}](1 + d z + c z^2)^n$, and thus
$$\label{eq:trinomial}
\size{\mathcal{P}_{n,k}(c,d)}=
\frac{k+1}{n+1}\sum_{l=0}^{n-k}\binom{n+1}{l}\binom{l}{2l-n+k}d^{2l-n+k}c^{n-k-l}.$$
To count matchings, set partitions or permutations according to crossings (modified or not), the only sets of parameters that we need to consider are $(c,d)=(1,0)$, $(c,d)=(1,2)$ and $(c,d)=(0,1)$. Curiously, these are precisely the values for which Equation allows a closed form, [*i.e.*]{} can be written as a linear combination of hypergeometric terms. A (sketch of a) justification of this fact is given in Appendix \[sec:trinomial\].
Matchings
---------
For matchings, we have $(c,d)=(1,0)$ and we obtain the ballot numbers:
\[dyck\_pre\] The cardinality of $\mathcal{P}_{n,n-2k}(1,0)$, [*i.e.*]{} the number of left factors of Dyck paths of length $n$ and final height $n-2k\geq
0$ is $$\binom nk - \binom n{k-1}.$$
Set partitions and permutations
-------------------------------
For set partitions and permutations, we have $(c,d)=(y,1+y)$ and obtain the following:
\[motz\_pre\] The generating function for $\mathcal{P}_{n,k}(y,1+y)$ is: $$\label{motz_pre_eq}
\sum\limits_{j=0}^{n-k}
\left( \binom{n}{j}\binom{n}{j+k} -
\binom{n}{j-1}\binom{n}{j+k+1}\right) y^j.$$
The elements of $\mathcal{P}_{n,k}(y,1+y)$ have weight $1+y$ on each level step. However, it is again more convenient to pretend that there are two different kinds of level steps, with weight $1$ and $y$ respectively. Let $P$ be a left factor of a Motzkin path with weight $y^j$. We then use the following step by step translation to transform it into a pair $(C_1,C_2)$ of non-intersecting paths taking north and east steps, starting at $(0,1)$ and $(1,0)$ respectively (see Figure \[paths\_r\] for an example):
$i$^th^ step of $P$ $i$^th^ step of $C_1$ $i$^th^ step of $C_2$
--------------------------- ----------------------- -----------------------
$\nearrow$ $\uparrow$ $\rightarrow$
$\rightarrow$, weight $1$ $\uparrow$ $\uparrow$
$\rightarrow$, weight $y$ $\rightarrow$ $\rightarrow$
$\searrow$, weight $y$ $\rightarrow$ $\uparrow$
(0,-1)(8,4) (0,0)(8,4) (0,0)(4,0)(5,1)(6,1)(7,1)(8,2) (0.5,0.5)[$1$]{} (1.5,0.5)[$y$]{} (2.5,0.5)[$y$]{} (3.5,0.5)[$1$]{} (4.5,1.5)[$1$]{} (5.5,1.5)[$y$]{} (6.5,1.5)[$1$]{} (7.5,2.5)[$1$]{}
(-2,-1)(9,7) (0,0)(7,0) (0,0)(0,7) (0,0)(7,0) (0,0)(0,7) (0,1)(7,1) (1,0)(1,7) (0,2)(7,2) (2,0)(2,7) (0,3)(7,3) (3,0)(3,7) (0,4)(7,4) (4,0)(4,7) (0,5)(7,5) (5,0)(5,7) (0,6)(7,6) (6,0)(6,7) (1,0)(0,1)(3,6)(6,3) (1,0)(1,1)(3,1)(3,2)(5,2)(5,3)(6,3) (0,1)(0,2)(2,2)(2,4)(3,4)(3,6) (-1,-1)[0]{}(1,-1)[1]{}(3,-1)[$j$]{}(8,-1)[$j+k+1$]{} (-1,1)[1]{}(-3,3)[$n-k-j$]{}(-3,6)[$n-j+1$]{}
The condition that a Motzkin path does not go below the $x$-axis translates into the fact that $C_1$ and $C_2$ are non-intersecting. Since $P$ has $j$ steps weighted by $y$, the path $C_1$ ends at $(j,n-j+1)$. Since $P$ ends at height $k$, the number of up steps $\nearrow$ and the number of level steps $\rightarrow$ with weight $y$ add up to $j+k$, so $C_2$ ends at $(j+k+1, n-j-k)$.
By the Lindström-Gessel-Viennot Lemma [@GeVi85], these pairs of non-intersecting paths can be counted by a $2\times 2$-determinant, which gives precisely Formula .
For $y=1$ the sum in Equation can be simplified using Vandermonde’s identity. Thus, the number of left factors of Motzkin paths of length $n$ and final height $k$, with weight $2$ on every level step is $$\binom{2n}{n-k} - \binom{2n}{n-k-2}.$$
For $(c,d)=(0,1)$, that is, $y=0$, we obtain what we need to count modified crossings in set partitions, namely the binomial coefficient $\binom{n}{k}$.
counting $\mathcal{M}^*_k(a,b,c;q)$ {#sec:paths2}
===================================
In this section we use a continued fraction to find the generating function for the Motzkin paths in $\mathcal{M}^*_k(a,b,c;q)$ (these paths are described in Definition \[defsMP\]). It turns out that this continued fraction can be expressed as a basic hypergeometric series, which allows us to compute the coefficients corresponding to paths with given length. Let $K(a,b,c;q)$ be $$\cfrac{1}
{1 + c - (a+b) - \cfrac{(c-ab)(1-q)}
{1 + c - (a+b)q - \cfrac{(c-abq)(1-q^2)}
{1 + c - (a+b)q^2- \cfrac{(c-abq^2)(1-q^3)} {\ddots} } }}.$$
Let us first give a combinatorial interpretation of $K(at,bt,ct^2;q)$ in terms of weighted Motzkin paths. This result is close to those given by Roblet and Viennot [@RoVi96], who developed a combinatorial theory of $T$-fractions. These are continued fractions of the form $1/(1-a_0t-b_0t/(1-a_1t-b_1t/\dots))$, and they are generating functions of Dyck paths with some weights on the peaks.
\[cfrac2\] The coefficient of $t^k$ in the expansion of $K(at,bt,ct^2;q)$ is the generating function of $\mathcal{M}^*_k(a,b,c;q)$.
The continued fraction $K(at,bt,ct^2;q)$ equals: $$\cfrac{1}
{1+ ct^2 - (a\hspace{-0.3mm}+\hspace{-0.3mm}b)t - \cfrac{t^2(c - ab)(1-q)}
{1+ ct^2 - (a\hspace{-0.3mm}+\hspace{-0.3mm}b)qt - \cfrac{t^2(c - abq)(1-q^2)}
{1+ ct^2 - (a\hspace{-0.3mm}+\hspace{-0.3mm}b)q^2t-
\cfrac{t^2(c\hspace{-0.3mm} -\hspace{-0.3mm} abq^2)
(1\hspace{-0.3mm}-\hspace{-0.3mm}q^3)} {\ddots} } }}.$$
Using the ideas introduced by Flajolet [@Fla82 Theorem 1], we thus obtain paths with four types of steps, denoted up $\nearrow$, down $\searrow$, level $\rightarrow$ and double-level $\longrightarrow$, the last type of step simply being twice as long as the usual level step. Moreover, the weight of
- an up step $\nearrow$ starting at height $h$ is either $1$ or $-q^{h+1}$,
- a level step $\rightarrow$ starting at height $h$ is $(a+b)q^h$,
- a down step $\searrow$ starting at height $h$ is either $c$ or $-abq^{h-1}$,
- a double-level step $\longrightarrow$ is $-c$.
To prove the statement, it suffices to construct a involution on the paths, such that
- its fixed points are precisely the elements of $\mathcal{M}^*_k(a,b,c;q)$, [*i.e.*]{} paths without double-level steps $\longrightarrow$ and without peaks $\nearrow\searrow$ such that the up step has weight $1$ and the down step has weight $c$,
- the weight of a path that is not fixed under the involution and the weight of its image add to zero.
Such an involution is easy to find: a path that is not in $\mathcal{M}^*_k(a,b,c;q)$, we look for the first occurrence of one of the two forbidden patterns, [*i.e.*]{} a double level step $\longrightarrow$ or a peak $\nearrow\searrow$ with steps weighted $1$ and $c$ respectively. We then exchange one of the patterns for the other – since the double level step $\longrightarrow$ has weight $-c$, the weights of the two paths add up to zero.
As mentioned above, $K(at,bt,ct^2;q)$ can be expressed as a basic hypergeometric series. We use the usual notation for these series, as for example in [@GaRa90].
\[cfrac1\] For $A\neq 1$, $B\neq 0$, we have
$$\label{cfrac_hyp}
K(A,B,C;q) =
\frac{1}{1-A} \cdot {}_2\phi_1
\left( \left. \begin{matrix} CB^{-1}q, q \\ Aq \end{matrix} \, \right| q,B \right).$$
For $A\neq 1$, $B=0$, we have
$$\label{cfrac_hypb0}
K(A,0,C;q) =
\frac{1}{1-A} \cdot {}_1\phi_1
\left( \left. \begin{matrix} q \\ Aq \end{matrix} \, \right| q,Cq \right).$$
Consider the following more general continued fraction, containing a new variable $z$: $$M(z) =
\cfrac{1}
{1+ C - (A+B)z - \cfrac{(C - ABz)(1-qz)}
{1+ C - (A+B)qz - \cfrac{(C - ABqz)(1-q^2z)}
{1+ C - (A+B)q^2z- \cfrac{(C - ABq^2z)(1-q^3z)} {\ddots} } }}.$$ Following Ismail and Libis [@IsLi89] (see also Identity 19.2.11a in the Handbook of Continued Fractions for Special Functions [@CPCWJ08]), we have:
$$M(z) = \frac{1}{1-z} \cdot
{}_2\phi_1 \left( \left. \begin{matrix} A,B \\ Cq \end{matrix} \, \right| q,qz\right) \cdot
{}_2\phi_1 \left(\left. \begin{matrix} A,B \\ Cq \end{matrix} \, \right| q,z \right)^{-1}.$$ To be able to specialise $z=1$, we can use one of Heine’s transformations [@GaRa90 p.13]. For $B\neq 0$ we obtain $$\begin{aligned}
M(z) = & \;
\frac{1}{1-z} \cdot
\frac{ (Aqz,B,Cq,z;q)_\infty}{(Az,B,Cq,qz ;q)_\infty } \cdot
{}_2\phi_1\left(\left. \begin{matrix} CB^{-1}q, qz \\ Aqz \end{matrix} \, \right| q,B \right) \cdot
{}_2\phi_1
\left(\left. \begin{matrix} CB^{-1}q,z \\ Az \end{matrix} \, \right| q,B\right)^{-1} \\[3mm]
= & \;
\frac{1}{1-Az} \cdot
{}_2\phi_1
\left(\left. \begin{matrix} CB^{-1}q,qz \\ Aqz \end{matrix} \,\right| q,B \right)
\cdot {}_2\phi_1
\left(\left. \begin{matrix} CB^{-1}q,z \\ Az \end{matrix} \,\right| q,B \right)^{-1}.
\end{aligned}$$ In case $B=0$, we have $$\begin{aligned}
M(z) = & \;
\frac{1}{1-z} \cdot
\frac{ (Aqz,q,Cq,z;q)_\infty}{(Az,q,Cq,qz ;q)_\infty } \cdot
{}_1\phi_1\left(\left. \begin{matrix} qz \\ Aqz \end{matrix} \, \right| q,Cq \right) \cdot
{}_1\phi_1
\left(\left. \begin{matrix} z \\ Az \end{matrix} \, \right| q,Cq\right)^{-1} \\[3mm]
= & \;
\frac{1}{1-az} \cdot
{}_1\phi_1
\left(\left. \begin{matrix} qz \\ Aqz \end{matrix} \,\right| q,Cq \right)
\cdot {}_1\phi_1
\left(\left. \begin{matrix} 1 \\ A \end{matrix} \,\right| q,Cq \right)^{-1}.
\end{aligned}$$
Although the symmetry in $A$ and $B$ is not apparent in Equation , it can be seen using one of Heine’s transformations [@GaRa90 p.13].
In the following, we will always use
$$\label{K_phi}
K(at,bt,ct^2;q)=
\frac{1}{1-at} \cdot {}_2\phi_1
\left( \left. \begin{matrix} cb^{-1}qt,q \\ aqt \end{matrix}\, \right| q,bt \right).$$
Besides, it is also possible to use a method giving $M(z)$ as a quotient of basic hypergeometric series without knowing [*a priori*]{} which identity to use. This method was employed in [@CJPR09], following Brak and Prellberg [@PrBr95]. Namely, note that the continued fraction expansion of $M(z)$ is equivalent to the equation: $$\label{eq_fun}
M(z) = \frac{1}{ 1 - c + (a+b)z - (c-abz)(1-qz)M(qz) }.$$ By looking for solutions of the form $M(z) = (1-az)^{-1}
\frac{H(qz)}{H(z)} $, we obtain a linear equation in $H(z)$, which gives a recurrence for the coefficients of the Taylor expansion of $H(z)$, which is readily transformed into the explicit form of $H(z)$ as a basic hypergeometric series.
Matchings {#sec:matchings}
---------
For matchings, we have $(a,b,c)=(0,0,1)$ and obtain:
\[qhermite\] $$K(0,0,t^2;q) = \sum_{k=0}^{\infty} (-t^2)^k q^{\binom{k+1}{2}}.$$
Essentially, this was shown by Penaud [@Pen95], who enumerated $\mathcal{M}^*_{2k}(0,0,1;q)$ by first constructing a bijection with parallelogram polyominoes, passing through several intermediate objects with beautiful names like cherry trees. On the polyominoes he was finally able to construct a weight-preserving, sign-reversing involution, with the only fixed point having weight $(-1)^k q^{\binom{k+1}{2}}$, corresponding to weighted Dyck paths with a single peak, and all weights maximal.
Set partitions {#sec:set-partitions}
--------------
For set partitions, we have $(a,b,c)=\big(0,-1,y(1-q)\big)$ and can use the following lemma:
\[qchar\] $$\label{eq:qchar}
K(0,-t,ct^2;q) =
\sum_{i=0}^{\infty} \sum_{j=0}^{i} t^{i+j} c^j (-1)^i
q^{\binom{j+1}2} {\genfrac{[}{]}{0pt}{}{i}{j}_q}.$$
Using Equation we find: $$K(0,-t,ct^2;q) =
{}_2\phi_1
\left(\left. \begin{matrix} -cqt,q \\ 0 \end{matrix} \,\right| q,-t \right)
= \sum_{i=0}^\infty (-cqt;q)_i (-t)^i.$$ The proof follows by plugging in the elementary expansion $$(-cqt;q)_i = \prod_{j=1}^i (1+q^jct) = \sum_{j=0}^{i}
q^{\binom{j+1}2} {\genfrac{[}{]}{0pt}{}{i}{j}_q} c^jt^j.$$
In the appendix, we give a bijective proof of this lemma.
Set partitions, modified crossings {#sec:set-partitions-*}
----------------------------------
When using the modified definition of crossings in set partitions, we have $(a,b,c)=\big(-1,y(1-q),0\big)$ and can use the following lemma:
\[qchar\*\] $$\label{qchar*_eq} K\big(-t,bt,0;q\big) =
\sum_{i=0}^\infty\sum_{j=0}^i t^i b^j (-1)^{i-j} {\genfrac{[}{]}{0pt}{}{i}{j}_q}.$$
Using Equation we find: $$K(-t,bt,0;q) =\frac{1}{1+t} {}_2\phi_1 \left(\left. \begin{matrix}
0,q \\ -qt \end{matrix} \,\right| q,bt \right) =
\sum_{i=0}^\infty \frac{1}{(-t;q)_{i+1}} (bt)^i.$$ The proof follows by plugging in the elementary expansion $$\frac{1}{(-t;q)_{i+1}} = \prod_{j=1}^i \frac{1}{1+q^jt} =
\sum_{j\geq 0}
{\genfrac{[}{]}{0pt}{}{i+j}{j}_q} (-t)^j.$$
Permutations {#sec:permutations}
------------
In the case of permutations, we have $(a,b,c)=(-1,-yq,y)$ and find:
\[An\_K\] $$K(-t,-yqt,yt^2;q) = \sum_{k=0}^{\infty} (-t)^k
\left( \sum_{i=0}^k y^i q^{i(k+1-i)} \right).$$
Using Equation , we have: $$\begin{aligned}
K(-t,-yqt,yt^2;q) = & \;
\frac{1}{1+t} \cdot {}_2\phi_1
\left(\left. \begin{matrix} -t,q \\ -qt \end{matrix} \, \right| q,-yqt \right)
= \sum_{i=0}^\infty \frac{ (-yqt)^i }{ 1 + tq^i } \\
=& \; \sum_{i=0}^\infty \sum_{j=0}^\infty (-yqt)^i (-tq^i)^j
= \sum_{i=0}^\infty \sum_{j=0}^\infty (-t)^{i+j} y^i q^{i(j+1)}.
\end{aligned}$$ To finish the proof it only remains to substitute $k$ for $i+j$.
Conclusion
==========
Let us briefly summarize how the four theorems announced in the introduction can be proved using the previous sections. In each case, the enumeration of crossings in combinatorial objects is linked with the enumeration of the weighted Motzkin paths in $\mathcal{M}_n(a,b,c,d;q)$. The bijection $\Delta$ shows that the generating function of crossings can be decomposed into the generating functions of the sets $\mathcal{P}_{n,k}(c,d)$ and $\mathcal{M}^*_n(a,b,c;q)$, which in turn have been obtained in the previous two sections. This fulfills our initial objective as stated in the introduction.
Inverse relations
=================
We would like to mention an interesting non-bijective point of view of the path decomposition given in Section \[sec:penaud-decomp\], using inverse relations. Given two sequences $\{a_n\}$ and $\{b_n\}$, an inverse relation is an equivalence such as, for example: $$\forall n\geq0, \quad a_n = \sum_{k=0}^n \binom n k b_k
\quad \Longleftrightarrow \quad
\forall n\geq0, \quad b_n = \sum_{k=0}^n \binom n k (-1)^{n-k} a_k.$$ This particular relation is easily proved by checking that the (semi-infinite) lower triangular matrix $(\binom i j )_{i,j\in\mathbb{N}}$ has an inverse, which is $( (-1)^{i+j} \binom i j )_{i,j\in\mathbb{N}}$. Other relations of this kind can be found in Chapters 2 and 3 of Riordan’s book ‘Combinatorial Identities’ [@Rio68]. To prove the Touchard-Riordan formula , let $a_{2n}=(1-q)^n\mu_{2n}^{\tilde H}$ and $a_{2n+1}=0$. We then use the following inverse relation [@Rio68 Chapter 2, Equation (12)]: $$\label{inverse_pair1}
a_n = \sum_{k=0}^{\lfloor \frac n2 \rfloor}
\left(\tbinom{n}{k} - \tbinom{n}{k-1}\right) b_{n-2k}
\quad \Longleftrightarrow \quad
b_n = \sum_{k=0}^{\lfloor \frac n2 \rfloor}
(-1)^k \tbinom{n-k}{k} a_{n-2k}.$$ Again, this can be proved by inverting a lower-triangular matrix. It remains to prove that $b_{2k}=(-1)^k q^{\binom{k+1}2}$ and $b_{2k+1}=0$. To this end, we relate the sequence $b_n$ to Schröder paths:
A [*Schröder path*]{} of length $2n$ is a path in $\mathbb{N}\times\mathbb{N}$ starting at $(0,0)$, arriving at $(2n,0)$ with steps $(1,1)$, $(1,-1)$, or $(2,0)$.
Suppose that $a_{2n+1}=0$ and $a_{2n}$ is the generating function of Dyck paths of length $2n$, with weight $1-q^{h+1}$ on each north-east step starting at height $h$ (this is to say $a_{2n}=(1-q)^n\mu_{2n}^{\tilde H}$). Suppose that $a_n$ and $b_n$ are related by . Then $b_{2n+1}=0$, and $b_{2n}$ is the generating function of Schröder paths of length $2n$, with weight $-1$ on each level step, and $1-q^{h+1}$ on each north-east step starting at height $h$.
For any even $n$, consider a Schröder path of length $n$ with $k$ level steps, weighted as described above. This path has $n-k$ non-level steps, and can thus be obtained from a Dyck path of length $n-2k$ by inserting the level steps. There are $\binom {n-k}{k}$ ways to do so, which implies that the generating function of Schröder paths is indeed equal to $\sum_{k=0}^{\lfloor \frac n2 \rfloor} (-1)^k
\tbinom{n-k}{k} a_{n-2k}$, and therefore equal to $b_n$.
With the bijective decomposition of paths in Section \[sec:penaud-decomp\], we showed that we have to obtain the generating function of the set $\mathcal{M}^*_k(0,0,1;q)$ to prove Touchard-Riordan formula. With the inverse relations, we showed that we have to count certain weighted Schröder paths. The fact that both sets have the same generating function follows from the involution given in the proof of Proposition \[cfrac2\].
It is also possible to use an inverse relation to obtain the formula in Theorem \[th:Laguerre\] for the $q$-Laguerre moments. By inverting a lower triangular matrix, one can check that $$a_n = \sum_{k=0}^n \left( \sum_{j=0}^k y^j \Big( \tbinom{n}{j}\tbinom{n}{j+k} -
\tbinom{n}{j-1}\tbinom{n}{j+k+1}\Big) \right) b_k$$ for all $n$ is equivalent to $$\label{inv22}
b_n = \sum_{k=0}^n \left( \sum_{j=0}^{\lfloor \frac{n-k}2\rfloor} \binom{n-j}{n-k-j}
\binom{n-k-j}j (-y)^j (-1-y)^{n-k-2j} \right) a_k$$ for all $n$.
Equation can be interpreted as follows: given that $a_k$ counts elements of $\mathcal{M}_n(1,-yq,y,1+y;q)$, then $b_n$ count paths of length $n$, with the same weights as in $\mathcal{M}_n(1,-yq,y,1+y;q)$, where we insert some level steps $\rightarrow$ with weight $-1-y$, and some double level steps $\longrightarrow$ with weight $-y$.
Indeed, suppose that we inserted $j$ double-level steps $\longrightarrow$, and hence $n-k-2j$ level steps $\rightarrow$, starting with a path of length $k$. This yields the weight $(-y)^j (-1-y)^{n-k-2j}$ for the inserted steps. The first binomial coefficient, $\binom{n-j}{n-k-j}$, is the number of ways to insert the $n-k-j$ level steps among the $n-j$ steps (the total number of steps being $n-j$ because the length is $n$, and $j$ steps of double length).
The second binomial coefficient, $\binom{n-k-j}{j}$, is the number of combinations of the $j$ inserted double-level steps and the $n-k-2j$ inserted level steps. Using the involution given in the proof of Proposition \[cfrac2\], we see that $b_n$ counts elements in the set $\mathcal{M}_n^*(1,-yq,y;q)$.
A bijective proof of Lemma \[qchar\] {#sec:biject-proof-lemma}
====================================
We show in this appendix that Penaud’s bijective method of proving Lemma \[qhermite\] can be generalised to prove also Lemma \[qchar\]. Namely, we construct a sign-reversing involution on a set of weighted Motzkin paths, whose fixed points are enumerated by the right-hand side of . This involution was essentially given by the first author in [@MJV] in a different context. We take the opportunity to correct some mistakes in this reference.
In the following we fix integers $j,k\geq0$ and consider the set $\mathcal{C}_{j,k}$ of Motzkin paths of length $k+j$ with $k-j$ level steps $\rightarrow$, (and hence $j$ up steps $\nearrow$ and $j$ down steps $\searrow$), satisfying the following conditions:
- the weight of all up steps $\nearrow$ is $1$,
- the weight of a level step $\rightarrow$ at height $h$ is $-q^h$,
- the weight of a down step $\searrow$ starting at height $h$ is either $1$ or $-q^h$,
- there is no peak $\nearrow\searrow$ such that both the up step and the down step have weight $1$.
The generating function of $\mathcal{C}_{j,k}$ is then the coefficient of $a^kt^{k+j}$ in $K(0,-t,at^2;q)$. It thus suffices to prove the following:
\[invol\] There is an involution $\theta$ on the set $\mathcal{C}_{j,k}$ such that:
- the fixed points are the paths that
- start with $j$ up steps $\nearrow$,
- and contain no down steps $\searrow$ of weight $1$,
- the weight of a path that is not fixed under the involution and the weight of its image add to zero.
Moreover, the sum of weights of fixed points of the involution is $(-1)^kq^{\binom{j+1}{2}}{\genfrac{[}{]}{0pt}{}{k}{j}_q}$.
Penaud’s method consists in introducing several intermediate objects as described in Section \[sec:matchings\]. However, in the case at hand we will not use intermediate objects, but rather construct the involution directly on the paths. What we give is a generalisation of Penaud’s construction, which we recover in the case $k=j$.
Following Penaud [@Pen95], we use in this proof a word notation for elements in $\mathcal{C}_{j,k}$. The letters $x$, $z$, $y$, and $\bar{y}$ will respectively denote the steps $\nearrow$, $\rightarrow$, $\searrow$ with weight $1$, and $\searrow$ with weight $-q^h$. For any word $c\in\mathcal{C}_{j,k}$, we define:
- $u(c)$ as the length of the last sequence of consecutive $x$,
- $v(c)$ as the starting height of the last step $y$, if $c$ contains a $y$ and there is no $x$ after the last $y$, and $j$ otherwise.
See Figure \[invol\_cores\] for an example. The fixed points of $\theta$ will be $c\in\mathcal{C}_{j,k}$ such that $u(c)=v(c)=j$, which correspond to the paths described in Proposition \[invol\].
now, suppose $c$ is such that $u(c)<j$ or $v(c)<j$. We will build $\theta$ so that $v(c)\leq u(c)$ if and only if $u(\theta(c)) < v(\theta(c))$. Thus it suffices to define $\theta(c)$ in the case $v(c)\leq u(c)$, and to check that we have indeed $u(\theta(c)) < v(\theta(c))$. So let us suppose $v(c)\leq u(c)$, hence $v(c)<j$.
Since $v(c)<j$, there is at least a letter $y$ in $c$ having no $x$ to its right. Let $\tilde{c}$ be the word obtained from $c$ by replacing the last $y$ with a $\bar{y}$. There is a unique factorisation $$\tilde{c}=f_1x^{u(c)} ay^{\ell}f_2$$ such that:
- $a$ is either $z$, or $\bar{y}$,
- $f_2$ begins with $z$ or $\bar{y}$, contains at least one letter $\bar{y}$, but contains no $x$.
Let us explain this factorisation. By definition of $u(c)$, we can write $\tilde{c}=f_1x^{u(c)} c'$, where $c'$ does not contain any $x$. In a word $c\in\mathcal{C}_{j,k}$, an $x$ cannot be followed by a $y$. So we can write $\tilde{c}=f_1x^{u(c)} a c''$ where $a$ is either $z$, or $\bar{y}$. Then, we write $c''=y^\ell f_2$ with $\ell\geq0$ maximal, and $f_2$ satisfy the conditions ($f_2$ contains indeed a $\bar y$ because we transformed a $y$ into a $\bar y$). Uniqueness is immediate.
We set: $$\label{def_theta}
\theta(c) = f_1x^{u(c)-v(c)}ay^\ell x^{v(c)}f_2.$$ See Figure \[invol\_cores\] for an example with $u(c)=4$, $v(c)=2$, $j=9$ and $k=12$. We can check that $w(c)=-q^{19}= - w(\theta(c))$, and $u(\theta(c))=2$, $v(\theta(c))=3$.
(0,-1)(21,8) (0,0)(21,6) (0,0)(1,1)(2,2)(3,1)(4,2)(5,2)(6,1)(7,2)(8,3)(9,2)(10,3)(11,4)(12,5)(13,6) (14,6)(15,5)(16,4)(17,3)(18,2)(19,1)(20,1)(21,0) (2,2)(3,1) (8,3)(9,2) (16,4)(17,3)(20,1)(21,0) (4,2)(5,2) (19,1)(20,1) (13,6)(14,6) (-2,3)[$c=$]{} (0.5,-1)[$x$]{}(1.5,-1)[$x$]{}(2.5,-1)[$\bar y$]{}(3.5,-1)[$x$]{}(4.5,-1)[$z$]{} (5.5,-1)[$y$]{}(6.5,-1)[$x$]{}(7.5,-1)[$x$]{}(8.5,-1)[$\bar y$]{}(9.5,-1)[$x$]{} (10.5,-1)[$x$]{}(11.5,-1)[$x$]{}(12.5,-1)[$x$]{}(13.5,-1)[$z$]{} (14.5,-1)[$y$]{}(15.5,-1)[$y$]{}(16.5,-1)[$\bar y$]{}(17.5,-1)[$y$]{} (18.5,-1)[$y$]{}(19.5,-1)[$z$]{}(20.5,-1)[$\bar y$]{} (12.9,6.5)(9.1,6.5) (11.5,7.3)[$u(c)=4$]{} (23,3.2)[$v(c)=2$]{}(21.5,2.8)(18.6,1.6)
\
(0,-1)(21,8) (0,0)(21,6) (0,0)(1,1)(2,2)(3,1)(4,2)(5,2)(6,1)(7,2)(8,3)(9,2)(10,3)(11,4)(12,5)(13,6) (14,6)(15,5)(16,4)(17,3)(18,2)(19,1)(20,1)(21,0) (2,2)(3,1) (8,3)(9,2) (16,4)(17,3)(20,1)(21,0) (4,2)(5,2) (19,1)(20,1) (13,6)(14,6) (18,2)(19,1) (-2,3)[$\tilde c=$]{} (0.5,-1)[$x$]{}(1.5,-1)[$x$]{}(2.5,-1)[$\bar y$]{}(3.5,-1)[$x$]{}(4.5,-1)[$z$]{} (5.5,-1)[$y$]{}(6.5,-1)[$x$]{}(7.5,-1)[$x$]{}(8.5,-1)[$\bar y$]{}(9.5,-1)[$x$]{} (10.5,-1)[$x$]{}(11.5,-1)[$x$]{}(12.5,-1)[$x$]{}(13.5,-1)[$z$]{} (14.5,-1)[$y$]{}(15.5,-1)[$y$]{}(16.5,-1)[$\bar y$]{}(17.5,-1)[$y$]{} (18.5,-1)[$\bar y$]{}(19.5,-1)[$z$]{}(20.5,-1)[$\bar y$]{} (8.9,6.5)(0.1,6.5) (12.9,6.5)(9.1,6.5) (15.9,6.5)(13.1,6.5) (20.9,6.5)(16.1,6.5) (5,7.3)[$f_1$]{} (11.5,7.3)[$x^{u(c)}$]{} (14.5,7.3)[$ay^\ell $]{} (18,7.3)[$f_2$]{}
\
(0,-2)(21,8) (0,0)(21,6) (0,0)(1,1)(2,2)(3,1)(4,2)(5,2)(6,1)(7,2)(8,3)(9,2)(10,3)(11,4)(12,4)(13,3) (14,2)(15,3)(16,4)(17,3)(18,2)(19,1)(20,1)(21,0) (2,2)(3,1) (8,3)(9,2) (16,4)(17,3)(18,2)(19,1) (20,1)(21,0) (4,2)(5,2) (19,1)(20,1) (11,4)(12,4) (-2,3)[$\theta(c)=$]{} (0.5,-1)[$x$]{}(1.5,-1)[$x$]{}(2.5,-1)[$\bar y$]{}(3.5,-1)[$x$]{}(4.5,-1)[$z$]{} (5.5,-1)[$y$]{}(6.5,-1)[$x$]{}(7.5,-1)[$x$]{}(8.5,-1)[$\bar y$]{}(9.5,-1)[$x$]{} (10.5,-1)[$x$]{}(11.5,-1)[$x$]{}(12.5,-1)[$x$]{}(13.5,-1)[$z$]{} (14.5,-1)[$y$]{}(15.5,-1)[$y$]{}(16.5,-1)[$\bar y$]{}(17.5,-1)[$y$]{} (18.5,-1)[$\bar y$]{}(19.5,-1)[$z$]{}(20.5,-1)[$\bar y$]{} (8.9,6.5)(0.1,6.5) (10.9,6.5)(9.1,6.5) (13.9,6.5)(11.1,6.5) (15.9,6.5)(14.1,6.5) (20.9,6.5)(16.1,6.5) (4,7.3)[$f_1$]{} (9,7.6)[$x^{u(c)-v(c)}$]{} (12.8,7.3)[$ay^\ell $]{} (15.3,7.6)[$x^{v(c)}$]{} (19,7.3)[$f_2$]{}
We show the following points:
- The path $\theta(c)$ is a Motzkin path. Indeed, the factor $ay^j$ in $c$ ends at height at least $v(c)$, since the factor $f_2$ contains a step $\bar y$ starting at this height and contains no $x$. We can thus shift this factor $\tilde c$ so that the result is again a Motzkin path.
- The path $c$ and its image $\theta(c)$ have opposite weights. To begin, between $c$ and $\tilde c$, the weight is multiplied by $-q^{v(c)}$, since we have transformed a $y$ into a $\bar y$ starting at height $v(c)$. Between $\tilde c$ and $\theta(c)$, the height of the factor $ay^j$ has decreased by $v(c)$, so the weight has been divided by $q^{v(c)}$. A factor $-1$ remains, which proves the claim.
- The path $\theta(c)$ is such that $u(\theta(c))<v(\theta(c))$. From the definition (\[def\_theta\]) we see that $u(\theta(c))=v(c)$. Besides, $v(c) < v(\theta(c))$ since the last step $y$ of $c$ has been transformed into a $\bar{y}$ to obtain $\tilde c$ and $\theta(c)$.
- Every path $c'$ with $u(c')<v(c')$ is obtained as a $\theta(c)$ for some other path $c$ with $u(c)\geq v(c)$. Indeed, let $\tilde{c}'$ be the word obtained from $c'$ after replacing the last $\bar{y}$ at height $u(c')$ with a $y$. There is a unique factorisation $\tilde{c}'=f_1ay^jx^{u(c')}f_2$, where $a$ is $z$ or $\bar{y}$, and $f_2$ contains no $x$. Then by construction, $c=f_1x^{u(c')}ay^jf_2$ has the required properties.
Thus, $\theta$ is indeed an involution with the announced fixed points.
It remains only to check that the sum of weights of the fixed points is equal to $(-1)^kq^{\binom{j+1}{2}}{\genfrac{[}{]}{0pt}{}{k}{j}_q}$. A fixed point of $\theta$ is specified by the heights $h_1,\dots,h_{k-j}$ of the $k-j$ steps $\rightarrow$. These heights can be any set of values provided that $j\geq h_1 \geq \dots \geq h_{k-j} \geq 0$. The weight of such a fixed point is: $$(-1)^jq^{\frac{j(j+1)}2} \prod_{i=0}^{k-j} (-q^{h_i})
= (-1)^kq^{\frac{j(j+1)}2} q^{\sum_{i=0}^{k-j}h_i}.$$ Indeed, the $j$ steps $\nearrow$ have respective weights $-q,-q^2,\dots,-q^j$, which gives a factor $(-1)^jq^{\frac{j(j+1)}2}$. Besides, we have: $$\sum_{j\geq h_1\geq\dots\geq h_{k-j}\geq 0} q^{\sum h_i} = {\genfrac{[}{]}{0pt}{}{k}{j}_q},$$ by elementary property of $q$-binomial coefficients. This ends the proof.
Closed forms for $\size{\mathcal{P}_{n,k}(c,d)}$ {#sec:trinomial}
================================================
In this appendix we give a justification of the fact that there is no (hypergeometric) closed form (in the sense of Petkov[š]{}ek, Wilf and Zeilberger [@AeqB]) for $$\size{\mathcal{P}_{n,k}(c,d)}=
\frac{k+1}{n+1}\sum_{l=0}^{n-k}\binom{n+1}{l}\binom{l}{2l-n+k}d^{2l-n+k}c^{n-k-l}.$$ except when $(c,d)$ is one of $(1,0)$, $(0,1)$ or $(y^2,2y)$. More precisely, we claim that $\size{\mathcal{P}_{n,k}(c,d)}$ cannot be written as a linear combination (of a fixed finite number) of hypergeometric terms except in the specified cases. In the following we sketch a straightforward way to check this is using computer algebra.
First we convert the summation into a polynomial recurrence equation. This can be done by using Zeilberger’s algorithm (which also proves that the recurrence is correct), for example. Writing $p_n = \size{\mathcal{P}_{n,k}(c,d)}$ we obtain $$(4c-d^2)(n+1)(n+2)p_n + d(n+2)(2n+5)p_{n+1}-(n+2-k)(n+4+k)p_{n+2}=0.$$ Alternatively, one can also find a recurrence for $q_k =
\size{\mathcal{P}_{n,k}(c,d)}$, which is $$(k+3)(k-n)q_k + d(k+1)(k+3) q_{k+1} + c(k+1)(k+n+4)q_{k+2}=0.$$
It now remains to show that both equations admit no hypergeometric solutions, except for the values of $(c,d)$ mentioned above. To this end we use Petkov[š]{}ek’s algorithm [**hyper**]{}, as described in Chapter 8 of ‘A=B’ [@AeqB]. Unfortunately, this time we cannot use the implementation naively. Namely, [*a priori*]{} [ **hyper**]{} decides only for fixed parameters $(c,d)$ whether a hypergeometric solutions exists or not.
However, it is possible to trace the algorithm, and, whenever it has to decide whether a quantity containing $c$ or $d$ is zero or not, do it for the computer. (Of course, it should be possible to actually program this, but that is outside the scope of this article.) We refrain from giving a complete proof, but rather give only a few details to make checking easier.
First of all, let us assume that $c$, $d$ and $d^2-4c$ are all nonzero. Then the degrees of the coefficient polynomials in both recurrence equations are all the same. From now on, the procedure is the same for both recurrence equation, so let us focus on the one for $p_n$. According to the remark in Example 8.4.2 in [@AeqB], we have to consider all monic factors $a(n)$ of the coefficient of $p_n$, and also the monic factors $b(n)$ of the coefficient of $p_{n+2}$, such that the degree of $a(n)$ and $b(n)$ coincide. In this case, the characteristic equation, Equation (8.4.5) in [@AeqB] one has to solve turns out to be $z^2-2dz+d^2-4c$. For each of the two solutions in $z$, one has to check that there is no polynomial solution of the recurrence $$P_0(n)c_n + zP_1(n)c_{n+1} + z^2P_2(n) c_{n+2}=0,$$ where the coefficient polynomials $P_0(n),P_1(n)$ and $P_2(n)$ are polynomials derived from the coefficient polynomials of the original recurrence by multiplying with certain shifts of $a(n)$ and $b(n)$.
This can be done with the algorithm [**poly**]{}, described in Section 8.3 of [@AeqB]. Namely, depending on the degrees of yet another set of polynomials derived from $P_0(n),P_1(n)$ and $P_2(n)$, it computes an upper bound for the degree of a possible polynomial solution. Indeed, the algorithm decides that the degree of such a solution would have to be negative, provided that $c$, $d$ and $d^2-4c$ are nonzero, which is what we assumed.
Acknowledgement {#acknowledgement .unnumbered}
===============
This article develops ideas presented in the extended abstract [@CJPR09]. We thank Sylvie Corteel and Thomas Prellberg, who contributed considerably to our curiosity about generalising Penaud’s construction.
[999]{}
S. Corteel, Crossings and alignments of permutations, Adv. in App. Math. 38(2) (2007), 149–163.
S. Corteel, M. Josuat-Vergès, T. Prellberg and R. Rubey, Matrix Ansatz, lattice paths and rook placements, Proc. FPSAC 2009.
A. Cuyt, A.B. Petersen, B. Verdonk, H. Waadeland and W. B. Jones, Handbook of Continued Fractions for Special Functions, Springer, 2008.
A. de Médicis, D. Stanton and D. White, The combinatorics of $q$-Charlier polynomials, J. Combin. Theory Ser. A 69 (1995), 87–114.
R. Ehrenborg and M. Readdy, Juggling and applications to $q$-analogues, Discrete Math. 157 (1996), 107–125
P. Flajolet, Combinatorial aspects of continued fractions, Discrete Math. 41 (1982), 145–153.
G. Gasper and M. Rahman, Basic hypergeometric series, Second edition, Cambridge University Press, 2004.
I.M. Gessel and X.G. Viennot, Binomial determinants, paths and hook length formulae, Adv. in Math. 58 (1985), 300–321.
H.W. Gould, The $q$-Stirling numbers of first and second kinds, Duke Math. J. 28(2) (1961), 281–289.
M.E.H. Ismail and C. A. Libis, Contiguous relations, basic hypergeometric functions, and orthogonal polynomials, I. J. Math. Anal. Appl. 141(2) (1989), 349–372.
M.E.H. Ismail, D. Stanton and X.G. Viennot, The combinatorics of $q$-Hermite polynomials and the Askey-Wilson integral, European J. Combin. 8 (1987), 379–392.
M. Josuat-Vergès, Rook placements in Young diagrams and permutation enumeration, to appear in Adv. Appl. Math. (2010).
A. Kasraoui, D. Stanton and J. Zeng, The combinatorics of Al-Salam-Chihara $q$-Laguerre polynomials, arXiv$:$0810.3232v1 \[math.CO\].
A. Kasraoui and J. Zeng, Distribution of crossings, nestings and alignments of two edges in matchings and partitions, Electron. J. Combin. 13(1) (2006), R33.
D. Kim, D. Stanton and J. Zeng, The combinatorics of the Al-Salam-Chihara $q$-Charlier polynomials, Sém. Lothar. Combin. 54 (2006), Article B54i.
R. Koekoek, P.A. Lesky and R.F. Swarttouw, Hypergeometric Orthogonal Polynomials and Their $q$-analogues, Springer, 2010.
J.-G. Penaud, A bijective proof of a Touchard-Riordan formula, Discrete Math. 139 (1995), 347–360.
M. Petkov[š]{}ek, H. Wilf and D. Zeilberger, $A=B$, Peters Ltd. 1996.
T. Prellberg and R. Brak, Critical Exponents from Non-Linear Functional Equations for Partially Directed Cluster Models, J. Stat. Phys. 78 (1995), 701–730.
J. Riordan, The distribution of crossings of chords joining pairs of 2n points on a circle, Math. Comput. 29(129) (1975), 215–222.
J. Riordan, Combinatorial identities, Wiley, 1968.
E. Roblet and X.G. Viennot, Théorie combinatoire des T-fractions et approximants de Padé en deux points, Discrete Math. 153(1-3) (1996), 271–288.
R. Simion and D. Stanton, Octabasic Laguerre polynomials and permutation statistics, J. Comp. Appl. Math. 68 (1996), 297–329.
R. Stanley, Enumerative Combinatorics Vol. 2, Cambridge University Press, 1999.
J. Touchard, Sur un problème de configurations et sur les fractions continues, Can. J. Math. 4 (1952), 2–25.
X.G. Viennot, Une théorie combinatoire des polynômes orthogonaux, Lecture notes, UQÀM, Montréal, 1984.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We present molecular dynamics simulations of pseudo hard sphere fluid (generalized WCA potential with exponents (50, 49) proposed by Jover [*[et al.]{}*]{} [*J. Chem. Phys*]{} [**137**]{}, (2012)) using GROMACS package. The equation of state and radial distribution functions at contact are obtained from simulations and compared to the available theory of true hard spheres (HS) and available data on pseudo hard spheres. The comparison shows agreements with data by Jover [*[et al.]{}*]{} and the Carnahan-Starling equation of HS. The shear viscosity is obtained from the simulations and compared to the Enskog expression and previous HS simulations. It is demonstrated that the PHS potential reproduces the HS shear viscosity accurately.'
author:
- 'Faezeh Pousaneh$^{\scriptscriptstyle{}}$[^1]'
- 'Astrid S. de Wijn'
bibliography:
- 'ref.bib'
title: ' Shear viscosity of pseudo hard-spheres'
---
Introduction
============
Hard Sphere (HS) models have been widely used as a basic approximation of a spherical atom or molecule (See, for example [@pippo]), because of the simplicity of the interaction potential and the instantaneous elastic collision dynamics. Although the HS model is an idealized model, it still captures the essential physics of macroscopic behavior of real fluids, both in and out of equilibrium. Consequenly, HS-based models are often used to study and understand the thermodynamics and transport properties of liquids. Nevertheless, because HS models are highly idealised, it is difficult to make quantitative predictions for more complex molecules based on purely theoretical calculations. This is why theory based hard sphere models are often used as a basis for empirical approaches to fluid properties that go far beyond spherical molecules [@VW; @viscvogelmethane]. This type of approach, however, requires analytical theory that can be challenging to develop, especially for more complicated models.
Because of the simple instantaneous dynamics, there are a great deal of analytical results for HS models for properties of liquids, such as equations of state [@carnahanstarling] and transport coefficients [@pippo; @Chapman:52:0]. These kinds of results can still be challenging to obtain, but they provide a powerful basis for continuing development of fluid theory [@safths]. There are also a number of extensions of HS model that still retain the instantaneous collision dynamics and therefore are still somewhat tractable when it comes to analytical approaches. Examples are sperocylinders [@spherocylinders], rough hard spheres [@Chapman:52:0], loaded hard spheres [@loadedspheres], and dipolar hard spheres [@dhsshear; @dhsrelaxation].
In supporting this development molecular dynamics (MD) simulations are tools that have become much more commonplace, for example as numerical experiments to verify the theoretical results. Although there are available MD studies on transport coefficients of HS fluids, [ for example a comprehensive one by Sigurgeirsson [*[et al.]{}*]{} [@heyes:09:0],]{} it is not possible to directly use the state of art simulations packages like LAMMPS, GROMACS or NAMD which provide high-speed simulations of complex and large systems [@Plimpton:95:0; @Spoel:05:0; @Phillips:05:0]. These simulation packages rely on approximately smooth dynamics, and thus do not support the HS model’s instantaneous collision dynamics.
To get around this, Jover [*[et al.]{}*]{} proposed a pseudo hard sphere (PHS) model which is a cut and shifted version of a Mie potential with exponents (50,49) [@jover:12:0]. Using this nearly hard, but smooth potential, Jover [*[et al.]{}*]{} have been able to reproduce structural and thermodynamic properties accurately compared with available simulation data for the original HS system. It has also been shown to provide reliable results for studying liquid-solid coexistence [@vega:13:0].
The focus of the PHS research mentioned above has been primarily on equilibrium properties. However, non-equilibrium properties such as transport coefficients are much more sensitive to changes in the dynamics than equilibrium properties. Because of this, a good reproduction of equilibrium properties does not directly imply that non-equilibrium properties would be described as well. So far, only the self-diffusion was briefly tested in [@jover:12:0]. The performance of the PHS potential for non-equilibrium properties has therefore not been sufficiently established.
In this work, we test in detail the reliability of the (50,49) PHS potential for calculating the viscosity. [ We investigate the density dependence of the viscosity, and compare it to both analytical results for HS and previous MD simulations of true HS [@heyes:09:0]. ]{}
Simulation setup
================
[ The Mie ($\lambda_r, \lambda_a$) potential (the generalized Lennard-Jones) can be changed to a repulsive potential by shifting it by its well-depth value and equating the larger distances interactions to zero. WCA potential is one of this cut-and-shift potential which has the form $$U_{WCA}(r)=
\begin{cases}
4 \epsilon \big[ \big(\frac{\sigma}{r}\big)^{12}-\big(\frac{\sigma}{r}\big)^{6 } \big]+\epsilon; \hspace{2cm} r< 2^{1/6} \sigma \\
\hspace{3.5cm} 0 ; \hspace{2cm} r \geq 2^{1/6}\sigma
\label{eqWCA}
\end{cases}.$$ $$$$ The WCA potential can be used to present a HS system in an approximate manner. However, the reliability of the approximation depends on ($\lambda_r, \lambda_a$) and on the temperature of the system. Jover [*[et al.]{}*]{} [@jover:12:0] have studied the effect of exponents $\lambda_r$ and $\lambda_a$ on the steepness and the shape of the potential. They have chosen the pair (50,49) as a compromise between fidelity of the representation of the HS and the size of the time step in MD simulations. The steeper the potential the greater the fidelity, but the smaller the time step that is required. The PHS potential proposed by Jover [*[et al.]{}*]{} [@jover:12:0] has the following form $$U_{50,49}(r)=
\begin{cases}
50 ~ (\frac{50}{49})^{49} \epsilon \big[ \big(\frac{\sigma}{r}\big)^{50 }-\big(\frac{\sigma}{r}\big)^{49 } \big]+\epsilon; \hspace{2cm} r< \frac{50}{49} \sigma \\
\hspace{5.cm} 0 ; \hspace{2cm} r \geq \frac{50}{49} \sigma
\label{eqPHS}
\end{cases}.$$ $$$$ In addition the effect of temperature on $U_{50,49}(r)$ potential is studied in ref. [@jover:12:0] and is concluded that at a reduced temperature of $ T^*=\frac{\epsilon}{k_B T}=2/3$ the potential produces better agreement with the true HS. They have examined this in terms of thermodynamics and structural properties.]{}
\[sec3\] In order to simulate the PHS model, we have used GROMACS version 5. The potential is implemented as a tabular form. The first goal was to obtain the equation of state (pressure versus density). We have simulated a box of $N=1000$ particles with LJ parameters $\sigma$, $\epsilon$ at different pressures corresponding to different densities, at reduced temperature $ T^*=2/3$. We limit ourselves to this temperature because according to Refs. [@vega:13:0; @jover:12:0] the equilibrium properties of the PHS and HS models are similar at this temperature. In what follows, all quantities are given in reduced units: $t^*=t\sqrt{ \frac{k_B T}{\sigma^2 m}}$, $r^*= \frac{r}{\sigma}$, $\rho^*= \rho \sigma^3$ and $P^*= \frac{P \sigma^3}{k_B T}$, $\eta^*=\eta (\frac{\sigma ^2 }{\sqrt {mk_B T}})$, where $\rho$, $P$ and $\eta$ denote the number density, the pressure, and the viscosity respectively. The equilibrated systems are simulated in NPT ensemble for $t^*=20000 $ with time-step of $\delta t^*=0.0011$ using a Parrinello-Rahman barostat and a velocity-rescale thermostat.
![(Left) Reduced pressure versus reduced number density from MD simulation of PHS in this work (squares) is compared to earlier work by Jover [*[et al.]{}*]{} [@jover:12:0] (circles) and theoretical Carnahan-Starling equation of state (solid line).[ (Right) The same as left plot, but rescaled with the Carnahan-Starling pressure, in order to enhance details. ]{} The errors in the current work data of $P^*$ are less than 4%, so are not well visible. []{data-label="PHS"}](Graph_comparison.eps "fig:") ![(Left) Reduced pressure versus reduced number density from MD simulation of PHS in this work (squares) is compared to earlier work by Jover [*[et al.]{}*]{} [@jover:12:0] (circles) and theoretical Carnahan-Starling equation of state (solid line).[ (Right) The same as left plot, but rescaled with the Carnahan-Starling pressure, in order to enhance details. ]{} The errors in the current work data of $P^*$ are less than 4%, so are not well visible. []{data-label="PHS"}](Graph_comparison2.eps "fig:")
Simulation results for the equation of state are shown in Fig. \[PHS\]. The relative errors in the pressure $P^*$ are less than 4%. The results agree with the results by Jover [*[et al.]{}*]{} [@jover:12:0]. They are also in good agreement with the Carnahan-Starling equation of state (solid line) [@Carnahan:69:00] $$\begin{aligned}
P^*= (6/\pi)\frac{\xi(1+\xi^*+\xi^{*2}-\xi^{*3})}{(1-\xi^*)^3}, \\
\xi^*=(\pi/6) \rho^*.\hspace{1.5cm}\end{aligned}$$ Ref. [@vega:13:0] also reports similar results using GROMACS package.
Shear viscosity
---------------
The shear viscosity can be determined both by equilibrium molecular-dynamics (EMD) or non-equilibrium molecular-dynamics (NEMD) simulations. EMD methods are based on pressure or momentum fluctuations, for example the Transverse-Current Autocorrelation Function (TCAF) method or Green-kubo method. The TCAF method is the easiest to implement and has several advantages over the Green-kubo relation. The first advantage is that any non-hydrodynamic behavior is easy to identify in the TCAF. The second advantage of the TCAF method is that it provides a natural way to estimate the magnitude of finite-size effects. This can be done in a single simulation and the results extrapolated to the infinite system limit in a straightforward calculation, (see [@palmer:94:0] and reference therein). In NEMD methods such as periodic perturbation (PP) method, instead of measuring intrinsic fluctuations an external force is applied to the system. The magnitude of this force is chosen such that the effects are much easier to detect than the internal fluctuations [@zhao:08:00; @hess:02:0; @Gregori:12:0; @Gaskell:74:00].
There are many works studying shear viscosity using MD simulations, both EMD and NEMD methods [@zhao:08:00; @Gaskell:74:00; @palmer:94:0; @plathe:02:00; @sunda:13:00; @Gezeltr:10:0; @Rowley:07:0; @Sun:08:0; @hess:02:0]. Here we follow the work by B. Hess [@hess:02:0] which studies shear viscosity determination using GROMACS. We obtain shear viscosities of PHS model from two methods: TCAF and PP. We explain both methods shortly below.
### Transverse-current autocorrelation function
Consider an incompressible liquid with an initial velocity field generated for example by equilibrium thermal fluctuations. The velocity field can be decomposed into plane waves of the form ${\bold u} (x,0,0)=u_0 \cos(kz)$. The solution to the Navier-Stokes equation for these components is then of the form $$u_x (z,t) = u_0 e^{-t/\tau_r} \cos (kz); \hspace{2cm} \tau_r=\frac{\rho}{\eta k^2}.
\label{NS}$$ The solution indicates that the plane waves decay exponentially with a time constant which is inversely proportional to the viscosity $\eta$. However, at microscopic level and short time scales the behavior is not purely exponential. To account for this, a phenomenological correction can be applied by incorporating a relaxation time, leading to different solution to Eq. \[NS\] (for details see ref. [@hess:02:0]); $$u_x (z,t) = u_0 e^{-t/(2\tau_m)} \big( \cosh (\Omega t/(2\tau_m))+ \frac{1}{\Omega} \sinh (\Omega t/(2\tau_m)) \big) \cos (kz),
\label{NS_2}$$ where $$\Omega = \sqrt{1-4 \tau_m \frac{\eta}{\rho} k^2}.$$ For large $k$ equation (\[NS\_2\]) leads to a similar solution as equation (\[NS\]). The viscosity from this method is given by [@palmer:94:0] $$\eta (k)= \eta(0)(1-ak^2)+O(k^4).
\label{eta_TCAF}$$
### The periodic perturbation method
As mentioned earlier, in PP method an external force is applied to the system. The external field leads to development of a velocity field $\bold {u}$ throughout the system according to the Navier-Stokes equation. For a force only in the $x$ direction, the applied acceleration $a_x$ for a periodic system is given as $$a_x(z)= \alpha \cos(kz); \hspace{2cm} k=\frac{2 \pi}{l_z},$$ where, $l_z$ is the height of the box and $\alpha$ is the amplitude of the acceleration. With the initial value of $u_z(x)=0$, the resulting velocity profile has amplitude $$\nu (1-e^{-t/\tau_r}),$$ where $$\nu=\alpha \frac{\rho}{\eta k^2}.
\label{nu}$$ Thus, at each time step the average velocity can be measured and viscosity can be obtained from Eq. \[nu\]. In order to obtain accurate results efficiently, the parameters of the periodic external force must be chosen carefully. If the velocity profile does not have large fluctuations, less statistics needs to be collected, and thus the simulation time is shorter. This can be achieved with a large amplitude. However, the shear rate should also not be so high that the system moves too far from equilibrium. For more details about the method and estimation of parameters see ref. [@hess:02:0].
Shear viscosity from simulations
================================
We first discuss the TCAF method. We take the final configurations of the NPT simulations explained in previous section and simulate the system in NVT ensemble for $t^*= 3000$ with $\delta t^*= 0.0011$ time step. In the TCAF method the correct values of viscosity are obtained when the velocity profile is not coupled to a heat bath. Therefore, a Berendsen thermostat with a long coupling time of $t^*=11$ is used in order to minimize the influence of the thermostat, see details in ref. [@hess:02:0]. The velocity and coordinates are stored every $\delta t^*= 0.011$. The neighbor list was updated every $\delta t^*= 0.005$. The resulting k-dependent viscosities from the simulations are fitted to Eq. \[eta\_TCAF\] to obtain the viscosities. Fig. \[tcaf\_example\] shows an example of simulation result at density $\rho^*=0.8$. The resulting viscosity from the fit gives $\eta^*=2.11 \pm .043 $ in reduced unit, $\eta^*=\eta (\frac{\sigma ^2}{\sqrt {mk_B T} })$. The obtained values of $\eta^* $ from this method are given in Table. \[vis\_PHS\] for various pressures/densities (the last two columns), and are shown in Figure. \[vis\_TCAF\] as crosses. [ The errors in the given viscosities from this method are obtained from block averaging over k-dependent viscosities plus uncertainty of the fittings to Eq. \[eta\_TCAF\]. ]{}
![[ Shear viscosity of PHS obtained from simulation at $\rho^* =0.80$ (data). The solid line shows the fit to Eq. \[eta\_TCAF\] (TCAF method). $k^* =k\sigma$. ]{}[]{data-label="tcaf_example"}](kvis.eps)
For the PP method, we simulate the system in NVT ensemble for $t^*= 3000$ with an addition of an externally imposed acceleration [@hess:02:0]. The coupling time of the Berendsen thermostat is set to $t^*= 2.84$ in oder to remove the energy introduced in the system by the external force more rapidly. The optimal acceleration amplitude was estimated (see Eqs. 22, 26, 27 in ref. [@hess:02:0]) to be around $ \alpha^*= 0.054$, which is big enough to develop sufficient shear, but also not so big that the system moves too far from equilibrium. [ One disadvantage of this method is that the obtained viscosities depend on the chosen amplitude $\alpha$ [@hess:02:0], moreover, the chosen amplitude should be changed with the density in the systems. ]{} We have started the analysis at time $300$ reduced units after the start of applying the force, so that there is enough time to develop a steady shear amplitude. The obtained values of $\eta^* $ from this method are given in Table. \[vis\_PHS\] (the fifth and sixth columns). The errors in the viscosities are obtained from block averaging. The results are included in Figure. \[vis\_TCAF\], along with the results of the TCAF methods and the MD results of the true HS model reported by Sigurgeirsson [*[et al.]{}*]{} [@heyes:09:0]. We give the simulation results for $\rho^*>0.1$, because when the system is dilute the mean-free-path of the particles becomes large and the simulation box should be large enough to make the collisions to occur enough. That is computationally expensive.\
The results from to the TCAF method show better agreement than the PP method. The reason is that the chosen amplitude $\alpha$ has effect on the viscosity, as studied comprehensively in ref. [@hess:02:0] and in order to obtain more accurate estimations from this method one should try several amplitudes for each density, which is time-consuming.
![ (Left) Shear viscosity of PHS model from TCAF method (crosses) and from PP method (squares). The red circles are results of Sigurgeirsson [*[et al.]{}*]{} [@heyes:09:0] and the solid line is the Enskog expression, Eq. \[Enskog\]. [ (Right) The same as the left plot with the Enskog equation as a base function. ]{} []{data-label="vis_TCAF"}](AR_both.eps "fig:") ![ (Left) Shear viscosity of PHS model from TCAF method (crosses) and from PP method (squares). The red circles are results of Sigurgeirsson [*[et al.]{}*]{} [@heyes:09:0] and the solid line is the Enskog expression, Eq. \[Enskog\]. [ (Right) The same as the left plot with the Enskog equation as a base function. ]{} []{data-label="vis_TCAF"}](AR_both2.eps "fig:")
Comparison with theory
======================
The Enskog expression for the viscosity of a fluid of hard spheres is [@Chapman:52:0; @Santos:16:0; @Viswanath:07:0; @Sengers:00:0; @Lucas:79:0] $$\eta= \eta_0 \bigg [ g^{-1}(\sigma)+0.8 ~V_{excl}~ \rho + 0.776~ V^2_{excl} ~\rho^2 ~g(\sigma)\bigg],
\label{Enskog}$$ where $$\hspace{2cm} \eta_0=\frac{5}{16\sigma^2}\sqrt{\frac{mk_B T}{\pi}},
\label{ensk}$$ and $V_{excl} $ is the excluded volume of HS, $V_{excl}=\frac{2\pi\sigma^3}{3}$, and $g(\sigma)$ is the radial distribution function (rdf) at contact. $\eta_0$ is the zero-density viscosity. The rdf at contact can be obtained directly from the Carnahan-Starling equation of state, which yields, $$g(\sigma) = \frac{1-\xi/2}{(1-\xi)^3},
\label{CS}$$ where $ \xi=\frac{\pi \rho }{6}$ is the volume fraction. It can also be found from the Percus-Yevick approximation, [@yelash:01:00; @yevick:58:00]. $$g(\sigma) = \frac{1+\xi/2}{(1-\xi)^2}.
\label{PY}$$ The latter is more accurate at high density metastable fluid region [@heyes:09:0].
In Figure. \[vis\_TCAF\] we compare the simulation results to the Enskog theory and to the previous MD simulations of true HS simulations by Sigurgeirsson [*[et al.]{}*]{} [@heyes:09:0]. The figures present a qualitative agreement with both. Similar to the ref. [@heyes:09:0] the Enskog theory produces good agreements only for low- to mid-density ranges and fails at high densities, since it does not take into account correlated collisions. That is the reason for deviations at high densities in Fig. \[vis\_TCAF\] [@Santos:16:0; @Viswanath:07:0; @Sengers:00:0; @Lucas:79:0].
![Radial distribution functions of PHS obtained from the simulations at several densities. []{data-label="rdf_ar"}](RDF_AR.eps)
![Radial distribution functions of PHS at contact obtained from the simulations (diamonds) along with the theoretical expressions of Carnahan-Starling (dot-dashed line) and Percus-Yevick (solid line). The circles are MD simulations by Jover [*[et al.]{}*]{}.[]{data-label="rdf_contact"}](RDF_contact.eps)
We also compare the radial distribution function at contact obtained from simulations (only for several densities) to the theoretical prediction from using the Carnahan-Starling equation of state and the Percus-Yevick approximation. In order to obtain the contact values of the radial distribution functions from simulations, we first obtain the rdf profiles for each density, see Figure. \[rdf\_ar\] (the results are taken from the simulations explained in Section. \[sec3\]). The contact values then are collected at $r^*=1$. The values of the rdf at contacts are given in black diamonds in Fig. \[rdf\_contact\]. Simulation results from ref. [@jover:12:0] are also included as circles. Eqs. \[CS\]-\[PY\] are represented in this figure by dot-dashed and solid lines, respectively. As seen in the figure, the simulation results agree well with the Percus-Yevick expression and with results produced by Jover [*[et al.]{}*]{} [@jover:12:0]. As expected Carnahan-Starling’s expression deviates at higher densities, because it does not accurately capture higher virial coefficients.
Conclusions
===========
We have tested the reliability of the pseudo hard sphere (PHS) potential for calculating the shear viscosity of hard spheres. We have used molecular-dynamics simulations (with GROMACS) for this purpose and obtained shear viscosities of PHS from two different methods; transverse-current autocorrelation function and the periodic perturbation method.
We have run simulations under the same conditions that have previously been confirmed to represents the hard-sphere equilibrium properties as demonstrated by Jover [*[et al.]{}*]{} [@jover:12:0]. We have compared the shear viscosities from simulations to the Enskog theory and to available MD simulation data [@heyes:09:0]. Similar to [@heyes:09:0], the comparison with Enskog equation shows a good agreement for densities up to $\rho\sigma^3\le 0.65$. In addition, contact values of the radial distribution functions agree with both Carnahan-Starling (at low densities) and Percus-Yevick expression (up to $\rho\sigma^3\approx0.85$).
The validation of shear viscosity of PHS model helps to use it directly into the state of art simulations packages like LAMMPS, GROMACS or NAMD (which provide high-speed simulations) in order to study complex and large systems. This in turn greatly simplifies its use to support development of analytical theory for similar models based on hard spheres, such as dipolar hard spheres, which will aid the development of transport theory as well as empirical approaches to more complex liquids.
Acknowledgments
===============
The work has been supported by National Infrastructure for Computational Science in Norway (UNINETT Sigma2) with computer timed for the Center for High Performance Computing (NN9573K and NN9572K). The authors acknowledge The Research Council of Norway for NFR project number 275507 and The Faculty of Engineering, Norwegian University of Science and Technology (NTNU), for financial support. FP acknowledges Prof. Erich M[ü]{}ller for providing data for radial distribution of PHS.
[^1]: Corresponding author. Electronic address: `[email protected]`.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'content/abstract.tex'
bibliography:
- 'hashnn.bib'
- 'kilian.bib'
---
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Many biological examples of fluid-structure interaction, including the transit of red blood cells through the narrow slits in the spleen and the intracellular trafficking of vesicles into dendritic spines, involve the near-contact of elastic structures separated by thin layers of fluid. Motivated by such problems, we introduce an immersed boundary method that uses elements of lubrication theory to resolve thin fluid layers between immersed boundaries. We demonstrate 2^nd^-order accurate convergence for simple two-dimensional flows with known exact solutions to showcase the increased accuracy of this method compared to the standard immersed boundary method. Motivated by the phenomenon of wall-induced migration, we apply the lubricated immersed boundary method to simulate an elastic vesicle near a wall in shear flow. We also simulate the dynamics of a vesicle traveling through a narrow channel and observe the ability of the lubricated method to capture the vesicle motion on relatively coarse fluid grids.'
address:
- |
Paulson School of Engineering and Applied Sciences, Harvard University\
29 Oxford St., Cambridge, MA 02138
- |
Mathematics Group, Lawrence Berkeley Laboratory\
1 Cyclotron Rd., Berkeley, CA 94720
author:
- 'Thomas G. Fai'
- 'Chris H. Rycroft'
bibliography:
- 'mybiblio.bib'
title: Lubricated Immersed Boundary Method in Two Dimensions
---
immersed boundary method,lubrication theory,fluid-structure interaction,eccentric rotating cylinders,wall-induced migration
Introduction
============
The immersed boundary method is a widely-used numerical method for fluid-structure interaction that has been applied to problems including blood clotting [@crowl], osmotic swelling due to thermal fluctuations [@wu2015], sperm locomotion through viscoelastic fluids [@sperm], and insect flight [@millercaf]. Several extensions of the immersed boundary method have been developed to incorporate realistic structural properties such as added mass [@kim2007penalty], permeability [@Kim2006; @Kim2012], intrinsic twist [@slim], nonlinear constitutive laws [@griffith2012hybrid], and variable viscosity and density [@vc_ib; @vc_ib2]. The chief idea of the immersed boundary method is to use an Eulerian description of the fluid and a Lagrangian description of the structure, while coupling these descriptions through integral operators involving delta functions. In the continuum formulation, these are the usual Dirac delta functions, and in practice much of the efficiency and accuracy of the immersed boundary method depends on how these Dirac delta functions are regularized [@acta; @moridf].
One of the features of the immersed boundary method is that it is formulated in a manner that avoids the contact problem; since all structures move in the same global velocity field, they cannot in principle cross themselves or one another [@lim2010dynamics]. Though it is useful for many purposes, advecting all structures in the same regularized velocity field presents challenges when multiple immersed boundaries are in near contact, which is typical in biological phenomena such as red blood cell motion through the microcirculation [@Pivkin2016] and membrane receptor trafficking in dendrites [@Kusters2014]. In simulations using fixed fluid grids, the flow in the thin fluid layer between immersed boundaries may not be sufficiently resolved, resulting in boundaries that are effectively stuck together. This is sometimes dealt with by the addition of repulsive forces, as done by Krishnan et al. to prevent “unphysical overlap” of rigid particles [@krishnan2017fully], and by Lim et al. in their simulations of flagellar bundling in *E. Coli* to “prevent filaments from crossing each other” [@lim2012]. We wish to overcome this difficulty by directly using the governing equations instead.
The immersed boundary method takes the viewpoint that the entire domain, including any immersed elastic structures, is filled with fluid having a continuously varying velocity field, without large jumps in the velocities of structures that are nearby relative to the grid spacing $\Delta x$. Given the difficulties observed above, we propose a modification to the immersed boundary method that makes use of lubrication theory to resolve the flow in thin layers. Unlike the standard immersed boundary method, which becomes less accurate as two boundaries approach each other, the lubrication approximation actually *improves* as the gap size decreases. The grid spacing $\Delta x$ provides a natural cutoff at which to apply the lubrication approximation. In the proposed method, the lubrication approximation is automatically turned off when the gap size becomes larger than the grid spacing, as described in detail later on.
We illustrate this method by applying it to two test problems of increasing complexity for which the lubricated immersed boundary method has significantly smaller errors than the standard method. The primary additional cost of the lubricated immersed boundary method involves the construction of a height function that gives the distance between the immersed boundaries. We describe a method to compute this height function accurately and efficiently. Finally, we apply the lubricated immersed boundary method to study the motion of elastic vesicles near walls and observe the significantly increased accuracy of simulations of wall-induced migration and channel flow on coarse fluid grids.
Our work builds on several previous efforts to merge direct numerical simulation with lubrication theory. In the context of particle suspensions, Nguyen and Ladd and Janoschek et al. worked out lubrication corrections to lattice-Boltzmann simulations of particle suspensions [@nguyen2002lubrication; @janoschek2013accurate], while Seto et al. performed simulations of hard spheres with regularized lubrication forces [@seto2013discontinuous]. Thomas et al. performed a multiscale computation of a fluid drop interacting with a wall in which lubrication theory was used to replace the no-slip boundary condition by a specified wall-shear stress [@thomas2010multiscale]. Our approach is most closely related to this last work, since we also combine a grid-based fluid solver with analytical results from lubrication theory to model subgrid effects. Moreover, there are common features in terms of the mathematical formulation; we also use a height function to describe the size of the lubricating layer between boundaries, and relate the flow in the gap to the stress on the boundaries. However, whereas Thomas et al. solve an evolution equation for the height and flux through the gap, the solution to which gives a boundary condition for the fluid, our boundaries are defined in a Lagrangian manner with the height function constructed based on an explicit representation of the boundaries as piecewise polynomial curves. In our approach, we use lubrication theory to modify the advection velocity. Further, whereas Thomas et al. considered a multiphase fluid in which a drop of fluid having no surface elasticity and with ten times greater density and viscosity than the ambient fluid falls down a slope, in the present study we restrict attention to elastic structures immersed in fluids with constant density and viscosity.
We note that an alternative approach for ensuring that the fluid dynamics is sufficiently resolved throughout the domain is to use adaptive mesh refinement (AMR). The idea of AMR is to use a hierarchy of fluid grids with different levels of refinement, with the finer grids placed in specific regions of space to obtain a desired accuracy [@berger1984adaptive]. This allows for highly-resolved computations at drastically reduced computational cost compared to uniform grids. However, implementing AMR requires additional data structures to keep track of the grid hierarchy, and typical implementations of AMR require the user to specify a finest grid level so that issues with near-contact are not definitively resolved. This finest grid size can also lead to severe timestep restrictions for numerical stability. The advantage of using a subgrid model, as we do here through lubrication theory, is that the asymptotic results hold for infinitesimally small gaps, and does not introduce additional timestep constraints. Further, relatively little additional machinery is required beyond what is already needed for uniform grid simulations; as we will discuss later on, the main extra cost of the lubricated immersed boundary method is the construction of the height function that gives the distance between the immersed boundaries.
The movies referenced throughout the manuscript are available at author TGF’s website at <http://scholar.harvard.edu/tfai/home>.
Formulation {#sec:form}
===========
First, we review the standard immersed boundary method, so that the new features of the lubricated method will be clear when they are introduced later on.
Classical Immersed Boundary Method {#sec:form_fl}
----------------------------------
Consider a domain $\Omega$ filled with incompressible fluid of density $\rho$ and viscosity $\mu$, having velocity ${\mathbf{u}}({\mathbf{x}},t)$ and pressure $p({\mathbf{x}},t)$ defined in terms of cartesian coordinates ${\mathbf{x}}$. For many relevant problems, the nonlinear term in the Navier-Stokes equations may be neglected, and we will restrict attention to this case in which the fluid is described by the unsteady Stokes equations: $$\begin{aligned}
\rho\frac{\partial{\mathbf{u}}}{\partial t}+\nabla p &= \mu \Delta {\mathbf{u}}+{\mathbf{f}}\\
\nabla \cdot {\mathbf{u}} &= 0,\end{aligned}$$ where ${\mathbf{f}}({\mathbf{x}},t)$ is the applied force density on the fluid.We are interested in the case in which an immersed elastic body exerts forces on the fluid and thereby determines ${\mathbf{f}}$.
The elastic body is described using Lagrangian coordinates ${\mathbf{q}}$, e.g. arclength in the undeformed frame for a one-dimensional immersed boundary, and the function ${\mathbf{X}}({\mathbf{q}},t)$ gives the cartesian position at time $t$ of the material point labeled by ${\mathbf{q}}$ along the boundary. We assume that the Lagrangian force density ${\mathbf{F}}({\mathbf{q}},t)$ is calculated in terms of ${\mathbf{X}}({\mathbf{q}},t)$; for instance, if there exists an energy functional $E = E[{\mathbf{X}}(\cdot,t)]$, ${\mathbf{F}}$ may be calculated via ${\mathbf{F}}=-\varrho E/\varrho {\mathbf{X}}$, where $\varrho /\varrho {\mathbf{X}}$ denotes the first variation or Fréchet derivative. The Dirac delta function is used to convert the Eulerian force density from Lagrangian to cartesian coordinates via $$\label{eqn:spread}
{\mathbf{f}}({\mathbf{x}},t)=\int_{\mathbf{q}} {\mathbf{F}}({\mathbf{q}},t)\delta({\mathbf{x}}-{\mathbf{X}}({\mathbf{q}},t)){\mathrm{d}}{\mathbf{q}}.$$ The immersed boundary moves at a velocity ${\mathbf{U}}^\text{IB}({\mathbf{q}},t)$ equal to the local fluid velocity. This condition can be expressed in terms of the Dirac delta function as follows: $$\label{eqn:interp}
{\mathbf{U}}^\text{IB}({\mathbf{q}},t)=\int_\Omega {\mathbf{u}}({\mathbf{x}},t)\delta({\mathbf{x}}-{\mathbf{X}}({\mathbf{q}},t)){\mathrm{d}}{\mathbf{x}}.$$ As noted in [@acta], the above equation can be simplified using the definition of the Dirac delta function to obtain ${\mathbf{U}}^\text{IB}({\mathbf{q}},t)={\mathbf{u}}({\mathbf{X}}({\mathbf{q}},t),t)$, but the expression is particularly convenient since this integral form is the basis of the numerical method used, in which the Dirac delta functions are replaced by regularized delta functions and ${\mathbf{U}}^\text{IB}({\mathbf{q}},t)$ is evaluated by summation over the fluid grid. Regularization gives the immersed boundaries an effective width or blurriness on the order of the grid spacing $\Delta x$.
Lubrication Corrections {#sec:form_lube}
-----------------------
Now, consider two immersed boundaries separated by a thin layer of fluid with height $h \ll 1$. Note that throughout this article $h$ will be used to represent height and should *not* be assumed equal to the grid spacing $\Delta x$ (or $\Delta y$). In the limit $h \to 0$, the two immersed boundaries coincide and move at identical velocities given by . By the same logic, we expect that when boundary velocities are interpolated using regularized delta functions, boundaries separated by a gap smaller than the grid spacing $\Delta x$ will move at similar velocities and stick together. A thought experiment illustrates the issue: imagine two parallel lines separated by a distance $h$ and pulled in opposite directions. The Lagrangian forces on the lines are equal and opposite, and if the regularization length scale is large compared to the spacing between the lines, the delta function layers of force will be blurred so that they nearly cancel. In the limit $h \to 0$, this leads to a vanishing Eulerian force density in , and therefore vanishing velocity fields ${\mathbf{u}}({\mathbf{x}},t)$ and ${\mathbf{U}}^\text{IB}({\mathbf{q}},t)$. In practice, similar difficulties are encountered whenever $h \lesssim \Delta x$, since in this case the spread forces partially cancel and the thin lubricating layers separating immersed boundaries are not sufficiently resolved. Instead of computing the distinct velocities of the two boundaries, underresolved simulations assign the averaged velocity to both immersed boundaries, and the lines fail to move unless an extremely large force is exerted.
![Schematic of the variables used in the lubrication immersed boundary method.[]{data-label="fig:test1_schem1"}](lube_fig1-eps-converted-to.pdf){width="3"}
Motivated by this observation, we propose an alternative formulation for the advection velocity ${\mathbf{U}}({\mathbf{q}},t)$ with the property ${\mathbf{U}}({\mathbf{q}},t) \neq {\mathbf{U}}^\text{IB}({\mathbf{q}},t)$. We use the average velocity computed by the standard immersed boundary method, but compute the velocity difference in an alternative way. Assume that there are two immersed boundaries close to touching, and let ${\mathbf{U}}_0$ and ${\mathbf{U}}_h$ be the velocities on the lower and upper surfaces, respectively, with corresponding force densities ${\mathbf{F}}_0$ and ${\mathbf{F}}_h$ and positions ${\mathbf{X}}_0$ and ${\mathbf{X}}_h$ satisfying ${\mathbf{X}}_h = {\mathbf{X}}_0+h{\mathbf{N}}_0$, where ${\mathbf{N}}_0$ is the unit normal to the lower surface at position ${\mathbf{X}}_0$ (Figure \[fig:test1\_schem1\]). In our formulation, ${\mathbf{U}}_0$ and ${\mathbf{U}}_h$ are determined by the following equations specifying their sum and difference: $$\begin{aligned}
&{\mathbf{U}}_h+{\mathbf{U}}_0 = {\mathbf{U}}_h^\text{IB}+{\mathbf{U}}_0^\text{IB} \label{eq:lubesum}\\
( &{\mathbf{U}}_h-{\mathbf{U}}_0)\cdot{\mathbf{T}}_0 = \frac{h}{2}\Bigg(\frac{({\mathbf{N}}_0\cdot{\mathbf{N}}_h)^2{\textcolor{black}{(F_h^\parallel/J_h)}}-{\textcolor{black}{(F_0^\parallel/J_0)}}}{\mu}\notag\\
&\qquad\qquad\qquad\qquad\qquad+{\textcolor{black}{\left(({\mathbf{N_0}} \cdot \nabla){\mathbf{u}}\right)}}\Big |_{0^-}\mkern-18mu\cdot{\mathbf{T}}_0+{\textcolor{black}{\left(({\mathbf{N_0}} \cdot \nabla){\mathbf{u}}\right)}}\Big |_{h^+}\mkern-18mu\cdot{\mathbf{T}}_0\Bigg) \label{eq:lubediff}\\
( &{\mathbf{U}}_h-{\mathbf{U}}_0)\cdot{\mathbf{N_0}}= {\textcolor{black}{({\mathbf{T_0}} \cdot \nabla)}}\left(\frac{h^3}{24\mu} {\textcolor{black}{({\mathbf{T_0}} \cdot \nabla)}}\left( {\textcolor{black}{\frac{F_0^\perp}{J_0}}}-{\textcolor{black}{\frac{F_h^\perp}{J_h}}}+p |_{h^+}+p |_{0^-}\right)\right)\notag\\
&\qquad \qquad \qquad -{\textcolor{black}{({\mathbf{T_0}} \cdot \nabla)}}\left(h{\mathbf{T}}_0\cdot\left(\frac{{\mathbf{U}}_h+{\mathbf{U}}_0}{2}\right)\right) +\left({\mathbf{U}}_h\cdot{\mathbf{T}}_0\right) ({\textcolor{black}{({\mathbf{T_0}} \cdot \nabla)}} h), \label{eq:lubediffperp}\end{aligned}$$ where ${\mathbf{T}}$ is the unit tangent vector and $J = \det\left(\partial X_i/ \partial q_j\right)$ is the Jacobian, so that ${\mathbf{F}}/J$ has units of an Eulerian force density. We write $F^\parallel = {\mathbf{T}}\cdot{\mathbf{F}}$ and $F^\perp = {\mathbf{N}}\cdot{\mathbf{F}}$ to denote the tangential and normal components of the force density, respectively.
Together, and – define ${\mathbf{U}}_h$ and ${\mathbf{U}}_0$; to summarize, the standard immersed boundary formulation is used for the average velocity and a sharp formulation is used for the velocity difference. It is reasonable to compute the mean velocity using the standard immersed boundary method as in , since smoothing tends to yield the net motion as in the thought experiment above. The distinction between our approach and the standard immersed boundary method is the formulation for the difference ${\mathbf{U}}_h-{\mathbf{U}}_0$, which has several practical advantages compared to the standard method. Consider , for instance: the Lagrangian forces ${\mathbf{F}}_0$ and ${\mathbf{F}}_h$ are quantities supported on the immersed boundary, and it follows that $F_0^\parallel$ and $F_h^\parallel$ are never smoothed by convolution against a delta function kernel. Moreover, the quantities and are *outer* derivatives, which implies they are typically well-resolved and/or small in magnitude. Therefore, the terms on the right-hand side of are sharply defined and convenient to use, and is similarly formulated in terms of quantities supported on the boundary or outside the lubrication layer. Next, we show how – may be derived from lubrication theory.
Lubrication Equations {#sec:deriv}
---------------------
In the thought experiment presented above involving two parallel lines pulled in opposite directions with $x$-velocities $U_0$ and $U_h$, the exact solution ${\mathbf{u}}=(u,v)$ inside the gap between the lines is linear shear flow, with $u(y) = (U_h-U_0) y/h+U_0$ and $v \equiv 0$. We wish to use this formula for $u(y)$ to obtain the difference $U_h-U_0$ in terms of known quantities, and it will turn out that the derivative $\partial u/\partial y$ normal to the lines is a convenient choice. By the fundamental theorem of calculus, $$U_h-U_0 = \int_0^h \frac{\partial u}{\partial y} {\mathrm{d}}y,$$ and because the derivative $\partial u/\partial y$ is constant for linear shear, it follows that $$\label{eq:diff_ss}
U_h-U_0=h \frac{\partial u}{\partial y} \Bigg |_{y^*},$$ where $y^*$ is any value of $y$ in $(0,h)$. Although the above equation is trivial, as it simply states that the flow in a linear shear is completely determined by its shear rate, surprisingly there is a similar statement for flows through narrow lubrication layers in general.
To prove this more general statement, we first review some classical results from lubrication theory, although we will use these results in a nonstandard manner. In most applications of lubrication theory, velocity boundary conditions are prescribed and the pressure is calculated by solving the Reynolds equation [@Acheson1990], whereas in our case the relative velocity of the boundaries is the unknown.
Assume a flat lower surface aligned with the $x$-direction is separated by height $h$ from a (possibly curved) upper surface with characteristic horizontal length scale $L$, where $h \ll L$. The unsteady Stokes equations can be simplified substantially in this case by neglecting terms of order $h/L$ and smaller to obtain [@Acheson1990] $$\begin{aligned}
\frac{\partial p}{\partial x} &= \mu \frac{\partial^2 u}{\partial y^2} \label{eq:lube1}\\
\frac{\partial p}{\partial y} &=0\\
\partial u/\partial x+\partial v/\partial y &= 0.\end{aligned}$$ Since $p$ is $y$-independent, may be integrated explicitly to yield $$\begin{aligned}
\label{eq:dudy_lube}
\frac{\partial u}{\partial y} = \frac{1}{2\mu}\frac{\partial p}{\partial x} (2y-h) + \frac{U_h-U_0}{h}.\end{aligned}$$ Because the first term of the right-hand side has opposite signs for $y=0$ and $y=h$, we may eliminate the pressure gradient and invert the above equation for $U_h-U_0$ to get $$\begin{aligned}
\label{eq:diff_lu1}
U_h-U_0=\frac{h}{2}\left(\frac{\partial u}{\partial y}\Big|_{0^+}+\frac{\partial u}{\partial y}\Big|_{h^-}\right).\end{aligned}$$ This implies that the observation made in the special case of linear shear that the velocity difference may be computed by knowledge of $\partial u/\partial y$ at the boundaries applies to lubrication flows in general. In fact, the pressure through the gap can also be recovered using $$\begin{aligned}
\label{eq:diff_lu2}
\frac{\partial p}{\partial x} = \frac{\mu}{h}\left(\frac{\partial u}{\partial y}\Big|_{h^-}-\frac{\partial u}{\partial y}\Big|_{0^+}\right).\end{aligned}$$ Knowledge of the pressure gradient and boundary velocities is sufficient to reconstruct the full tangential velocity profile through the gap by the formula $$\begin{aligned}
\label{eq:u_lube}
u(y) = \frac{1}{2\mu}\frac{\partial p}{\partial x}y(y-h)+\frac{U_h-U_0}{h}y+U_0,\end{aligned}$$ which follows from integrating . (The normal velocities may be computed in terms of the pressure gradient and tangential velocities, as shown later on.)
We have now derived how the boundary velocities and flow through the gap may be calculated in terms of normal derivatives on the boundaries. Next, we will show these normal derivatives may themselves be calculated in terms of more convenient quantities.
Jump Conditions {#sec:jump}
---------------
Here, we review the classical jump conditions across an immersed boundary [@lai2001remark; @xu2006systematic], closely following the exposition of [@lecjump]. Let $[\phi]$ denote the jump in a quantity $\phi$ across an immersed boundary, so that e.g. $[\sigma_{ij}]$ for $i,\,j = 1,\,2$ gives the jump in the components of the stress tensor. Since immersed boundaries are considered massless, all stresses must balance, i.e. $$[\sigma_{ij}]N_j+F_i/J=0,$$ where ${\mathbf{F}}$ is the Lagrangian force density, ${\mathbf{N}}$ is the unit normal vector, and $J$ is Jacobian. In a Newtonian fluid the stress tensor is given by $$\sigma_{ij}=-p\delta_{ij}+\mu\left(\partial u_i/\partial x_j+\partial u_j/\partial x_i \right).$$ It follows that $$-[p]N_i+\mu \left[ \partial u_i/\partial x_j \right]N_j +\mu \left[ \partial u_j/\partial x_i \right]N_j +F_j/J=0,$$ which can be simplified [@lecjump] to $$\label{eq:jump}
[p]{\mathbf{N}}-\mu \left[ ({\mathbf{N}} \cdot \nabla) {\mathbf{u}}\right] = {\mathbf{F}}/J.$$ Separating into tangential and normal components, we have $$\begin{aligned}
-\mu \left[ ({\mathbf{N}} \cdot \nabla) {\mathbf{u}}\right] &= \left({\mathbf{F}}-({\mathbf{F}}\cdot{\mathbf{N}}){\mathbf{N}}\right)/J
\label{eq:jump1}\\
[p]&= \left({\mathbf{F}}\cdot{\mathbf{N}}\right)/J, \label{eq:jump2}\end{aligned}$$ where we have used that the normal component of $\left[ ({\mathbf{N}} \cdot \nabla) {\mathbf{u}}\right]$ vanishes [@lecjump]. The first jump condition relates forces on the immersed boundary to the normal derivative in velocity, which by is just the quantity needed to find $U_h-U_0$. Substituting into the expressions and yields $$\label{eq:lubediff_simple}
U_h-U_0 = \frac{h}{2}\left(\frac{({\mathbf{N}}_0\cdot{\mathbf{N}}_h)^2{\textcolor{black}{(F_h^\parallel/J_h)}}-{\textcolor{black}{(F_0^\parallel/J_0)}}}{\mu}+\frac{\partial u}{\partial y}\Bigg |_{0^-}+\frac{\partial u}{\partial y}\Bigg |_{h^+}\right),$$ To this point, we have not considered the normal velocities $V_0$ and $V_h$ on the lower and upper boundaries, respectively. In order to obtain an expression for $V_h-V_0$, we invoke the incompressibility constraint $\partial u/\partial x+\partial v/\partial y=0$ to write $$\begin{aligned}
V_h-V_0 &= \int_0^h \frac{\partial v}{\partial y} {\mathrm{d}}y \notag\\
&= -\int_0^h \frac{\partial u}{\partial x} {\mathrm{d}}y \notag \\
&= -\left(\frac{\partial q}{\partial x}-U_h \frac{\partial h}{\partial x} \right) \label{eq:diff_n1},\end{aligned}$$ where $q = \int_0^h u {\mathrm{d}}y$ is the flow rate and we have used the fundamental theorem of calculus in the first and third equalities. Integrating directly yields $$q = \frac{-h^3}{12 \mu}\frac{\partial p}{\partial x}+h\frac{U_h+U_0}{2},$$ and substituting this expression into results in $$V_h-V_0 = \left(\frac{h^3}{12\mu} \frac{\partial p}{\partial x}\right)_x-\left(h\left(\frac{U_h+U_0}{2}\right)\right)_x+U_h \frac{\partial h}{\partial x}.
\label{eq:u_lube2_int}$$
As was the case for the tangential velocities, we would like to be able to compute $V_h-V_0$ in terms of convenient quantities such as the boundary forces. This is possible using the above formula , which involves tangential velocities already known through , the pressure gradient $\partial p/\partial x$ already known through , and derivatives of these quantities. In practice, to compute the pressure gradient we have found it preferable to use the jump condition , according to which $$\begin{aligned}
p |_{0^+} &= {\textcolor{black}{\frac{F_0^\perp}{J_0}}}+p |_{0^-}\\
p |_{h^-} &= -{\textcolor{black}{\frac{F_h^\perp}{J_h}}}+p |_{h^+},\end{aligned}$$ where $F^\perp = ({\mathbf{F}}\cdot{\mathbf{N}})$. We average $p |_{0^+}$ and $p |_{h^-}$ to get a pressure in the gap and obtain the following formula: $$V_h-V_0 = \left(\frac{h^3}{24\mu} \left( {\textcolor{black}{\frac{F_0^\perp}{J_0}}}-{\textcolor{black}{\frac{F_h^\perp}{J_h}}}+p |_{h^+}+p |_{0^-}\right)_x\right)_x-\left(h\left(\frac{U_h+U_0}{2}\right)\right)_x+U_h \frac{\partial h}{\partial x}.
\label{eq:u_lube2}$$ Thus we have an equation for $V_h-V_0$ that involves boundary forces, horizontal velocities that are themselves computed in terms of boundary forces, and the outer pressures $p |_{0^-}$ and $p |_{h^+}$.
, we replace $x$- and $y$-velocities by velocities tangential and normal to the lower surface, respectively, and $x$-derivatives by tangential derivatives to the lower surface, resulting in the equations and for the lubricated immersed boundary method. The equations analogous to and in the general case are and of Section \[sec:form\_lube\], respectively. The equation analogous to for the tangential velocity through the gap in the case of curved surfaces is $$\begin{aligned}
{\mathbf{u}}({\mathbf{x}})\cdot{\mathbf{T}}_0 = \frac{1}{2\mu} w(w-h) ({\mathbf{T}}_0 \cdot \nabla)p+w\left(\frac{{\mathbf{U}}_h-{\mathbf{U}}_0}{h}\right)\cdot{\mathbf{T}}_0+{\mathbf{U}}_0\cdot{\mathbf{T}}_0, \label{eq:u_lube3}\end{aligned}$$ where $w \in [0,h]$ and where ${\mathbf{x}}={\mathbf{X}}_0+w{\mathbf{N}}_0$ for fixed ${\mathbf{X}}_0$, so that ${\mathbf{x}}$ travels through the gap in the direction normal to the lower curve. Given this explicit representation for the tangential velocity, the normal component of the gap velocity may be computed as before via incompressibility: $$\begin{aligned}
{\mathbf{u}}({\mathbf{x}})\cdot{\mathbf{N}}_0 &= \int_0^{\left({\mathbf{x}}\cdot{\mathbf{N}}_0\right)} \frac{\partial\left({\mathbf{u}}\cdot{\mathbf{N}}_0\right)}{\partial {\mathbf{N}}_0} {\mathrm{d}}w+{\mathbf{U}}_0\cdot{\mathbf{N}}_0 \notag \\
&= - \int_0^{\left({\mathbf{x}}\cdot{\mathbf{N}}_0\right)} \frac{\partial\left({\mathbf{u}}\cdot{\mathbf{T}}_0\right)}{\partial {\mathbf{T}}_0} {\mathrm{d}}w+{\mathbf{U}}_0\cdot{\mathbf{N}}_0, \label{eq:u_lube4}\end{aligned}$$ where the integrand can be computed through .
Numerical Method
================
Spatial discretization
----------------------
We next specialize to the case of a rectangular unit domain $\Omega$ of size $L_x$ by $L_y$ and periodic boundary conditions in both directions. We discretize $\Omega$ by overlaying an $N_x \times N_y$ grid with grid spacing $\Delta x=L_x/N_x$ and $\Delta=L_y/N_y$ in the $x$- and $y$-directions, respectively. The $(i,j)$^th^ grid cell has its center at ${\mathbf{x}}_{i,j} = ((i+1/2)\Delta x, (j+1/2)\Delta y)$ for $i=0,\dots,N_x-1$ and $j=0,\dots,N_y-1$.
We will consider a one-dimensional boundary immersed in $\Omega$ and parameterized by the Lagrangian coordinate $q \in [0, L_q]$. The immersed boundary is discretized using $N_q$ points so that the Lagrangian grid spacing is $\Delta q = L_q/N_q$, and the positions of Lagrangian points in cartesian coordinates are given by ${\mathbf{X}}_k(t) := {\mathbf{X}}(k\Delta q, t)$ for $k=0,\dots,N_q-1$.
The grid velocities ${\mathbf{u}}_{i,j}$ are located at cell centers ${\mathbf{x}}_{i,j}$, as are the external forces. The pressures are located at cell corners, with $p_{i,j}$ defined at the bottom-right corner of the $(i,j)$^th^ grid cell. The unsteady Stokes equations are discretized in space in a centered, 2^nd^-order accurate manner: $$\begin{aligned}
\rho \frac{{\mathrm{d}}{\mathbf{u}}_{i,j}}{{\mathrm{d}}t} + ({\mathbf{G}}p)_{i,j} &= \mu (L{\mathbf{u}})_{i,j} +{\mathbf{f}}_{i,j}\\
\left({\mathbf{D}} \cdot {\mathbf{u}}\right)_{i,j} &= 0,\end{aligned}$$ where $$\begin{aligned}
\
({\mathbf{G}}p)_{i,j} &= {\frac{1}{2}}\parenVectorstack{{(p_{i+1,j}-p_{i,j}+p_{i+1,j+1}-p_{i,j+1})/\Delta x} {(p_{i,j+1}-p_{i,j}+p_{i+1,j+1}-p_{i+1,j})/\Delta y}}, \\
(L{\mathbf{u}})_{i,j} &= \parenVectorstack{{(u_{i-1,j}-2u_{i,j}+u_{i+1,j})/(\Delta x)^2} {(v_{i,j-1}-2v_{i,j}+v_{i,j+1})/(\Delta y)^2}},\\
\left({\mathbf{D}} \cdot {\mathbf{u}}\right)_{i,j} &={\frac{1}{2}}\left(\frac{u_{i,j}-u_{i-1,j}+u_{i,j-1}-u_{i-1,j-1}}{\Delta x}\right.\notag\\
&\qquad \left. +\frac{v_{i,j}-v_{i,j-1}+v_{i-1,j}-v_{i-1,j-1}}{\Delta y}\right),
$$ and where all index arithmetic is to be interpreted modulo $N_x$ or $N_y$. We also have the equation of motion for the immersed boundary points $$\begin{aligned}
\frac{{\mathrm{d}}{\mathbf{X}}_{k}}{{\mathrm{d}}t} = {\mathbf{U}}_k.\end{aligned}$$ To close the system, we must provide discretized equations for the spreading and interpolation operators used to compute ${\mathbf{f}}_{i,j}$ and ${\mathbf{U}}_k$. They are as follows: $$\begin{aligned}
{\mathbf{f}}_{i,j}(t) &= \sum_{k=0}^{N_q-1} {\mathbf{F}}_k(t) \delta_h\left({\mathbf{x}}_{i,j}-{\mathbf{X}}_k(t)\right)\Delta q \label{eq:spread_discrete}\\
{\mathbf{U}}_k(t) &= \sum_{i=0}^{N_x-1}\sum_{j=0}^{N_y-1} {\mathbf{u}}_{i,j}(t) \delta_h\left({\mathbf{x}}_{i,j}-{\mathbf{X}}_k(t)\right)\Delta x\Delta y, \label{eq:interp_discrete}\end{aligned}$$ where we use the standard 4-point immersed boundary delta function $\delta_h({\mathbf{x}}) = \phi(x/\Delta x)\phi(y/\Delta y)$ [@acta], in which $\phi$ is given by $$\begin{aligned}
\phi(r)&=
\begin{cases}
\frac{1}{8}\left( 5+2r-\sqrt{-7-12r-4r^2} \right), & -2 \le r < -1, \\
\frac{1}{8}\left( 3+2r+\sqrt{1-4r-4r^2} \right), & -1 \le r < 0, \\
\frac{1}{8}\left( 3-2r+\sqrt{1+4r-4r^2} \right), & 0 \le r < 1, \\
\frac{1}{8}\left( 5-2r-\sqrt{-7+12r-4r^2} \right), & 1 \le r \le 2, \\
0, & |r| > 2.
\end{cases}
\label{eq:phi}\end{aligned}$$ We assume that the Lagrangian force ${\mathbf{F}}_k(t)$ can be calculated in terms of the positions ${\mathbf{X}}_k(t)$, specific examples of which will be given later on.
Timestepping
------------
Upon discretizing in space, we have a semi-discrete problem with unknowns ${\mathbf{u}}_{i,j}(t)$, $p_{i,j}(t)$, and ${\mathbf{X}}_k(t)$ that depend on a continuous time variable. The problem is fully discretized by introducing a timestep $\Delta t$ and letting ${\mathbf{u}}^n_{i,j} := {\mathbf{u}}_{i,j}(n\Delta t)$, $p^n_{i,j} := p_{i,j}(n\Delta t)$, and ${\mathbf{X}}^n_k := {\mathbf{X}}_k(n\Delta t)$. We use a second-order accurate predictor-corrector timestepping scheme. A preliminary approximation ${\mathbf{X}}^*$ to ${\mathbf{X}}^{n+1}$ is first computed through a first-order accurate forward Euler step: $${\mathbf{X}}^* = {\mathbf{X}}^n+\Delta t {\mathbf{U}}^n,$$ where ${\mathbf{U}}^n$ is computed in terms of the grid velocity ${\mathbf{u}}^n$ via . We calculate ${\mathbf{u}}^{n+1}$ and $p^{n+1/2}$ by solving a time-centered system of equations: $$\begin{aligned}
\rho \frac{{\mathbf{u}}^{n+1}-{\mathbf{u}}^n}{\Delta t} + {\mathbf{G}}p^{n+1/2} &= \frac{\mu}{2} L\left({\mathbf{u}}^n+{\mathbf{u}}^{n+1}\right) +{\frac{1}{2}}\left({\mathbf{f}}^n+{\mathbf{f}}^*\right) \label{eq:us_corr_mom}\\
{\mathbf{D}} \cdot {\mathbf{u}}^{n+1} &= 0, \label{eq:us_corr_div}\end{aligned}$$ where ${\mathbf{f}}^*$ is calculated from ${\mathbf{X}}^*$. Finally, we update the positions in a time-centered manner to obtain ${\mathbf{X}}^{n+1}$: $${\mathbf{X}}^{n+1} = {\mathbf{X}}^n+\frac{\Delta t}{2}\left( {\mathbf{U}}^n+{\mathbf{U}}^*\right),$$ where ${\mathbf{U}}^*$ is computed in terms of ${\mathbf{X}}^*$. Although this scheme involves the first-order accurate preliminary guess ${\mathbf{X}}^*$, the approach is similar to the broader class of Runge-Kutta schemes and it can be shown to be second-order accurate (see for instance the related scheme in [@vc_ib; @vc_ib2]).
To actually solve for the velocity and pressure in , we take the discrete divergence of both sides of the equation and use to eliminate the discretely divergence-free velocity field (we assume the initial velocity satisfies ${\mathbf{D}} \cdot {\mathbf{u}}^0 = 0$). This results in a Poisson problem for the pressure $p^{n+1/2}$, and once this problem has been solved may be rewritten as an explicit equation for ${\mathbf{u}}^{n+1}$. We solve this linear system using geometric multigrid [@brandt1977multi; @briggs2000multigrid].
Computing the height function {#sec:height}
-----------------------------
Given the sensitivity of the underlying equations to the height, e.g. the dependence in of $V_h-V_0$ on $h^3$, the accurate computation of the height function is essential to this method. This additional computational cost compared to the standard immersed boundary method, in which computing the height is not necessary, is the price paid for resolving velocity gradients across thin fluid layers.
![[]{data-label="fig:height1"}](height1-eps-converted-to.pdf){width="3"}
To compute the height $h_i$ above a point on the lower surface with index $i$ and position ${\mathbf{X}}_{0,i}$ (Figure \[fig:height1\]), we approximate the upper surface by a piecewise cubic curve. On an arbitrary section of the upper surface with endpoints ${\mathbf{X}}_{h,j}$ and ${\mathbf{X}}_{h,j+1}$, the curve is given by ${\mathbf{X}}_{h,j}(s) = {\mathbf{a}}_0+{\mathbf{a}}_1 s+{\mathbf{a}}_2 s^2 +{\mathbf{a}}_3 s^3$ for $s \in [0,1]$. The four unknown constants $\left\{{\mathbf{a}}_k \big| k=0,\dots,3\right\}$ are determined by the conditions $$\begin{aligned}
{\mathbf{X}}_h(0) &= {\mathbf{X}}_{h,j} \label{eq:cub1}\\
{\mathbf{X}}_h(1) &= {\mathbf{X}}_{h,j+1}\\
\frac{{\mathrm{d}}{\mathbf{X}}_h}{{\mathrm{d}}s}(0) &= {\mathbf{T}}_{h,j}\\
\frac{{\mathrm{d}}{\mathbf{X}}_h}{{\mathrm{d}}s}(1) &= {\mathbf{T}}_{h,j+1} \label{eq:cub4},\end{aligned}$$ where ${\mathbf{T}}_{h,j}$, the tangent vector to the upper surface at point $j$, is defined as the weighted average of the two edge vectors connected to vertex ${\mathbf{X}}_{h,j}$: $${\mathbf{T}}_{h,j} := \frac{\left({\mathbf{X}}_{h,j+1}-{\mathbf{X}}_{h,j}\right){\| {\mathbf{X}}_{h,j+1}-{\mathbf{X}}_{h,j} \|}+\left({\mathbf{X}}_{h,j}-{\mathbf{X}}_{h,j-1}\right){\| {\mathbf{X}}_{h,j}-{\mathbf{X}}_{h,j-1} \|}}{{\| {\mathbf{X}}_{h,j+1}-{\mathbf{X}}_{h,j} \|}+{\| {\mathbf{X}}_{h,j}-{\mathbf{X}}_{h,j-1} \|}}$$ Solving the system – results in the coefficients $$\begin{aligned}
{\mathbf{a}}_0 &= {\mathbf{X}}_{h,j}\\
{\mathbf{a}}_1 &= {\mathbf{T}}_{h,j}\\
{\mathbf{a}}_2 &= 3\left({\mathbf{X}}_{h,j+1}-{\mathbf{X}}_{h,j}\right)-2{\mathbf{T}}_{h,j}-{\mathbf{T}}_{h,j+1}\\
{\mathbf{a}}_3 &= {\mathbf{T}}_{h,j}+{\mathbf{T}}_{h,j+1}-2\left({\mathbf{X}}_{h,j+1}-{\mathbf{X}}_{h,j}\right).\end{aligned}$$
Defining components of the tangent vector by ${\mathbf{T}} = (T^1,T^2)$, the unit normal ${\mathbf{N}}$ is given by $${\mathbf{N}} = (-T^2,T^1)/{\| {\mathbf{T}} \|}.$$ To calculate the height above ${\mathbf{X}}_{0,i}$, we find the pair $(s,h)$ and the index $j$ such that $${\mathbf{X}}_{0,i}+h {\mathbf{N}}_{0,i} = {\mathbf{X}}_{h,j}(s).
\label{eq:lower_height}$$ On a section $j$ of the piecewise cubic upper curve, we find $s$ by taking the dot product of both sides with ${\mathbf{T}}_{0,i}$ to obtain $$\begin{aligned}
{\mathbf{X}}_{0,i}\cdot {\mathbf{T}}_{0,i} &= {\mathbf{X}}_{h,j}(s)\cdot {\mathbf{T}}_{0,i} \notag \\
&= \left({\mathbf{a}}_0+{\mathbf{a}}_1 s+{\mathbf{a}}_2 s^2 +{\mathbf{a}}_3 s^3\right)\cdot {\mathbf{T}}_{0,i},\end{aligned}$$ and the resulting nonlinear equation for $s$ is solved using Newton’s method. Having found $s$, the height $h$ is given by $$h = \left({\mathbf{X}}_{h,j}(s)-{\mathbf{X}}_{0,i}\right) \cdot {\mathbf{N}}_{0,i}.$$ To take advantage of the quadratic convergence of Newton’s method, it is crucial to have a good initial guess. We use a piecewise linear (rather than piecewise cubic) approximation to the upper surface to obtain the initial guess $s_0$. This has the virtue of being easy to solve: finding $s_0$ such that ${\mathbf{X}}_{0,i}+h_0 {\mathbf{N}}_{0,i} = {\mathbf{X}}_{h,j}+s_0({\mathbf{X}}_{h,j+1}-{\mathbf{X}}_{h,j})$ yields $$s_0 = \frac{\left({\mathbf{X}}_{0,i}-{\mathbf{X}}_{h,j}\right)\cdot{\mathbf{T}}_{0,i}}{\left({\mathbf{X}}_{h,j+1}-{\mathbf{X}}_{h,j}\right)\cdot{\mathbf{T}}_{0,i}}.$$
The above procedure is done for each $j$ corresponding to a section of the upper surface. For some applications, there is a unique section that yields $s \in [0,1]$ and $h > 0$. However, in some cases there are multiple sections with positive heights above ${\mathbf{X}}_{0,i}$ and $s$ in the appropriate range, such as when the upper surface is a closed curve. In that case, the height is defined to be the minimum of the eligible values. We note that a naïve implementation of this algorithm to compute heights would involve a nested loop over $i$ and $j$ and result in an $\mathcal{O}(N^2)$ operation. However, looping over all $j$ can be avoided since the lubricated method is only applicable when two immersed boundaries are separated by a distance comparable to the grid spacing. Only nearby points need to be checked, and this allows for fast methods based on binning points into containers of size on the order of $\Delta x$ by $\Delta y$.
We also need to define heights $h_j$ below each point on the upper surface in order to compute velocities there. The procedure used to do so is similar to the one just described for the lower surface, although since the height is defined relative to the lower surface normal, the equation that needs to be solved is slightly more complicated. We describe this procedure in detail in Appendix \[app:height\].
Convergence Results
===================
Simple shear {#sec:simp_shear}
------------
As a first demonstration of the lubricated immersed boundary method, we simulate the situation described in Section \[sec:form\_lube\] of two parallel lines separated by a distance $h$ and pulled in opposite directions.
\
See Figure \[fig:test1\_schem2\] for an illustration. We implement this scenario by using tether points, i.e. we prescribe the motion of virtual lines ${\mathbf{Z}}_0(t)$ and ${\mathbf{Z}}_h(t)$ and connect immersed boundaries to these lines by stiff springs with spring constant $k_\text{tether}$. Tether points are widely used in immersed boundary simulations to prescribe either rigid boundaries, as in Wiens et al. [@wiens2015simulating], or moving boundaries with prescribed kinematics, as in the simulations of jellyfish in Hamlet et al. [@hamlet2011numerical]. The Lagrangian forces are given by $$\begin{aligned}
{\mathbf{F}}_h &= -k_\text{tether}({\mathbf{X}}_h-{\mathbf{Z}}_h(t)) \\
{\mathbf{F}}_0 &= -k_\text{tether}({\mathbf{X}}_0-{\mathbf{Z}}_0(t)).\end{aligned}$$
In our simulations we use a periodic domain of size $L_x = L_y = 2$, with a prescribed shear rate of $\dot{\gamma} = 0.15$ and a height $h = 1/24$ between the two lines. The fluid has a density of $\rho = 1$ and a dynamic viscosity of $\mu = 0.02$, which together with the characteristic velocity of $0.004$ gives a Reynolds number of $\text{Re} \approx 0.2$. The units are taken to be dimensionless. The timestep is refined according to $\Delta t = 0.01 \Delta x$. Each line is parameterized in the Lagrangian domain $[0,2\pi]$ and discretized using $N_r$ points so that $\Delta q = 2\pi/N_r$. We choose $N_r = 2N_x$ to satisfy the rule of thumb that, to prevent leakage through boundaries, the spacing between points on immersed boundaries should be approximately half the grid spacing [@acta]. Further, we increase the tethering constant together with the grid resolution according to $k_\text{tether} = 12.5\cdot(N_r/16)^2$, so that the lines become rigid in the limit $\Delta x \to 0$. Simulations are run up to a total time of $T = 10$.
\
Given the tether point kinematics prescribed in our case, as $k_\text{tether} \to \infty$ we expect to recover ${\mathbf{X}}_0 = {\mathbf{Z}}_0$ and ${\mathbf{X}}_h = {\mathbf{Z}}_h$ with a linear shear flow throughout the domain. The results obtained using the standard immersed boundary method are illustrated in Figure \[fig:test1\_ib\], with . The standard method yields accurate results provided that $h \gg \Delta y$. However, as would be reasonably expected it . Because of the underresolved relative velocity between the two lines, extreme spring forces are generated that lead to large overshoots in the fluid velocities near boundaries. See Table \[tab:test1\_ib\] for the errors obtained on different grid sizes,
\
The lubricated immersed boundary method, on the other hand, resolves the fluid flow through the thin gap exactly in the special case of linear shear and therefore converges rapidly to the correct velocities. As shown in Figure \[fig:test1\_lube\], the resulting fluid velocities do not exhibit the large overshoots that appear in the standard immersed boundary method. The errors computed by the lubricated immersed boundary method are significantly smaller. For instance, on a coarse $16 \times 16$ grid for which $\Delta y = 3h$, the errors made by the lubricated method are over an order of magnitude smaller (see Table \[tab:test1\_lube\] for $\textrm{L}^1$ and $\textrm{L}^\infty$ errors on different grid sizes and Movies `flat_ib64.gif` and `flat_lube64.gif` for a visual comparison of the two methods).
\
On coarse grids that underresolve the fluid, using the lubricated immersed boundary method results in accuracies that would require a $2-4$ times finer grid if the standard immersed boundary method were used. Since the number of degrees of freedom scales with the inverse grid spacing squared, this implies that the lubricated method can obtain the same accuracy approximately $8-64$ times faster, taking into account both the number of degrees of freedom and the CFL stability restriction $\Delta t \lesssim \Delta x$.
Note that on a $128 \times 128$ grid for which $\Delta y = 3h/8$, the standard immersed boundary method gives nearly the same error as the lubricated method. This implies that when the gap contains more than 2 gridpoints, the linear shear flow is sufficiently resolved and the lubricated method is no longer necessary.
\
Eccentric rotating cylinders
----------------------------
The test problem of two parallel lines is a relatively simple case, of course. The assumptions of lubrication theory are actually valid for linear shear flow independent of the gap size, and the absence of curvature hides approximations made by the lubricated immersed boundary method. Therefore, we next consider the more demanding problem of eccentric rotating cylinders.
This problem is appealing because it is simple to describe and implement but gives rise to non-trivial fluid dynamics. Unlike the case of linear shear flow, the lubrication approximation is not exact in this case. Moreover, interesting phenomena emerge such as lift on the inner cylinder and counter-rotation of the fluid on the wide side of the gap, as we describe next. In discussing the setup of this well-known problem we closely follow the exposition of Acheson [@Acheson1990].
On a domain of size $L_x=L_y=2$, we use inner and outer cylinders of radii $r_1 = 3/4$ and $r_2 = 3/4(1+1/24)$ respectively, so that the nondimensional thickness $\epsilon:=(r_2-r_1)/r_1=1/24$. The outer cylinder is centered at the origin, whereas the inner cylinder is centered at $(x_0,0)$ with $x_0=3/128$ so that the cylinders are not concentric. With this choice of parameters, the dimensionless eccentricity $\lambda:=x_0/(r_1 \epsilon)$ satisfies $\lambda= 3/4$. This value is chosen partly because whenever $\lambda > (\sqrt{13}-3)/2 \approx 0.30$, the exact solution exhibits a counter-rotating vortex on the wide side of the gap [@Acheson1990]. Moreover, there is a lift $F$ on the cylinder equal to $F=12 \pi \mu U \lambda/(\epsilon^2 \sqrt{1-\lambda}(2+\lambda^2))$.
Next, we investigate which of these features can be recovered on coarse grids using the lubricated immersed boundary method. To simulate this problem, we again use tether points with specified kinematics. The outer cylinder is taken to be stationary, while the inner cylinder is tethered to points rotating counterclockwise with prescribed angular velocity $U=8.33 \cdot 10^{-4}$. The simulation is run up to a total time of $T = 100$. We use a fourfold stronger spring constant of $k_\text{tether} = 50\cdot(N_r/16)^2$ compared to the linear shear simulations of Section \[sec:simp\_shear\]. All other parameters are unchanged.
\
On the coarsest $16 \times 16$ grid used, the two cylinders come to within a minimal distance of $\Delta x/16$, and attain a maximal distance of about $\Delta x/2$. Given these subgrid separations, it is no great surprise based on the results of the previous section that using the standard immersed boundary method leads to large overshoots near the boundaries that are significantly improved by the lubricated method (Figures \[fig:test2\_ib\] and \[fig:test2\_lube\]). Further, by computing the tangential velocity through the gap using , we are able on a coarse, $16 \times 16$ grid to detect reverse flow indicating the existence of a counter-rotating vortex, whereas using the standard method we cannot observe this even using a significantly finer, $128 \times 128$ grid. See Figures \[fig:test2\_ib\] and \[fig:test2\_lube\] for a comparison of the two methods on a $32 \times 32$ grid.
The lubricated method accurately resolves the pressure through the gap (Figure \[fig:test2\_lube\]), allowing for the accurate calculation of lift. The lift is calculated by summing the Lagrangian force density over the inner immersed boundary and multiplying by $\Delta q$ to obtain a force. The lift values obtained using the lubricated immersed boundary method on $16^2$, $32^2$, $64^2$, and $128^2$ grids have corresponding relative errors 67%, 27%, 2.1%, and 0.58%, consistent with 2^nd^-order accurate convergence. The lifts calculated using the standard immersed boundary method, on the other hand, actually have the wrong sign for grids smaller than $128^2$, and converge to the correct value so slowly that there is still a $14\%$ error on a $512^2$ grid. These significant errors are not entirely surprising given that the standard method does not accurately capture discontinuities at the fluid-structure interface [@bg].
Interestingly, the advantages of the lubricated method are not made readily apparent by the errors in certain other quantities calculated at several grid sizes (Tables \[tab:test2\_ib\] and \[tab:test2\_lube\]). The standard immersed boundary method is able to separate the cylinders nearly as well as the lubricated method and does not seem to exhibit a significantly underresolved relative velocity between the cylinders. Indeed, the mean shear rate is the only quantity displayed in Tables \[tab:test2\_ib\] and \[tab:test2\_lube\] that results in an error at least an order of magnitude smaller for the lubricated method. However, as in the linear shear test problem, there are large overshoots in the fluid velocities (see Movies `ecc_ib64.gif` and `ecc_lube64.gif`) and other quantities such as the gap pressure are not computed accurately. This shows the importance for assessing accuracy of investigating auxiliary quantities, such as the lift and flow reversal, that are not controlled directly through the tether points.
\
\
\
Applications
============
Wall-induced migration
----------------------
Many fluid-structure interaction problems in biology, including blood flow, intracellular trafficking, and leaf aerodynamics involve the near-contact of multiple objects. Blood is a non-Newtonian fluid that exhibits several counter-intuitive behaviors, and one interesting such phenomenon is the emergence of a cell-free layer at the vessel walls in which no red blood cells are present. The cell-free layer has its origins in wall-induced migration, which causes deformable objects such as liquid droplets and red blood cells to acquire lift and move away from walls (see Figure \[fig\_onewall\] and [@chaffey1965particle; @uijttewaal1993droplet; @cantat1999lift; @coupier2008noninertial]). Because blood in the microcirculation is at low Reynolds numbers ($\text{Re} \approx 10^{-4}-10^{-2}$) so that the Stokes flow approximation is valid, directed motion away from the wall may seem to be at odds with the well-known scallop theorem that asserts no net displacement may be achieved by a time-symmetric motion in Stokes flow [@purcell1977life]. The paradox is resolved in this case since symmetry is broken as the object deforms with the flow [@uijttewaal1993droplet; @cantat1999lift].
\
\
Simulating wall-induced migration by the standard immersed boundary method is challenging because of the issues discussed thusfar in resolving the lubricating layer that separates the deformable object from the wall. However, by applying the lubricated immersed boundary method, motion from the wall can be seen even on coarse grids. We have studied this phenomenon by simulating an elastic ring moving through a channel. The elastic ring is made up of immersed boundary points ${\mathbf{X}}_i$ for $i=0,\,1,\,\dots,\,N_q-1$ arranged in a loop and connected by springs having total energy $$E_\text{spring} = \frac{k_\text{spring}}{2} \sum_{i=0}^{N_q-1} \left(\frac{{\| {\mathbf{X}}_{i+1}-{\mathbf{X}}_{i} \|}}{\Delta q}\right)^2\Delta q,$$ and by torsional springs with total bending energy $$E_\text{bend} = \frac{k_\text{bend}}{2} \sum_{i=0}^{N_q-1} \left(\frac{{\| {\mathbf{X}}_{i-1}-2{\mathbf{X}}_{i}+{\mathbf{X}}_{i+1} \|}}{\Delta q^2} \right)^2\Delta q,$$ where the indices are computed modulo $N_q$. We simulate the motion away from the wall using the standard and lubricated immersed boundary methods and compare the qualitative results (see Figure \[fig\_onewall\]). The elastic ring has radius $R_0 = \pi/4$, $k_\text{spring} = 1$, and $k_\text{bend} = 0.004$, and it is immersed in a periodic domain of unit cell size $L_x = 4 \pi$ by $L_y= 2 \pi$ that is filled with fluid having density $\rho = 1$ and viscosity $\mu = 0.2$. The viscosity is taken to be spatially uniform. This is the appropriate assumption for systems such as lipid-bilayer endosomes and red cell ghosts in which the internal and external fluids share the same material properties.
The channel walls are built by putting down a single line of tether points at $y = 0$. One wall of tether points is sufficient to define the channel because the periodic boundary conditions in $y$ imply that a particle will reach the wall regardless of whether it moves up or down. Flow is established in the channel by applying a body force of the form $(f_x = \mu \sin(y),f_y=0)$. In the absence of an elastic ring, this establishes a unidirectional flow $(u = \sin(y),v=0)$ in the domain. We use a $64 \times 64$ grid (in which case $\Delta x \neq \Delta y$) and timestep $\Delta t = 0.005 \Delta x$, and 32 discrete points around the circumference of the ring. The elastic ring is initialized nearby the wall with its center of mass at $(x,y) = (0, R_0+\pi/192)$. This implies that at the beginning of our simulation the ring is at a distance of $\Delta y/6$ to the wall. The simulations are run up to a time of $T = 28$.
In these simulations of elastic vesicles, we use a modified fluid solver that has been used previously to simulate inkjet printers [@yu2003coupled] and is based on the finite element projection of [@almgren1996numerical]. A marker-and-cell (MAC) fluid velocity is constructed at each time step in which the horizontal and vertical velocities are defined at cell edges and are not co-located.
Figure \[fig\_onewall\] shows that, at a qualitative level, the lubricated immersed boundary method is able to capture the sequence of deformations during wall-induced migration. The elastic ring simulated by the lubricated immersed boundary method deforms and moves away from the wall, whereas the ring simulated by the standard immersed boundary method becomes stuck and cannot travel with the flow (see Movies `onewall_ib64.gif` and `onewall_lube64.gif`).
Channel flow
------------
\
\
\
\
\
Finally, we perform a simulation of an elastic ring flowing through a narrow channel. This example is inspired by the fluid-structure interaction that takes place in the microcirculation when red blood cells squeeze through capillaries as small as a few m in diameter.
The size of the elastic ring in the following simulation is chosen so that it fits just within a channel defined by two walls of tether points. We use an elastic ring of radius $R_0 = 9\pi^2/40$ that is initialized at a distance of $\pi/192$ from each wall. As in the previous section, we use a $64 \times 64$ grid on a domain of size $L_x = 4 \pi$ by $L_y= 2 \pi$ so that, in terms of the grid spacing, the ring is initialized at a minimal distance of $\Delta y/6$ from each wall. The ring is forced in the $+x$-direction by applying a uniform force to each immersed boundary point. Here we discretize the ring with $64$ immersed boundary points around its perimeter and use the timestep $\Delta t = 0.01 \Delta x$. The simulation is run up to $T = 40$. Note that for this problem, we must compute two height functions, one for the distance of the vesicle to each wall. When the lubricated method is used, the ring flows along in the channel, adopting bullet-like shapes reminiscent of red blood cells in capillaries. This is in contrast to results obtained by the standard immersed boundary method, in which the ring becomes stuck at its starting point (Figure \[fig\_channel\]). The horizontal velocity field computed by the standard method displays grid artifacts in the velocities in the narrow gap separating the elastic ring from the walls, and small velocities throughout the rest of the channel, indicating a clog. The horizontal velocities computed by the lubricated immersed boundary method, on the other hand, do not display these grid artifacts in the lubrication layers, and are over an order of magnitude greater in the rest of the channel (Figure \[fig\_channel\_u\]). See Movies `channel_ib64_p.gif` and `channel_lube64_p.gif`, which show the evolution of the pressure field, and Movies `channel_ib64_u.gif` and `channel_lube64_u.gif`, which show the evolution of the horizontal velocity, for a comparison of the results obtained by the standard and lubricated methods.
Conclusions
===========
We have proposed an immersed boundary method that uses aspects of lubrication theory to resolve thin fluid layers between elastic structures. The two key ingredients of the method are (i) the observation that across a lubrication layer the normal derivatives of velocity at the boundaries determine the relative velocity, and (ii) the well-established jump conditions that relate these normal derivative of velocity to the Lagrangian force on the boundary. We have applied the lubricated immersed boundary method to problems of increasing complexity and have found that the method outperforms the standard immersed boundary method when the lubrication layer is small compared to the fluid grid spacing. The lubricated method makes it possible on coarse fluid grids to capture interesting phenomena, such as the lift and reverse flow in the problem of eccentric rotating cylinders, and wall-induced migration in the problem of an elastic vesicle near a wall in shear flow.
Note that, although in the problems considered here the immersed boundaries were initialized to be close to touching, immersed boundaries that are initially well-separated can also come within subgrid distances over the course of a simulation. This effect is particularly apparent in underresolved 3D computations with significant differences in velocities from gridpoint to gridpoint, as occurs for instance in simulations of red cells moving through capillaries [@vc_ib] and flagellar bundling in E. Coli [@lim2012]. As noted in the Introduction, in this last case initially separated immersed boundaries have even been observed to cross, making it a good candidate to which to apply the lubricated method.
Although we have demonstrated the effectiveness of the present formulation in several settings, much work remains to be done to make the lubricated immersed boundary method readily applicable to real-world problems. The results presented here are restricted to 2D, and although much of the mathematical formulation extends to 3D, it is not clear to what extent the specific numerical approach (e.g. the definition of the height function in terms of piecewise cubic curves) will extend to the challenges of 3D problems. Further investigation may indicate that alternative formulations are preferable to the one presented here; for example, the current formulation is not symmetric with respect to upper and lower surfaces, a property that would be desirable for some applications. In future work, we plan to find appropriate lubrication corrections in these types of situations as well.
Acknowledgments
===============
We thank Charles Peskin for useful conversations . TGF acknowledges funding through National Science Foundation grant DMS-1502851. CHR was supported by the Applied Mathematics Program of the U.S. Department of Energy (DOE) Office of Advanced Scientific Computing Research under contract DE-AC02-05CH11231.
Computing the height on the upper surface {#app:height}
=========================================
Consider a point ${\mathbf{X}}_{h,j}$ on the upper surface (Figure \[fig:height\_upper\]). We approximate the lower surface by a piecewise cubic curve as before. That is, given a section of the lower surface with endpoints ${\mathbf{X}}_{0,i}$ and ${\mathbf{X}}_{0,i+1}$, the curve between these endpoints is ${\mathbf{X}}_{0,i}(s) = {\mathbf{a}}_0+{\mathbf{a}}_1 s+{\mathbf{a}}_2 s^2 +{\mathbf{a}}_3 s^3$ for $s \in [0,1]$, where the coefficients are defined in an analogous manner to –.
![[]{data-label="fig:height_upper"}](height_upper-eps-converted-to.pdf){width="3"}
We wish to find the pair $(s,h)$ and the index $i$ such that $${\mathbf{X}}_{0,i}(s)+h {\mathbf{N}}_{0,i}(s) = {\mathbf{X}}_{h,j}, \label{eq:upper_height}$$ where as before ${\mathbf{N}}_{0,i}(s)$ is defined by ${\mathbf{N}} = (-T^2,T^1)/{\| {\mathbf{T}} \|}$ through the tangent vector, which in this case is given by differentiating the cubic surface, i.e. $${\mathbf{T}}_{0,i}(s) = \frac{{\mathrm{d}}{\mathbf{X}}_{0,i}(s)}{{\mathrm{d}}s} = {\mathbf{a}}_1+2 {\mathbf{a}}_2 s +3 {\mathbf{a}}_3 s^2.$$ Since the normal vector itself depends on the unknown $s$, the formulation is subtly different from the analogous formulation of the height above a point on the lower surface. To solve for $s$, we take the dot product of with ${\mathbf{T}}_{0,i}(s)$ to obtain the 5^th^ order polynomial equation $$\begin{aligned}
&\,{\mathbf{X}}_{0,i}(s)\cdot {\mathbf{T}}_{0,i}(s) = {\mathbf{X}}_{h,i}\cdot {\mathbf{T}}_{h,i}(s) \notag \\
\iff &\left({\mathbf{a}}_0+{\mathbf{a}}_1 s+{\mathbf{a}}_2 s^2 +{\mathbf{a}}_3 s^3-{\mathbf{X}}_{h,j}\right)\cdot \left({\mathbf{a}}_1+2 {\mathbf{a}}_2 s +3 {\mathbf{a}}_3 s^2\right) = 0\end{aligned}$$ and once more use Newton’s method to compute the solution. Given $s$, it is straightforward to solve for the height via $h_j = \left({\mathbf{X}}_{h,j}-{\mathbf{X}}_{0,i}(s)\right) \cdot {\mathbf{N}}_{0,i}(s)$. As before, it is important to have a good initial guess to ensure the efficient convergence of Newton’s method. To obtain this initial guess, we approximate the lower surface by a piecewise linear function so that ${\mathbf{X}}_{0,i}(s) = {\mathbf{X}}_{0,i}+s\left({\mathbf{X}}_{0,i+1}-{\mathbf{X}}_{0,i}\right)$, but we use the continuously varying normal vector $${\mathbf{N}}_{0,i}(s) = {\mathbf{N}}_{0,i}+s\left({\mathbf{N}}_{0,i+1}-{\mathbf{N}}_{0,i}\right).$$ The continuously-varying normal is required to ensure there is a region of finite width around the lower surface above which heights can be computed. Now, we wish to find $s_0$ such that $$\begin{aligned}
{\mathbf{X}}_{h,j}&={\mathbf{X}}_{0,i}(s_0)+h {\mathbf{N}}_{0,i}(s_0) \notag \\ &={\mathbf{X}}_{0,i}+s_0\left({\mathbf{X}}_{0,i+1}-{\mathbf{X}}_{0,i}\right)+h\left({\mathbf{N}}_{0,i}+s_0\left({\mathbf{N}}_{0,i+1}-{\mathbf{N}}_{0,i}\right)\right).
\label{eq:upper_height_guess}\end{aligned}$$ Taking dot products with ${\mathbf{T}}_{0,i}(s_0)$ yields the quadratic equation $$a s^2+b s+c=0,$$ where $$\begin{aligned}
a &=\left({\mathbf{X}}_{0,i+1}-{\mathbf{X}}_{0,i}\right)\cdot\left({\mathbf{T}}_{0,i+1}-{\mathbf{T}}_{0,i}\right)\\
b &=\left({\mathbf{X}}_{0,i+1}-{\mathbf{X}}_{0,i}\right)\cdot{\mathbf{T}}_{0,i}+\left({\mathbf{X}}_{0,i}-{\mathbf{X}}_{h,j}\right)\cdot\left({\mathbf{T}}_{0,i+1}-{\mathbf{T}}_{0,i}\right)\\
c &= \left({\mathbf{X}}_{0,i}-{\mathbf{X}}_{h,j}\right)\cdot{\mathbf{T}}_{0,i}.\end{aligned}$$ Of the two roots $(-b\pm\sqrt{b^2-4ac})/(2a)$, the negative root is the relevant one since it remains finite in the limit of zero curvature in which $a \to 0$. We have found that using the quadratic formula to actually compute the roots can lead to a loss of precision; instead we use the numerically robust method described in [@press2007numerical].
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'A fall is an abnormal activity that occurs rarely; however, missing to identify falls can have serious health and safety implications on an individual. Due to the rarity of occurrence of falls, there may be insufficient or no training data available for them. Therefore, standard supervised machine learning methods may not be directly applied to handle this problem. In this paper, we present a taxonomy for the study of fall detection from the perspective of availability of fall data. The proposed taxonomy is independent of the type of sensors used and specific feature extraction/selection methods. The taxonomy identifies different categories of classification methods for the study of fall detection based on the availability of their data during training the classifiers. Then, we present a comprehensive literature review within those categories and identify the approach of treating a fall as an abnormal activity to be a plausible research direction. We conclude our paper by discussing several open research problems in the field and pointers for future research.'
address: |
David R. Cheriton School of Computer Science, University of Waterloo,\
200 University Ave W, Waterloo, ON N2L 3G1, Canada
author:
- 'Shehroz S. Khan'
- Jesse Hoey
bibliography:
- 'references.bib'
title: 'Review of Fall Detection Techniques: A Data Availability Perspective'
---
Fall Detection ,One-Class Classification ,Outlier Detection ,Anomaly Detection ,Cost-Sensitive Learning 00-01,99-00
Introduction
============
Research in activity recognition has led to the successful realization of intelligent pervasive environments that can provide context, assistance, monitoring and analysis of a subject’s activities that are usually backed up by advanced machine learning and vision algorithms [@Lara2013Survey; @ke2013review; @Incel2013Review]. However, a lot of this research is centred around developing techniques to identify normal Activities of Daily Living (ADL) either at an atomic level (e.g. walking, running, cycling) or at a higher level (e.g. preparing breakfast, washing hands). These techniques are generally applied to monitor a subject’s movements, assess physical fitness and provide feedback. Though this research is useful, there can be scenarios where detection of abnormal activities become important, challenging and relevant. Missing out such abnormal activities can impose health and safety risks on an individual. Falling is one of the most common type of abnormal activity and the most studied [@Mubashir:2013:SFD; @igual2013challenges]. In real life, most falls are caused by a sudden loss of balance due to an unexpected slip or trip, or loss of stability during movements such as turning, bending, or rising [@Robinovitch20041329].
Falls are the major cause of both fatal and non-fatal injury among people and create a hindrance in living independently. According to the report by [SMARTRISK]{} [@website:execsum2009], in Canada in $2004$, falls constituted $25\%$ of all the unintentional injuries besides transport injuries or suicides, resulting in $2225$ deaths, $105,565$ hospitalizations and $883,676$ non-hospitalizations. The report also suggests that falls accounted for $50\%$ of all injuries that resulted in hospitalization, and was the leading cause of permanent partial disability ($47\%$) and total permanent disability ($50\%$). Falls were the leading cause of overall injury costs in Canada in $2004$, accounting for $\$6.2$ billion or $31\%$ of total costs besides other unintentional injuries. According to the WHO report [@website:whofall], the frequency of falls increase with an increase in age and frailty. Older people living in nursing homes fall more often than those living in the community (around $30-50\%$) and $40\%$ of them experience recurrent falls [@website:whofall]. The reason is that most of the older adults living in the nursing homes are more frail and these facilities report fall incidences more accurately [@rubenstein1994falls]. According to the Public Health Agency of Canada [@website:phacfall], older adults in Canada who were hospitalized due to a fall spent up to three weeks in the hospital, which is three times more than the average hospital stay among other age groups. Falls can impact a person both economically and psychologically. Experiencing a fall may lead to a fear of falling [@igual2013challenges], which in turn can result in lack of mobility, social isolation, less productivity and can increase the risk of a fall.
Falls occur infrequently and diversely. The rarity of occurrence of falls lead to a lack of sufficient data for them for training the classifiers [@khan2016classification]. More than one type of fall may also occur and their unexpectedness make it difficult to model them in advance. Collecting fall data can be cumbersome because it may require the person to actually undergo a real fall which may be harmful and unsafe. Alternatively, artificial fall data can be collected in controlled laboratory settings; however, that may not be the true representative of actual falls [@kangas2012comparison]. Analyzing artificially induced fall data can be good from the perspective of understanding and developing insights into falls as an activity but it does not simplify the difficult problem of detecting falls. Moreover, the classification models built with artificial falls are more likely to suffer from over-fitting on them and may poorly generalize on actual falls. The approaches that exclusively collect fall data still suffer from their limited quantity and ethics clearances. The Centers for Disease Control and Prevention, USA [@website:cdc] suggests that on an average, nursing home residents incur $2.6$ falls per person per year. If an experiment is to be set up to collect real falls and assuming an activity is monitored every second by a sensor, then we get around $31.55$ million normal activities per year in comparison to only $2.6$ falls. The data for real falls may be collected by running long-term experiments in nursing homes or private dwelling using wearable sensors and/or video camera. However, the fall data generated from such experiments will still be skewed towards normal activities [@Stone2015Fall] and it is difficult to develop generalizable classifiers to identify falls efficiently. In addition to very few or no labelled data, the diversity and types of falls further make it difficult to model them efficiently. Most of the previous review papers on fall detection assume sufficient data for falls, and survey methods and techniques based on different types of sensors and specific feature extraction/selection methods. We argue that since falls are rare events, standard supervised machine learning methods may not be well-posed to identify them efficiently. Keeping this view in mind, we present a taxonomy for the study of fall detection methods that depends on the availability of fall data present during training the classifiers. This taxonomy is independent of the type of sensors used to capture human activities and specific type of feature extraction/selection methods. The taxonomy envisages the problem of fall detection from a real-world perspective where falls are not abundant but the normal ADL can be easily gleaned.
The rest of the paper is outlined as follows. In Section \[sec:survey\], we survey the existing review papers on fall detection, present their contributions, highlight their limitations and analyze their cumulative outcomes. In Section \[sec:taxonomy\], we present the proposed taxonomy for the study of fall detection methods based on the availability of fall data. Section \[sec:litrev\] presents a comprehensive review of the current and significant research in the field of fall detection using the proposed taxonomy. In Section \[sec:abnormal\], we exclusively review research on methods that treat falls as abnormal activity or can be adapted for this task. We conclude the paper with open research questions and future direction in this field in Section \[sec:conc\].
Survey of Existing Literature Review on Fall Detection {#sec:survey}
======================================================
In the last decade, several review papers on fall detection are published that discuss different aspects of the fall detection problem involving various classification techniques, types of sensors and specific feature engineering methods. In this section, we survey major review papers on fall detection and highlight their focus of research, contributions and limitations.
Noury et al. [@noury2007fall] report a short review on fall detection with an emphasis on the physics behind a fall, methods used to detect a fall and evaluation criteria based on statistical analysis. They discuss several analytical methods to detect falls by incorporating thresholds on the velocity of sensor readings, detecting no-movements, intense inversion of the polarity of the acceleration vector resulting from impact shock and suggest that such methods will result in high false positive rates. They mention, since falls are rare, unsupervised machine learning techniques are likely to fail to identify the first fall event because it was not observed earlier. Supervised algorithms can only classify ‘known classes’ on which they are trained and such techniques may label a rare activity, like a fall, as ‘Others’ along with other activities e.g. to stumble, to slip etc. Yu [@Yu_2008] presents a survey on approaches and principles of fall detection for elderly patients. Yu first identifies the characteristics of falls from sleeping, sitting and standing and categorize fall detection methods based on wearable, computer vision and ambient devices. These approaches were further broken down into specific techniques such as based on motion analysis, posture analysis, proximity analysis, inactivity detection, body shape and $3$D head motion analysis. Yu mentions that a fall is a rare event and it is important to develop techniques to deal with such scenarios. Yu further addresses the need for generic fall detection algorithms and fusion of different sensors such as wearable and vision sensors for providing better fall detection solutions. Perry et al. [@Perry2009Survey] present a survey on real-time fall detection methods based on techniques that measure only the acceleration, techniques that combine acceleration with other methods, and techniques that do not measure acceleration. They conclude that the methods measuring acceleration are good at detecting falls. They also comment that placement of a sensor at the right position on the body can impact the accuracy of fall detection techniques.
Hijaz et al. [@Hijaz2010Survey] present a survey on fall detection and monitoring ADL and categorize them into vision based, ambient-sensor based and kinematic-sensor based approaches. They identify kinematic-sensor based approaches that use accelerometer and/or gyroscopes as the best among them because of its cost effectiveness, portability, robustness and reliability. Mubashir et al. [@Mubashir:2013:SFD] present another survey on fall detection methods with an emphasis on different systems for fall detection and their underlying algorithms. They categorize fall detection approaches into three main categories: wearable device based, ambience device based and vision based. Within each category they review literature on approaches using accelerometer data, posture analysis, audio and video analysis, vibrational data, spatio-temporal analysis, change of shape or posture. They conclude that wearable and ambient devices are cheap and easy to install; however, vision based devices are more robust for detecting falls. Delahoz and Labrador [@Delahoz2014Survey] present a review of the state-of-the-art in fall detection and fall prevention systems along with qualitative comparisons among various studies. They categorize fall detection systems based on wearable devices and external sensors that includes vision based and ambient sensors. They also discuss general aspects of machine learning based fall detection systems such as feature extraction, feature construction and feature selection. They also summarize various classification algorithms such as Decision Trees, Naive Bayes, K-Nearest Neighbour and SVM; compare their time complexities and discuss strategies for model evaluation. They further discuss several design issues for fall detection and prevention systems including obtrusiveness, occlusion, multiple people in a scene, aging, privacy, computational costs, energy consumption, presence of noise, and defining appropriate thresholds. They also present a three-level taxonomy to describe the falling risks factors associated with a fall that includes physical, psychological and environmental factors and review several fall detection method in terms of design issues and other parameters. Schwickert et al. [@Schwickert2013] present a systematic review of fall detection techniques using wearable sensors. One of the major focus of their survey is to determine if the prior studies on fall detection use artificially recorded falls in a laboratory environment or natural falls in real-world circumstances. They observe that around $94\%$ studies use simulated falls. This is an important finding because it highlights the difficulty in obtaining real fall data due to their rarity. They also discuss that accelerometers along with other sensors such as gyroscopes, photo-diodes or barometric pressure sensors can help obtain better accuracy, and the placement of sensors on the body can be of importance in detecting falls.
Zhang et al. [@zhang2015_petra_fall] present a survey of research papers that exclusively use vision sensors, where they introduce several public datasets on fall detection and categorize vision based techniques that uses single or multiple RGB cameras and 3D depth cameras. Pannurat et al. [@pannurat2014automatic] present another review for automatic fall detection by categorizing the existing platforms based on either wearable and ambient devices, and the classification methods are divided into rule-based and machine learning techniques. They present a detailed overview of different aspects of fall detection, including sensor types and placement, subject details, ADLs and fall protocols, extracting features, classification methods, and performance evaluation. They also compare several fall detection products based on size, weight, sensor type, battery, transmission range, features and comment on future trends in the area of fall detection. Igual et al. [@igual2013challenges] review $327$ research papers on fall detection and categorize them as either context-aware systems or wearable devices (including smartphones). The context-aware systems are further categorized as based on cameras, floor sensors, infrared sensors, microphones and pressure sensors. They point out that despite the use of many feature extraction and machine learning techniques adopted by researchers, there is no standardized context-aware technique widely accepted by the research community in this field. The major contributions of their survey are the identification of emerging trends, ensuing challenges and outstanding issues in the field of fall detection. They point out that the limited availability of real life fall data is one of the significant issue which could hinder the system performance. Ward et al. [@ward2012fall] present a review of fall detection methods from the perspective of use and application of technology designed to detect falls and alert for help from end-user and health and social care staff. They categorize the technologies for fall detection based on manually operated devices, body worn automatic alarm systems and devices that detect changes which may increase the risk of falling. They also comment that the users of fall detection technologies are concerned with privacy, lack of human contact, user friendliness and appropriate training, but they identify the importance and benefits of such systems within the community. In their study, health and social care staff appear less informed and less convinced on the benefits of fall detection technologies. There are several other survey papers on fall detection [@el2013fall; @hegde2013technical; @Patsadu2012Survey; @spasova2014survey; @chaudhuri2014fall; @kulkarni2013review] that address similar ideas and issues already covered in this section.
In the past few years, smart phones have becomes very popular as they are non-invasive, easy to carry, work both indoors and outdoors and are equipped with sensors that are useful for activity recognition. There has been a considerable amount of research work done for general activity recognition and fall detection using smartphones. Luque et al. [@luque2014comparison] present a review of comparison and characterization of fall detection systems based on android smart phones. They mention that most of the techniques for fall detection based on smart phones either use machine learning (pattern matching) techniques or fixed threshold(s). They conduct experiments with simulated falls, compare them with several algorithms and observe that the accuracy of the accelerometer based techniques to identify falls depend strongly on the fall pattern. They also find difficulty in setting acceleration thresholds that allow achieving a good trade-off between false alarms and missed alarms. They further mention the hardware limitations of the memory and real-time processing capabilities of the smart phones that may not support complex fall detection algorithms. Another major problem raised is the rate of battery consumption when mobile application for continuous monitoring are used. Casilari et al. [@Casilari_2015] present another survey on the analysis of android based smart phones solutions for fall detection. They systematically classify and compare many algorithms from the literature taking into account different criteria such as the system architecture, sensors used, detection algorithms and the response in case of a false alarms. Their study emphasizes the analysis of the evaluation methods that are employed to assess the effectiveness of the detection process.
Analysis
--------
We observe several recurring themes that consistently appear among all the review papers we discussed previously:
- There exists no standard methodology for fall detection in terms of type of sensors, feature engineering or machine learning techniques that supersedes other methods or perform consistently better than others.
- It is noted in many of the above survey papers that techniques based on fixed thresholds on sensor readings, though simple to implement and computationally inexpensive, are very hard to generalize across different persons and does not provide a good trade-off between false positives and false negatives [@igual2013challenges].
- Many of these survey papers reveal the complete lack of a reference framework, publicly available datasets and almost no access to real fall data to validate and compare to other methods.
- Most of the above discussed survey papers review research on fall detection that assume sufficient data for falls and or adequate prior knowledge and understanding of falls. A fall is a rare event that can occur in diverse ways [@Debard2012Camera]; therefore, collecting sufficient fall data is very difficult. A long term experiment is required to glean real falls; however, such an experiment may only result in very few samples for real falls [@Stone2015Fall; @Debard2012Camera].
- Some of the review papers we discussed above acknowledge the rarity of real falls and the difficulty in generalizing results obtained from artificial or simulated falls [@igual2013challenges; @noury2007fall; @Yu_2008; @Schwickert2013]; however, they did not review techniques that may be capable of identifying falls when their training data is very limited or not present.
- Since most of the above discussed research papers assume sufficient falls collected from laboratories, we could not get useful insights about setting up long term experiments for collecting real fall data.
Taxonomy for the Study of Fall Detection {#sec:taxonomy}
========================================
Based on the inferences drawn from the recent review work on fall detection, we present a taxonomy for the study of fall detection methods that depends on the availability of training data for falls (see Figure \[fig:taxonomy\]). This taxonomy is independent of the type of sensors to capture human motions and specific feature engineering methods employed to tackle this problem. The proposed taxonomy focuses on investigating classification methods based on availability of data for falls. Specific sensors (e.g. wearable, vision, ambient etc) can be used for the task for fall detection. Similarly, specific feature extraction and selection methods can be used based on the type of data captured with the sensors.[^1]. The taxonomy has two high level categories:
1. Sufficient training data for falls
2. Insufficient or no training data for falls
The category \[cat1\] of the taxonomy shows the case when sufficient data for falls is available for training the classifiers. Even though real falls occur rarely, sufficient amount of fall data can be collected using simulated falls. In this category, due to the presence of sufficient falls and normal ADL data, different algorithms based on supervised machine learning, threshold(s) and one-class classifiers (trained only on sufficient fall data) can be employed. This category represents an optimistic view of the difficult problem of fall detection. However, even for these methods, the generalization of fall detection results across different persons is challenging [@medrano2016effect], mostly because of significant variations in the properties between real and simulated falls [@klenk2011comparison]. In a real world scenario, we may expect either too few falls or none to begin with due to their rarity and difficulties in the data collection process. In these highly skewed data scenarios, heuristic and traditional supervised classification algorithms may not work as desired and other classification frameworks based on over/under-sampling, semi-supervised learning, cost-sensitive learning, outlier/anomaly detection and One-Class Classification (OCC) are needed; these techniques are mentioned in category \[cat2\] of the taxonomy. Both the categories \[cat1\] and \[cat2\] assume sufficient training data for normal ADL and only differ in the amount of fall data available during training; however, both of these categories give rise to different types of approaches for fall detection. The approaches in category \[cat1\] attempt to detect a fall directly given their training data, whereas the approaches in category \[cat2\] either manipulate little available fall data or try to indirectly detect a fall as an abnormal activity given (almost) no training data for them.
1. **Sufficient data for falls** \[cat1\]
1. Apply supervised machine learning techniques.\[1a\]
2. Apply threshold based techniques to detect falls from normal activities.\[1b\]
3. Apply OCC techniques to filter out normal activities as outliers.\[1c\]
2. **Insufficient data for falls** \[cat2\]
1. If some fall data is available, apply over/under-sampling techniques.\[2a\]
2. If some fall data is available along with a lot of unlabelled data, apply semi-supervised techniques.\[2b\]
3. If some fall data is available, apply cost sensitive classification algorithms \[2c\]
4. If no fall data is available, apply outlier / anomaly / density based detection techniques.\[2d\]
5. If no fall data is available, apply OCC techniques.\[2e\]
The OCC approaches for fall detection appear in both category (I) and (II); however their underlying principle to detect falls is different. The category (I) assumes sufficient amount of fall data; therefore, an OCC can be trained on only fall data and can be used to identify a test sample as a fall or not-fall (or normal activity in our case, see Taxonomy \[1c\]). Similarly, category (II) assumes insufficient or no training data for falls; therefore an OCC can be trained on the sufficiently available normal data. This classifier can be used to identify a test sample as a normal or not-normal activity (or fall in our case, see Taxonomy \[2e\]).
The techniques mentioned in category \[cat1\] that directly attempt to detect falls are mostly discriminative and may use domain knowledge about the falling event (e.g. sudden change in acceleration or its short duration) [@Debard2012Camera] . There are three major drawback of these approaches (i) they cannot handle a realistic scenario when there is no training data for falls and the classifier will always make mistake in prediction because falls class was never observed, (ii) at the best, they can classify falls as members of some ‘Others’ category; however, there should be some samples available for that category as well, and (iii) a lot of time and effort may be spent on labelling the data.
The techniques mentioned in category \[cat2\] can be discriminative if they use some falls or generative if they detect falls as abnormal activity as an indirect evidence on the occurrence of falls [@Debard2012Camera]. This evidence may include prolonged inactivity, unusual locations, sudden change from normal behaviour and unknown or unseen behaviours. However, a major challenge is such techniques is to identify the first occurrence of a fall. This primarily depends on the definition of “what is a normal behaviour?”, which needs to be defined carefully as it can vary across different persons, specially for different age groups. These techniques only need to learn the normal behaviour; therefore, the inherent data imbalance between normal activities and falls is not an issue because they do not need samples for falls (or their different types) during training of the classifier. However, if the normal behaviour is not properly learned, these systems can result in large amount of false alarms because any slight variation from the normal behaviour would be classified as a fall. Finding an optimal threshold that can minimize both false alarms and missed alarms in these techniques is very challenging [@DBLP:journals/tkde/YinYP08]. It is important to note that every abnormal behaviour or deviation from the normal behaviour does not imply the occurrence of a fall incident. As a result of these problems, such techniques require a lot of training data to effectively capture the normal behaviour over a long duration. The techniques that treat a fall as an abnormal activity does not require labelled data (models are only trained for normal activities) and a lot of time and effort can be saved.
In the next sections, we review the literature based on the taxonomy for fall detection described above. We will not further discuss many supervised methods for fall detection and interested readers may find those references in the survey papers on fall detection discussed in Section \[sec:survey\]. In the literature review, we present algorithms that
- Work with different types of sensors.
- Use a variety of machine learning algorithms, especially those that are known to model temporal and sequential data (both for falls and ADL).
- Can train models using only fall data, and
- May work with a small amount of (or none) training data for falls using sampling techniques, cost-sensitive learning and semi-supervised learning approaches.
Literature Review {#sec:litrev}
=================
Sufficient data for falls
-------------------------
Most of the research work in fall detection deals with applying supervised classification methods. In this paper, we are not reviewing such techniques and interested readers may refer to the papers cited in Section \[sec:survey\]. Several research works in fall detection are based on thresholding techniques [@Bourke200884; @dai2010perfalld; @Jung2015Inertial; @Chen1617246; @Abba1101:Recognition; @Lan:2009:SAF; @DBLP:conf/bsn/LiSHBLZ09; @Jung2015Inertial], wherein raw or processed sensor data is compared against a single or multiple pre-defined thresholds to detect a fall (see Taxonomy \[1b\] in Figure \[fig:taxonomy\]). Most of these techniques need training data for falls and employ either domain knowledge or data analysis techniques to compute threshold(s) for their identification. The problem with thresholding techniques for fall detection is that it is very difficult to adapt thresholds to new types of falls and generalize them across different people [@bagala2012evaluation; @igual2013challenges]. More information on these methods can be found in the survey papers discussed in Section \[sec:survey\].
### One-Class Classification
If sufficient fall data are available, one-class classifiers can be trained on falls to reject normal activities (see Taxonomy \[1c\] in Figure \[fig:taxonomy\]). However, in realistic settings such a strategy is highly unlikely because the availability of sufficient real fall data is difficult . Zhang et al. [@hunag:li:irwin:2006] train OSVM from falls and outliers from non-fall ADL and show that falls can be detected effectively. Yu et al. [@Miao:Naqvi:2011] propose to train Fuzzy OSVM on fall activity captured using video cameras and to tune parameters using fall and some non-fall activities. Their method assigns fuzzy membership to different training samples to reflect their importance during classification and is shown to perform better than OSVM. Yu et al. [@Yu2012One] introduce a video-based fall detection system for elderly people. They extract several video features and apply OCC techniques to determine whether the new instances lie in the ‘fall region’ or outside it to distinguish a fall from other activities such as walking, sitting, standing, crouching or lying. They test four OCC methods; $K$-center, $K$ nearest neighbour, OSVM and single class minimax probability machine (SCMPM) and find that SCMPM achieves the overall best performance among them. Han et al. [@han2014wifall] propose to use wireless signal propagation by employing the time variability and special diversity of Channel State Information as the indicator of human activities. Firstly a local outlier factor algorithm [@breunig2000lof] is used to filter out dynamic activities such as walking, sitting, standing up and falling and then an OSVM is trained on fall activities to distinguish it from other normal activities.
Insufficient data for falls {#sec:insufficient}
---------------------------
### Sampling Techniques
In some situations, there may be few real falls available during training phase (see Taxonomy \[2a\] in Figure \[fig:taxonomy\]). In these cases, either the minority ‘fall’ class be over-sampled or the majority normal activity class be under-samples to train supervised classifiers. Stone and Skubic [@Stone2015Fall] present a two-stage fall detection system. In the first stage, a person’s vertical state is characterized in individual depth image frames followed by an ensemble of decision trees to compute a confidence on the occurrence of a fall. The data they collected has a fall ratio w.r.t. normal activities of $1:400$. They under-sample the normal activities s.t. the ratio is reduced to $1:40$ and create a decision tree ensemble. The activities studied in their analysis are standing, sitting, and lying down positions, near (within $4$ m) versus far fall locations, and occluded versus not occluded fallers and report better results in comparison to the state-of-the-art methods. Debard et al. [@Debard2012Camera] use a weighted SVM to handle the imbalance in the dataset obtained for real world falls and normal activities from camera. The weights were computed using cross-validation and a grid search maximizing the area under the curve of a Receiver Operating Characteristic (ROC) curve. However, oversampling techniques may suffer from over-fitting if too many artificial data points are generated that do not represent actual falls, or under-fitting if the normal class is too much under-sampled to match the size of falls class.
### Semi-Supervised Techniques
Labelling of activities manually can be very cumbersome, exhausting and time-consuming. In most of the cases, there is plenty of unlabelled data along with some labelled data; such a case calls for the implementation of semi-supervised methods for fall detection (see Taxonomy \[2b\] in Figure \[fig:taxonomy\]). Liu et al. [@Liu2008Vision] present a vision based semi-supervised learning mechanism for detecting falls and other ADL to overcome the exhaustive labelling of human activities. They use spatial field constraint energy to assist SVM-based activity decision with a Bayesian inference model, followed by semi-supervised step to retrain the classifier by automatic annotating the activities with the highest confidence. However, their method is sensitive to changes in environment, and needs to be retrained in new situations. Fahmi et al. [@Fahmi2012Semi] present a semi-supervised fall detection method using smartphones by first training a supervised algorithm using decision trees, then using fall profiles to develop a semi-supervised algorithm based on multiple thresholds. Medrano et al. [@Medrano2014Personalizable] present a nearest neighbour based semi-supervised fall detection method for smartphones that can be personalized and updated easily as a new user records new ADL and the system is retrained on the fly. Makantasis et al. [@Makantasis20153D] present a $3$D semi-supervised fall detection system that uses a monocular camera and uses an expert to refine an initially created small subset of labelled activity samples. Yang et al. [@Yang2016] present a semi-supervised approach to detect near-miss falls for ironworkers. They emphasize that collecting and labelling data for near-miss falls is very difficult and use abnormal signals as stand-ins for near misses. They collect data through IMU device and train OSVM on normal walking patterns; any observation that lies beyond the classification boundary is considered as a near-fall.
### Cost Sensitive Classification and Decision Theory {#sec:costsensitive}
Learning from imbalanced data pertains to situations where data from one class is available in abundance, but the data from the other class is rare, difficult to collect or not readily available. The problem of fall detection is a good example for imbalanced learning because the data for normal activities are easy to collect and sufficiently available in comparison to falls. Traditional supervised algorithms expect balanced datasets with equal misclassification costs for different classes. However, these algorithms fail to represent characteristics of the data and provide unfavourable accuracies when presented with imbalanced dataset [@he2009Learning], and their predictions may be dominated by the majority class [@klement2011classifying]. To handle imbalanced classes, cost-sensitive classification is performed [@he2009Learning], where the main idea is to treat different costs in a classification problem differently. This can be done by either presenting different cost matrix to a cost-insensitive classifier or by changing the inner workings of a classification algorithm such that it uses a cost function to build a cost-sensitive classifier. Let us take the case of fall detection, where the dataset is mostly imbalanced in favour of normal activities. Identifying a rare activity (such as a fall) is important from the health and safety perspective, but the cost of a false alarm and missing to report a fall should be very different and must not be treated equally. Therefore, when some data for falls is present for analysis along with abundant data for normal activities, techniques based on cost-sensitive classification can be used for fall detection (see Taxonomy \[2c\] in Figure \[fig:taxonomy\]). However, the incorporation of cost of classification while reporting or not-reporting a fall is absent in most of the studies on fall detection. It is important to note that the cost of errors in this problem should be domain-specific and not vary across different datasets. However, such costs are mostly unknown and hard to compute [@khan2016classification]. Huang et al. [@huang2011enhanced] perform cost-sensitive analysis for fall detection using Bayesian minimum risk and the Neyman-Pearson method. They vary the ratio of the cost of a missed alarm to a false alarm to find an optimal region of operation using the ROC curve. On the contrary, this ratio is generally fixed and must not depend on the dataset. The technique presented in the paper to estimate cost ratio can overfit the dataset without providing any intuitive interpretation about it. However, traditional approaches for cost-sensitive learning for imbalanced datasets may not be directly applicable in the OCC problems. The reason is that in OCC case, the data for one of the class is absent during training; therefore, the probability estimates for the unseen or abnormal class is hard to compute and estimating the costs of errors in such cases are even harder. This problem can also be stated from a decision-theoretic perspective because costs and utilities are related concepts. Decision theory pertains to rational decision making by agents and can be employed to compute the expected cost of (not)reporting a fall. There is very sparse literature on decision-theoretic methods for identification of outliers or rare events/activities. Decision-theoretic approaches have been applied in some tasks, including detecting anomalies in internet path [@fida2011internetpath], intrusion detection [@maloof2006machine], fraud detection [@torgo2011utility] and fault detection in wireless sensor networks [@nandi2014hypothesis]. For the fall detection application, Khan and Hoey [@khan2016dtfall] present a decision-theoretic framework to report unseen falls given an arbitrary amount (possibly zero) of training data for falls, and given little or no information about the costs associated with falls. They present a novel method to parameterize unseen falls to accommodate situations when no fall data is available during training. Their results show that the difference in the cost of between a reported and non-reported fall is not that useful. However, the central question to derive exactly those values of cost remain open.
Table \[tab:study1\] shows the summary of methods discussed in Section \[sec:insufficient\] to detect falls with insufficient data that includes sampling techniques, semi-supervised techniques and cost-sensitive methods. We observe that different machine learning methods are used to handle this problem for both wearable and vision sensors and many of the methods incorporate domain knowledge.
[|l|l|l|]{}**Type of Sensor** & **Authors** & **Technique Used**\
& Fahmi et al. [@Fahmi2012Semi] & Decision Trees, thresholds\
& Medrano et al. [@Medrano2014Personalizable] & Semi-supervised classification\
& Yang et al. [@Yang2016] & OSVM\
& Huang et al. [@huang2011enhanced] & Sensitivity Optimization\
& Khan and Hoey [@khan2016dtfall] & GMM, Decision Theory\
& Stone and Skubic [@Stone2015Fall] & Random Forest, Under-sampling\
& Debard et al. [@Debard2012Camera] & Weighted-SVM\
& Liu et al. [@Liu2008Vision] & SVM, Bayesian Inference\
& Makantasis et al. [@Makantasis20153D] & Visual cues\
Detecting Falls in the Absence of their Training Data {#sec:abnormal}
=====================================================
In a realistic setting, due to the lack of availability of sufficient data for falls and the lack of knowledge and understanding of what those falls might be, approaches based on outlier/anomaly detection (see Taxonomy \[2d\] in Figure \[fig:taxonomy\]) and one-class classification (see Taxonomy \[2e\] in Figure \[fig:taxonomy\]) can be employed for detecting falls. These techniques may not identify falls directly because fall samples are not available to the classifiers during the training phase; however, they can identify falls indirectly as an abnormal (or not-normal) activity. In these approaches, deviations from normal behaviour are flagged as an abnormal activity. Falling is one of the most common abnormal activities; therefore, such techniques can be adapted for fall detection. However, the concept of normal activities must be clearly defined because every abnormal activity may not be a fall. If the normal activities data is not sufficient, then such techniques may produce excessive false alarms. Recent research works [@medrano2014detecting; @micucci2015falls; @khan2014iwaal; @Khan:2012:TDU:2370216.2370444] show that falls can be identified without seeing them in the past or specific domain knowledge about them. Below, we review techniques that build classification models using only the normal activities data and treat fall as an anomaly.
Studies on Detecting Falls as Anomalies
---------------------------------------
Zhou et al.[@Zhou2012An] present a method to detect falls using transitions between the activities as a cue to model falls. They train supervised classification methods using normal activities collected from a mobile device, then extract transitions among these activities and use them to train an OSVM and show that it performs better than an OSVM trained with only normal activities. Yu et al. [@DBLP:journals/titb/YuYRNWC13] present an online OSVM learning algorithm to detect falls captured through a single video source. They extract three types of features: ellipse, shape-structure and position features to build the normal model by an online OSVM which can be updated to new emerging postures. Additional rules were added to the system to report fewer false alarms and to improve fall detection performance. Popescu [@Popescu2009] presents a fall detection technique that uses acoustic signals of normal activities for training and detecting fall sounds from it. They train OSVM, one-class nearest neighbour approach (OCNN) and One-class Gaussian Mixture Model (GMM) (that uses a threshold) to train models on normal acoustic signals and find that OSVM performs the best. However, it is outperformed by its supervised counterpart. Khan et al. [@SalmanKhan2015199] propose an unsupervised acoustic fall detection system with interference suppression that makes use of the features extracted from the normal sound samples, and constructs an OSVM model to distinguish falls from non-falls. They show that in comparison to Popescu [@Popescu2009], their interference suppression technique makes the fall detection system less sensitive to interferences by using only two microphones. Medrano et al. [@medrano2014detecting] propose to identify falls using a smartphone as a novelty from the normal activities and find that OCNN performs better than OSVM but is outperformed by supervised SVM. Micucci et al. [@micucci2015falls] evaluate fall detection methods that do not require fall data during training on different datasets collected using smartphone accelerometers. Their results show that in most of the cases, one-class K-nearest neighbour approach performs better or equivalent to the supervised SVM and KNN classification approaches that require data for both the normal and fall activities. The main contribution of this study is the finding that to design an effective fall detection method, prior understanding of falls patterns is not necessary. Khan et al [@khan2014iwaal] present two ‘X-Factor’ HMM approaches that are like normal HMMs, but have inflated output covariances that can be used as alternative models to estimate the parameters of unseen falls. The inflated covariance parameter can be estimated using a cross-validation on the set of ‘outliers’ in the normal data that serve as proxies for the (unseen) fall data that allows the XHMMs to be learned from only normal activity data. Their results show high detection rates for falls on two activity recognition datasets, albeit with an increase in the number of false alarms. In an extended work, Khan et al. [@khan2015detecting] show empirically that ‘X-Factor’ approach with no fall data can perform better than supervised classification data with few falls. When the number of fall data increases, the supervised classifiers’ performance improve, but collecting sufficient fall data can take lot of time (in years), effort, resources and extensive setup to conduct such experiment. Taghvei and Kosuge [@taghvaei2014hmm] propose a method for a real-time visual state classification of a user with a walking support system. They extract visual features using principal component analysis and use an HMM for identifying real-time falls and state-recognition. In their experiment, an HMM is trained on walking activity and a threshold is separately calculated using a safe distance from the minimum probability distribution value of normal walking, which is adjusted with a constant. If the log-likelihood of a new activity is more than the calculated threshold, then it is identified as either a fall, sit or stand. A multi-class step is performed on top of the one-class classifier to identify one of the states of the HMM that corresponds to different activities including falls. A major drawback of this method is that the threshold is chosen only based on one activity, but sitting and standing are also normal activities. Taghvaei et al. [@Taghvaei2012Comparative] present another fall detection method from a walker that uses features obtained from a depth camera and uses one-class GMM to identify non-walking states and a continuous HMM for identifying different types of falls. The threshold for one-class GMM that represents normal activity is set experimentally. However, it is difficult to generalize this type of threshold across different people. Moreover, the threshold for a one-class GMM is only obtained using walking activity and it classifies sitting and falls as outlier activities. Rougier et al. [@Rougier2011] present a fall detection method using video feeds by tracking the person’s silhouette and performing shape analysis. They use GMM to distinguish falls from the normal data by manually setting a threshold on the log-likelihood. The above survey is presented in tabular form in Table \[tab:study2\].
[|l|l|l|]{}**Type of Sensor** & **Authors** & **Technique Used**\
& Zhou et al.[@Zhou2012An] & OSVM\
& Medrano et al. [@medrano2014detecting] & OCNN, OSVM\
& Micucci et al. [@micucci2015falls] & OCNN\
& Khan et al [@khan2014iwaal; @khan2015detecting] & HMM\
& Yu et al. [@DBLP:journals/titb/YuYRNWC13] & OSVM\
& Taghvei and Kosuge [@taghvaei2014hmm] & HMM\
& Taghvaei et al. [@Taghvaei2012Comparative] & HMM\
& Rougier et al. [@Rougier2011] & GMM\
& Popescu [@Popescu2009] & OSVM, OCNN, GMM\
& Khan et al. [@SalmanKhan2015199] & OSVM\
Studies on Anomaly Detection not Specifically Focused on Fall Detection
-----------------------------------------------------------------------
We discussed in Section \[sec:taxonomy\] that fall detection methods in category \[cat2\] (see Taxonomy \[2d\] and \[2e\] in Figure \[fig:taxonomy\]) do not detect falls directly because either the training data for falls is insufficient or absent during training. In such scenario, a plausible approach is to identify a fall as an abnormal activity (or an anomaly), which we have discussed in the above section. Alternatively, several abnormal activity detection methods that are not specific for fall detection can also be adapted to detect a fall as one of the abnormal activity or an anomaly. Now, we review several abnormal activity recognition techniques that capture data from either vision-based or wearable sensors; only require training data from normal activities to learn the models and are able to identify abnormal activities. These methods can very well be adapted for the task of fall detection.
### Vision based {#sec:vision_based}
Several approaches have been proposed for abnormal activity recognition using the computer vision sensors. Xiang and Gong [@DBLP:conf/iccv/XiangG05a] propose a Dynamic Bayesian Network approach to model each normal video pattern and use a threshold to detect abnormal activity. This approach is simple; however, choosing a threshold remains challenging. Duong et al. [@DBLP:journals/ai/DuongPBV09] model the duration of activities using a Coxian distribution [@DBLP:books/daglib/0018540] and consider a hierarchy of activities to propose a Switching Hidden Semi-Markov Model and show its application to activity segmentation and abnormality detection in smart environments. Zhang et al. [@DBLP:conf/cvpr/ZhangGBM05] propose a semi-supervised adapted HMM framework for audio-visual data streams which comprises of supervised learning of normal data and unsupervised learning of unusual events using Bayesian adaptation. Their method has an iterative structure, where each iteration corresponds to a new detected unusual event. However, it is not clear from their work how many iterations are needed to terminate the process of outlier detection. Their model assumes that the normal activities data contains unusual events and guarantees one outlier per iteration that could result in high false alarm rate. Jiang et al. [@ieee:Jiang:4379786] mention that the HMM trained on small number of samples can overfit and propose a dynamic hierarchical clustering method based on a multi-sample based similarity measure. Their method starts with clustering the data in a few groups; the groups containing large number of samples are treated as normal patterns and HMMs are learned for each of them. This is followed by an iterative procedure of merging similar clusters (re-classifying) and re-training the remaining HMMs until no more merging occurs. An abnormal event is identified if its maximum log-likelihood from all normal events is below a threshold. They show their results on real surveillance video and point out that following the proposed method, the initial training and clustering errors due to overfitting will be sequentially corrected in later steps. Pruteanu-Malinici and Carin [@DBLP:journals/tip/Pruteanu-MaliniciC08] propose infinite HMM modelling to train normal video sequences; unusual events are detected if a low likelihood is observed. Zhang et al. [@DBLP:conf/pakdd/ZhangLGH09] propose an abnormal event detection algorithm from video sequences using a three-phased approach. First, they build a set of weak classifiers using Hierarchical Dirichlet Process Hidden Markov Model (HDP-HMM) and then use ensemble learning to identify abnormal events. Finally, they extract abnormal events from the normal ones in an unsupervised manner to reduce the false positive rates. Hu et al. [@DBLP:conf/ijcai/HuZYZY09] propose a refinement of the HDP-HMM method by incorporating Fisher Kernel into OSVM instead of ensemble learning and using sensor data instead of video data that can be discrete or continuous. The advantage of their method relies on using the HDP-HMM models that can decide on the appropriate number of states of the underlying HMM automatically. Antonakaki et al. [@DBLP:journals/sigpro/AntonakakiKP09] combine the use of HMM and OSVM to detect abnormal human behaviour using multiple cameras. They treat short term behaviour classification and trajectory classification as separate classification problems by providing different set of features to both HMM and OSVM. Two feature vectors are computed per instance to capture short term behaviour and trajectory information and fed to OSVM and HMM. Utilizing these two views of the data, they fuse the output of both the classifiers (logical OR) to identify abnormal activities. Matilainen et al. [@DBLP:conf/acivs/MatilainenBS09] present an unusual activity recognition method in noisy environments that uses a body part segmentation (BPS) algorithm [@DBLP:conf/icmcs/BarnardMH08], which gives an estimation of similarity between the current pose to the poses in the training data. The normal activities they considered are are walking and sitting down and everything else is considered abnormal. The BPS algorithm uses an HMM and a GMM which are trained through synthetic data created by motion capture. They use three sequences containing walking and falling over as the training set to find a statistically optimal threshold for unusual poses. They also propose to use a majority voting over large number of consecutive decisions for the actions that spans over a period of time and to mitigate the effect of single frames with incorrect decisions that helps in reducing false positive rates. However, the approach is based on carefully chosen optimal thresholds and training on a minimalist set of synthetic normal activities, which render this approach not very useful[^2].
Mahajan et al. [@mahajan04] propose an activity recognition framework based on multi-layered finite state machines (FSM) built on top of a low level image processing module for spatio-temporal detection and limited object identification. The FSM learns the model for normal activities over a period of time in an unsupervised manner and can identify deviant activities as abnormal. Parisi and Wermter [@Parisi2013Hierarchical] propose a hierarchical Self Organizing Maps (SOM) based architecture for the detection of novel human behaviour in indoor environments by learning normal activities in an unsupervised manner using SOM and report novel behaviour as abnormal. To handle tracking errors, a first SOM is used to remove outliers from the training motion vectors. The pre-processed motion vectors are encoded by three types of descriptors – trajectories, body features and directions. A separate SOM is then trained for each type of descriptor. If a new observation deviates from the normal behaviour it is flagged as abnormal. Two thresholds are empirically defined to remove outliers from the first SOM and detect abnormal behaviour from the other three SOMs.
### Wearable Sensor Based {#sec:sensor_based}
Several recent works focus on using sensor networks to detect abnormal activities. Yin et al. [@DBLP:journals/tkde/YinYP08] propose a two-stage abnormal activity detection method in which an OSVM is first trained on normal activities and the abnormal activities are filtered out and passed on to a kernel nonlinear regression routine to derive abnormal activity models from a general normal activity model in an unsupervised manner. The method iteratively detects different types of abnormal activities based on a threshold. They claim that this method provides a good trade-off between false alarms and abnormal activity detection without collecting and labelling the abnormal data. The data is collected by using wearable sensors attached to a user and abnormal instances were collected by simulating ‘falls’ and ‘slipping’ in different positions. Quinn et al. [@quinn] present a general framework of Switched Linear Dynamical Systems (SLDS) for condition monitoring of a premature baby receiving intensive care. They introduce the ‘X-factor’ to deal with unmodelled variation from the normal events that may not have been seen previously. The general principle to identify an unusual event is to vary the covariance of the model of normal events to determine the interval with the highest likelihood where events can be classified as ‘not normal’. To model dynamic detection of abnormal events, they add a new factor to the existing SLDS model by inflating the system noise covariance of the normal dynamics. The sensor data is collected using various probes connected to each baby. The computation of the factor related to increasing the covariance remains challenging and is critical in this application. Cook et al. [@cook2012act] present a method for activity discovery (AD) for smart-homes to identify behavioural patterns that do no belong to the pre-defined classes. Their algorithm scans the data and find the patterns that may represent similar activities and their variations. The algorithm then reports the best patterns that were found and the sensor event data can be compressed using the best pattern. The process is repeated several times until no new patterns can be found that compress the data. The search is carried out using Minimum Description Length principle [@Rissanen:1989:SCS:534258] and final set of discovered patterns are clustered using quality threshold clustering method [@heyer1999exploring] in which the final number of clusters are not need to be specified apriori. This procedure can generate ‘Other’ or unknown activity cluster; however, it may contain some already discovered activities. The AD algorithm is performed on the ‘Other’ cluster and already discovered activities can be separated out from it, thereby reducing the ‘Other’ cluster size and the overall recognition of the system reportedly improves. Luca et al. [@luca2014detecting] present a method for detecting rare hypermotor seizures in children by attaching accelerometer on the body. They use all the observed data to learn normal behaviour and use extreme value theory to detect deviations from this model.
The above survey on abnormal activity detection using wearable and vision sensors that can be adapted for fall detection is presented in tabular form in Table \[tab:study3\].
[|l|l|l|]{}**Type of Sensor** & **Authors** & **Technique Used**\
& Yin et al. [@DBLP:journals/tkde/YinYP08] & OSVM, Kernel Nonlinear Regression\
& Quinn et al. [@quinn] & SLDS\
& Cook et al. [@cook2012act] & Clustering\
& Luca et al. [@luca2014detecting] & Extreme Value Theory\
& Xiang and Gong [@DBLP:conf/iccv/XiangG05a] & Dynamic Bayesian Network\
& Duong et al. [@DBLP:journals/ai/DuongPBV09] & Switching HMM\
& Zhang et al. [@DBLP:conf/cvpr/ZhangGBM05] & Semi-supervised HMM\
& Jiang et al. [@ieee:Jiang:4379786] & HMM\
& Pruteanu-Malinici and Carin [@DBLP:journals/tip/Pruteanu-MaliniciC08] & Infinite HMM\
& Zhang et al. [@DBLP:conf/pakdd/ZhangLGH09] & HDP-HMM, Ensemble Learning\
& Hu et al. [@DBLP:conf/ijcai/HuZYZY09] & HDP-HMM, OSVM\
& Antonakaki et al. [@DBLP:journals/sigpro/AntonakakiKP09] & HMM, OSVM\
& Matilainen et al. [@DBLP:conf/acivs/MatilainenBS09] & HMM, GMM\
& Mahajan et al. [@mahajan04] & FSM\
& Parisi and Wermter [@Parisi2013Hierarchical] & SOM\
From Tables \[tab:study2\] and \[tab:study3\], we observe that for detecting a fall as an anomaly or adapting abnormal activity detection methods to identify falls, the most common techniques are based on OSVM, OCNN, HMM, GMM and their variants. Some other techniques are also used such as Extreme Value Theory, Clustering, Dynamic Bayesian Network and SLDS. There is a wide range of one-class classification algorithms [@Khan:KER:2014] that could be potentially used for detecting falls as anomaly.
Future Research Directions and Challenges {#sec:conc}
=========================================
In this paper, we shift our research focus from traditional approaches of fall detection that assume an overtly optimistic assumption that data for falls is sufficiently available. We discussed that a fall is an abnormal activity that occurs rarely; hence, its data is scarce. This idea is also supported by several recent studies and statistics available from nursing homes [@Stone2015Fall; @Debard2012Camera; @website:cdc]. In these situations with highly skewed or no data for falls, machine learning approaches other than traditional supervised classification must be investigated. Keeping this view, we presented a taxonomy for the study of fall detection that is based on the availability of data and reviewed literature that follows the taxonomy. We identify that considering a fall as an abnormal activity is an important research area, because this reflects the real situation of the world where falls happen infrequently. In the paper, we pointed to some recent studies where a few real fall data is collected over a period of a year or more. This type of experimental set up requires a long term facility, numerous computing resources, rigorous ethics clearances, extensive time and money investment, data analysis and data labelling effort after the experiment is finished. Under these circumstances, identifying falls as abnormal activities makes more sense; however, the overall concept of normal activity needs to be defined and a lot more data may be needed in comparison to the typical supervised classification.
It has been shown that personalized fall detection systems work better than a general fall detection system that is trained with data from one (or more) person(s) and be adapted to other persons [@medrano2016effect]. The generalization of results is problematic in the supervised classification of falls and further aggravates when falls are detected as anomalies by learning models using only data from the normal activities. It is very hard to develop personalized fall detection solution for each and every faller; therefore, research is needed to improve the generalization capabilities of fall detection systems across different persons.
There are several open research areas for fall detection techniques that may require few or no training data for falls. Auto-encoders can be trained directly from the sensor or vision data and can be used to identify anomalous activities based on their reconstruction error [@japkowicz1995novelty; @Marchi2015novel]. With Auto-encoders, a reduced set of features can be learned from the raw sensor data that can be used to train other standard one-class classifiers for detecting unseen falls as abnormal activities. Recurrent neural networks (RNN) are well-suited for this problem because the sensor data fror human activities is sequential in nature. RNN have been used for intrusion detection [@anyanwu2010scalable] and we expect further research to investigate their suitability for unseen fall detection to be very useful. Extreme Value Theory (EVT) has been successfully used to identify rare health conditions among patients in an unsupervised manner [@luca2014detecting; @clifton2014extending]. The methods based on EVT are mathematically sound to exclusively model rare events that occur in the tails of a distribution and have the potential to be adapated for the fall detection problem. The decision theoretic methods for fall detection in the absence of training data (e.g. Khan and Hoey [@khan2016dtfall]) assume that agents make rational choices under uncertainty and risks. When humans are involved in decision making under similar circumstances, those decisions are not rational anymore because humans over-estimate the probability of rare events and under-estimate the probability of more-likely events [@kahneman1979prospect]. This is the outcome of Prospect Theory and we believe that research in that direction can help in understanding the value of deployment of fall detection devices and their subsequent evaluation by humans.
A major unresolved question in fall detection methods is the incorporation of costs of errors; however, such costs are not well understood and are hard to compute. This cost also depends on the optimization of the decision threshold to give an optimal choice between false alarms and missed alarms. Since the data for falls is not available; parameter optimization is very challenging in OCC approaches [@liu2016modular]. Some approaches reject noisy data from the target class (or normal activities) and use them as validation set for estimating the parameters of the unseen class; however, they may be sensitive to false alarms [@khan2014iwaal]. Research efforts are required in this direction to find good strategies to optimize parameters in the absence of no or few data from abnormal activities (or falls in this context). The traditional way to reduce false alarms in a fall detection application is the use of domain knowledge and heuristic rules that are static and hard to generalize. Various researchers have identified different fall risks factors, such as variability in voluntary movement paths of older adults [@kearns2012path] as an independent predictor of fall risk. We believe that combining these fall risk scores with the probabilities (or scores) for the unseen falls will be helpful in reducing false alarms. Using the OCC strategies along with combining activity data from multiple data sources such as sensors, video camera, microphone and RFID can result in robust estimation of normal activities, from which falls may be identified effectively as abnormal events.
We also noted that there is a lack of standard framework to evaluate different fall detection methods and for data repositories. Researchers collect their own activity data in specific settings that may be hard to replicate in other settings. It has been found that the performance of classifiers decrease significantly when they are tested on a dataset different from the one used for training [@igual2015comparison]. As discussed earlier, for learning the concept of normal activities, a lot of data is required that should be collected across different persons, age groups and experimental settings. This type of variation may not be present in individual studies and the dataset created from them. An interesting idea would be to combine these publicly available datasets to understand the overall concept of normal activities and perform further research on identifying unseen falls.
Declaration {#declaration .unnumbered}
===========
Conflicts of Interest : None
Funding: None
Ethical Approval : Not required
References {#references .unnumbered}
==========
[^1]: Interested readers can consult the review papers discussed in Section \[sec:survey\] for sensor specific fall detection techniques along with dedicated feature extraction methods.
[^2]: A detailed review on several other types of video-based abnormality detection methods is provided by Popoola and Wang [@6129539]
| {
"pile_set_name": "ArXiv"
} |
---
abstract: |
Using the Eris zoom-in cosmological simulation of assembly of a Milky Way analog, we study the chemical enrichment of stars due to accretion of metal-enriched gas from the interstellar medium during the Galaxy’s development. We consider metal-poor and old stars in the Galactic halo and bulge through the use of stellar orbits, gas density and metallicity distributions in Eris. Assuming spherically symmetric Bondi-Hoyle accretion, we find that halo and bulge stars accrete metals at the rate of about $10^{-24}$ M$_\odot$yr$^{-1}$ and $10^{-22}$ M$_\odot$yr$^{-1}$, respectively, at redshifts $z\lesssim
3$, but this accretion rate increases roughly a hundred-fold to about $10^{-20}$ M$_\odot$yr$^{-1}$ at higher redshifts due to increased gas density. Bulge and halo stars accrete similar amounts of metals at high redshifts when kinematically distinct bulge and halo have not yet developed, and both sets of stars encounter a similar metal distribution in the ISM. Accretion alone can enrich main-sequence stars up to $[\mathrm{Fe}/\mathrm{H}]\sim -2$ in extreme cases, with the median enrichment level due to accretion of about $[\mathrm{Fe}/\mathrm{H}] \sim -6$ to $-5$. Because accretion mostly takes place at high redshifts, it is $\alpha$-enriched to $[\alpha/\mathrm{Fe}]\sim 0.5$. We find that accretive metal enrichment is sufficient to affect the predicted metallicity distribution function of halo stars at $[\mathrm{Fe}/\mathrm{H}]<-5$. This can hinder attempts to infer natal chemical environment of metal-poor stars from their observed enrichment. Peculiar enrichment patterns such as those predicted to arise from pair-instability supernovae could help in disentangling the natal and accreted metal content of stars.
bibliography:
- 'refs.bib'
date: 'Accepted —. Received —; in original form —'
title: 'Chemical enrichment of stars due to accretion from the ISM during the Galaxy’s assembly'
---
=1
\[firstpage\]
accretion, accretion disks – stars: abundances – stars: chemically peculiar – stars: Population III – Galaxy: abundances – Galaxy: stellar content
Introduction
============
Stars are thought to preserve a chemical record of their natal environment in their presently observed metal abundance [@2013pss5.book...55F]. This makes old, low-mass, metal-poor stars in the Galaxy an important probe of star formation at the earliest cosmological times . Stars with masses less than 1 M$_\odot$ have lifetimes longer than 12.7 Gyr . As a result, stars formed in this mass range at redshifts $z>6$ can survive until the present day. These stars probe conditions during the formation of the Galaxy and can likely reveal properties of the earliest, metal-free, so-called Population III stars. Some of today’s low-mass stars could themselves be Population III [@2015MNRAS.453.2771J] as recent simulations of the formation of Population III stars have increasingly favoured the view that these stars could have had masses less than 1 M$_\odot$ [@2011ApJ...727..110C; @2011ApJ...737...75G; @2013ApJ...766..103D; @2014ApJ...781...60H; @2014ApJ...785...73S; @2014ApJ...792...32S; @2015MNRAS.448.1405M; @2016MNRAS.462.1307S].
In the last few decades, significant effort has been made to find metal-poor stars in the Galactic halo [@1981ApJ...248..606B; @1985AJ.....90.2089B; @1991AJ....101.1865R; @1992AJ....103.1987B; @1996AJ....112..668C; @2003RvMA...16..191C; @2007ApJ...670..774N; @2009AJ....137.4377Y; @2011Natur.477...67C; @2014Natur.506..463K] as well as in the bulge [@2016MNRAS.460..884H; @2016PASA...33...22N; @2015Natur.527..484H; @2015ApJ...809..110C; @2015ApJ...807..171J]. These surveys have revealed about 400 stars with iron abundance $[\mathrm{Fe}/\mathrm{H}]<-3$ and eight stars with $[\mathrm{Fe}/\mathrm{H}]<-7.5$ to $-4.5$ , based on high-resolution spectroscopic observations of candidate extremely metal-poor stars. Abundances of a variety of other elements such as lithium, carbon, magnesium, and calcium have also been measured in these stars. None of these stars has the minimum observable metal abundance, which is set by the capabilities of spectrographs and currently corresponds to the metal abundance of a cool red giant with $T_\mathrm{eff}=4500$ K and $\log g=1.5$ with observed Ca II K line strength of 20 mÅ . Thus, no Population III star has yet been discovered.
Still, models of galaxy formation that account for metal production in stars and supernova-driven metal enrichment of the interstellar medium (ISM) can be used to infer constraints on early star formation, including Population III star formation, from the chemical properties of observed metal-poor stars . For example, @2007MNRAS.381..647S derived constraints on the typical mass of Population III stars $m_\mathrm{PopIII}$ and the critical gas metallicity Z$_\mathrm{cr}$ below which Population III stars can form. They argued that non-detection of a metal-free star constrains Z$_\mathrm{cr}>0$ and $m_\mathrm{PopIII}>0.9$ M$_\odot$. In their model, the observed metallicity distribution function (MDF) of metal-poor stars in the Galactic halo was well-fit by Z$_\mathrm{cr}=10^{-4}$ Z$_\odot$ and $m_\mathrm{PopIII}=200$ M$_\odot$. @2012MNRAS.423L..60S showed that the elemental abundance pattern of the metal-poor star SDSS J102915+172927, which has metallicity $Z=4.5\times 10^{-5} Z_\odot$ and mass less than 0.8 M$_\odot$, can be accounted for by the chemical yields of core-collapse supernovae with metal-free progenitor stars having masses 20 and 35 M$_\odot$. By tracking the production of dust in these supernovae, @2012MNRAS.423L..60S argued that the existence of SDSS J102915+172927 implied that dust cooling played an important role in the formation of the first low-mass Population II stars. Recently, @2015MNRAS.447.3892H argued that if no metal-free star is detected in a sample of $2\times 10^7$ halo stars then $m_\mathrm{PopIII}>0.8$ M$_\odot$ at 99% confidence level.
However, all of the above attempts at exploiting the metal abundance properties of metal-poor stars assume that these stars preserve the metal abundance properties of their natal star-forming clouds. This assumption can break down if stars are able to efficiently accrete metals from the ISM as they orbit in the potential well of the Galaxy throughout their lifetime. Due to their large ages, low-mass metal-poor stars can potentially interact with a large column of ISM gas, depending on the stellar orbits. During this time, the ISM itself becomes increasingly enriched due to metal production from other stars, as evidenced by the existence of an age-metallicity relation in high-redshift galaxies . Accretion of metals from this enriched ISM can pollute old metal-poor stars, and can result in metal abundance patterns that can be quite different from the original metal abundance patterns of these stars. As a result, depending on the magnitude of this effect, errors can be introduced in any inference about high-redshift star formation drawn from the chemical properties of old metal-poor stars.
Chemical enrichment of stars due to accretion from the ISM has been studied over the last several decades . Assuming that stars accrete only when they pass through the Galactic disk and assuming a static Galactic disk with height 100 pc and density $n\sim 5$ cm$^{-3}$, @2009MNRAS.392L..50F calculated metal accretion for 474 stars using their kinematic measurements from the Sloan Digital Sky Survey, and concluded that accretion is generally negligible. @2009ApJ...696L..79K [@2010ApJ...717..542K; @2015ApJ...808L..47K] considered metal accretion of metal-poor stars in a semi-analytical galaxy formation model built on halo merger trees obtained using the extended Press-Schechter formalism and concluded that Population III survivors can be enriched up to $[\mathrm{Fe}/\mathrm{H}]\sim -5$ due to accretion from the ISM. @2011MNRAS.413.1184J used a semi-analytical model of galaxy formation built on average halo growth histories [@2010MNRAS.406..896B] and assumptions about typical stellar orbits to consider the accretion of metals on stars and its suppression due to stellar winds. @2015MNRAS.453.2771J performed a similar analysis to study the effect of radiation pressure on metal accretion and argued that radiation pressure selectively inhibits dust accretion. This implies that stars enriched by accretion from the ISM will show selective reduction in refractory elements, which may explain the composition of at least some carbon-enhanced metal-poor stars without neutron-capture element overabundance (CEMP-no stars; , and references therein). By comparing the metallicities of low-mass stars with different spectral types, @2014ApJ...784..153H found possible observational evidence for metal accretion on the halo and thick-disk stars of the Milky Way, and suggested that metal accretion could have occurred for the stars in the satellite galaxies before they merged into the main halo, since the velocity dispersions of these dwarf systems are much smaller than the main progenitor of the Milky Way.
{width="\textwidth"}
Nevertheless, these studies of metal accretion by stars simplify one or both of two important aspects of the problem. First, the dynamical and chemical evolution of the ISM is often simplified, so that the density evolution [@2013ApJ...773...43B] of gas in the Galactic disk and halo as well as its chemical evolution [@2011ApJ...729...16K; @2012ApJ...760...50S; @2013ApJ...772...93K; @2013ApJ...765...89S; @2015ApJ...807..115S] are neglected. In reality, gas density not only evolves on average as $(1+z)^3$, but can also briefly increase accretion rates on stars during episodes of major galaxy mergers when the Galaxy has a disturbed morphology [@2013ApJ...773...43B]. Secondly, stellar orbits are also often simplified. As the Galaxy grows in mass via accretion and mergers, its potential well deepens and virial velocity increases. Major mergers can have a strong effect on stellar orbits, which can increase accretion rates for brief periods of time [@2013ApJ...773...43B]. A large fraction of stars in the bulge and halo can be formed in satellite progenitors rather than the main host halo [@2015ApJ...799..184P]. Also, before the Galaxy develops a distinct halo, stars that are in the halo today can be on radial orbits and can encounter denser-than-average gas, thus increasing the metal accretion rates. Our aim in this work is to achieve a more realistic picture for both of these ingredients of stellar metal-accretion models.
In this paper, we study the accretion of metals on halo and bulge stars using the high-resolution cosmological zoom-in simulation “Eris” [@Guedes11]. Eris creates a Milky Way analog at $z=0$ and represents a possible assembly history of the Galaxy. Due to its cosmological nature, Eris accounts for the cosmological inflow of metal-poor gas as well as a plausible merger history of the Galaxy. Due to its high resolution, Eris is able to evolve star particles that represent the earliest stars on realistic orbits in an evolving galactic potential. Star formation and supernova implementations lead to the chemical evolution of the ISM. These properties of Eris make it possible for us to investigate metal accretion on stars as they interact with an ISM with realistic chemical and dynamical evolution.
We describe the Eris simulation in Section \[sec:erissim\]. Section \[sec:acc\] details our metal-accretion model and our results are presented in Section \[sec:results\]. We end with a summary of our main conclusions in Section \[sec:conc\].
The Eris Simulation {#sec:erissim}
===================
We use the high-resolution, zoom-in cosmological hydrodynamical simulation “Eris” of a Milky Way Galaxy analog to investigate accretion of metals onto old and metal-poor stars. A detailed description of the Eris simulation is provided by @Guedes11. Here we briefly outline aspects relevant to this study. The simulation was performed with the parallel TreeSPH code <span style="font-variant:small-caps;">Gasoline</span> [@Wadsley04] in a WMAP-3 cosmology [@2007ApJS..170..377S]. The halo has total mass $M_{\rm vir} =
7.9 \times 10^{11} {{\rm M}_{\odot}}$ at $z = 0$, and was chosen to have a quiet merger history with no major merger (of mass ratio greater than 1/10) after $z$ = 3. The high-resolution region, centred around a Lagrangian subregion of 1 Mpc on a side, contains 13 million dark matter particles and an equal number of gas particles, for a final dark matter and gas particle mass of $m_{\rm DM} = 9.8 \times 10^{4} {{\rm M}_{\odot}}$ and $m_{\rm SPH} = 2.0 \times 10^{4} {{\rm M}_{\odot}}$, respectively. The gravitational softening length, $\epsilon_{\rm G}$, was fixed to 120 physical pc for all particle species from $z=9$ to the present, and evolved as $1/(1+z)$ from $z=9$ to the starting redshift of $z=90$. A total number of 400 snapshots were taken. The time spacing between two subsequent snapshots is about 30 Myr.
The simulation includes a uniform metagalactic UV background [@HM96], Compton cooling, atomic cooling and metallicity-dependent radiative cooling at T $<10^{4}$ K. Star formation is modelled by stochastically forming “star particles” out of gas that is sufficiently cold ($T <
3 \times 10^{4}$ K) and reaches a threshold density of $n_{\rm SF}=5$ atoms cm$^{-3}$. The local star formation rate follows: $$\frac{d\rho_*}{dt}=0.1 \frac{\rho_{\rm gas}}{t_{\rm dyn}} \propto \rho_{\rm
gas}^{1.5},$$ where $\rho_*$ and $\rho_{\rm gas}$ are the stellar and gas densities, respectively, and $t_{\rm dyn}$ is the local dynamical time. Each star particle has an initial mass $m_{\ast} = 6000 \ {{\rm M}_{\odot}}$ and represents a simple stellar population that follows a @Kroupa93 initial mass function (IMF). Note that this IMF is also applied for the first generation of stars (Population III). In the simulation, new star particles inherit the metallicity of their parent gas particles. Star particles inject energy, mass and metals back into the ISM through Type Ia and Type II SNe, and stellar winds [@Stinson06]. Eris’s high resolution enables the development of an inhomogeneous ISM, which allows clustered star formation and strong cumulative feedback from coeval supernova explosions. Large scale galactic winds are launched as a consequence of stellar feedback, which transports a substantial quantity of metals into the circumgalactic medium and enriches the subsequent gas accretion [@2013ApJ...765...89S]. At $z = 0$, Eris forms an extended, rotationally supported stellar disk with a small bulge-to-disk ratio and a radial scale length of 2.5 kpc. The structural properties, the mass budget in various components, and the scaling relations in Eris are simultaneously consistent with observations of the Galaxy [@Guedes11].
The simulation follows @Raiteri96 to model metal enrichment from Type Ia and Type II SNe ([SN Ia]{} and [SN II]{}). Metals are distributed to gas within the SPH smoothing kernel, which consists of 32 neighboring particles. For [SN II]{}, metals are released as the main-sequence progenitors die, and iron and oxygen are produced according to the following fits to the @Woosley95 yields: $$M_{\rm Fe} = 2.802 \times 10^{-4} \left({m_*\over
{{\rm M}_{\odot}}}\right)^{1.864}\,{{\rm M}_{\odot}},$$ and $$M_{\rm O} = 4.586 \times 10^{-4} \left({m_*\over {{\rm M}_{\odot}}}\right)^{2.721}\,{{\rm M}_{\odot}}.$$ For [SN Ia]{}, each explosion produces 0.63 ${{\rm M}_{\odot}}$ of iron and 0.13 ${{\rm M}_{\odot}}$ of oxygen [@Thielemann86]. The mass return from stellar wind feedback is computed for stars with masses between 1 and 8 ${{\rm M}_{\odot}}$. For each stellar particle, the amount of stars lost as supernovae during every timestep is computed following stellar lifetimes by @Raiteri96, and the returned mass is calculated using the initial-final mass relation of @Weidemann87. The mass is distributed to neighbouring gas particles in the same way as the SNe feedback but with no energy injection. The returned gas inherits the metallicity of the star particle. We adopt the @Asplund09 solar abundance scale for the elements other than O and Fe that are not tracked in the simulation.
It is worth noting that the simulation used a traditional SPH formalism where metals advect with the fluid perfectly without mixing due to microscopic motions, and this may cause an artificially inhomogeneous chemical distribution. Following @2015ApJ...807..115S, to model the unresolved mixing we post-process the simulation to assign each stellar particle with the <[*intrinsic*]{} average metallicity of its 128 neighbouring gas particles at the formation time. The corresponding mixing length is generally around 50–120 pc at all redshifts. At our star formation threshold density of 5 atoms cm$^{-3}$, this corresponds to a distance that gas can cross within the free-fall time, assuming the typical velocity dispersion of a molecular cloud. We have shown in @2015ApJ...807..115S that, with this smoothing model, the overall chemical evolution in Eris appears to reproduce the observed \[$\alpha$/Fe\] as a function of metallicity in the Milky Way. When we calculate the metal accretion rate (see Section \[sec:acc\] for details), the gas properties are also smoothed over 128 neighbours.
{width="1.7\columnwidth"}
Metal-Accretion Model {#sec:acc}
=====================
In Eris, stars are represented by collisionless star particles that are created when gas particles satisfy the temperature and density condition described in the previous section. Each star particle has an initial mass $m_{\ast} = 6000$ M$_\odot$, which reduces with time due to stellar evolution. While we cannot resolve individual stars with mass of the order of $1$–$100$ M$_\odot$ due to unavoidable limitations imposed by numerical resolution, star particles in Eris still trace the underlying gravitational potential well. Therefore, we assume that these star particles provide a fair sample of the true stellar orbits. To calculate accretion of gas from the ISM onto stars, we assume that stars follow the orbits of star particles in Eris, and model accretion as a spherically symmetric Bondi-Hoyle flow, given by [@1939PCPS...35..405H; @1952MNRAS.112..195B; @1985MNRAS.217..367S] $$\dot{M} = \frac{2\pi(GM_\ast)^{2}\rho}{(v_\mathrm{rel}^{2} + c_{s}^{2})^{3/2}},
\label{eqn:bh}$$ where $M_{\ast}$ is the stellar mass, $\rho$ is the gas density, $v_\mathrm{rel}$ is the relative velocity between the star and the gas, and $c_{s}$ is the gas sound speed. For an ideal gas $c_{s}=\sqrt{\gamma k_{B} T/\mu m_{H}}$, where $\gamma$ is the adiabatic index, $k_{B}$ is the Boltzmann constant, $\mu$ is the mean molecular weight and $m_{H}$ is the mass of the hydrogen atom. For simplicity, we use a constant $\gamma = 5/3$ corresponding to an adiabatic gas, and a constant $\mu = 1.2$, which corresponds to primordial neutral gas. In reality radiative cooling may reduce $\gamma$ to a value between 5/3 and 1, photo-ionisation may decrease $\mu$ by a factor of 2. These changes to $\gamma$ and $\mu$ may alter the sound speed by at most 40%.
We consider gas accretion onto four classes of stars, two of which are defined by age and two by intrinsic metallicity: (1) stars formed in the first 600 Myr of the simulation and classified as halo stars at redshift $z=0$; (2) stars formed in the first 600 Myr of the simulation and classified as bulge stars at redshift $z=0$; (3) stars with intrinsic iron abundance as calculated in Eris $[\mathrm{Fe}/\mathrm{H}]< -4$ and classified as halo stars at redshift $z=0$; and (4) stars with intrinsic iron abundance $[\mathrm{Fe}/\mathrm{H}]< -4$ and classified as bulge stars at redshift $z=0$. We thus consider accretion onto two classes of halo stars and two classes of bulge stars. Stars selected by metallicity will in general have smaller ages than stars selected by our age criterion. Observed low-mass metal-poor stars lie at the intersection of the two selections. From a theoretical point of view, the two selections may be expected to bracket extreme accretion scenarios, as age-selected stars will accrete gas longer. Note that the disk, bulge, and halo of the Eris Galaxy analog at $z=0$ are determined by kinematic decomposition as described by @Guedes11. Of all the Eris star particles falling in each group, we select a uniform random sample of 2000 star particles to compute metal accretion. Orbits of these star particles are traced to the current epoch using 400 snapshots evenly spaced in time with $\Delta t \simeq$ 30 Myr. For each star particle, we compute the average gas sound speed $c_{s}$ and metallicity over its 128 nearest neighbouring gas particles [@2015ApJ...807..115S], as described in Section \[sec:erissim\]. The relative velocity $v_\mathrm{rel}$ is the velocity difference between the star particle and the centre-of-mass velocity of the neighbouring gas particles. The iron or oxygen accretion rate is then given by multiplying the averaged iron or oxygen mass fraction of the neighbouring gas by the mass accretion rate from Equation (\[eqn:bh\]). Figure \[fig:orbits\] provides a visual impression of halo and bulge stars in Eris by showing the two-dimensional probability distribution functions derived from projected stellar orbits of a uniform random sample of halo and bulge stars. (Orbits of only metallicity-selected stars are shown.) The projection is over complete orbits so that information from all redshifts is combined.
We assume $M_*=0.8$ M$_\odot$ while evaluating the mass accretion rate using Equation (\[eqn:bh\]). This approximately corresponds to the main-sequence turnoff mass in a 12 Gyr old stellar population. The maximum stellar mass for which a star in our age-limited sample can survive up to the present day is 0.99 M$_\odot$. The lifetime of a star with mass 0.9 M$_\odot$ is 15 Gyr and that with mass 0.8 M$_\odot$ is 16.1 Gyr. Stars selected by metallicity can have much higher mass: metal-poor stars could form relatively recently in pristine pockets of the ISM that are yet to be polluted by metals; such stars can have higher mass and still survive to the present day. Still, we find that 36% of the bulge star particles selected by metallicity form in the first 600 Myr of the simulation and 80% form in the first Gyr. Of the halo stars selected by metallicity, 11% form in the first 600 Myr and 80% form in the first 1.8 Gyr. Metal mixing in the ISM is efficient enough that most metallicity-selected stars have large ages. We assume that accreted metals are mixed in surface convective zones with mass $3\times 10^{-3}$ M$_\odot$ for $M_*=0.8$ M$_\odot$ . For a fixed convective zone mass fraction, the accretion rate scales as $M_*^2$ and the relative metal abundance ratios \[X/H\] scale as $M_*$.
Results {#sec:results}
=======
With stellar orbits and gas metallicity predictions from Eris, and our metal accretion model described in the previous section, we can now consider metal enrichment of halo and bulge stars due to accretion.
Metal accretion in Eris
-----------------------
Figure \[fig:FeH\_PDF\] shows probability distribution functions (PDFs) of the relative abundance values \[Fe/H\] of accreted iron for the four categories of stars defined in Section \[sec:acc\]. The four PDFs are quite similar, perhaps surprisingly so. They peak around $[\mathrm{Fe}/\mathrm{H}] \sim -6$ to $-5$, well within the metallicity detection limit of modern surveys [@2013pss5.book...55F] and comparable to the metallicities of the most metal-poor stars known [@2011Natur.477...67C; @2014Natur.506..463K]. This indicates that accretion can have a significant impact on metal-poor main-sequence stars. The four PDFs also show a considerable scatter—about three decades around their average values. The remarkable agreement between the four distributions results from the fact that both halo and bulge stars encounter dense gas in the central regions of the Galaxy. This is evident in Figure \[fig:orbits\], which shows that both halo and bulge stars exhibit a significantly higher incidence in the central regions. This is understandable, as a kinematically distinct halo and bulge do not exist in Eris at redshift $z>2$, and the last major merger of the Galaxy is at $z\sim 3$. Additionally, the \[Fe/H\] PDFs of age- and metallicity-selected samples also agree quite well, because most metallicity-selected stars ($[\mathrm{Fe}/\mathrm{H}]<
-4$) are quite old. As mentioned above, 80% of metallicity-selected halo stars were formed within the first 1.8 Gyr of the simulation, while the same fraction of metallicity-selected bulge stars were formed within the first 1 Gyr. Thus, regardless of how they are selected at redshift $z=0$, halo as well as bulge stars in our analysis accrete to a level of $[\mathrm{Fe}/\mathrm{H}] \sim
-6$ to $-5$ on average.
While the four distributions in Figure \[fig:FeH\_PDF\] are remarkably similar, there are small differences. Age-selected bulge stars have the highest median accretion ($[\mathrm{Fe}/\mathrm{H}]=-4.9$ with a 68% spread of 1.7), followed by age-selected halo stars ($[\mathrm{Fe}/\mathrm{H}]=-5.2$ with a 68% spread of 1.8), and metallicity-selected bulge stars ($[\mathrm{Fe}/\mathrm{H}]=-5.4$ with a 68% spread of 2.3). Metallicity-selected halo stars have the lowest value, $[\mathrm{Fe}/\mathrm{H}]=-6.0$ with a 68% spread of 2.0. The age-selected sample of bulge stars has the highest accreted metallicity, while the metallicity-selected sample of halo stars shows the lowest accreted metallicity. The PDFs of age-selected halo stars and metallicity-selected bulge stars have intermediate values. This difference is still quite small compared to the overall spread of the PDFs. This ordering of the PDFs can also be understood from the small differences in ages and orbital distribution of stars in the four samples. Age-selected stars in our analysis are marginally older than the metallicity-selected stars, and therefore, on an average accrete metals for a longer duration, resulting in higher \[Fe/H\]. Likewise, bulge stars encounter marginally denser gas than halo stars, as seen in Figure \[fig:orbits\], which boosts the accretion rates onto bulge stars. Our finding that metal-poor bulge stars are generally older than metal-poor halo stars is along the lines of previous theoretical results [@2000fist.conf..327W; @2007ApJ...661...10B; @2010ApJ...708.1398T]. This explains the ordering of the histograms for metallicity-selected samples in Figure \[fig:FeH\_PDF\].
The result in Figure \[fig:FeH\_PDF\] describes accretion onto 0.8 M$_\odot$ main-sequence stars, but it is possible to understand how it will change for different stellar masses. For a fixed convective-zone mass fraction, the accreted metal abundance \[Fe/H\] scales with $M_*$. For smaller stellar masses, the enrichment levels will drop due to the reduced accretion rate and enhanced convective layer mass. For a 0.6 M$_\odot$ main-sequence star, the convective zone mass increases to $\sim
10^{-2}$ M$_\odot$ while the accretion rate drops by a factor of 1.8. As a result, the histograms in Figure \[fig:FeH\_PDF\] will shift towards lower values of \[Fe/H\] by about two decades. Old stars with masses greater than 0.8 M$_\odot$ will likely be on the giant branch today. Although accretion rates of these stars remain quite high, the resultant \[Fe/H\] ratios drop because of larger convective zone masses, which are of the order of $10^{-1}$ M$_\odot$ for giants . In this case the histograms in Figure \[fig:FeH\_PDF\] will shift towards lower values of \[Fe/H\] by almost three decades. Figure \[fig:FeH\_PDF\] thus presents as extreme case. Negative feedback from processes such as stellar winds and radiation feedback may also reduce accretion [@2015MNRAS.453.2771J; @2011MNRAS.413.1184J]. Our results from Figure \[fig:FeH\_PDF\] predict marginally higher accreted metallicity than the model of @2009MNRAS.392L..50F. Our results are closer to the extreme model of @2009MNRAS.392L..50F, in which it is assumed that the stars pass once through a dense cloud with density $n\sim
10^3$ cm$^{-3}$. This difference is mainly because the gas density increases quite rapidly towards high redshift, while @2009MNRAS.392L..50F only consider accretion from a static gas disk with height 100 pc and density $n\sim 5$ cm$^{-3}$. The median accreted \[Fe/H\] in their fiducial model is about $-7$. This increases to $-5$ in their extreme model. The latter value is closer to the median values of all four samples of stars in Figure \[fig:FeH\_PDF\]. Our median values are in agreement with predictions by @2015ApJ...808L..47K obtained from a semi-analytical galaxy formation model built on halo merger trees.
![Evolution of the metal-accretion rate in our four samples of stars. Solid curves show the median values and shaded regions show 68% scatter around the median. Blue and yellow curves and shaded regions describe properties of halo stars selected by age and metallicity, respectively. Bulge stars selected by age and metallicity are described, respectively, by purple and red curves. The sudden drop in the accretion rates onto halo stars at $z\sim
3.1$ is associated with the last major merger of the galaxy.[]{data-label="fig:accrate"}](accrate.pdf){width="\columnwidth"}
{width="\textwidth"}
Understanding metal-accretion rates
-----------------------------------
In order to understand the stellar metal-accretion levels in our model, Figures \[fig:accrate\] and \[fig:evolution\] show the evolution of various relevant quantities in Eris for the four samples of stars considered. Figure \[fig:accrate\] shows the evolution of the metal-accretion rate of iron. Yellow and blue curves correspond the metallicity- and age-selected samples, respectively, of halo stars. Red and purple curves correspond to bulge stars. Solid curves show the evolution of the median value; shaded regions in corresponding colours show ranges of 68% scatter around the medians. We see that the average metal accretion rate increases as we go towards higher redshifts. All four categories of stars have very similar accretion rates at redshifts $z>3$. It is only at lower redshifts ($z<3$), after the last major merger of the Eris Milky Way, that halo and bulge stars kinematically separate with halo stars occupying regions with low gas density and therefore low accretion rates. The average metal accretion rate at high redshifts is about $10^{-20}$ M$_\odot$yr$^{-1}$. By $z=0$, this drops by two orders of magnitude for bulge stars and by four orders of magnitude for halo stars. Note that the difference between the metal accretion rates of age- and metallicity-selected stars is relatively small. This is simply because our metallicity-selected samples are also predominantly composed of old stars, as explained above.
The relatively high values of accretion rates at high redshift are driven by the gas density evolution. This is evident in panel (a) of Figure \[fig:evolution\], which shows the evolution of the average physical density of gas along the stellar orbits in each of our four samples. The difference between the four curves is along the lines of what we expect—metallicity-selected halo stars encounter the lowest gas density, while the age-selected bulge stars encounter the highest gas density. This difference explains the trends in the accretion rates shown in Figure \[fig:accrate\] as well as in the histograms of accreted \[Fe/H\] in Figure \[fig:FeH\_PDF\]. The black curve in panel (a) shows the evolution of the average density within the virial radius of the halo. The virial density evolves at $(1+z)^3$. A comparison between the virial density curve and the median gas density curves shows that the density of gas encountered by stars in our samples grows roughly as $(1+z)^3$ from redshift $z=0$ to $6$, and then drops because the halo turns around. Importantly, the stars in our samples encounter a significantly higher gas density than the virial density as both halo and bulge stars are centrally concentrated compared to the virial radius, as in Figure \[fig:orbits\]. This can be understood further with help from the brown dot-dashed, dashed, and solid curves in panel (a) of Figure \[fig:evolution\]. These curves show the average gas density in Eris within 25, 2, and 0.5 physical kpc (pkpc), respectively. At high redshifts, $z>3$, almost all stars in all four of our samples are interacting with gas within in the inner 2 pkpc of the progenitor haloes of the Milky Way. This gas density is considerably higher than the gas density assumed in previous works [cf. @2009MNRAS.392L..50F; @2011MNRAS.413.1184J], which results in the high metal accretion seen in Figure \[fig:FeH\_PDF\]. This picture is reinforced by panel (b) of Figure \[fig:evolution\], which shows the evolution of the radial distribution of the stars in our four samples. The solid curves in this panel show the radial distance of stars from centres of mass of respective host halos with the scatter represented by the shaded areas. All four samples of stars are quite centrally concentrated. The halo stars generally have larger radial separations than the bulge stars, which explains the density trends seen in panel (a).
In panel (c), Figure \[fig:evolution\] shows the evolution of the relative velocity of stars with respect to ambient gas. The velocity decreases with increasing redshift due to the decreasing halo mass. This also increases the accretion rate. While accretion rates at high redshift are enhanced by increased gas density and reduced velocities, the gas metallicity evolution restricts the enhancement. The brown dashed curve in this panel shows the evolution of the circular velocity $v_\mathrm{circ}$ at the virial radius, while the black dashed curve shows the evolution of the maximum rotational velocity $v_\mathrm{max}$ in the main progenitor halo of the Eris galaxy. Stars in all four samples tend to have lower velocity than $v_\mathrm{max}$, as the rotation curve drops at small radii. The stellar velocities can also be compared with the evolution of the gas sound speed, which is shown in panel (d). The sound speed decreases towards higher redshift with reduced gas temperature. Panels (e) and (f) of Figure \[fig:evolution\] show the evolution of the iron and oxygen abundances of the gas along orbits of stars in our four samples. There is a metallicity gradient in the ISM with most metal-enriched gas present near the center of the halo. This metallicity gradient is also seen in other simulations of the Galaxy’s assembly [e.g., @2013MNRAS.436..625S]. As a result, bulge stars accrete relatively more metal-rich gas compared to halo stars. The metallicity of gas around bulge stars evolves from $[\mathrm{Fe}/\mathrm{H}]=-1.5$ at $z\sim 10$ to $[\mathrm{Fe}/\mathrm{H}]=0.0$ at $z=0$. The oxygen abundance evolves from $[\mathrm{O}/\mathrm{H}]=-1.0$ to $[\mathrm{O}/\mathrm{H}]=0.1$ in this redshift range. The metallicity of gas around halo stars is lower by about 0.7. The resultant $\alpha$-enhancement of ISM gas at high redshift by about 0.5 relative to low redshifts is seen in panel (e) of Figure \[fig:evolution\], which shows the evolution of $[\mathrm{O}/\mathrm{Fe}]$.
Figure \[fig:ofe\] shows the resultant \[O/Fe\] ratio at $z=0$ due to accretion in the four classes of stars considered here, as a function of the accreted \[Fe/H\]. This $\alpha$-enhancement is shown in comparison with observed \[$\alpha$/Fe\] ratios from disk and halo stars in the Galaxy . The accreted \[O/Fe\] ratios are in agreement with the \[O/Fe\] ratios seen in the most metal-poor stars (see also @2006ApJ...648..383P). At iron abundances of $[\mathrm{Fe}/\mathrm{H}]>-1$ the observed \[O/Fe\] drops sharply to less than $\sim 0.2$, signalling the effect of SN Ia [@1979ApJ...229.1046T]. The accreted $\alpha$-enhancements are much higher. The reason is clear from Figure \[fig:accrate\]—stars in all four of our samples accrete most of their metals at high redshifts before SN Ia start playing an important role. Thus, most accreted gas is $\alpha$-enhanced (cf.panels (e) and (f) of Figure \[fig:evolution\]). This also tells us that measurements of \[$\alpha$/Fe\] will not help in separating stars with intrinsic metal enrichment from stars that have accreted their metals. Perhaps the only way to disentangle these stars would be to observe a peculiar enrichment pattern that is unlikely to exist in the low-redshift ISM, such as the enhanced \[Si/C\] ratio or the large contrast between the abundances of odd and even element pairs that is predicted for pair-instability supernovae [@2002ApJ...567..532H; @2014Sci...345..912A].
![Median $\alpha$-enhancement \[O/Fe\] of accreted gas by stars in the four samples considered in this paper (solid curves) and 68% scatter around the median (shaded regions). Blue and yellow curves and shaded regions describe properties of halo stars selected by age and metallicity, respectively. Bulge stars selected by age and metallicity are described, respectively, by purple and red curves. Open circles show $[\alpha/\mathrm{Fe}]$ measurements for halo and disk stars from the compilation by @2015ApJ...807..115S. There is a good agreement between the $[\alpha/\mathrm{Fe}]$ for iron-poor stars and the accreted gas in our model.[]{data-label="fig:ofe"}](ofe.pdf){width="\columnwidth"}
Effect of accretion on the MDF
------------------------------
Most models of Galactic chemical enrichment—including our base Eris simulation—assume that enrichment of stars due to accretion of gas from the ISM is negligible. In Figure \[fig:mdf\_corrected\] we consider the error incurred due to this assumption. Black symbols in this figure show the measured metallicity distribution function (MDF) as compiled by @2017MNRAS.465..926D, shown here as a probability distribution function. The green histogram shows the MDF of halo stars from Eris without any correction due to accretion. This histogram refers to the intrinsic metallicity of halo stars in Eris. The red histogram in Figure \[fig:mdf\_corrected\] shows how the black histogram is modified when accretion is taken into account. In each \[Fe/H\] bin, stars are added from lower metallicity bins by gaining excess \[Fe/H\] due to accretion, in proportion to the PDF derived in Figure \[fig:FeH\_PDF\]. Similarly, each \[Fe/H\] bin loses stars to higher metallicity bins due to accretion. Only the metallicity-selected sample of halo stars (which has stars that have $[\mathrm{Fe}/\mathrm{H}]<-4$ at $z=0$) is used for this purpose. The predicted MDFs agree quite well with observations.
As described in previous sections, enrichment due to accretion is small ($[\mathrm{Fe}/\mathrm{H}]\sim -6$ to $-5$ on average). Therefore, accretion affects only the metal-poor tail of the MDF; the MDF at high metallicities remains unmodified. However, the effect on the MDF at $[\mathrm{Fe}/\mathrm{H}]<-5$ is considerable, suggesting that a large fraction of metal-poor stars in this regime could be potentially enriched due to accretion. This is emphasised by the bottom panel of Figure \[fig:mdf\_corrected\], which shows the ratios of the uncorrected and corrected MDFs from the top panel. Accretion truncates the low-metallicity end of the MDFs to about $[\mathrm{Fe}/\mathrm{H}]=-7$. Metal abundance measurements of these stars would then contain little or no information about the chemistry of their natal environments, thus affecting any inference of the Population III IMF from these stars [e.g., @2007MNRAS.381..647S; @2012MNRAS.423L..60S; @2015MNRAS.447.3892H; @2017MNRAS.465..926D]. It should also be noted that accretion may change the relative abundances of elements by a greater degree than the absolute abundance of a single element. @2015MNRAS.453.2771J has proposed that if stars do indeed accrete metals from the ISM, then radiation pressure would selectively inhibit accretion of dust grains. As a result, if most metal enrichment in these stars is due to accretion, then refractory elements such as iron would appear to be selectively depleted. This pattern of metal enrichment is tantalizingly similar to that seen in carbon-enhanced metal poor (CEMP) stars . Seven of the eight stars with $[\mathrm{Fe}/\mathrm{H}]<-4.5$ are carbon-enhanced. The one star with low carbon abundance [@2011Natur.477...67C] could still be a CEMP star with a carbon abundance that is yet to be detected, possibly because of the limited spectral S/N and warm temperature.[^1] The argument of @2015MNRAS.453.2771J would then suggest that the known metal-poor stars have been enriched mostly by accretion of metals from the ISM. This is in good agreement with what we find in Figure \[fig:mdf\_corrected\].
![The metallicity distribution function (MDF) of halo stars in Eris simulation with (red curve) and without (green curve) accounting for metal accretion by stars from the ISM. The data points from the compilation by @2017MNRAS.465..926D. The blue curve in the lower panel shows the ratio of the two MDFs, illustrating the truncation of the low-metallicity tail of the MDF due to accretion.[]{data-label="fig:mdf_corrected"}](mdf.pdf){width="\columnwidth"}
Conclusions {#sec:conc}
===========
We have studied the chemical enrichment of stars due to accretion of metal-enriched gas from the ISM during the Milky Way’s development, using the Eris zoom-in cosmological simulation of assembly of a Milky Way analog. We considered metal-poor and old stars in both the Galactic halo and bulge and make use of the stellar orbits and gas density and metallicity distributions predicted by Eris.
Our main conclusion is that both halo and bulge stars accrete metals from the ISM and, in extreme cases, can reach enrichment levels of $[\mathrm{Fe}/\mathrm{H}]\sim -2$ due to accretion alone. The median enrichment level is about $[\mathrm{Fe}/\mathrm{H}] \sim -6$ to $-5$ for age- as well as metallicity-selected stars in both the halo and bulge. There is little difference in the accreted enrichment levels of the four classes of stars, although metallicity-selected halo stars tend to accrete the least amount of metals. Metal-poor bulge stars are generally older than metal-poor halo stars. As a result, bulge stars show a marginally higher degree of effect of accretion. Halo and bulge stars accrete metals at the rate of about $10^{-24}$ M$_\odot$yr$^{-1}$ and $10^{-22}$ M$_\odot$yr$^{-1}$, respectively, at redshifts $z\lesssim 3$, but this accretion rate increases a hundred-fold to about $10^{-20}$ M$_\odot$yr$^{-1}$ at higher redshifts due to increased gas density. The drop in accretion rates with time is mainly because of decreasing gas density. Bulge and halo stars accrete similar amounts of metals at high redshifts as kinematically distinct bulge and halo are not yet developed at these redshifts and both sets of stars encounter similar metal distribution in the ISM on average.
Because accretion mostly takes place at high redshifts, it is $\alpha$-enriched to $[\alpha/\mathrm{Fe}]\sim 0.5$. This is comparable to the $\alpha$-enrichment levels seen in the most metal-poor stars. It also means that $\alpha$-enrichment does not act as a discriminant between intrinsically and extrinsically enriched stars. Accretive metal enrichment is sufficient to change the predicted MDF of halo stars at $[\mathrm{Fe}/\mathrm{H}]<-5$. This suggests that attempts to infer the natal chemical environment of the most metal-poor stars from their observed enrichment today can be hindered due to metal accretion. Our analysis assumes a stellar mass of 0.8 M$_\odot$. Accreted metal enrichment level drops rapidly at smaller stellar masses due to decreasing accretion rates and increased convective zone masses. This, faint dwarfs will be more likely to probe the first generation of stars. Peculiar enrichment patterns such as those predicted to arise from pair-instability supernovae could help in disentangling natal and accreted metal content of stars [@2014Sci...345..912A]. Most stars are rich in oxygen and carbon with $[\mathrm{O}/\mathrm{H}]<-3.2$ and $[\mathrm{C}/\mathrm{H}]<-3.5$. The accreted gas in our model only has $[\mathrm{O}/\mathrm{Fe}]=0.5$. Therefore, although the iron abundance is significantly enhanced by accretion the abundance in elements such as C and O should still remain unchanged and could be useful in probing Population III. Also, our findings suggest that lowest-metallicity Damped Lyman-$\alpha$ Systems would provide an important cross-check for Population III yields inferred from metal-poor stars.
Acknowledgements {#acknowledgements .unnumbered}
================
We thank the referee, Timothy Beers, for his positive comments and constructive suggestions, and also acknowledge helpful discussions with Andrew Casey, Ryan Cooke, Denis Erkal, Gerry Gilmore, Martin Haehnelt, Kohei Hattori, Joseph Hennawi, Sergey Koposov, Thomas Masseron, Max Pettini, Enrico Ramirez-Ruiz, Alberto Rorai, and Stefania Salvadori. S.S. gratefully acknowledges support by Science and Technology Facilities Council and by ERC Starting Grant 638707 ‘Black holes and their host galaxies: coevolution across cosmic time’. G.K. and S.S. acknowledge support from ERC Advanced Grant 320596 ‘The Emergence of Structure During the Epoch of Reionization’. Support for this work was provided by NASA through grant HST-AR-13904.001-A (P.M.). P.M. also acknowledges a NASA contract supporting the WFIRST-EXPO Science Investigation Team (15-WFIRST15-0004), administered by GSFC, and thanks the Préfecture of the Ile-de-France Region for the award of a Blaise Pascal International Research Chair, managed by the Fondation de l’Ecole Normale Supérieure. This work used the DiRAC/Darwin Supercomputer hosted by the University of Cambridge High Performance Computing Service (http://www.hpc.cam.ac.uk/), provided by Dell Inc. using Strategic Research Infrastructure Funding from the Higher Education Funding Council for England and funding from the Science and Technology Facilities Council.
\[lastpage\]
[^1]: This is the case, for instance, for the Group II CEMP-no stars G64-12 and G64-37, which were not recognised as CEMP until extremely high S/N spectra were obtained [@2016ApJ...829L..24P].
| {
"pile_set_name": "ArXiv"
} |
=-1cm
[**BU-HEPP-11-05**]{}\
[**Dec., 2011**]{}\
[Estimates of Radiation by Superluminal Neutrinos]{}
[**B.F.L. Ward**]{}\
[*Department of Physics, Baylor University, Waco, TX, USA*]{}\
[**Abstract**]{}
We show that the more energetic superluminal neutrinos with quadratically dispersed superluminalities $\delta=\beta^2-1$, for $\beta=v/c$ where $v$ is the neutrino velocity, also lose significant energy to radiation to the $\nu+e^-+e^+$ final state in traveling from CERN to Gran Sasso as has been shown to occur for those with constant superluminality by Cohen and Glashow if indeed $\delta \simeq 5\times 10^{-5}$. In addition, we clarify the dependence of such radiative processes on the size of the superluminality.
=cmssbx10 scaled 2
.1truein
In trying to understand the results of the OPERA Collaboration [@opera] for apparent faster than light velocity for neutrinos in transit from CERN to Gran Sasso, Cohen and Glashow [@cg] have made an estimate of the Cherenkov effect to the $\nu+e^++e^-$ final state using the standard methods adapted to Lorentz invariance-violating interacting quantum field theory. This effect is also studied in Ref. [@matt] in a general framework for such violations of Lorentz invariance with particular attention to the case where the superluminality is quadratically dispersed in contrast to the case considered by Cohen and Glashow with constant superluminality, where we define the superluminality by $\delta=\beta^2-1$ when $v$ is the neutrino velocity. Here, we extend the analysis of Cohen and Glashow to the quadratically dispersed case of Ref. [@matt] and we compare the analyses of Cohen and Glashow and of Ref. [@matt] with an eye as well toward the very interesting observations of the OPERA Collaboration.
We need to first state very clearly at the outset here that we do not claim to rule out the possibilities discussed in Refs. [@dsr1a; @dsr1b; @carcor] in which appropriate deformation of the Lorentz symmetry renders these Cherenkov processes kinematically forbidden just as one has for the usual non-superluminal case in the usual special theory of relativity. Thus, non-observation of the effects discussed herein does not rule-out such changes in the standard orthodoxy.
Specifically, Cohen and Glashow arrive at the results $$\begin{split}
\Gamma&=k'\frac{G_F^2}{192\pi^3}E^5\delta^3\\
\frac{dE}{dx}&=-k\frac{G_F^2}{192\pi^3}E^6\delta^3
\end{split}
\label{cg1}$$ for the Cherenkov radiation rate and energy loss per unit distance for the process $\nu\rightarrow \nu+e^++e^-$ for $\nu$’s of energy $E$ where the parameters $k',\;k$ are $1/14$ and $25/448$, respectively [@dnan]. In this way, assuming that $\delta$ is independent of $E$, they find [@cg] that the $\nu$’s should lose considerable energy between CERN and Gran Sasso.
On the other hand, Mattingly [*et al.*]{} in Ref. [@matt] have discussed the same transition for superluminal $\nu$’s via analogy with the so-called $\nu$-splitting process $\nu_A\rightarrow \nu_A\nu_B\bar\nu_B$ for which they find the result $$\Gamma_{\nu\nu\bar\nu}\sim \frac{3G_F^2}{64\pi^3}\frac{\eta_\nu^4E^{13}}{M_{Pl}^8}
\label{mattg1}$$ where the factor of $3$ multiplying $G_F^2$ sums over three neutrino species so, ignoring the masses of the neutrinos and the electron, Mattingly [*et al.*]{} assert (see Sect. 6 of Ref. [@matt]) that the rate for the $\nu+e^+e^-$ decay should be nearly the same as that for the $\nu$-splitting decay in (\[mattg1\])[^1]. The Lorentz-violating parameter $\eta_\nu$ is such that the dispersion relation for the $\nu$ is $$E_p^2=m_\nu^2+p^2+\eta_\nu p^4/M_{Pl}^2$$ where $M_{Pl}$ is the Planck mass, $m_\nu$ is the neutrino rest mass and $p$ is the $\nu$ momentum. The parameter $\eta_\nu$ is to be determined from experiment. Thus, using $v=\frac{dE_p}{dp}$ and ignoring the neutrino mass we see that $$\delta\cong 3\eta_\nu E^2/M_{Pl}^2$$ so that Mattingly [*et al.*]{} argue that the rate for the $\nu\rightarrow \nu+e^++e^-$process should be nearly given by $$\begin{split}
\Gamma_{\nu e^+e^-}&\simeq \Gamma_{\nu\nu\bar\nu}\\
&\cong \frac{3G_F^2}{64\pi^3}\frac{\eta_\nu^4E^{13}}{M_{Pl}^8}\\ & \cong \frac{3G_F^2}{64(81)\pi^3}E^5\delta^4.
\end{split}
\label{mattg2}$$ While the numerical factors and the explicit dependence on $E$ are consistent between the two estimates of $\Gamma$ for the $\nu\rightarrow \nu+e^++e^-$ process, the two dependencies on $\delta$ differ by one factor $\delta$: this factor would obviate the arguments of Cohen and Glashow, given that OPERA have found that $\delta\cong 5\times 10^{-5}$. Which dependence on $\delta$ is actually correct? The literature is beginning to feature use of both dependencies, as we see in Ref. [@carcor] where the corresponding dependence of Mattingly [*et al.*]{} is used and in Ref. [@matt2] where both dependencies are featured [^2]. Secondly, what effect does quadratic dispersion for $\delta$ have on the results of Cohen and Glashow? We address these two questions here in inverse order.
We see that, unlike the case discussed by Cohen and Glashow, we have to consider pronounced dependence of $\delta$ on momentum: for a neutrino of initial energy $E_0\sim 20$GeV, $\delta(E)\cong\delta_0(E/E_0)^2$ in an obvious notation. Moreover, the factor of 3 between the deviation between the squared group velocity and 1 here and the squared incoming $\nu$ four-momentum vector in units of its squared laboratory energy $E$ means that we should identify the quantity $\delta$ in the Cohen-Glashow formulas in (\[cg1\]) with our $\delta/3$. Thus, the dependence of $dE/dx$ on $E$ here is given by $$\frac{dE}{dx}=-k\frac{G_F^2}{192\pi^3}\frac{E^{12}\delta_0^3}{E_0^6}.
\label{eqnw1}$$ Integrating this relation gives us $$\begin{split}
\frac{1}{E^{11}}-\frac{1}{E_0^{11}}&=11k\frac{G_F^2}{192\pi^3}\frac{\delta_0^3L\
}{E_0^6}\\
&=\frac{11}{5}\frac{1}{E_T^5E_0^6}
\end{split}
\label{eqnew2}$$ where we introduced the distance $L\cong730$km from CERN to Gran Sasso and the Cohen-Glashow quantity ${1/E_T^5}\equiv \frac{5kG_F^2\delta_0^3L}{192\pi^3}$ so that for $\delta_0=5\times 10^{-5}$ $E_T\cong 12.8GeV$. This should be compared with Eq.(21) in Ref. [@matt2], wherein only one power of the analog of $\eta_\nu$ appears versus three powers here. If we define the average value of $\delta$ via[^3] $$\begin{split}
<\delta>&=\frac{1}{L}\int_0^Ldx \delta(x)\\
&= \frac{1}{L}\int_0^Ldx (E^2(x)/E_0^2)\delta_0,
\end{split}
\label{eqv1}$$ we arrive at the relation $$<\delta>=\frac{5\delta_0}{9}\frac{E_T^5}{E_0^5}\left(\left(1+\frac{11}{5}\frac{\
E_0^5}{E_T^5}\right)^{\frac{9}{11}}-1\right).
\label{eqv2}$$ This means that, if we identity $3<\delta>$ with the OPERA result $5\times 10^{-5}$, we can solve this latter relation (\[eqv2\]) to get $3\delta_0=5.3\times 10^{-5}$ when the initial energy is $E_0=20$GeV. Thus, even though $\delta$ varies quadratically with momentum, it does not deviate very much from its initial value and a $20$ GeV $\nu$ now loses about 5.8% of its energy in travelling from CERN to Gran Sasso compared to 37.4% for the case discussed by Cohen and Glashow. The two groups of average velocity data in Ref. [@opera] have average energies 13.8 and 40.7 GeV with superluminalities of $\beta-1$ given by $(2.25\pm .81)\times 10^{-5}$ and $(2.79^{+.84}_{-.83})\times 10^{-5}$ respectively. Solving for the respective values of $3\delta_0$ we get $5.1\times 10^{-5}$ and $1.2\times 10^{-4}$ with energy losses of 1.1% and 41.8% respectively, where we use again $3<\delta>\cong 5\times 10^{-5}$ in view of the errors on the data without loss of content – the respective energy losses in the Cohen-Glashow analysis are 16.6% and 68.6%. The factor of order 2 variation of $\delta_0$ with initial creation energy is not excluded by the data. In any case, when the $\nu$ energy is in the 10 MeV regime, our values of $\delta$ are well below the limits from SN1987a [@sn1987a], $4\times 10^{-9}$, due to the quadratic dependence on energy in principle. We say ’in principle’ because there is, presuming the OPERA result is true, an unknown source of this energy dependence and possibly of the variation of $\delta_0$ with initial energy and we cannot make definitive statements until this source is known. We need more experimental input.
Concerning the detailed relationship between our calculation and that in Ref. [@rao], we note that the authors in Ref. [@rao] make a numerical estimate for the $\nu\rightarrow \nu +e^++e^-$ process with the quadratic dispersion for the $\nu$ superluminality and with the usual special relativity dispersion for the $e^-,\; e^+$ with the result that the coefficient of $G_F^2E^5\delta^3(1-4\sin^2\theta_w+8\sin^4\theta_w)$ in the Cohen-Glashow process is 1/45197.8 vs the exact result [@cg; @dnan] 1/41672.4, neglecting the lepton rest masses. This is acceptable in view of the current errors on the data. The authors in Ref. [@rao] then approximate the energy loss using a constant value of $\delta$ so that they give in their eq.(21) a mean decay time that scales as the 5th power of the $\nu$ energy when in fact the dependence of $\delta$ on energy shows that the decay time has to scale with the 11th power of this energy, as we have shown here. This results in significant deviation from the correct decay energy loss profile as we have shown above for the higher energies in the OPERA spectrum. However, for the 17.5 GeV energy $\nu$’s, we find that they lose 3.2% of their energy in travelling from CERN to Gran Sasso whereas the authors in Ref. [@rao] estimate this as a 4.2% loss if you take their result literally as written below their eq.(20). If you interpret their 4.2% as the mean number that decay via the $\nu\rightarrow \nu+e^-+e^+$ process so that we do not actually lose $\nu$’s but we lose their energy, then to get an estimate of the corresponding energy loss we have to multiply by the mean energy loss per decay, which is [@cg; @dnan] 78%. This would then agree with our 3.2% result – this is expected because as the energy does not change very much the assumption of a constant $\delta$ is not a bad one. We note as well that the authors in Ref. [@rao] do not address the issue of the difference between the power of $\delta$ in the decay rate formulas in (\[cg1\]) on the one hand and in (\[mattg1\]) and (\[mattg2\]) on the other. We now turn to this issue.
To answer the question on the proper power law for the dependence of the radiation rate on $\delta$, we revisit the analysis in Ref. [@matt] and observe that, in their Eq.(4.10), there is an energy conserving delta function, $\delta(E_p-E'_p-E_q-E'_q)$, where for definiteness we reproduce this equation here as $$\Gamma_{AB}=\frac{g^4}{16(2\pi)^5M_Z^4cos^4\theta_w}\int\delta(E_p-E'_p-E_q-E'_q)d^3p'd^3qF^2$$ for the width for the $\nu_A(p)\rightarrow \nu_A(p')\nu_B(q')\bar\nu_B(q)$ splitting process in a standard notation, with $E'_a=E_{a'},\; a=p,\; q$. Here, $q'(q)$ is the 4-momentum of the $\nu_B(\bar\nu_B)$, respectively, and the $SU_{2L}\times U_1$ [@gsw] couplings and parameters are standard by now – the rest mass of the $Z^0$ is $M_Z$, the $SU_{2L}$ coupling constant is $g$, the Weinberg angle is $\theta_w$. The function $F$ contains the dependence of the differential decay rate on the various angular phase space variables [@matt]. Upon arguing correctly that the excess energy from Lorentz-violation goes to transverse motion, Mattingly [*et al.*]{} then estimate the size of this transverse phase space and argue that the only effect of the delta function constraint on energy is to remove one factor of the energy from their result for the integral over the two phase spaces for $p'$ and $q$. However, there is a Jacobian required to implement this removal, by the standard methods, given by the absolute value of the derivative of the argument of the delta function with respect to $p'_{\parallel}$ for example[^4] due to the result $$\delta(f(x))=\sum_{x_a\in\{\text{zeros of {\it f}}\}}\frac{\delta(x-x_a)}{|\frac{df(x_a)}{dx}|}$$ for any sufficiently well-behaved function $f(x)$ : This provides the missing factor of $\delta$ that should divide the expressions on the RHS of (\[mattg2\]) and of (\[mattg1\]) to bring them into agreement with Cohen and Glashow. To see this note that the absolute value of the derivative of the argument of the delta function with respect to $p'_\parallel$ can be represented as follows: $$\begin{split}
\left|\frac{d(E_p-E'_p-E_q-E'_q)}{dp'_\parallel}\right|&=\left|\frac{d(E'_p+E'_q)}{dp'_\parallel}\right|\\
&=\left|(p'_\parallel+\frac{2p'_\parallel\eta_\nu({p'_\parallel}^2+{p'_\perp}^2)}{M_{Pl}^2})/E'_p+\frac{p'_\parallel+q_\parallel-p}{E'_q}\right|\\
&\cong \left|\frac{3\eta_\nu {p'_\parallel}^2 }{2M_{Pl}^2}+\frac{(\vec{q}_\perp+\vec{p'}_\perp)^2}{2(p-p'_\parallel-q_\parallel)^2}-\frac{{p'_\perp}^2}{2{p'_\parallel}^2}\right|\\
&\cong \frac{3\eta_\nu E^2}{M_{Pl}^2}\left|\left((\frac{1}{2}-\frac{\lambda_{p'}^2}{6x^4})x^2+\frac{\lambda_{q'}^2}{6(1-x-y)^2}\right)\right| \\
&\cong \delta\left|\left((\frac{1}{2}-\frac{\lambda_{p'}^2}{6x^4})x^2+\frac{\lambda_{q'}^2}{6(1-x-y)^2}\right)\right|,
\end{split}
\label{eqfnl}$$ where we follow Ref. [@matt] and use the longitudinal momentum fractions $x,\; y$ for $p'$ and $q$ respectively and where we parametrize the transverse momenta as $\vec{r}_\perp\equiv \hat{r}_\perp\sqrt{\eta_\nu}\frac{p^2}{M_{Pl}}\lambda_r$ for the unit vector $\hat{r}_\perp$ in their direction. By going to the “preferred frame” of the incoming $\nu$ we see that the parameters $2\lambda_a$, defined to be non-negative here, are bounded by constants close to $1$. We have used the usual dispersion relations $E_q=\sqrt{q^2+m_e^2},\;E'_q=\sqrt{{q'}^2+m_e^2} $ for the $e^+$ and $e^-$ respectively, where $m_e$ is the rest mass of the electron and have neglected all electron and neutrino masses[^5]. We stress that this factor of $\delta$ in (\[eqfnl\]) should divide the RHS of (\[mattg1\]) also when one uses it for the process $\nu\rightarrow \nu\nu\bar\nu$ by the standard methods.
We conclude that the estimates of Cohen and Glashow that the very energetic $\nu$’s should lose a large amount of their energy due to Cherenkov $e^+e^-$ radiation hold true for the observations of OPERA in Ref. [@opera] for superluminal $\nu$’s with quadratically dispersed values of $\delta$ as well. While showing this result, we have clarified the issue of the dependence on $\delta$ of this and the attendant related radiative processes.
[**Note Added in proof.**]{} – We understand [@matt3] that the difference noted in the text between our Eq.(\[eqnew2\]) and Eq.(21) in Ref. [@matt2] did not enter the respective results given in Ref. [@matt2].
[99]{} OPERA Collaboration, T. Adam [*et al.*]{}, arXiv:1109.4897. A.G. Cohen and S.L. Glashow, arXiv:1109.6562. D.M. Mattingly [*et al.*]{}, J. Cos. Astropart. Phys.[**1002**]{} (2010) 007. G. Amelino-Camelia [*et al.*]{}, arXiv:1110.0521 and references therein. Y. Huo [*et al.*]{}, arXiv:1111.4994, and references therein. J. M. Carmona and J.L. Cortes, arXiv:1110.0430, and references therein. Y. Huo [*et al.*]{}, arXiv:1112.0264, independently derive the values of $k$ and $k'$ in the Cohen-Glashow analysis. L. Maccione, S. Liberati and D.M. Mattingly, arXiv:1110.0783. J.D. Bjorken and S.D. Drell, [*Relavistic Quantum Mechanics*]{}, (McGraw-Hill, San Francisco, 1964) p.263. K. Hirata [*et al.*]{}, Phys. Rev. Lett. [**58**]{} (1987) 1490; R. M. Bionta [*et al.*]{}, [*ibid.*]{} [**58**]{} (1987) 1494; E. N. Alekseev [*et al.*]{}, Phys. Lett. B[**205**]{} (1988) 209; M. J. Longo, Phys. Rev. D[**36**]{}(1987) 3276; L. Stodolsky, Phys. Lett. B[**201**]{}(1988) 353. S. Mohanty and S. Rao, arXiv:1111.2725. S.L. Glashow, Nucl. Phys. [**22**]{} (1961) 579; S. Weinberg, Phys. Rev. Lett. [**19**]{} (1967) 1264; A. Salam, in [*Elementary Particle Theory*]{}, ed. N. Svartholm (Almqvist and Wiksells, Stockholm, 1968), p. 367; G. ’t Hooft and M. Veltman, Nucl. Phys. B[**44**]{},189 (1972) and [**50**]{}, 318 (1972); G. ’t Hooft, [*ibid.*]{} [**35**]{}, 167 (1971); M. Veltman, [*ibid.*]{} [**7**]{}, 637 (1968). L. Maccione [*et al.*]{} (private communication).
[^1]: The sum over the three neutrino species mildly over compensates for the almost equal coupling of the Z$^0$ boson to left and right handed electrons so that the rates would become almost equal if we multiply latter one by 2/3 – we ignore such factors here.
[^2]: By going to the ’preferred frame’ of the incident $\nu$, we see from the textbook results in Ref. [@bj-d] that, when we neglect all final state particle masses, the total rate to the $\nu e^-e^+$ final state is in fact independent of the handedness of the outgoing electron when $v_e$, the SM vector coupling of $e^-$ to the $Z^0$, is set to zero. Thus the arguments in Ref. [@matt2] as to why their splitting process rate has a different dependence on $\delta$ are not valid.
[^3]: Note that the strict definition of the average velocity, $<v>=\int dt v(t)/T\equiv L/\int dx/v(t)$, reduces to this result in (\[eqv1\]) in the limit in which we work wherein all velocities are close to 1.
[^4]: Here, $\parallel$ denotes the direction along that of the 3-momentum of the initial state $\nu$.
[^5]: Note that if one considers the case in which the $e^-, e^+$ also have the superluminal quadratic dispersive property, with the corresponding parameter $\eta_e$, then our analysis in eq.(\[eqfnl\]) still holds, with the shift of the RHS to the factor $\delta\left|\left((\frac{1}{2}-\frac{\lambda_{p'}^2}{6x^4})x^2+\frac{\lambda_{q'}^2}{6(1-x-y)^2}-(\eta_e/\eta_\nu)\frac{(1-x-y)^2}{2}\right)\right|$, so that in this case also the RHS of (\[mattg2\]) should be divided by a factor of $\delta$.
| {
"pile_set_name": "ArXiv"
} |
---
author:
- 'R. Sturm'
- 'F. Haberl'
- 'A. Rau'
- 'E. S. Bartlett'
- 'X.-L. Zhang'
- 'P. Schady'
- 'W. Pietsch'
- 'J. Greiner'
- 'M. J. Coe'
- 'A. Udalski'
bibliography:
- '../auto.bib'
- '../general.bib'
date: 'Received 22 February 2012 / Accepted 29 April 2012'
title: 'Discovery of the neutron star spin and a possible orbital period from the Be/X-ray binary IGR J05414-6858 in the LMC. [^1] '
---
Introduction {#sec:introduction}
============
[lcccccrrrr]{} & & & & & & & & &\
& & & & & & & & &\
XMM 2001 & 0094410101 & 2001-10-19 & 17:49 – 20:30 & EPIC-pn & ff–medium & 8.2 & 8.2 & $<$26 & 30\
& & & 17:09 – 20:29 & EPIC-MOS1 & ff–medium & 7.2 & 11.7 & $<$5 & 30\
& & & 17:09 – 20:29 & EPIC-MOS2 & ff–medium & 6.9 & 11.7 & $<$14 & 30\
XMM 2011 & 0679380101 & 2011-08-13 & 07:45 – 14:07 & EPIC-pn & ff–thin & 1.1 & 7.2 & 2018 & 21\
& & & 07:22 – 14:07 & EPIC-MOS1 & ff–medium & 1.2 & 9.4 & 863 & 29\
& & & 07:22 – 14:07 & EPIC-MOS2 & ff–medium & 1.2 & 9.4 & 954 & 31\
Swift 2010 a & 00031745001 & 2010-06-25 & 05:35 – 09:10 & XRT & pc & 5.3 & 3.9 & 221 & 40\
Swift 2010 b & 00031745002 & 2010-06-30 & 01:14 – 23:59 & XRT & pc & 2.0 & 5.2 & 497 & 50\
Swift 2011 a & 00045428001 & 2011-08-05 & 15:28 – 22:26 & XRT & pc & 10.5 & 3.3 & 40 & 40\
Swift 2011 b & 00031745003 & 2011-08-09 & 23:53 – 04:55 & XRT & pc & 1.8 & 3.7 & 120 & 40\
Swift 2011 c & 00031745004 & 2011-08-12 & 00:03 – 22:53 & XRT & pc & 2.4 & 4.1 & 100 & 40\
Swift 2011 d & 00031745005 & 2011-08-20 & 00:25 – 16:53 & XRT & pc & 3.1 & 2.1 & $<$6 & 40\
Swift 2011 e & 00031745006 & 2011-08-24 & 02:44 – 04:29 & XRT & pc & 2.5 & 1.2 & $<$5 & 40\
\[tab:xray-obs\]
Depending on the donor star, high mass X-ray binaries (HMXBs) are divided into super-giant systems and Be/X-ray binaries . In the latter case, a not well understood mechanism causes matter ejection of the Be star in the equatorial plane, leading to the build up of an equatorial decretion disc around the Be star [@2001PASJ...53..119O]. This disc dominates the emission of the system in infrared and some emission lines like H$\alpha$. The variability of this emission points to the instability of such discs. Due to a supernova kick, a neutron star (NS) can have an eccentric orbit around the Be star. During periastron passage the NS can accrete matter from the decretion disc, causing a so-called type I X-ray outburst, enduring several days at typical luminosities of $10^{36}$ erg s$^{-1}$. During disc instabilities, the NS can accrete a large fraction of the decretion disc, resulting in type II outbursts with luminosities up to $10^{37}$ erg s$^{-1}$ for several weeks.
The transient behaviour of Be/X-ray binaries and the wide extent of the Large Magellanic Cloud (LMC) on the sky, which imposes a low observational coverage by X-ray missions, complicate the discovery and investigation of Be/X-ray binaries in this galaxy. In contrast to that, the SMC was monitored with RXTE for about 14 years [@2008ApJS..177..189G]. Therefore, only nine HMXB X-ray pulsars are known to date in the LMC, which inhibits a statistical comparison of this sample with those of the Galaxy and the Small Magellanic Cloud (SMC). In the Galaxy and the SMC $\sim$66 and $\sim$55 HMXB pulsars are known, respectively. A major fraction of the pulsars is found in Be/X-ray binaries [e.g. @2010ASPC..422..224C]. Population studies of these systems are important to understand the stellar evolution, as they e.g. allow to estimate supernova kick velocities [@2005MNRAS.358.1379C] or the star formation history [@2010ApJ...716L.140A; @2011AN....332..349M]. Recently, a bimodal NS spin period distribution for the Galactic and SMC samples has been associated with two different types of supernovae [@2011Natur.479..372K]. To enable such statistical studies for the LMC, it is necessary to successively build up a larger sample of X-ray measurements of pulsars in outburst.
In 2010, was discovered serendipitously within INTEGRAL observations of SN1987A [@2010ATel.2695....1G] and later localised [@2010ATel.2696....1L] and identified as Be-X-ray binary [@2010ATel.2704....1R] with [*Swift*]{} and GROND follow-up observations. In 2011, [*Swift*]{} performed a UV survey of the LMC (PI: S. Immler). This provided a shallow coverage of this galaxy with the [*Swift*]{} X-ray telescope (XRT) and allowed the detection of bright X-ray transients. In an observation on 2011 Aug. 5 an outburst of was detected [@2011ATel.3537....1S], which allowed us to request an [XMM-Newton]{} target of opportunity (ToO) observation.
In this study, we report our analysis of the [XMM-Newton]{} observation of . The detection of the NS spin period adds the tenth X-ray pulsar in the LMC sample and a detailed X-ray spectral and temporal analysis allows a characterisation of the system. We compare our new results to those from archival [*Swift*]{} data and discuss complementary optical data to characterise the optical counterpart and the circum-stellar disc.
Observations and data reduction {#sec:observations}
===============================
XMM-Newton
----------
The [XMM-Newton]{} ToO observation was performed on 2011 Aug. 13. The source was observed on-axis, placed on CCD4 of EPIC-pn and CCD1 of both EPIC-MOS detectors. We used [XMM-Newton]{} SAS 11.0.0[^2] to process the data. Unfortunately, the observation was affected by an increased background caused by soft protons. During the first $\sim$11.5 ks of the $\sim$17 ks observation, the background was at a moderately elevated level, allowing the selection of time intervals where the background rate in the (7.0–15.0) keV band was below 50 cts ks$^{-1}$ arcmin$^{-2}$ for EPIC-pn and below 4 cts ks$^{-1}$ arcmin$^{-2}$ for EPIC-MOS. The detailed observation setup is recorded in Table \[tab:xray-obs\]. Here, we also list an [XMM-Newton]{} observation from 2001 covering the position of , in which the source was not detected. We used this observation to derive an upper limit for the flux. For EPIC-pn, we used single and double pixel events and single to quadruple events in the case of EPIC-MOS, all having [FLAG=0]{}. Background events were selected from a point source free area on the same CCDs as the source. Source events were extracted from a circle, with radius optimised for the signal-to-noise ratio by the SAS task [eregionanalyse]{}. We created spectra and response matrices with [especget]{} and used a binning to have at least a signal to noise ratio of 5 for each bin. For time series, the photon arrival times were randomised within the CCD frame time and calculated for the solar system barycentre.
Swift
-----
We re-analysed archival [*Swift*]{}/XRT observations. The spectra were created by using the [ftool]{}[^3] [xselect]{} to select events in the cleaned level 3 event files within a circle, placed on the source with radii given in Table \[tab:xray-obs\]. Background spectra were created from a circular extraction region with radius of 200. The spectra were binned to have $\geq$20 cts bin$^{-1}$. The ancillary response files were calculated with [xrtmkarf]{}. [*Swift*]{} observations of including non-detections, are also listed in Table \[tab:xray-obs\].
NIR, optical, and UV photometry
-------------------------------
Optical photometry of IGR J05414–6858 was obtained with OGLE, GROND and [*Swift*]{}/UVOT. The optical counterpart was monitored regularly during the Optical Gravitational Lensing Experiment (OGLE) of phase III [@2008AcA....58...69U] between October 2001 and April 2009 in the $I$-band. The source identification is OGLEIIILMC175.4.21714.
The Gamma-ray Burst Optical Near-ir Detector [GROND; @2008PASP..120..405G] at the MPG/ESO 2.2m telescope in La Silla, Chile, observed the source at three epochs in June 2010 and January 2012. Preliminary results of the 2010 observations were already presented in @2010ATel.2704....1R. GROND is a 7-channel imager that observes in four optical and three near-IR channels simultaneously. The IGR J05414–6858 data were reduced and analysed with the standard tools and methods described in [@2008ApJ...685..376K]. The photometry was obtained using point-spread-function (PSF) fitting taking into account the contamination from the two nearby sources (see Fig. \[fig:fc\]). Calibration was performed against observations of an SDSS standard star field ($g^{\prime}r^{\prime}i^{\prime}z^{\prime}$) or against selected 2MASS stars [@2006AJ....131.1163S] ($J H K_S$). This resulted in 1$\sigma$ accuracies of 0.04mag ($g^\prime z^\prime$), 0.03mag ($r^\prime i^\prime$), 0.05mag ($JH$), and 0.07mag ($K_S$).
The Ultraviolet/Optical Telescope (UVOT) aboard [*Swift*]{} has three optical ($v$ $b$ $u$) and three UV filters ($uvw1$ $uvm2$ $uvw2$). was observed in all UVOT filters during the pointed X-ray observations. For the LMC UV-survey observation 00045428001, the source is not in the field of view of the UVOT. Photometry was carried out on pipeline processed sky images downloaded from the [*Swift*]{} data centre[^4], following the standard UVOT procedure [@2008MNRAS.383..627P].
Optical spectroscopy
--------------------
Optical spectroscopy was taken with the ESO Faint Object Spectrograph (EFOSC2) mounted at the Nasmyth B focus of the 3.6m New Technology Telescope (NTT), La Silla, Chile on the nights of 2011 December 8 and 10. The EFOSC2 detector (CCD\#40) is a Loral/Lesser, Thinned, AR coated, UV flooded, MPP chip with 2048$\times$2048 pixels corresponding to 4.1$\times$4.1on the sky. The instrument was in longslit mode with a slit width of 1.5. Grisms 14 and 20 were used for blue and red end spectroscopy respectively. Grism 14 has a wavelength range of $\lambda\lambda3095$–$5085$ Å and a grating of 600 lines mm$^{-1}$ and a dispersion of 1 Å pixel$^{-1}$. The resulting spectra have a spectral resolution of $\sim12$ Å. Grism 20 is one of the two new Volume-Phase Holographic grisms recently added to EFOSC2. It has a smaller wavelength range, from 6047–7147 Å, but a superior dispersion of 0.55 Å pixel$^{-1}$ and 1070 lines mm$^{-1}$. This produced a spectral resolution for our red end spectra of $\sim$6 Å. Filter OG530 was used to block second order effects. The data were reduced using the standard packages available in the Image Reduction and Analysis Facility . Wavelength calibration was implemented using comparison spectra of Helium and Argon lamps taken throughout the observing run with the same instrument configuration. The spectra were normalized to remove the continuum and a redshift correction applied corresponding to the recession velocity of the LMC [-280 km s$^{-1}$, @2002LEDA.........0P].
Analyses and results of X-ray data {#sec:analyses}
==================================
X-ray coordinates {#sec:analyses:coord}
-----------------
We created X-ray images from all three EPIC cameras in the [XMM-Newton]{} standard energy sub-bands. A simultaneous source detection was performed on these images with [edetect\_chain]{}. The best-fit source position is RA (J2000) = 05412662 and Dec (J2000) = $-$6901230. The $1\sigma$ uncertainty of the position is 0.52, where we assume a systematic error of 0.5, which is quadratically added to the statistical error. The angular separation to the optical counterpart is 0.52 for the 2MASS position and 0.69 for the GROND position [star A, @2010ATel.2704....1R]. The distance to the [*Swift*]{} position of @2010ATel.2696....1L is 2.0 with an uncertainty in the [*Swift*]{} measurement of $\sim$3. A finding chart obtained from GROND data is presented in Fig. \[fig:fc\]. The white circle in the zoom-in gives the [XMM-Newton]{} position. The improved X-ray coordinates from the [XMM-Newton]{} observation further confirm the identification of the X-ray source with the optical counterpart.
Spectral analysis {#sec:analyses:spec}
-----------------
Spectral analysis was performed with [xspec]{} [@1996ASPC..101...17A] version 12.7.0. The three [XMM-Newton]{} EPIC spectra were fitted simultaneously and we included constant factors in the models to consider instrumental differences. For all models, we obtained consistent values of $C_{\rm MOS1}=1.11\pm0.08$ and $C_{\rm MOS2} = 1.10\pm0.08$ relative to EPIC-pn ($C_{\rm pn}=1$). Therefore, the fluxes for all instruments are consistent within uncertainties, as EPIC-MOS is known to derive $\sim$5% higher values compared to EPIC-pn. All other model parameters were forced to be the same for all instruments. The spectra are well described by an absorbed power-law. The photoelectric absorption was modelled by a fixed Galactic foreground column density of $N_{\rm H, Gal} = 6\times 10^{20}$ cm$^{-2}$ with abundances according to @2000ApJ...542..914W. An additional column density $N_{\rm H, LMC}$ was determined in the fit. It accounts for the interstellar medium of the LMC and source intrinsic absorption and the abundances were set to 0.5 for elements heavier than helium [@1992ApJ...384..508R].
We further tested the spectra for the existence of typical features of BeXRBs. A possible soft excess was modelled by a black-body (Model: PL+BB) or a multi-temperature disc black-body model (PL+DiscBB). The additional model component only improves the fit marginally, but demonstrates model-dependent uncertainties for the power-law parameters. The f-test probability for these additional components is around 13%. We use the black-body model to derive an upper limit for a soft component contribution. A soft excess can account only for $<$1% of the detected flux and up to 22.0% of the unabsorbed luminosity in the (0.2–10.0) keV band. The derived radius of the emission region from the black body is too large for an NS. With the disc model we obtain a similar inner radius for an inclination of $\Theta=0$ ($R_{\rm in} \propto (\cos\Theta)^{-1/2}$). Also according to @2004ApJ...614..881H, we obtain an inner disc radius of $R_{\rm in} = ( L_{\rm X} / 4\pi\sigma T^4)^{1/2} = 52$ km. Fluorescent iron line emission at 6.4 keV was modelled with a Gaussian line, having a fixed central energy and no measurable broadening. Also here, we receive only a marginal improvement of the fit and a line flux of $2.8\pm2.6 \times 10^{-6}$ photons cm$^{-2}$ s$^{-1}$. The spectrum and best-fit model are presented in Fig.\[fig:spec\_xmm\] and the best-fit results are listed in Table \[tab:spectra\]. All uncertainties and limits correspond to a 90% confidence level ($\Delta \chi^2 = 2.71$).
To investigate the spectral variability, we compared the EPIC spectra to the archival [*Swift*]{} spectra. The lower statistics only allowed to fit the power-law model. For the two spectra in 2010, $N_{\rm H, LMC}$ and $\Gamma$ are consistent within errors. Also, no significant evolution is found in the three [*Swift*]{} spectra of 2011. We assumed that the spectral shape is time-independent during the individual outbursts, and fitted the model simultaneously to the two 2010 and three 2011 [*Swift*]{} spectra (see Fig. \[fig:spec\_swift\]). The fluxes were determined for all observations individually. The results are also listed in Table \[tab:spectra\]. There is significant spectral variability observed between the outbursts of in 2010 and 2011, where the recent outburst exhibits a harder X-ray spectrum, consistent with the [XMM-Newton]{} 2011 observation. If we fit the [*Swift*]{} 2010 data with the spectral model derived from the [XMM-Newton]{} 2011 observation (allowing a re-normalisation), the fit degrades from $\chi^2_{\rm red} = 1.03$ to 1.75. In contrast to that, the $\chi^2_{\rm red}$ improves from 1.18 to 1.06 ($\chi^2=10.6$, dof$=10$), if we fit the Swift 2011 spectra with the [XMM-Newton]{}-derived model. This is caused by the increase of the degrees of freedom.
Since the lower photon index in 2011 is seen with [XMM-Newton]{} and [*Swift*]{}, this is unlikely caused by the high background during the [XMM-Newton]{} observation or by instrumental differences.
[rlccccccccc]{} & & & & & & & & & &\
& & & & & & & & & &\
XMM 2011 & PL & $<$0.89 & $0.32_{-0.05}^{+0.03}$ & – & – & – & $3.01_{-0.32}^{+0.24}$ & 0.91 & 0.92 & 118\
& PL+BB & $4.98_{-3.8}^{+3.4}$ & $0.40_{-0.09}^{+0.07}$ & $112_{-23}^{+56}$ & $62_{-46}^{+66}$ & – & $2.98_{-0.35}^{+0.20}$ & 0.98 & 0.91 & 116\
& PL+BB+Fe & $4.98_{-3.6}^{+3.6}$ & $0.41_{-0.09}^{+0.07}$ & $111_{-23}^{+49}$ & $62_{-45}^{+73}$ & $78_{-70}^{+71}$ & $2.97_{-0.47}^{+0.37}$ & 0.98 & 0.88 & 115\
& PL+DiscBB & $3.68_{-1.4}^{+2.6}$ & $0.39_{-0.08}^{+0.08}$ & $128_{-33}^{+85}$ & $49_{-44}^{+134}$ & – & $2.98_{-0.33}^{+0.16}$ & 1.01 & 0.91 & 116\
Swift 2010 a & & & & & & & $5.6_{-1.6}^{+1.1}$ & 1.77 & &\
Swift 2010 b & & & & & & & $9.2_{-1.9}^{+1.3}$ & 2.91 & &\
Swift 2011 a & & & & & & & $5.4_{-5.4}^{+2.6}$ & 1.62 & &\
Swift 2011 b & PL & $<$2.6 & $0.25_{-0.22}^{+0.14}$ & – & – & – & $5.1_{-4.5}^{+1.3}$ & 1.51 & 1.18 & 8\
Swift 2011 c & & & & & & & $3.5_{-3.1}^{+9.8}$ & 1.04 & &\
\[tab:spectra\]
Pulsations {#sec:analyses:pulsation}
----------
A strong signal at $\omega = 0.2262$ Hz and its first harmonic appeared in a Fast Fourier Transformation (FFT) of the EPIC-pn time series in the (0.2–10.0) keV band. The power density spectrum is plotted in Fig. \[fig:pds\]. The signal is also clearly present in the (0.2–2.0) keV and (2.0–10.0) keV sub-bands. The period is not resolved by EPIC-MOS (2.6 s frame time), since the period is shorter than twice the frame time of the instrument, i.e. $\omega$ is above the Nyquist frequency. The same holds for the [*Swift*]{} data (2.5 s frame time). A $\chi^2$ test, a Bayesian odds ratio [@1996ApJ...473.1059G], and a Rayleigh Z$^2$ test for one harmonic around the periodicity signal are shown in Fig. \[fig:period\_sta\]. All tests independently confirm the pulse period. Following , we used the Bayesian detection method to determine the pulse period and a 1$\sigma$ uncertainty of 4.420866(2) s on 2011-08-13.
Fig. \[fig:pp\] shows the folded background-subtracted light curves from EPIC-pn in the total (0.2 – 10.0) keV band and the standard sub-bands (0.2 – 0.5) keV, (0.5 – 1.0) keV, (1.0 – 2.0) keV, (2.0 – 4.5) keV, and (4.5 – 10.0) keV, where we merged the first two bands, to increase the statistics. Also hardness-ratio (HR) variations are presented. HRs are defined by HR$_{i}$ = (R$_{i+1}$ $-$ R$_{i}$)/(R$_{i+1}$ + R$_{i}$) with R$_{i}$ denoting the background-subtracted count rate in the standard energy band $i$ (with $i$ from 1 to 4). In the light curves, two narrow peaks are seen within one period having only small variations in energy. By modelling the (0.2 – 10.0) keV curve with a non-pulsating contribution and one Gaussian for each peak, we estimate a pulsed fraction of (48$\pm$7)% in the total flux, and a flux ratio for both peaks of 2.3$\pm$0.5.
A search for pulsations in the INTEGRAL ISGRI observations of 2010 was performed. We detect the source at $0.28\pm0.04$ cts s$^{-1}$ in the (20–40) keV band, but could not find a significant period in the power density spectrum or variability in the 4.4208 s folded light curve. This might be caused by binary orbital modulations, as the INTEGRAL observations cover a long time.
X-ray flux variability {#sec:analyses:time}
----------------------
In addition to the fluxes as measured in Sec. \[sec:analyses:spec\], we calculated upper limits for non-detections of . The field was observed by [XMM-Newton]{} in 2001, but no source was detected by [edetect\_chain]{} (analogous to Sec. \[sec:analyses:coord\]). Spectra were extracted in the same manner as described in Sec. \[sec:observations\] from a 30 source region and a 50 background region. Using C statistics and the spectral shape as determined with [XMM-Newton]{} in 2011, we derived a 90% upper limit for the flux of 9.0$\times 10^{-14}$ erg cm$^{-2}$ s$^{-1}$. This lowest upper limit results in a variability of at least a factor of 100, compared to the maximum flux measured in the [*Swift*]{} 2010b observation. Prior to that, we could not find any corresponding X-ray detection. There is no ROSAT source within 1 listed in the literature.
The [*Swift*]{} monitoring of the recent outburst, determined the turn-off between MJD 55786 and 55793. For the two non-detections, we extracted spectra in the same way as above and determined 90% confidence upper limits of $9.5\times 10^{-13}$ erg cm$^{-2}$ s$^{-1}$ and $2.3\times 10^{-12}$ erg cm$^{-2}$ s$^{-1}$. The long term evolution of the system is presented in Fig. \[fig:lc\].
The variability during the [XMM-Newton]{} observation in 2011 was at a moderate level. We created a background corrected light curve, merged using all EPIC instruments and binned to 70 s, corresponding to 30 cts bin$^{-1}$ on average. A $\chi^2$ test of this light curve against a constant resulted in $\chi^2/{\rm dof} = 130/121$.
Analysis and results of optical data
====================================
Using GROND, the counterpart of the [*Swift*]{} X-ray source was resolved into three point sources. Star A of @2010ATel.2704....1R was originally classified as of spectral type B1-2 III from GROND and [*Swift*]{}/UVOT photometry and is the most likely counterpart within the improved [XMM-Newton]{} X-ray position uncertainty of .
The possible northern counterpart at 05412657 $-$6901217 has a somewhat larger angular separation to the X-ray position of 1.3 (2.6$\sigma$). The $I$-band light curve derived from OGLE is flat. GROND photometry ($g^\prime = 17.54 \pm 0.08$, $r^\prime = 17.55 \pm 0.05$, $i^\prime = 17.12 \pm 0.05$, $z^\prime = 16.89 \pm 0.05$, $J = 16.49 \pm 0.05$, $H = 16.37 \pm 0.05$, and $K_s = 16.62 \pm 0.08 $) can be modelled by emission from a K star. This is too faint and too red for a Be star. Depending on its luminosity class, the star can be located in the LMC or the Galaxy. In the latter case, coronal X-ray emission would be significantly softer than observed for thus we reject this star as possible optical counterpart in the following. Since there is no soft source detected in the 2001 [XMM-Newton]{} observation to a limit of 0.005 cts s$^{-1}$, X-ray emission of this star is unlikely contributing to the X-ray spectrum of .
Photometry
----------
The OGLE III $I$-band light curve of this star is presented in Fig. \[fig:lc\]. It exhibits two different brightness states with a transition phase in between. Before MJD 54400 (left dotted line in Fig. \[fig:lc\]), the $I$-band emission is at (15.51–15.43) mag, followed by an impetuous increase to the high state. After MJD 54650 (right dotted line in Fig. \[fig:lc\]), the $I$-band emission is in between (15.24 – 15.13) mag.
A Lomb-Scargle periodogram of the low and high state of the OGLE III light curve is shown in Fig. \[fig:ls\]. The high state reveals a periodicity with 19.9 days. The folded light curve of the high state is presented in Fig. \[fig:opt\_lc\_conv\]. Dotted and dashed lines mark the phases of X-ray detections and non detections of 2011, respectively.
-------- ------------------ ------------------
$v$ 2010-06-25 06:09 15.47 $\pm$ 0.06
07:51 15.51 $\pm$ 0.06
2011-08-10 00:36 15.11 $\pm$ 0.04
02:15 15.20 $\pm$ 0.05
03:53 15.10 $\pm$ 0.05
2011-08-13 23:07 15.13 $\pm$ 0.04
$b$ 2010-06-25 06:02 15.36 $\pm$ 0.04
07:45 15.43 $\pm$ 0.04
09:28 15.34 $\pm$ 0.04
2011-08-10 00:24 15.08 $\pm$ 0.03
02:06 15.08 $\pm$ 0.03
03:48 15.07 $\pm$ 0.04
2011-08-13 22:55 15.07 $\pm$ 0.03
2011-08-20 17:13 15.10 $\pm$ 0.03
$u$ 2010-06-25 06:00 15.41 $\pm$ 0.03
07:43 15.38 $\pm$ 0.03
09:26 15.35 $\pm$ 0.03
2011-08-10 00:21 15.06 $\pm$ 0.03
02:05 15.04 $\pm$ 0.03
03:46 15.05 $\pm$ 0.03
05:17 15.05 $\pm$ 0.05
2011-08-13 22:52 15.10 $\pm$ 0.03
2011-08-20 17:11 15.12 $\pm$ 0.03
$uvw1$ 2010-06-25 05:58 15.45 $\pm$ 0.03
07:41 15.44 $\pm$ 0.03
09:24 15.44 $\pm$ 0.03
2011-08-10 00:17 15.20 $\pm$ 0.03
02:02 15.25 $\pm$ 0.03
03:45 15.24 $\pm$ 0.03
05:15 15.26 $\pm$ 0.03
2011-08-13 22:49 15.31 $\pm$ 0.03
2011-08-20 17:08 15.22 $\pm$ 0.03
$uvm2$ 2010-06-25 06:19 15.56 $\pm$ 0.02
07:59 15.57 $\pm$ 0.03
2010-07-01 00:13 15.70 $\pm$ 0.03
2011-08-10 00:41 15.29 $\pm$ 0.03
02:18 15.41 $\pm$ 0.03
03:55 15.38 $\pm$ 0.03
2011-08-13 00:31 15.33 $\pm$ 0.03
02:06 15.33 $\pm$ 0.02
23:12 15.43 $\pm$ 0.03
$uvw2$ 2010-06-25 06:05 15.60 $\pm$ 0.03
07:48 15.63 $\pm$ 0.03
09:30 15.61 $\pm$ 0.03
2010-06-30 01:40 15.59 $\pm$ 0.02
03:21 15.97 $\pm$ 0.02
04:49 15.61 $\pm$ 0.02
17:56 15.71 $\pm$ 0.02
19:32 15.70 $\pm$ 0.02
22:35 15.77 $\pm$ 0.03
2011-08-10 00:30 15.36 $\pm$ 0.02
02:11 15.48 $\pm$ 0.02
03:50 15.42 $\pm$ 0.03
2011-08-13 23:01 15.54 $\pm$ 0.02
2011-08-20 01:00 15.41 $\pm$ 0.02
17:15 15.38 $\pm$ 0.04
2011-08-24 03:10 15.44 $\pm$ 0.02
04:46 15.45 $\pm$ 0.02
-------- ------------------ ------------------
: Swift/UVOT photometry.
\[tab:uvot\_log\]
The [*Swift*]{}/UVOT photometry is presented in Table \[tab:uvot\_log\]. Within uncertainties most values are constant. For the 2010 measurements, we see an indication for a flux decrease in the $uvm2$ and $uvw2$ magnitudes of $\sim$0.13$\pm$0.04 mag and $\sim$0.17$\pm$0.04 mag, respectively. Other magnitudes in 2011 are constant, except a possible short increase in the UV on 2011-08-13 between 22:49 and 23:12 by $\sim$0.1 mag. These are of the same order as the variations observed by OGLE during high and low state.
By comparing the averaged magnitudes of 2010-06-25 and those of 2011, we find a stronger flux increase as observed with OGLE III in 2008, with $\Delta v = -0.36$ mag, $\Delta b = -0.30$ mag, $\Delta u = -0.31$ mag, $\Delta uvw1 = -0.20$ mag, $\Delta uvm2 = -0.20$ mag, and $\Delta uvw2 = -0.18$ mag. This indicates a further transition from low to high state between June 2010 and August 2011.
The GROND magnitudes are summarised in Table \[tab:grond\_log\]. Due to pointing constraints, the first observation was performed only in the $J$, $H$, and $K_s$ bands, while the remaining epochs cover all seven channels. Note the increase in magnitude between 2010-06-26 and 2010-06-29 of $\Delta J = (0.25\pm0.07)$ mag, $\Delta H = (0.45\pm0.07)$ mag, and $\Delta K_s = (0.42\pm0.11)$ mag within three days. This affirms the observed drop of the flux as observed with [*Swift*]{}/UVOT.
The fit to the spectral energy distribution composed from the first GROND epoch together with averaged UVOT photometry from 2010 June 25th, is shown in Fig. \[fig:grond+uvotsed\]. Here, all magnitudes were corrected for Galactic reddening of E$_{\rm B-V}=0.075$mag [@1998ApJ...500..525S] using the @1989ApJ...345..245C extinction law and for the LMC-intrinsic reddening of E$_{\rm B-V}=0.15$ mag (see Sec. \[specclass\]) using the @1992ApJ...395..130P extinction law. The data are best fit with a hot ($\approx31\,000$ K) black-body spectrum, consistent with the B0-1 III stellar classification suggested by the optical spectroscopy (Sec. \[specclass\]). We note the existence of a clear excess in the near-IR bands.
------------------ ---------------- ---------------- ---------------- ---------------- ---------------- ---------------- ----------------
UT Date
$g^\prime$ $r^\prime$ $i^\prime$ $z^\prime$ $J$ $H$ $K_s$
2010-06-26 10:51 – – – – $15.93\pm0.05$ $16.07\pm0.05$ $16.29\pm0.07$
2010-06-29 10:29 $15.34\pm0.04$ $15.60\pm0.03$ $15.82\pm0.03$ $15.98\pm0.04$ $16.17\pm0.05$ $16.53\pm0.05$ $16.71\pm0.08$
2012-01-25 04:14 $15.18\pm0.04$ $15.41\pm0.03$ $15.57\pm0.03$ $15.72\pm0.04$ $15.96\pm0.05$ $16.28\pm0.05$ $16.54\pm0.08$
------------------ ---------------- ---------------- ---------------- ---------------- ---------------- ---------------- ----------------
\[tab:grond\_log\]
![UV-near-IR SED composed of [*Swift*]{}/UVOT observations obtained on 2010 June 25th and GROND data taken on 2010 June 26th. The dotted line shows the simplified best-fit black-body model indicating that the UVOT photometry is consistent with the B0-1 stellar classification derived from the spectroscopy.[]{data-label="fig:grond+uvotsed"}](sed_grondUVOT.eps){width="50.00000%"}
Spectral classification {#specclass}
-----------------------
OB stars in our own galaxy are classified using the ratio of certain metal and helium lines [@1990PASP..102..379W] based on the Morgan-Keenan [MK; @1943QB881.M6.......] system. However, this is unsuitable in lower metallicity environments as the metal lines are either much weaker or not present. As such, the optical spectrum of IGR J05414-6858 was classified using the method developed by for B-type stars in the SMC and implemented for the SMC, LMC and Galaxy by . This system is normalized to the MK system such that stars in both systems show the same trends in their line strengths. The luminosity classification method from @1990PASP..102..379W was assumed in this work.
{height="180mm"}
Fig. \[fig:blue\] shows the unsmoothed optical spectrum of IGR J04514-6858. The spectrum is dominated by the hydrogen Balmer series and neutral helium lines. The line at $\lambda4143$ Å is stronger than the $\lambda4200$Å, which means the star is later than type O9. $\lambda4686$ Å is also clearly present implying that the optical counterpart of IGR J04514-6858 is earlier than type B1.5, although the He<span style="font-variant:small-caps;">ii</span> $\lambda4541$ Å line is not visible above the noise level of the data. There is also evidence for the $\lambda4116$ Å line - consistent with a spectral classification of B1. However @1990PASP..102..379W present spectra of B0 type stars with clear Si lines indicating that a spectral classification of B0 is not ruled out by their presence. We note that there does not appear to be any evidence for the $\lambda4088$ Å line. Be stars are characterised by their rapid rotation velocity and it could be that this line is being concealed by the rotationally broadened H$\delta$ line in close proximity.
![Spectrum of IGR J04514-6858 in the wavelength range $\lambda\lambda$6000–7000Å with the NTT on 2011-12-10. The spectrum has been smoothed with a boxcar average of 3, normalized to remove the continuum and shifted by -280 km s$^{-1}$.[]{data-label="fig:red"}](red_spec.eps){height="90mm"}
The luminosity class of the system was determined using the ratios of $\lambda4116$/ $\lambda4121$, $\lambda4121$/ $\lambda4143$ and $\lambda4686$/ $\lambda4713$. The first two ratios strengthen with decreasing luminosity class (i.e. with increasing luminosity), whereas the latter ratio decreases with increasing luminosity. The relative strengths of these lines suggest a luminosity class III, making our spectral classification of B0-1 III consistent with that obtained photometrically with the GROND and [*Swift*]{}/UVOT data. To check the spectral identification we can compare the observed optical magnitudes with that predicted for a B0-1 III star in the LMC. Taking the faintest, least disc-contaminated $V$-band magnitude from Table \[tab:uvot\_log\] of $V=(15.51\pm0.06)$ mag, a distance modulus of 18.5$\pm$0.1 [@2009AJ....138....1K] and a reddening of E$_{\rm B-V}=0.15$ mag reveals an absolute magnitude for the star of $M_V= (-3.5\pm0.2)$. Such a value would be consistent with a B0.5 III star [@2006MNRAS.371..185W] and hence confirms the classification deduced from the spectra reported here. We note that we cannot use the derived X-ray absorption reported here to refine the column to the star because of the large uncertainties in that value.
Fig. \[fig:red\] shows the red end of the spectrum of IGR J04514-6858 taken near-simultaneously. The H$\alpha$ equivalent width, considered an indicator for circumstellar disc size, is relatively small at $-(3.2\pm0.6)$Å which is consistent with the lack of H$\beta$ in emission in Fig. \[fig:blue\]. The double peaked, asymmetric nature of the line profile shows a *V/R* pattern consistent with global one armed oscillations (GOAO) and suggests that the circumstellar disc of the star is inclined to the line of sight. This is not uncommon in the circumstellar discs of Be stars.
Discussion and conclusions {#sec:discussion}
==========================
We performed an [XMM-Newton]{} ToO observation of in August 2011, allowing us to measure the X-ray spectrum and discover the spin period. This adds the tenth known HMXB pulsar to the LMC sample and confirms the neutron star nature of the compact object.
The source was found in INTEGRAL observations performed on 2010 May 13–22 (130 ks) and June 6–14 (400 ks) at an average flux of $8\times10^{36}$ erg s$^{-1}$ in the (20–40) keV band [@2010ATel.2695....1G]. In the Swift follow-up observation on June 30 we see the source still in outburst at a luminosity of $\sim$$3\times10^{36}$ erg s$^{-1}$ in the (0.2–10.0) keV band. If these detections correspond to the same outburst, the duration of X-ray bright state would suggest a type II outburst. The luminosity, derived from INTEGRAL is at the lower limit for a classical type II outburst, but might have been higher at maximum. The luminosity and observed duration of the 2011 outburst are in agreement with a type I outburst, but since the time of the beginning of the outburst is unknown, we cannot exclude a type II outburst.
Furthermore, the spectrum of the current outburst was found to be significantly harder than in 2010. The power-law photon index of the 2011 outburst of $\Gamma$ = 0.3–0.4 is also relatively low, compared to the distribution known from the SMC sample ($\Gamma\sim1$). We note that also for a few BeXRBs in the SMC hard spectra were observed, as for the pulsars XTEJ0103-728 and XMMUJ004814.0-732204 . The first one of these also was detected in a type II outburst with INTEGRAL [@2010MNRAS.403.1239T]. Moreover, for both SMC pulsars an indication of a soft excess with comparable emission radius was found and suggested to originate from the accretion disc. The system intrinsic absorption strongly depends on the modelling ($0-5\times 10^{21}$ cm$^{-2}$) and the location of the system in the LMC. The total LMC column density along the line of sight at the position of is $3.6\times 10^{21}$ cm$^{-2}$ [@2003ApJS..148..473K].
According to the Corbet relation , we expect the orbital period of the system in the range of (1 – 100) days for the measured spin period of . The periodic variations seen in the $I$-band are therefore likely caused by binarity. Assuming masses of 21.5[M$_{\odot}$]{} and 1 [M$_{\odot}$]{} for the Be star and the NS, respectively, the orbital period implies a semi-major axis of the binary system of 0.41 AU, corresponding to $\sim$6 stellar radii [@1996ApJ...460..914V]. The X-ray detections in 2011 occurred during the bright phase of the folded $I$-band light curve (see dotted lines in Fig. \[fig:opt\_lc\_conv\]). In 2011, the X-ray luminosity follows the $I$-band emission and the non-detections were during low $I$-band emission. A correlation of X-ray and optical outbursts was e.g. reported for AX J0058-720 . In the case of the 2010 outburst, the two [*Swift*]{} X-ray detections are during low $I$-band emission at phase 0.80 and 0.04. This is either due to the fact, that type II outbursts are not correlated to the orbital phase, or that the optical period is not caused by binarity.
Long-term optical variation with different variability patterns are typical for BeXRBs [@2011MNRAS.413.1600R]. An explanation for the transition from low to high state by $\Delta I \sim 0.3$ mag might be the build up of a decretion disc around the Be star. The [XMM-Newton]{} upper limit (X-ray faint state) in 2001 was during the optical low state, where probably no decretion disc was present. From other BeXRBs , a correlation between NIR and optical magnitudes is seen. Therefore it is likely that the long term variability seen in the OGLE $I$-band extends from NIR to UV. As indicated by the GROND and [*Swift*]{}/UVOT observations, the system still undergoes strong variations in the NIR, optical and UV. A rapid strong drop in the NIR emission was observed between June 2010 26th and 29th. At this time, the source was still in a presumable type II X-ray outburst, which can be followed by a disc loss phase. The NIR-flux decrease further supports a type II outburst. Also, the $K_s$-band magnitude was high, compared to the GROND January 2012 observation, while $J$ was at the same level again pointing to the presence of a circumstellar disc. This is further supported by the H$\alpha$ line emission in December 2011. Unfortunately, the end of the 2010 X-ray outburst is not constrained. During the 2011 outburst, the [*Swift*]{}/UVOT observations do not indicate any strong changes in the optical. Forthcoming OGLE IV data will allow to extend the light curve and to confirm the periodicity.
IGRJ05414-6858 is the tenth known HMXB pulsar in the LMC with $P_{\rm spin} = 4.4208$ s. The optical counterpart was classified as of spectral type B0-1 IIIe and shows double-peaked H$\alpha$ emission and a variable NIR excess. A likely orbital period of $P_{\rm orb} = 19.9$ d was found. The two observed X-ray outbursts demonstrate the importance of optical monitoring during outbursts, to better understand the accretion process in these systems. To further increase the sample of explored HMXB in the LMC, further X-ray observations triggered during an outburst are necessary.
We thank the [XMM-Newton]{} team for scheduling the ToO observation. The XMM-Newton project is supported by the Bundesministerium für Wirtschaft und Technologie/Deutsches Zentrum für Luft- und Raumfahrt (BMWI/DLR, FKZ 50 OX 0001) and the Max-Planck Society. Part of the funding for GROND (both hardware as well as personnel) was generously granted from the Leibniz-Prize to Prof. G. Hasinger (DFG grant HA 1850/28-1). The OGLE project has received funding from the European Research Council under the European Community’s Seventh Framework Programme (FP7/2007-2013) / ERC grant agreement no. 246678 to AU. We acknowledge the use of public data from the [*Swift*]{} data archive. X-LZ acknowledges financial support by DLR FKZ 50 OG 0502. RS acknowledges support from the BMWI/DLR grant FKZ 50 OR 0907.
[^1]: Based on observations with XMM-Newton, an ESA Science Mission with instruments and contributions directly funded by ESA Member states and the USA (NASA)
[^2]: Science Analysis Software (SAS), http://xmm.vilspa.esa.es/sas/
[^3]: http://heasarc.nasa.gov/ftools/
[^4]: http://www.swift.ac.uk/swift\_portal
| {
"pile_set_name": "ArXiv"
} |
---
abstract: |
In a balls-in-bins process with feedback, balls are sequentially thrown into bins so that the probability that a bin with $n$ balls obtains the next ball is proportional to $f(n)$ for some function $f$. A commonly studied case where there are two bins and $f(n) = n^p$ for $p > 0$, and our goal is to study the fine behavior of this process with two bins and a large initial number $t$ of balls. Perhaps surprisingly, Brownian Motions are an essential part of both our proofs.\
For $p>1/2$, it was known that with probability $1$ one of the bins will lead the process at all large enough times. We show that if the first bin starts with $t+\lambda\sqrt{t}$ balls (for constant $\lambda\in \R$), the probability that it always or eventually leads has a non-trivial limit depending on $\lambda$.\
For $p\leq 1/2$, it was known that with probability $1$ the bins will alternate in leadership. We show, however, that if the initial fraction of balls in one of the bins is $>1/2$, the time until it is overtaken by the remaining bin scales like $\Theta({t^{1+1/(1-2p)}})$ for $p<1/2$ and $\exp(\theita{t})$ for $p=1/2$. In fact, the overtaking time has a non-trivial distribution around the scaling factors, which we determine explicitly.\
Our proofs use a continuous-time embedding of the balls-in-bins process (due to Rubin) and a non-standard approximation of the process by Brownian Motion. The techniques presented also extend to more general functions $f$.
author:
- 'Roberto Oliveira[^1]'
bibliography:
- 'binb\_brown.bib'
title: 'Balls-in-bins with feedback and Brownian Motion'
---
Introduction {#sec:intro}
============
The process
-----------
If $f:\N\to (0,+\infty)$ is a positive function, a [*balls-in-bins process with feedback function $f$*]{} a discrete-time Markov process with $B$ bins, each one of which containing $I_i(m)>0$ balls at time $m$ for each $m\in \{0,1,2,\dots\}$ and $i\in\{1,\dots,B\}$. Its evolution is as follows: at each time $m>0$, a ball is added to a bin $i_m$, so that $I_{i_m}(m)=I_{i_m}(m-1)+1$ and $I_j(m)=I_j(m-1)$ for all $i\in\{1,\dots,B\}\backslash\{i_m\}$, and the random choice of bin $i_m$ has distribution $$\label{eq:recipe}\Pr{i_{m}=i\mid \{I_j(m-1)\,:\,{1\leq j\leq B}\}}=
\frac{f(I_i(m-1))}{\sum_{j=1}^B f(I_j(m-1))}\;\; (1\leq i\leq
B),$$We will commonly refer to this recipe by saying that [*bin $i$ receives a ball at time $m$ (i.e. $i_m=i$) with probability proportional to $f(I_m(i))$*]{}.
Such processes[^2] were introduced to the Discrete Mathematics community by Drinea, Frieze and Mitzenmacher [@Drinea02], where they were motivated by economical problems of competition and mathematically related preferential attachment models for large networks[^3]. That paper treats only the case where $f(x)=x^p$ for some parameter $p>0$. In this case $f$ is increasing, and therefore [*the rich get richer*]{}: the more balls a bin has, the more likely it is to receive the next ball. In economic terms, one could think of bins as competing products and balls as customers; in that case, the more popular a product is, the more likely it is to obtain a new customer.
The main question addressed in that paper is whether this phenomenon ensures large advantages [*in the long run*]{} for some bin. The authors show that if $p>1$, there almost surely exists one bin that gets all but a negligible fraction of the balls in the large-time limit; whereas for $p<1$, the asymptotic fractions of balls in each bin even out with time. The $p=1$ case is the classic Pólya Urn model, for which it has been long known that the number of balls in each bin converges almost surely to a non-degenerate random variable, and thus the process has different [*regimes*]{} depending on the choice of parameter $p$.
The three regimes
-----------------
However, a much stronger result is available. A paper by Khanin and Khanin [@Khanin01] introduced what amounts to the same process as a model for neuron growth, and proved that if $p>1$, there almost surely is some bin that gets [*all but finitely many balls*]{}, an event that we call [*monopoly*]{}. They also show that for $1/2<p\leq 1$, monopoly has probability $0$, but there almost surely will be some bin which will lead the process from some finite time on (we call this [*eventual leadership*]{}), whereas this cannot happen if $0<p\leq 1/2$. Therefore, the balls-in-bins process has [*three regimes*]{} of behavior corresponding to three ranges of $p$.
In fact, the result of [@Khanin01] generalizes to any $f$ with $\min_{x\in\N}f(x)>0$. Consider the following events, which we call [*monopoly by bin $i\in[B]$*]{}, $$\begin{aligned}
\label{eq:defmon}\mon_i &\equiv& \{\exists M\in \N\;\forall m\geq M\;\;\forall j\in [B]\;\;j\neq i\Rightarrow I_m(j)=I_M(j)\}\\ \nonumber &=&\{\exists M\in \N\;\forall m\geq M\;\;i_{m}=i\};\end{aligned}$$ and [*eventual leadership by bin $i\in[B]$*]{}: $$\label{eq:defelead}\elead_i \equiv \{\exists M\in \N\;\forall m\geq M\;\;\forall j\in [B]\;\;j\neq i\Rightarrow I_m(j)<I_m(i)\}.$$ Clearly, $\elead_i\supset \mon_i$. A not-too-difficult extension of [@Khanin01] (proven in [@Spencer??; @tese]) says that
\[thm:Khanin\]If $\{I_m\}_{m=0}^{+\infty}$ is a balls-in-bins process with $B$ bins and feedback function $f=f(x)\geq c$ for some $c>0$, then there are three mutually exclusive possibilities.
1. $\sum_{n\geq 1}f(n)^{-1} < +\infty$, in which case $\Pr{\cup_{i=1}^B\,\mon_i} = \Pr{\cup_{i=1}^B\,\elead_i}=1$ (we call this the [*monopolistic*]{} regime);
2. $\sum_{n\geq
1}f(n)^{-1} = +\infty$ but $\sum_{n\geq 1}f(n)^{-2} < +\infty$, in which case $\Pr{\cup_{i=1}^B\,\mon_i} = 0$ but $\Pr{\cup_{i=1}^B\,\elead_i}=1$ (this is the [*eventual leadership regime*]{});
3. $\sum_{n\geq 1}f(n)^{-2} =+\infty$, in which case $\Pr{\cup_{i=1}^B\,\mon_i} =
\Pr{\cup_{i=1}^B\,\elead_i}=0$ (this is the [*almost-balanced regime*]{}).
This holds irrespective of the initial conditions of the process.
Notice that the three cases of the Theorem applied to the $f(x)=x^p$ family correspond to $p>1$, $1/2<p\leq
1$ and $0\leq p<1/2$; in other words, this family of $f$ has [*phase transitions*]{} at $p=1$ and $p=1/2$. We sketch a proof of this result in , both for completeness and to give readers a better acquaintance with the techniques in the present paper.
The present work
----------------
This paper is part of a series by the present author in collaboration with Michael Mitzenmacher and Joel Spencer [@MitzenmacherOS04; @Oliveira05; @avoid] that attempts to quantify different aspects of the three qualitative regimes presented in . Our specific purpose in the present paper is to prove two not-quite-related results about these processes in different regimes, when the initial number of balls on both bins is large. What brings these two results is that both proofs use Brownian Motion in an unexpected and surprising way.
Our first result is a [*scaling result for the eventual leadership and monopoly regimes*]{}. Suppose, for simplicity, that $f(x)=x^p$ with $p>1/2$. Recall the definition of eventual leadership by bin $i$ , and let $\lead_i$ be the event that bin $1$ leads the process at [*all times*]{}: $$\label{eq:deflead}\lead_i \equiv \{\forall m\geq 0\;\; \forall j\in[B]\;\; j\neq i\Rightarrow I_m(j)<I_m(i)\}.$$
If the initial number of bins is $t\gg 1$ and $I_0(1)\approx t/2$, then $\Pr{\lead_i}\approx 0$ and $\Pr{\elead_i}\approx 1$ for $i=1,2$. On the other hand, if $I_0(1)$ is much larger than $I_0(2)$, one think that $\Pr{\lead_1},\Pr{\elead_1}\approx 1$. Our question is [*how large is large enough*]{}? That is to say, at what [*scale*]{} do these two probabilities grow from $0$ to $1$? We show that the answer is in fact $\theita{\sqrt{t}}$, and give an exact asymptotic result.
\[thm:scaling\_cor\]Let $\{\lambda_t\}_{t\in\N}\subset\R$ form a sequence such that such that $$\lambda\equiv \lim_{t\to +\infty}\lambda_t \in\R\mbox{ exists}$$ and $$\forall t\in\N,\;\;\frac{t}{2}\pm \lambda_t\sqrt{\frac{t}{4p-2}}\in \N.$$ Assume that for each $t$, $\{I^{(t)}_m(j)\,:\,j\in[2],m\geq 0\}$ is a balls-in-bins process with two bins, feedback function $f(x)=x^p$, $1/2<p<+\infty$ and initial conditions $$(x(t),y(t))\equiv
\left(\frac{t}{2}+\lambda_t\sqrt{\frac{t}{4p-2}},\frac{t}{2}-\lambda_t\sqrt{\frac{t}{4p-2}}\right).$$ Then $$\begin{aligned}
\lim_{t\to +\infty}\Prp{(x(t),y(t))}{\elead_1} &=& \gamma(\lambda)\equiv \sqrt{\frac{1}{2\pi}}\int_{-\infty}^{\lambda}e^{-x^2/2}\,dx,\\
\lim_{t\to +\infty}\Prp{(x(t),y(t))}{\lead_1} &=&
\Gamma(\lambda)\equiv\sqrt{\frac{2}{\pi}}\int_{0}^{\max\{\lambda,0\}}e^{-x^2/2}\,dx.\end{aligned}$$
This theorem is an extension of a result by Mitzenmacher, Oliveira and Spencer [@MitzenmacherOS04], who showed a similar scaling for $\mon_1$ when $p>1$. That paper used Esséen’s inequality for approximation by Gaussians together with a continuous-time embedding of the balls-in-bins process; we shall also use the latter device together with approximation by Brownian Motion, especially to estimate $\Prp{(x(t),y(t))}{\mon_i}$. Notice that as $p\searrow 1/2$ the scaling term $\sqrt{\frac{t}{4p-2}}$ becomes bigger; i.e. near the $p=1/2$ phase transition, it becomes harder to bias the process towards (eventual) leadership by either bin.
The second result we prove is about the almost-balanced case. Suppose, again for simplicity, that there are two bins ($B=2$) and $f(x)=x^p$, $0<p\leq 1/2$, . In this case says that for any initial conditions $I_0(1),I_0(2)$ with $I_0(1)<I_0(2)$, there is a time $m\geq 0$ such that $I_m(1)>I_m(2)$. Call the first such time the [*overtaking time*]{} $V$. By the above, $V<+\infty$, but we have no idea of the distribution of $V$, and thus we don’t know how long the overtaking might take. We show that if the initial number of balls is large and bin $2$ has a non-negligibly bigger fraction of the initial balls, then $V$ can actually be quite large; moreover, it has an explicit asymptotic distribution.
\[thm:almost\_cor\]Let $V_{t,\alpha}$ be the overtaking time in a balls-in-bins process with feedback function $f(x)=x^p$ (with $p\in(0,1/2]$ constant) and initial conditions $(\lceil \alpha t\rceil,t-\lceil \alpha t\rceil)$ for $0<\alpha<1/2$. Then there exist random variables $\{U_{t,\alpha,p}\}_{t\in\N}$ such that $$V_{t,\alpha} =\left\{\begin{array}{l} 2\left\{\frac{(1-2p)}{(1-p)^2}[(1-\alpha)^{1-p} -\alpha^{1-p}]^2 + \frac{(1-\alpha)^{1-2p}\,U_{t,\alpha,p}^2}{t}\right\}^{\frac{1}{1-2p}}\frac{t^{1+\frac{1}{1-2p}}}{U_{t,\alpha,p}^{1+\frac{2p}{1-2p}}} - (t+1)\\
\;\;\;\;\;\;\;\;\;\;\;\;\;\mbox{ if }0<p<\frac{1}{2}\commaeq\\
2(1-\alpha)\, t\,
\exp\left\{4[1-2\sqrt{\alpha(1-\alpha)}]\,\frac{t}{U_{t,\alpha,\frac{1}{2}}^2}\right\}
-(t+1)\\ \;\;\;\;\;\;\;\;\;\;\;\;\;\mbox{ if }p
=\frac{1}{2}\commaeq\end{array}\right.$$ with probability tending to $1$ as $t\to +\infty$, and $$U_{t,\alpha,p} \weakto |N|\commaeq$$where $N$ is a standard Gaussian random variable.
This means that $V$ becomes larger and larger as $p\nearrow 1/2$, culminating with the exponential behavior at the phase transition point $p=1/2$. The economically-inclined might wish to deduce from this theorem that, under appropriate initial conditions, a product’s leadership might last a long time even in markets with no propensity for breeding monopolies or “eternal leaders".
Techniques and outline
----------------------
Our results in this paper are actually more general: they extend to a broader (though not entirely general) class of functions $f$ and, in the case of , to more than two bins. All proofs below are done for this more general case and then specialized for $f(x)=x^p$.
Our proofs have their first two steps in common. The first step has been employed in [@Khanin01; @Spencer??] and other works, and seems to have originated in Davis’ work on reinforced random walks [@Davis90]. We shall [*embed*]{} the discrete-time process we are interested in into a continuous-time process built from exponentially distributed random variables, so that interarrival times at different bins are independent and have an explicit distribution, which is very helpful in calculations. We call this the [*exponential embedding*]{} of the process.
The second technique we use is [*approximation by Brownian motion*]{} via Donsker’s Invariance Principle. While neither technique is novel, their conjunction in the way presented here yields surprising explicit results in the asymptotic regime, once the appropriate calculations are done.
The remainder of the paper is organized as follows. We discuss preliminary material in . rigorously introduces the exponential embedding process and discusses its key properties. In we detail the assumptions we make on our feedback functions $f$, while also deriving some consequences of those assumptions. proves the general version of , whereas contains the proof of the generalization of . Finally, discusses extensions to our results and some open problems.
Acknowledgements
----------------
Michael Mitzenmacher and my Ph.D. advisor Joel Spencer introduced me to this topic and stimulated me with several interesting questions and feedback. I also thank Eleni Drinea for useful discussions.
Preliminaries {#sec:prelim}
=============
[*General notation.*]{} Throughout the paper, $\N=\{1,2,3,\dots\}$ is the set of non-negative integers, $\R^+ =
[0,+\infty)$ is the set of non-negative reals, and for any $k\in\N\backslash\{0\}$ $[k]=\{1,\dots,k\}$.\
[*Asymptotics.*]{} We use the standard $O/o/\Omega/\Theta$ notation. The expressions “$a_n\sim b_n$ as $n\to n_0$" and “$a_n\ll b_n$ as $n\to n_0$" mean that $\lim_{n\to n_0}(a_n/b_n) = 1$ and $\lim_{n\to n_0}(a_n/b_n) = 0$, respectively.\
[*Balls-in-bins.*]{} Formally, a balls-in-bins process with feedback function $f:\N\to (0+\infty)$ and $B\in\N$ bins is a discrete-time Markov chain $\{(I_1(m),\dots,I_B(m))\}_{m=0}^{+\infty}$ with state space $\N^B$ and transition probabilities given in the Introduction (see ). We will usually refer to the index $i_m\in[B]$ as [*the bin that receives a ball at time $m$*]{}. For any $B$, if $E$ is an event of the process and $u\in\N^B$, $\Prp{u}{E}$ is the probability of $E$ when the initial conditions are set to $u$.\
[*Exponential random variables.*]{} $X\eqdist \exp(\lambda)$ means that $X$ is a random variable with exponential distribution with rate $\lambda>0$, meaning that $X\geq 0$ and $$\Pr{X>t} = e^{-\lambda t}\;\; (t\geq 0).$$ The shorthand $\exp(\lambda)$ will also denote a generic random variable with that distribution. Some elementary but extremely useful properties of those random variables include
1. [*Lack of memory.*]{} Let $X\eqdist \exp(\lambda)$ and $Z\geq 0$ be independent from $X$. The distribution of $X-Z$ conditioned on $X>Z$ is still equal to $\exp(\lambda)$.
2. [*Minimum property.*]{} Let $\{X_i\eqdist \exp(\lambda_i)\}_{i=1}^{m}$ be independent. Then $$X_{\min}\equiv \min_{1\leq i\leq
m}X_i\eqdist \exp(\lambda_1+\lambda_2+\dots \lambda_m)$$ and for all $1\leq i\leq m$ $$\Pr{X_i = X_{\min}} = \frac{\lambda_i}{\lambda_1+\lambda_2+\dots
\lambda_m}$$
3. [*Multiplication property.*]{} If $X\eqdist\exp(\lambda)$ and $\eta>0$ is a fixed number, $\eta X
\eqdist\exp(\lambda/\eta)$.
4. [*Moments and transforms.*]{} If $X\eqdist\exp(\lambda)$, $r\in\N$ and $t\in\R$, $$\begin{aligned}
\Ex{X^r} & = & \frac{r!}{\lambda^r}\commaeq \\
\Ex{e^{t X}} & = & \left\{\begin{array}{ll}\frac{1}{1 - \frac{t}{\lambda}} & (t<\lambda)\\
+\infty & (t\geq
\lambda)\end{array}\right.\end{aligned}$$
[*Weak convergence.*]{} $X_n\weakto Y$ means that the sequence $\{X_n\}$ of random variables converges weakly to $Y$ as $n\to
+\infty$.
[*Gaussians and cumulative distribution functions.*]{} Finally, we restate the definitions of $\Phi$ and $\Gamma$ in :
$$\begin{aligned}
\label{eq:defphi}\Phi(\lambda)&\equiv &\sqrt{\frac{1}{2\pi}}\int_{-\infty}^{\lambda}e^{-x^2/2}\,dx,\\
\label{eq:defgamma}\Gamma(\lambda)&\equiv&\sqrt{\frac{2}{\pi}}\int_{0}^{\max\{\lambda,0\}}e^{-x^2/2}\,dx.\end{aligned}$$
If $N$ is a standard Gaussian random variable, $\Gamma$ is the cdf (cumulative distribution function) of $N$ and $\Gamma$ is the cdf of $|N|$.
The exponential embedding {#sec:exp_embed}
=========================
Definition and key properties
-----------------------------
Let $f:\N\to(0,+\infty)$ be a function, $B\in\N$ and $(a_1,\dots,a_B)\in\N^B$. We define below a continuous-time process with state space $(\N\cup\{+\infty\})^B$ and initial state $(a_1,\dots,a_B)$ as follows. Consider a set $\{X(i,j)\,:\,i\in
[B],\, j\in \N\}$ of independent random variables, with $X(i,j)\eqdist \exp(f(j))$ for all $(i,j)\in [B]\times \N$, and define
$$\label{eq:exp_contproc}N_i(t)\equiv \sup\left\{n\in
\N\,:\,\sum_{j=a_i}^{n-1} X(i,j) \leq t\right\}\;\;\;(i\in
[B],t\in\R^+ = [0,+\infty))\commaeq$$
where by definition $\sum_{j=i}^{k}(\dots)=0$ if $i>k$. Thus $N_i(0)=a_i$ for each $i\in [B]$, and one could well have $N_i(T)=+\infty$ for some finite time $T$ (indeed, that [*will*]{} happen for our cases of interest); but in any case, the above defines a continuous-time stochastic process, and in fact the $\{N_i(\cdot)\}_{i=1}^B$ processes are independent. Each one of this processes is said to correspond to [*bin*]{} $i$, and each one of the times $$X(i,a_i),X(i,a_{i})+X(i,a_{i}+1),X(i,a_{i})+X(i,a_{i}+1)+X(i,a_{i}+2),\dots$$ is said to be an [*arrival time at bin $i$*]{}. As in the balls-in-bins process, we imagine that each arrival correspond to a ball being placed in bin $i$.
In fact, we [*claim*]{} that this process is related as follows to the balls-in-bins process with feedback function $f$, $B$ bins and initial conditions $(a_1,\dots,a_B)$.
\[thm:exp\_embed\]Let the $\{N_i(\cdot)\}_{i\in[B]}$ process be defined as above. One can order the arrival times of the $B$ bins in increasing order (up to their first accumulation point, if they do accumulate) so that $T_1<T_2<\dots$ is the resulting sequence. The distribution of $$\{I_m = (N_1(T_m),N_2(T_m),\dots,N_B(T_m))\}_{m\in\N}$$ is the same as that of a balls-in-bins process with feedback function $f$ and initial conditions $(a_1,a_2,\dots,a_B)$.
One can prove this result[^4] as follows. First, notice that the [*first arrival time $T_1$*]{} is the minimum of $X(j,a_j)$, ($1\leq
j\leq B$). By the minimum property presented above, the probability that bin $i$ is the one at which the arrival happens is like the first arrival probability in the corresponding balls-in-bins process with feedback: $$\label{eq:min}\Pr{X(i,a_i) = \min_{1\leq j\leq
B}X(j,a_j)} = \frac{f(a_i)}{\sum_{j=1}^B f(a_j)}.$$ More generally, let $t\in\R^+$ and condition on $(N_i(t))_{i=1}^B
= (b_i)_{i=1}^B \in\N^B$, with $b_i\geq a_i$ for each $i$ (in which case the process has not blown up). This amounts to conditioning on $$\forall i\in[B]\;\; \sum_{j=a_i}^{b_i-1}X(i,b_i)\leq t < \sum_{j=a_i}^{b_i}X(i,b_i).$$ From the lack of memory property of exponentials, one can deduce that the first arrival after time $t$ at a given bin $i$ will happen at a $\exp(f(b_j))$-distributed time, independently for different bins. This almost takes us back to the situation of , with $b_i$ replacing $a_i$, and we can similarly deduce that bin $i$ gets the next ball with the desired probability, $$\frac{f(b_i)}{\sum_{j=1}^Bf(b_j)}.$$
On the three regimes {#sec:exp_proof}
--------------------
Let us now briefly point out some of the key steps in the proof of via the exponential embedding, in the case $B=2$. Reading this sketch might help the reader to become acquainted with an important part of our methods.
We use the same notation and random variables introduced above.
Assume we start the process from state $(x,y)\in\N^2$. First, we note that $$\label{eq:monopoly_cond}\Ex{\sum_{j=x}^{+\infty}X(1,j)}=\sum_{j=x}^{+\infty}
\frac{1}{f(j)},$$ hence if the RHS is finite, $\sum_{j=x}^{+\infty}X(1,j)<+\infty$ almost surely, and similarly for $\sum_{j=y}^{+\infty}X(2,j)$. Moreover, the two random series are independent, and neither has point-masses in their distribution. Therefore, with probability $1$, $$\label{eq:dicmon}\mbox{either
}\sum_{j=x}^{+\infty}X(1,j)<\sum_{j=y}^{+\infty}X(2,j) \mbox{ or }
\sum_{j=x}^{+\infty}X(1,j)>\sum_{j=y}^{+\infty}X(2,j).$$ If the first alternative holds, there exists a finite $M>y$ such that $$\sum_{j=x}^{+\infty}X(1,j)<\sum_{j=y}^{M-1}X(2,j).$$ Now notice that the sequence $\{T_{n_k}=\sum_{j=x}^{x+k}X(1,j)\}_{k\in\N}$ is an infinite subsequence of the ball arrival times $\{T_n\}_{n\in\N}$, and at those times $$\forall k\in\N\;\; T_{n_k} =\sum_{j=x}^{x+k}X(1,j)<\sum_{j=y}^{M-1}X(2,j)\Rightarrow I_{n_k}(2)<M.$$ Since $\{I_m(2)\}_m$ is an increasing sequence, this means that $I_m(2)<M$ for all $m\in\N$; that is to say, bin $1$ [*must achieve*]{} monopoly. On the other hand, if the second alternative in holds, the same argument shows that bin $2$ must achieve monopoly. Thus the condition implies that with probability $1$, one of the two bins achieves monopoly. It is not too hard to prove that if does [*not*]{} hold, then almost surely $\sum_{j=x}^{+\infty}X(1,j)=\sum_{j=y}^{+\infty}X(2,j)=+\infty$; in fact, it suffices to show that, for some $\rho>0$ $$\Ex{\exp(-\rho\sum_{j=x}^{x+k}X(1,j))} = \prod_{j=x}^{x+k}\Ex{\exp(-\rho X(1,j))}\to 0\mbox{ as }k\to +\infty$$ and similarly for $\sum_{j=y}^{+\infty}X(2,j)$. In this case one can show, by reversing the above reasoning, with probability $1$ no bin will achieve monopoly.
Now assume that $x>y$ (for simplicity) and $$\label{eq:elead_cond}\sum_{j=1}^{+\infty}\frac{1}{f(j)^2}<+\infty.$$ In this case, even if $\sum_{j}f(j)^{-1}=+\infty$, the series $$\sum_{j=x}^{+\infty} (X(1,j)-X(2,j))$$ is made of independent, centered random variables whose variances satisfy $$\sum_{j=x}^{+\infty}\Var{(X(1,j)-X(2,j))} = \sum_{j=1}^{+\infty}\frac{2}{f(j)^2}<+\infty.$$ Hence Kolmogorov’s Three Series Theorem implies that $\sum_{j=x}^{+\infty} (X(1,j)-X(2,j))$ converges. Following the reasoning developed above, we deduce that almost surely $$\mbox{either }\sum_{j=x}^{+\infty} (X(1,j)-X(2,j))-\sum_{j=y}^{x-1}X(2,j)<0\mbox{ or }\sum_{j=x}^{+\infty} (X(1,j)-X(2,j))-\sum_{j=y}^{x-1}X(2,j)>0.$$ In the first case, for all large enough $M$ $$\sum_{j=x}^{M-1}X(1,j)<\sum_{j=y}^{M-1} X(2,j)$$ and one can check that this means that for all large enough $M$, bin $1$ reaches level $M$ [*before*]{} bin $2$ does (in the embedded and continuous-time processes): that is, bin $1$ achieves eventual leadership. Otherwise, if $\sum_{j=x}^{+\infty}
(X(1,j)-X(2,j))-\sum_{j=y}^{x-1}X(2,j)>0$, bin $2$ is the one that achieves eventual leadership. In either case, what we have discussed up to now proves items $1.$ and $2.$ of .
Finally, if $$\label{eq:almost_cond}\sum_{j=1}^{+\infty}\frac{1}{f(j)^2}=+\infty,$$ then for any $x$, as $k\to +\infty$, $$\sum_{j=x}^{x+k}\frac{1}{f(j)^3}\leq \frac{1}{\min_{j\geq 1}f(j)}\, \sum_{j=x}^{x+k}\frac{1}{f(j)^2}\ll \left(\sum_{j=x}^{x+k}\frac{1}{f(j)^2}\right)^{3/2}.$$ Checking the moments of the $X(i,j)$’s and using the results in shows that the sums $\sum_{j=x}^{x+k}X(1,j)$, $\sum_{j=y}^{y+k}X(2,j)$ $(k\in\N)$ are in the domain of attraction of Brownian Motion for any $x$ and $y$. This implies that there is a sequence of random numbers $M_1<M_2<M_3<\dots$ and a constant $0<\alpha<1$ such that for all $n\in\N$ $$\alpha<\Pr{\left.\sum_{j=x}^{k_n}X(1,j)<\sum_{j=y}^{k_n}X(2,j)\right| \{k_1,\dots,k_{n-1}\}\cup \{X(1,\ell),X(2,\ell):\ell\leq k_{n-1}\}}<1-\alpha.$$ This implies that both $\sum_{j=x}^{k_n}X(1,j)<\sum_{j=y}^{k_n}X(2,j)$ and $\sum_{j=x}^{k_n}X(1,j)>\sum_{j=y}^{k_n}X(2,j)$ must occur infinitely often almost surely. In this case, there are infinitely many $k$ for which bin $1$ reaches level $k$ before bin $2$ does, and vice-versa. It follows that implies that with probability $1$ neither bin will achieve eventual leadership, and this proves $3.$ and the theorem.
\[rem:fictitious\]Assume that bin $1$ achieves monopoly. Then all arrivals of the [*continuous-time*]{} process at bin $2$ after time $\sum_{j=x}^{+\infty}X(1,j)$ do not actually happen in the embedded [*discrete-time*]{} process $\{I_m=(I_m(1),I_m(2))\}$. We call these “ghost events" a [*fictitious continuation*]{} of our process. This very useful device is akin to the continuation of a Galton-Watson process beyond its extinction time (see e.g. [@AlonSpencer_Method]) and is equally useful in calculations and proofs.
Assumptions on feedback functions {#sec:technical}
=================================
The purpose of this rather technical section is two-fold. First, we spell out the technical assumptions on the feedback function $f$ that we need in our proofs. Nothing seems to actually [*require*]{} these assumptions, but they facilitate certain estimates that we employ in the proofs.
Some readers might wish to skip the proofs in this section on a first reading.
Valid feedback functions {#sec:valid}
------------------------
The feedback functions we allow in our results satisfy the following definition.
\[def:est\_smooth\] A function $f:\N\to (0,+\infty)$ with $f(1)=1$[^5] is said to be a [*valid feedback function*]{} if it can be extended to a piecewise $C^{1}$ function $g:\R^+\cup\{0\}\to(0,+\infty)$ with the following property: if $(\ln g(\cdot))'$ is the right-derivative of $\ln g$, and $h(x)\equiv x(\ln g(x))'$ (for $x\in\R^+\cup\{0\}$),
1. $\liminf_{x\to +\infty}h(x) = h_{min} >0$;
2. $\lim_{x\to +\infty}x^{-{1/4}}h(x) = 0$;
3. there exist $C>0$ and $x_0\in\R^+$ such that for all $\eps\in(0,1)$ and all $x\geq x_0$ $$\label{eq:est_hslow}\sup_{x\leq t\leq
x^{1+\eps}}\left|\frac{h(t)}{h(x)}-1\right|\leq
C\,\eps\periodeq$$ If in addition $h_{min}>1/2$, then we say that $f$ is ELM (ELM stands for “eventual leadership or monopoly"). If on the other hand $h(x)\leq 1/2$ for all large enough $x$, we say $f$ is AB (“almost-balanced").
With slight abuse of notation, we will always assume that $f$ is defined over $\R^+\cup\{0\}$ and is piecewise $C^1$. We will also call $h$ the [*characteristic exponent*]{} of $f$.
Functions with exponential growth (such as $f(x)=2^x$) or with oscillations fail to satisfy . On the other hand, requiring that $f$ be increasing seems natural, and the smoothness assumption still leaves us with plenty of interesting examples of feedback functions; some examples are given in Table \[tab:scaling\] The “canonical case” where $f(x)=x^p$ ($x\geq
1$) explains the terminology for the characteristic exponent: in that case, $h(x)\equiv p$ for all $x>1$.
Consequences of the definition
------------------------------
Let us now define the quantity $$S_r(n,m) \equiv \sum_{j=n}^{m-1}\frac{1}{f(n)^r}\;\;\;(r\in\N\backslash\{0\};\,n\in\N, m\in\N\cup\{+\infty\})$$ for some $f:\N\to (0,+\infty)$, and also let $S_r(n)\equiv
S_r(n,+\infty)$. If $f(x)=x^p$, then for $m-n,n\gg 1$ a simple shows that $$S_r(n,m)\sim\int_{n}^{m} \frac{dx}{f(x)^r} = \frac{n^{1-rp} - m^{1-rp}}{(rp-1)}.$$ The main content of the following lemmata (the first one proven in [@Oliveira05]) is that a similar result holds for any valid $f$, if $p$ is replaced by the characteristic exponent $h$. In particular, any valid $f$ satisfies the monopoly condition in . These lemmas are used in the two main proofs in the paper.
\[lem:est\_moment\]Assume that $r$ is an integer and $f$ is a valid feedback function with characteristic exponent $h$ satisfying $h_{\min}>1/r$. Define $$M_r(n) = \int_n^{+\infty} \frac{dx}{f(x)^r}\;\;\;\;\;(r\in\N\backslash\{0\},n\in\N)\periodeq$$ Then, as $n\to +\infty$ $$S_r(n) \sim M_r(n) \sim \frac{n}{(rh(n)-1)f(n)^r}\periodeq$$
Approximation by Brownian Motion
================================
The Invariance Principle – setup {#sec:brown}
--------------------------------
This is the last section in which technical preliminaries are discussed. In it, we review a form of Donsker’s Invariance Principle that shows that under suitable normalization, “nice" partial sums of random variables are close to Brownian Motion. All results in this section are quite standard and can be found in many books on Brownian Motion, e.g. [@BillingsleyBook]
Consider the vector space $C=C([0,1],\R)$ of all real-valued continuous functions on the unit interval, with the sup norm $$\|\phi(\cdot)\|_{\sup}\equiv \sup_{0\leq s\leq 1}|\phi(s)|,\;\;\;\phi(\cdot)\in C.$$ This gives $C$ a metric and a topology, and from now on we shall think of $C$ as a measurable space with the Borel $\sigma$-field. Brownian Motion is simply a probability measure on this measurable space, or rather a random variable $B(\cdot)$ taking values on $C$, whose defining properties are:
- $\Pr{B(0)=0} = 1$;
- for all $0\leq s_0<s_1< \dots<
s_k\leq 1$, the random variables $$\left\{\frac{B(s_i)-B(s_{i-1})}{\sqrt{s_i-s_{i-1}}}\right\}_{i=1}^k$$ are i.i.d. standard Gaussians.
We will also use the following distributional equalities below: $$\begin{aligned}
\max_{0\leq s\leq 1}B(s),\, -\min_{0\leq t\leq 1}B(t)\eqdist |N|\mbox{ where $N$ is standard Gaussian;} \\
\label{eq:diff}\mbox{if $B'(\cdot)$ is an independent copy of
$B$}, \frac{B(\cdot)-B'(\cdot)}{\sqrt{2}} \eqdist B.\end{aligned}$$
Now consider a “triangular sequence" $\{\xi_{n,t}\}_{t\in\N,1\leq
n\leq M_t}$ of independent, $0$-mean, square-integrable random variables,. Letting $$\label{eq:scaling_sigma}\sigma^2_{k,t} \equiv
\Var{\sum_{j=1}^{k}\xi_{j,t}} =
\sum_{j=1}^{k}\Var{\xi_{j,t}}\;(k\in[M_t])\mbox{ and
}\sigma_{0,t}^2=0\commaeq$$ we define a random element $\Xi_t(\cdot)$ ($t\in\N$) of $C$ as follows. $$\begin{aligned}
\Xi_t(s) &\equiv &\frac{\sum_{j=1}^{k(s)}\xi_{j,t} + \left(\frac{s-\sigma_{k(t),m}^2}{\Var{\xi_{k(s)+1,t}}} \right)\xi_{k(s)+1,t}}{\sigma_{M_t,t}}\commaeq \\
& & \mbox{where $s\in[0,1]$ and} \\
& & k(s) \equiv \max\left\{k\in[M_t]\cup\{0\}\suchthat
\frac{\sigma_{k,t}^2}{\sigma_{M_t,t}^2}\leq
t\right\}\periodeq\end{aligned}$$ Thus $\Xi_t(\sigma_{k,t}^2/\sigma_{M_t,t}^2)$ is the sum of the $k$ first $\xi_{j,t}$’s, divided by a normalizing factor; and for $s\in[\sigma_{k,t}^2/\sigma_{M_t,t}^2,\sigma_{k+1,t}^2/\sigma_{M_t,t}^2]$, $\Xi_t(s)$ is defined by linear interpolation of the values of $\Xi_t(\sigma_{k,M_t}^2/\sigma_{t,M_t}^2)$ and $\Xi_t(\sigma_{k+1,M_t}^2/\sigma_{t,M_t}^2)$. One can check that this is indeed a measurable element of $C$.
The Invariance Principle states that if the sequence $\{\xi_{n,M_t}\}$ satisfies certain conditions, the distribution of the $\Xi_t(\cdot)$’s [*converges weakly*]{} to a standard Brownian motion $B(\cdot)$. What this means is that if $A\subset
C$ is measurable with boundary $\partial A$ and $\Pr{B(\cdot)\in\partial A}=0$, then $\Pr{\Xi_t(\cdot)\in A}\to
\Pr{B(\cdot)\in A}$ as $t\to +\infty$. A sufficient condition for this is given by
\[thm:invariance\]If $$\frac{\sum_{n=1}^{M_t} \Ex{|\xi_{n,t}|^3}}{\sigma_{M_t,t}^{3/2}}\to 0 \mbox{ as }t\to +\infty,$$ then the sequence $\Xi_t(\cdot)$ converges weakly to $B(\cdot)$.
Application to the continuous-time process {#sec:apply_brown}
------------------------------------------
Our typical application of this invariance principle will be to the random variables in the exponential embedding. In the notation of the , let $i,i'\in [B]$ be fixed, $\{x_t\}_t,\{M_t\}_t\subset \N$ be sequences, and consider the triangular array of random variables $$\label{eq:ourtriangle}\left\{\xi_{n,t}\equiv X(i,x_t+n-1) -
X(i',x_t+n-1)\right\}_{m\in\N,1\leq n\leq
M_t}.$$ In this case, $\sigma_{M_t,t}^2 =
2S_2(x_t,x_t+M_t)$ and the condition in can be seen to be equivalent to $$\label{eq:check}\lim_{m\to
+\infty}\frac{S_3(x_t,x_t+M_t)}{(S_2(x_t,x_t+M_t))^{3/2}}=0.$$ That is, equation is the only condition we have to check in order to apply the Invariance Principle to the terms in . Notice also (by a simple limiting argument) that if $S_2(1)<+\infty$, we can also take $M_t=+\infty$ in the above.
The scaling result for leadership {#sec:scaling}
=================================
The general statement
---------------------
Recalling , let $f$ be a ELM function. There exists an $x_0$ such that $h(x) = xf'(x)/f(x)>1/2$ for all $x\geq x_0$, which means that for such large $x$ $$\label{eq:q0}q_0(x)\equiv \sqrt{\frac{x}{4h(x)-2}}\;\;(x\geq
x_0)$$ is a well-defined, positive function. Our generalization of shows that the quantity $q_0(t)$ plays the same role that as the map $t\mapsto
\sqrt{t/(4p-2)}$ in the specific case $f(x)=x^p$, $p>1/2$. That is, in order to bias a balls in bins process started with $t$ balls towards leadership by a given bin, the difference between the initial numbers of balls should be $\theita{q_0(t)}$.
\[thm:scaling\_main\]Let $f$ be a ELM function and define $q_0$ as above. Let $\lambda\in\R$ be a constant, and assume that $q=q(n)$ ($n\in\N$) is such that
- $t/2 \pm \lambda \,q(n)\in \N$ for all $n\in\N$;
- $q(n)\sim q_0(n)$ for $n\gg 1$.
Now consider the $2$-bin balls-in-bins process started from initial state $$(x(t),y(t))=\left(\frac{t}{2}+\lambda\, q(t),\frac{t}{2}-\lambda\, q(t)\right) \in \N^2\periodeq$$ Then $$\begin{aligned}
\lim_{t\to +\infty}\Prp{(x(t),y(t))}{\elead_1} &=& \Phi(\lambda)\commaeq\\
\lim_{t\to +\infty}\Prp{(x(t),y(t))}{\lead_1} &=&
\Gamma(\lambda).\end{aligned}$$
Table \[tab:scaling\] presents estimates of $q_0(n)$ for $n$ large, for several choices of feedback functions in the ELM regime. In particular, the case $f(x)=x^p$ ($p>1/2$) of the Theorem implies . The remainder of this section contains the proof of the general result.\
$f(x)=$ $h(x)\sim $ $q_0(x)\sim$ conditions
---------------------- ------------------------------ ------------------------------------------------ -----------------------
$x^p\ln^{q}(x+e-1)$ $p$ $\sqrt{\frac{x}{4p-2}}$ $p>1/2$, $q\in\R$
$x^{q\ln^\alpha(x)}$ $(\alpha+1)q\ln^{\alpha}(x)$ $\sqrt{\frac{x}{4(\alpha+1)q\ln^{\alpha}(x)}}$ $p>1/2$, $q,\alpha>0$
$e^{x^{p}}$ $px^p$ $\frac{x^{\frac{1-p}{2}}}{2\sqrt{p}}$ $0<p<1/4$
: \[tab:scaling\]
We start by discussing how one can write the event $\elead_1$ and lead in terms of the exponential embedding. We only prove the result for $\lambda>0$: the case $\lambda=0$ is a simple extension, and the case $\lambda<0$ reduces to the one we discuss below.
Let $(x,y)\in \N^B$ be the (for the time being arbitrary) initial conditions with $x>y$. The event $\elead_1$ holds whenever bin $1$ reaches reach level $M$ before bin $2$ does for all large enough $M$. This requires that the time it takes for bin $1$ to reach level $M$ in the continuous-time process is smaller than the corresponding time for bin $2$. In the exponential embedding, this corresponds to $$\exists M_0\,\forall M\geq M_0\,\, \sum_{j=x}^{M}X(1,j) < \sum_{j=y}^{M} X(2,j).$$ The above event can be rewritten as $$\exists M_0\,\forall M\geq M_0\,\, \sum_{j=x}^{M}(X(1,j) - X(2,j))< \sum_{j=y}^{x-1} X(1,j).$$ As noted in the proof sketch for in , $\sum_{j=x}^{M}(X(1,j) - X(2,j))$ converges as $M\to +\infty$. If follows that, except for a null event, the above holds if and only if $$\sum_{j=x}^{+\infty}(X(1,j) - X(2,j))< \sum_{j=y}^{x-1} X(2,j).$$ Thus we deduce that $$\label{eq:rewriteelead}\Prp{(x,y)}{\elead_1} = \Pr{\sum_{j=x}^{+\infty}(X(1,j) - X(2,j))< \sum_{j=y}^{x-1} X(2,j)}.$$ What about the probability of $\lead_1$? Using the above notation, $\lead_1$ holds if for [*all*]{} $M\geq x$ bin $1$ reaches level $M$ before bin $2$ does. That corresponds to $$\forall M\geq x\,\, \sum_{j=x}^{M}X(1,j) < \sum_{j=y}^{M} X(2,j),$$ or $$\forall M\geq x_{\ell_i}\,\, \sum_{j=x}^{M}(X(1,j) - X(2,j))< \sum_{j=y}^{x-1} X(2,j).$$ It follows that $$\label{eq:rewritelead}\Prp{(x,y)}{\lead_1} = \Pr{\sup_{M\geq x}\sum_{j=x}^{M}(X(1,j) - X(2,j))< \sum_{j=y}^{x-1} X(2,j)}.$$ Recall now the choice $x=x(t) = t/2 + \lambda q(t)$ and $y=y(t) =
t/2-\lambda q(t)$, where $$q(n) \sim \sqrt{\frac{n}{4h(n)-2}}\mbox{ for }n\gg 1\periodeq$$ As discussed in , $\{X(1,x(t)+n-1)-X(2,x(t)+n-1)\}_{t\in\N,n\in\N}$ is a doubly-infinite array of centered, square-integrable random variables with $$\sigma_{n,t}^2\equiv \sum_{j=1}^t\Var{X(1,x(t)+n-1)-X(2,x(t)+n-1)} = 2S_2(x(t),x(t)+n) \;\; (1\leq n\leq +\infty).$$ One can construct a random continuous path $\Xi_t(\cdot)$ defined by setting $$\Xi_t\left(\frac{S_2(x(t),x(t)+n)}{S_2(x(t))}\right) = \frac{\sum_{j=x}^{x(t)+n-1}X(1,j)-X(2,j)}{\sqrt{2S_2(x)}}, (1\leq n\leq +\infty)$$ and completing the remaining values by linear interpolation. As discussed in , the fact that we have infinite terms here poses no problems. Moreover, it is easy to check that $$\begin{aligned}
\sup_{0\leq s\leq 1}\Xi_t(s) &=& \frac{1}{\sqrt{2S_2(x)}}\sup_{M\geq x}\sum_{j=x}^{M}{X(1,j)-X(2,j)},\\
\Xi_t(1) &=&
\frac{1}{\sqrt{2S_2(x)}}\sum_{j=x}^{+\infty}{X(1,j)-X(2,j)}.\end{aligned}$$ Thus one can rewrite $$\begin{aligned}
\label{eq:rewriteelead2}\Prp{(x,y)}{\elead_1} &=& \Pr{\Xi_t(1) < \frac{\sum_{j=y}^{x-1} X(2,j)}{\sqrt{2S_2(x)}}},\\
\label{eq:rewritelead2}\Prp{(x,y)}{\lead_1} &=& \Pr{\sup_{0\leq
s\leq
1}\Xi_t(s)<\frac{\sum_{j=y}^{x-1}X(2,j)}{\sqrt{2S_2(x)}}}.\end{aligned}$$
Suppose we show that as $t\to +\infty$ $$\begin{aligned}
\label{eq:final1}\Xi_t(\cdot) &\weakto& \mbox{a standard Brownian Motion }B(\cdot),\\
\label{eq:final2}\frac{\sum_{j=y}^{x-1}X(2,j)}{\sqrt{2S_2(x)}}&\weakto
&\lambda.\end{aligned}$$ Since $\sum_{j=y}^{x-1}X(2,j)$ is independent of $\Xi_t(\cdot)$, this means that $$(\Xi_t(\cdot),\sum_{j=y}^{x-1}X(2,j))\weakto (B(\cdot),\lambda),$$ which implies $$\begin{aligned}
\label{eq:limitelead2}\Prp{(x,y)}{\elead_1} &\to & \Pr{B(1)<\lambda},\\
\label{eq:limitlead2}\Prp{(x,y)}{\lead_1} &\to& \Pr{\sup_{0\leq
s\leq 1}B(s) <\lambda}.\end{aligned}$$ These probabilities can be evaluated via standard formulae for Brownian motion in , yielding the final result. We thus concentrate on proving equations and .\
[*Proof of .*]{} By , it suffices to show that $S_3(x(t))\ll S_2(x(t))^{3/2}$. But this follows directly from the formulae in , the fact that $x=x(t)\to +\infty$, and the assumption that $h(x)\ll
\sqrt{x}$: $$S_3(x) \sim \frac{x}{(3h(x)-1)f(x)}\ll \left(\frac{x}{(2h(x)-1)f^2(x)}\right)^{3/2} = S_2(x)^{3/2}.$$\
[*Proof of .*]{} Let us first establish a few facts about $f$, $x$, $y$ and $q$.
1. [*For $t\gg 1$, $q(t) \sim
f(t/2)\sqrt{S_2(t/2)/2}=\bigoh{t}$*]{}. Indeed, for $t$ large, implies that $S_2(t/2)\sim
t/(4h(t/2)-2)f(t/2)^2$, and, because of in , $h(t/2)\sim h(t)$ . Moreover, since $\liminf_{n\to
+\infty}h(n)>1/2$, $q(t) = \sqrt{t/\ohmega{1}} =
\bigoh{\sqrt{t}}$.
2. [*For $t\gg 1$, $f(x)=f(t/2 + \lambda
q(t))\sim f(t/2)$*]{}. In this case $$\begin{aligned}
\left|\ln \frac{f(x)}{f\left(\frac{t}{2}\right)}\right| &=&
\left|\int_{\frac{t}{2}}^{\frac{t}{2} + \lambda q(t)} (\ln
f(u))'\,du\right| \\ \nonumber &=&
\left|\int_{\frac{t}{2}}^{\frac{t}{2} + \lambda q(t)}
\frac{h(u)}{u}\,du \right|\\ \nonumber & \leq &
\sup\limits_{\frac{t}{2}\leq u\leq \frac{t}{2} + \lambda q(t)}
|h(u)|\, \ln\left(1 + \frac{2
q(t)}{t}\right)\periodeq\end{aligned}$$ Now notice that equation implies that the $\sup$ of $h(u)$ above is $\sim h(t/2)\sim h(t)$. Moreover, $q(t)/t = \sqrt{1/t(2h(t)-1)}\ll
1$, since $\liminf_{n\to +\infty}h(n)>1/2$. Therefore, $$\left|\ln
\frac{f(x)}{f\left(\frac{t}{2}\right)}\right| =
\bigoh{\frac{h(t)q(t)}{t}} = \bigoh{\frac{1}{\sqrt{t}}} =
\liloh{1}\periodeq$$
3. [*For $t\gg 1$, $f(y)\sim f(t/2)$*]{}. The proof is almost identical to the one above.
4. [*For $t\gg 1$ and $r\geq 2$, $S_r(x),S_r(y)\sim
S_r(t/2)$*]{}. Indeed, because $q(t)=\bigoh{\sqrt{t}}$, $x,y\to
+\infty$,, and the formulae in apply. Thus $S_2(x)\sim x/[(rh(x)-1)f(x)^2]$, and by $1.$ and $3.$, $x\sim
t/2$, $h(x) \sim h(t/2) $ and $f(x)\sim f(t/2)$, which implies $S_2(x) \sim (t/2)/[(rh(t/2)-1)f(t/2)^2]\sim S_r(t/2)$. The same argument proves the desired result for $S_r(y)$.
We now apply the estimates to the problem at hand. For $t$ large enough (so that $x$ and $y$ are also large), we can ensure that $f$ is increasing on $[y,x]$, so that $$\frac{x-y}{f(x)} \leq \Ex{\sum_{\ell=y}^{x-1} X(2,\ell)} = \sum_{j=y}^{x-1}\frac{1}{f(j)} \leq
\frac{x-y}{f(y)}\periodeq$$ By items $2.$ and $3.$ above and the definition of $x,y$, this implies that $$\Ex{\sum_{\ell=y}^{x-1} X(2,\ell)} \sim \frac{2\lambda
q(t)}{f(t/2)}\;\;(t\gg 1)\periodeq$$ Similarly, one can show that $$\Var{\sum_{\ell=y}^{x-1} X(2,\ell)} \sim \frac{2\lambda
q(t)}{f(t/2)^2}\;\;(t\gg 1)\periodeq$$ Therefore, using $4.$, $$\Ex{\sum_{\ell=y}^{x-1} X(2,\ell)}^2 \geq
(1-\liloh{1})2\lambda\,q(t)\,\Var{\sum_{\ell=y}^{x-1}
X(2,\ell)}\;\;(t\gg 1)\periodeq$$ By Chebyshev’s Inequality, it follows that $$\frac{{\sum_{\ell=y}^{x-1} X(2,\ell)}}{\left(\frac{2\lambda
q(t)}{f(t/2)}\right)}\weakto 1\mbox{ as $t\to +\infty$.}$$ Finally, since $$\frac{2q(t)}{f(t/2)}\sim \sqrt{\frac{2t}{(2h(t)-1)f(t/2)^2}}\sim \sqrt{2S_2(x)},$$ we have $$\frac{{\sum_{\ell=y}^{x-1} X(2,\ell)}}{\lambda
\,\sqrt{2S_2(x)}}\weakto 1\mbox{ as $t\to +\infty$,}$$ which is the desired result.
The almost balanced regime {#sec:almost}
==========================
This section proves our result on the overtaking time, a generalization of . To recapitulate: tells us that, when the feedback function $f$ satisfies $$\label{eq:almost_ab}\sum_{j=1}^{+\infty}\frac{1}{f(j)^2}=+\infty\commaeq$$ each of the two bins will be the one with more balls infinitely many times. Our main interest in this chapter will be in determining [*how long*]{} it takes for bin $1$ to have more balls than bin $2$, given that the latter bin has more balls at the start. More specifically, assume the process $(I_1(\cdot),I_2(\cdot))$ is started from state $$(\lceil \alpha t\rceil,t - \lceil \alpha
t\rceil)\commaeq\;\; t\gg 1\mbox{ and } \alpha\in(0,1/2)\mbox{
fixed}\periodeq$$ As in the introduction, let $V$ be the [*overtaking time*]{} of the process: that is the first time when bin $1$ has more balls than bin $2$. $$\label{eq:almost_overtaking}V \equiv \min\{v\in\N\suchthat
I_1(v)>I_2(v)\}\periodeq$$ Under condition , this $\min$ exists and is finite with probability $1$. We will be interested in describing the asymptotic distribution of $V$.
To express our main result, let us introduce two mappings.
$$\begin{aligned}
\funcdef{F_{t,\alpha}}{((1-\alpha)t,+\infty)}{\R^+}{u}{\frac{\int_{\alpha t}^{(1-\alpha)
t}\frac{dx}{f(x)}}{\sqrt{\int_{(1-\alpha)
t}^u\frac{dx}{f(x)^2}}}}\commaeq\\
G_{t,\alpha}\equiv \mbox{the inverse of
}F_{t,\alpha}\periodeq\end{aligned}$$
Notice that $\lim_{u\searrow (1-\alpha)t}F_{t,\alpha}(u)=+\infty$ and $$\lim_{u\to +\infty}F_{t,\alpha}(u) =\frac{\int_{\alpha t}^{(1-\alpha)
t}\frac{dx}{f(x)}}{\sqrt{\int_{(1-\alpha)
t}^{+\infty}\frac{dx}{f(x)^2}}}=0$$ because (as a consequence of $\sum_{j}^{+\infty}f(j)^{-2}=+\infty$) the denominator in the RHS is infinite. Thus $F_{t,\alpha}$ is a monotone-decreasing function whose range is $\R^+$, and $G_{t,\alpha}$ is not only well-defined, but monotone-decreasing as well.
\[thm:almost\_main\]Assume that $f$ is a AB function (cf. ), and define $F_{t,\alpha}$, $G_{t,\alpha}$ as above. Let $V_{t,\alpha}$ be the random variable $V$ defined above, conditioned on the initial state $(\lceil
\alpha t\rceil,t-\lceil \alpha t\rceil)$ of the balls-in-bins process. Then, as $t\to +\infty$, $$\forall \lambda\in\R^+\commaeq\;\lim_{t\to +\infty}\Pr{V_{t,\alpha}\geq
2G_{t,\alpha}(\lambda) - (t+1)}=
\Gamma(\lambda)\commaeq$$ (where $\Gamma(\cdot)$ is defined as in ), or equivalently, $$F_{t,\alpha}\left(\frac{V_{t,\alpha}+ (t+1)}{2}\right)\weakto
|N|\commaeq$$ for a standard Gaussian random variable $N$. The latter expression makes sense because $V_{t,\alpha}\geq
(1-\alpha)t$, so $$\frac{V_{t,\alpha}+ (t+1)}{2}\geq \left(1-\frac{\alpha}{2}\right)t > (1-\alpha)t\periodeq$$
This result is quite general, but applying it to a specific situation requires a calculation. We do this for the case $f(x)=x^p$ below, and then prove below.
Proof of the special case
-------------------------
implies .
One way of interpreting in the $f(x)=x^p$ case is by saying that $$V_{t,\alpha} \equiv 2G_{t,\alpha}(U_{t,\alpha,p})-(t+1)\mbox{ with probability $\to 1$}\commaeq$$ where $U_{t,\alpha,p}\weakto |N|$ as $t\to +\infty$. Thus the Corollary follows from providing a formula for $G_{t,\alpha}$. We will first assume that $0<p<1/2$, in which case $$\begin{aligned}
\label{eq:almost_eta} e(t,\alpha) &\equiv& \int_{\alpha t}^{(1-\alpha)t}\frac{dx}{f(x)}\\
&=&[(1-\alpha)^{1-p} -
\alpha^{1-p}]\frac{t^{1-p}}{1-p}\commaeq\end{aligned}$$ and for all $u> (1-\alpha)t$ $$\begin{aligned}
\label{eq:almost_feta} f(u,t,\alpha) &\equiv& \int_{(1-\alpha) t}^{u}\frac{dx}{f(x)^2}\\
&=&\frac{u^{1-2p}- [(1-\alpha)
t]^{1-2p}}{1-2p}\periodeq\end{aligned}$$ Then, for $u\geq
(1-\alpha)t$, $$\begin{aligned}
F_{t,\alpha}(u) &=&
\frac{e(t,\alpha)}{\sqrt{f(u,t,\alpha)}} \\
&=&\frac{[(1-\alpha)^{1-p} -\alpha^{1-p}]\frac{t^{1-p}}{1-p}
}{\sqrt{\frac{u^{1-2p}- [(1-\alpha)
t]^{1-2p}}{1-2p}}}\end{aligned}$$ To compute $G_{t,\alpha}(\lambda)$ for some $\lambda\in\R^+$, we must solve the equation $F_{t,\alpha}(G_{t,\alpha}(\lambda))=\lambda$, which corresponds to $$G_{t,\alpha}(\lambda)^{1-2p} = [(1-\alpha) t]^{1-2p} + \frac{(1-2p)}{\lambda^2}\left\{[(1-\alpha)^{1-p} -\alpha^{1-p}]\frac{t^{1-p}}{1-p}\right\}^2\periodeq$$
Therefore, $$G_{t,\alpha}(\lambda) = \left\{\frac{(1-2p)}{(1-p)^2}[(1-\alpha)^{1-p} -\alpha^{1-p}]^2 + \frac{(1-\alpha)^{1-2p}\lambda^2}{t}\right\}^{\frac{1}{1-2p}}\frac{t^{1+\frac{1}{1-2p}}}{\lambda^{1+\frac{2p}{1-2p}}}\commaeq$$ and the result for $0<p<1/2$ follows.
For $p=1/2$, the above formula for $e(t,\alpha)$ still applies, but $$\begin{aligned}
\label{eq:almost_feta12} f(u,t,\alpha) &\equiv& \int_{(1-\alpha) t}^{u}\frac{dx}{f(x)^2}\\
&=&\ln\frac{u}{(1-\alpha)t}\commaeq\end{aligned}$$ and thus $$\begin{aligned}
G_{t,\alpha}(\lambda) &=& (1-\alpha)\, t\,
\exp\left\{\frac{e(t,\alpha)^2}{\lambda^2}\right\} \\ &=&
(1-\alpha)\, t\,
\exp\left\{4[1-2\sqrt{\alpha(1-\alpha)}]\,\frac{t}{\lambda^2}\right\}
\periodeq\end{aligned}$$ This finishes the proof.
Proof of {#sec:almost_proof}
---------
Define the [*overtaking number*]{} $N_{t,\alpha}$ to be $I_1(V_{t,\alpha})$, i.e. the number $I_1(v)$ of balls in bin $1$ at the first time $v$ when $I_1(v)>I_2(v)$, under initial conditions $(\lceil \alpha t\rceil,t-\lceil \alpha t\rceil)$. It follows from this definition that at time $v'=V_{t,\alpha}-1$, $I_1(v')=I_2(v') = N_{t,\alpha}-1$. Since $I_1(0)+I_2(0)=t$, this means that $$\begin{aligned}
\label{eq:almost_overtake}\nonumber V_{t,\alpha}-1 + t &=& I_1(v')+I_2(v') \\ \nonumber &=&\mbox{total $\#$ of balls at time $v'$}\\ \nonumber
&=& 2N_{t,\alpha} - 2 \\ \label{eq:almost_overtakelast}\Rightarrow
V_{t,\alpha} &=& 2N_{t,\alpha} - (t+1)\periodeq\end{aligned}$$ Thus results about the distribution of $N_{t,\alpha}$ translate immediately into results about $V_{t,\alpha}$. Since $N_{t,\alpha}$ is easier to analyze via our techniques, we shall spend most of our time considering this quantity, returning to the more significant $V_{t,\alpha}$ at the end of the proof.
We begin by showing that, in terms of the exponential embedding random variables, $$\label{eq:almost_expwrite}\forall M\in\N\;\; \{N_{t,\alpha}\geq M\} =
\left\{\sup_{y\leq m\leq M-1}\sum_{j=y}^{m-1}(X(2,j)-X(1,j))\leq
\sum_{\ell=x}^{y-1}X(1,\ell)\right\}\commaeq$$ where $x\equiv\lceil\alpha t\rceil$ (respectively $y\equiv
t-\lceil\alpha t\rceil$) is the initial number of balls in bin $1$ (resp. $2$). Indeed, $N_{t,\alpha}\geq M$ occurs if and only if for all $y\leq m\leq M-1$, the time it takes for bin $2$ to receive its $m$th ball (which is $\sum_{j=y}^{m-1}X(2,j)$) is smaller than or equal to the time it takes for bin $1$ to receive its $m$th balls (which is $\sum_{\ell=x}^{m-1}X(1,\ell)$). Symbolically, $$\label{eq:almost_expwrite2}\forall M\in\N\;\; \{N_{t,\alpha}\geq M\} =
\left\{\forall y\leq m\leq M-1\commaeq\;
\sum_{j=y}^{m-1}X(2,j)\leq \sum_{\ell=x}^{m-1}
X(1,\ell)\right\}\commaeq$$ from which follows.
We now wish to estimate the probability of the event at the RHS of . We begin by looking at $$\label{eq:partialsumsalmost}\sup_{y\leq m\leq M-1}\sum_{j=y}^{m-1}(X(2,j)-X(1,j))\periodeq$$ In particular, let us assume that $$\label{eq:almost_lambda}M \equiv \lfloor G_{t,\alpha}(\lambda)\rfloor$$ for some $\lambda\in\R^+$, so that $$\{N_{t,\alpha}\geq M\} = \{N_{t,\alpha}\geq G_{t,\alpha}(\lambda)\}\periodeq$$ We wish to apply the Invariance Principle to the random variables $$\{\xi_{n,t}\equiv X(2,y(t)+n-1)-X(1,y(t)+n-1)\}_{t\in\N,1\leq n\leq M-y(t)-1}.$$ Note that $$\sigma^2_{n,t}\equiv \sum_{j=1}^n \Var{X(2,y(t)+n-1)-X(1,y(t)+n-1)} = 2S_2(y(t),y(t)+n).$$ Following , we build the random path $\Xi_t(\cdot)$ that linearly interpolates the values $$\Xi_t\left(\frac{S_2(y(t),y(t)+n)}{S_2(y(t),M-1)}\right) \equiv \frac{\sum_{j=1}^n {X(2,y(t)+j-1)-X(1,y(t)+j-1)}}{\sqrt{2S_2(y,M-1)}}, \, 1\leq n\leq M-y(t).$$ As in the previous proof, we have $$\sup_{0\leq s\leq 1}\Xi_t(s) = \frac{\sup_{y\leq m\leq M-1}\sum_{j=y}^{m-1} {X(2,y(t)+j-1)-X(1,y(t)+j-1)}}{\sqrt{2S_2(y,M-1)}}.$$ Hence $$\Pr{N_{t,\alpha}\geq M} = \Pr{\sup_{0\leq s\leq 1}\Xi_t(s) \leq \frac{\sum_{j=x}^{y-1}X(1,j)}{\sqrt{2S_2(y,M-1)}}}.$$ We will eventually prove that as $t\to +\infty$, $$\begin{aligned}
\label{eq:almost_Cheby}\sum_{j=x}^{y-1}X(1,j)&\weakto& \lambda,\\
\label{eq:almost_Brown}\Xi_t(\cdot)&\weakto& \mbox{a standard
Brownian Motion }B(\cdot).\end{aligned}$$ It follows from this and the independence of $\Xi_t(\cdot)$ $\sum_{j=x}^{y-1}X(1,j)$ that $$\lim_{t\to +\infty}\Pr{N_{t,\alpha}\geq M} =\Pr{\sup_{0\leq s\leq 1}B(s) \leq \lambda} = \Gamma(\lambda),$$ which is the desired result. Thus we concentrate on proving and .\
[*Proof of .* ]{} The expectation of $\sum_{\ell =x}^{y-1} X(1,\ell)$ can be estimated as follows. $$\begin{aligned}
\Ex{\sum_{\ell =x}^{y-1} X(1,\ell)} &= &S_1(\lceil \alpha t\rceil, t-\lceil \alpha t\rceil)\\
\label{eq:integral}&\sim &\int_{\alpha
t}^{(1-\alpha)t}\frac{ds}{f(s)}\\ \label{eq:ohmega}&=&
\ohmega{\frac{t}{f(\alpha t)}}\gg 1.\end{aligned}$$Indeed, follows from the fact that $\sum_{j\geq
1}f(j)^{-1}=+\infty$, and follows from the assumption $h(s)\leq 1/2$ for $s$ large, which means that $f((1-\alpha)t)=\bigoh{f(\alpha t)}=\bigoh{\sqrt{t}}$. The variance of the is $$\Var{\sum_{\ell =x}^{y-1} X(1,\ell)} = S_2(\lceil\alpha t\rceil,t-\lceil\alpha
t\rceil) =\bigoh{\frac{t}{f(\alpha t)^2}} \ll \Ex{\sum_{\ell
=x}^{y-1} X(1,\ell)}^2 \;\;(t\gg 1)$$ because for $s$ large, $h(s)\geq 0$ and thus $f$ is decreasing. By Chebyshev’s Inequality, it follows that $$\label{eq:convCheby}\frac{\sum_{\ell =x}^{y-1} X(1,\ell)}{\int_{\alpha
t}^{(1-\alpha)t}\frac{ds}{f(s)}}\weakto 1.$$
On the other hand, notice that if $M'\equiv G_{t,\alpha}(\lambda)$ satisfies $$\label{eq:definitionFok}F_{t,\alpha}(M') =
\frac{\int_{\alpha
t}^{(1-\alpha)t}\frac{ds}{f(s)}}{\sqrt{2\int_{(1-\alpha)
t}^{M'}\frac{ds}{f(s)^2}}} = \lambda\periodeq$$ We wish to show that $$\label{eq:simisgood}F_{t,\alpha}(M)\sim
\lambda;$$ this will follow from $$\int_{(1-\alpha)t}^{M'}\frac{ds}{f(s)^2} \sim \int_{(1-\alpha)
t}^{M}\frac{ds}{f(s)^2}.$$ Since $M=\lfloor M'\rfloor$, $M\leq
M'\leq M+1$ and $$\left|\int_{(1-\alpha) t}^{M'}\frac{ds}{f(s)^2}-\int_{(1-\alpha)
t}^{M}\frac{ds}{f(s)^2}\right| \leq \frac{1}{f(M')^2}.$$ This follows from the fact that $h(s)>0$ (and thus $f$ is increasing) for $s$ large enough. On the other hand, we must have $M'-(1-\alpha)t\gg 1$, as for any constant $C$ $$\int_{(1-\alpha)t}^{(1-\alpha)t+C}\frac{ds}{f(s)^2}\leq \frac{C}{f((1-\alpha)t)}\ll 1\ll \left(\int_{\alpha
t}^{(1-\alpha)t}\frac{ds}{f(s)}\right)^2,$$ contradicting . Thus $$\int_{(1-\alpha)t}^{M'}\frac{ds}{f(s)^2}\geq \frac{M'-(1-\alpha)t}{f(M')^2}\gg \left|\int_{(1-\alpha) t}^{M'}\frac{ds}{f(s)^2}-\int_{(1-\alpha)
t}^{M}\frac{ds}{f(s)^2}\right|,$$ which proves . In particular, it follows from that $$\label{eq:almost_sec}\frac{\sum_{j=x}^{y-1}X(1,j)}{\int_{(1-\alpha)
t}^{M}\frac{ds}{f(s)^2}}\weakto \lambda.$$ Finally, since by the numerator of $$F_{t,\alpha}(M) = \frac{\int_{\alpha t}^{(1\alpha)t} \frac{ds}{f(s)^2}}{\sqrt{2\int_{(1-\alpha)t}^M \frac{ds}{f(s)^2}}}$$ diverges, we deduce that $\int_{(1-\alpha)t}^M
\frac{ds}{f(s)^2}\gg 1$, which implies that $$\int_{(1-\alpha)t}^M
\frac{ds}{f(s)^2}\sim S_2(y,M-1).$$ Plugging this into yields .\
[*Proof of .* ]{} We have already shown above that $S_2(y,M-1)\sim \int_{(1-\alpha)t}^M \frac{ds}{f(s)^2}\gg 1$. Since $f$ is bounded below, $$S_3(y,M-1) = \bigoh{S_2(y,M-1)}\ll S_2(y,M-1)^{3/2}.$$ Thus the condition for the Invariance Principlie in is satisfied, and this finishes the proof.
Extensions and open problems {#sec:last}
============================
Our admits an extension to the case of a general number $B\geq 2$ of bins.
\[thm:scaling\_main2\]Let $f$ be a ELM function, $B\geq 2$ and $\lambda_i\in\R$ ($i\in B$) be constants, and assume that $q=q(n)$ ($n\in\N$) is such that
- $t/B +\lambda_i \,q(n)\in \N$ for all $n\in\N$ and $i\in[B]$;
- $q(n)\sim q_0(n)$ for $n\gg 1$ ($q_0$ is defined in ).
Now consider the $B$-bin balls-in-bins process started from initial state $$(x_i(t))_{i=1}^B =\left(\frac{t}{B}+\lambda_i\, q(t)\right)_{i=1}^B \in \N^B\periodeq$$ Then $$\begin{aligned}
\lim_{t\to +\infty}\Prp{(x_i(t))_{i=1}^B}{\elead_1} & = & \Pr{\forall 2\leq i\leq B\, , \frac{N_1-N_i}{\sqrt{2}} < (\lambda_1 - \lambda_i)}\\
\lim_{t\to +\infty}\Prp{(x_i(t))_{i=1}^B}{\lead_1} & = &
\Pr{\forall 2\leq i\leq B\, , \sup_{0\leq t\leq 1}\frac{(B_1(t) -
B_i(t))}{\sqrt{2}}< (\lambda_1 - \lambda_i)},\end{aligned}$$ where $\{B_i(\cdot)\}_{i=1}^B$ ($\{N_i\}_{i=1}^B$) are i.i.d. standard Brownian Motions (resp. Gaussians).
The proof of this result, which we omit, follows essentially the same lines as that of . The only major differences is that the Invariance Principle is applied to the sequences $$\{\xi^{(i)}_{n,t}=X(i,x_i(t)+n-1) - f(x_i(t)+n-1)^{-1}\}_{t\in\N, n\in\N}\,\,(i\in[B])$$ and some extra care must be taken in looking at the maxima of the corresponding random continuous functions.
It is not entirely obvious how on the almost balanced regime should or could be generalized to more than $2$ bins. All we can say in general is that the addition of bins to the process will always make any overtaking take longer (to see this, notice that more bins just add more arrivals in the continuous-time exponential embedding between the start and overtaking times). In fact, there is a more basic question about such processes that we cannot answer; it was posed as a conjecture by Joel Spencer.
Consider a balls-in-bins process $(I_m(i))_{i\in[B],m\geq 0}$ with feedback function $f$ that is in the almost-balanced regime (cf. ). Then for all permutations $\Pi$ of $[B]$ and all initial conditions, with probability $1$ there are infinitely many $m\geq 0$ with $I_m(\Pi(1))<I_m(\Pi(2))<\dots<I_m(\Pi(B))$. That is, all permutations possible of the bins occur infinitely often almost surely.
Perhaps the techniques presented here could be used to settle this problem.
[^1]: IBM T.J. Watson Research Center, Yorktown Heights, NY 10598. `[email protected]`. Work mostly done while the author was a Ph.D. student at New York University under a CNPq scholarship.
[^2]: A longer background discussion is available from the author’s PhD thesis [@tese].
[^3]: More specifically, the model introduced by Krapivsky and Redner [@Krapivsky02] and independently by Driena, Enachescu and Mitzemacher [@DrineaEM]. This model generalizes the Barabási-Álbert model, which is discussed in the survey [@AlbertSurvey].
[^4]: The exact attribution of this result is somewhat confusing. Ref. [@Khanin01] cites the work of Davis [@Davis90] on reinforced random walks, where it is in turn attributed to Rubin.
[^5]: The requirement that $f(1)=1$ is just a normalization condition, as it does not change the process.
| {
"pile_set_name": "ArXiv"
} |
---
author:
- Alvaro Ballon Bordo
bibliography:
- 'references.bib'
title: '**Geometric and Thermodynamic Volume of Hairy Black Branes**'
---
Introduction
============
Since the publication of the extended first law of black hole thermodynamics in [@Kastor:2009wy], which allows for a varying cosmological constant, the addition of pressure and volume terms to obtain a full cohomogeneity first law has become widespread. In this perspective, the asymptotic spacetime energy $M$ is interpreted as its enthalpy and the cosmological constant acquires the meaning of pressure, with the geometric black hole volume as its conjugate potential. Explicitly, in a $D$-dimensional Schwarzschild-AdS spacetime, the Smarr relation $$\label{eqn:Smarr1}
(D-3)M= (D-2)TS+2 P V$$ and the thermodynamic law $$\label{eqn:FirstLaw1}
dM= TdS+ VdP$$ are obtained using purely geometric methods. These formulas can be extended to rotating and charged cases [@Cvetic:2010jb; @Dolan:2013ft], to Taub-NUT spacetimes [@Johnson:2014xza; @Kubiznak:2019yiu; @Ballon:2019uha; @Bordo:2019tyh], to different black hole topologies [@Altamirano:2014tva; @El-Menoufi:2013pza; @Dolan:2010ha] and even to higher curvature theories of gravity [@Kastor:2010gq; @Liberati:2015xcp], provided that the notion of thermodynamic volume is appropriately generalized. This quantity will only be equal to the naive spatial black hole volume in the simpler scenarios. A review of most of these developments can be found in [@Kubiznak:2016qmn]. Additionally, some explicit examples of extended first laws are known to hold in the presence of scalar fields [@Cvetic:2010jb; @Astefanesei:2019ehu], in which case the appropriate extension of the volume term is conjectured to be proportional to the average value of the potential inside the black hole region.
Our study will focus on the thermodynamics of planar black holes with a minimally coupled scalar field, which are asymptotically anti-de Sitter spacetimes that contain a horizon with the topology of a plane. It is then wise to first review the case in which the scalar field and any other matter fields are absent. We do this by summarizing some relevant results found in [@El-Menoufi:2013pza] and [@Traschen:2001pb]. The vacuum planar black hole solution is given by the metric $$\label{eqn:BBMetric}
ds^2=-\frac{r^2}{\ell^2}\left(1-\frac{{r_{+}}^{D-1}}{r^{D-1}}\right)dt^2+\frac{\ell^2}{r^2}\left(1-\frac{{r_{+}}^{D-1}}{r^{D-1}}\right)^{-1}dr^2+\frac{r^2}{\ell^2}\delta_{ij}dx^{i}dx^{j}$$ where $r={r_{+}}$ is the locus of the horizon and $\ell$ is the AdS radius. Unlike the spherical case, the transverse directions $x_{i}$, $i=1\ldots,D-2$, are not compact, but we can make them so by identifying $x^{i}\equiv x^{i}+L^{i}$. Such a spacetime possesses $(D-1)$ evident Killing vectors $\left\lbrace \partial_{t}, \partial_{i}\right\rbrace$ that generate the time and spatial translations symmetries in each of the $x^{i}$ directions. Associated to each of these symmetries, we can find $(D-1)$ ADM charges relative to the pure AdS background via the prescriptions given, for example in [@PhysRevD.31.283].
The mass $M$ is the asymptotic charge corresponding to the timelike Killing vector, which reads $$\label{eqn:MassBB}
M=(D-2)\frac{{r_{+}}^{D-1} v}{16 \pi G \ell^2}.$$ Here, we have defined the dimensionless quantity $v\equiv \prod_{i=1}^{D-2}\frac{L^{i}}{\ell}$ which is the volume of the box $\mathcal{B}$ with sides parallel to the transversal directions in units of $L$. Similarly, we can calculate the $(D-2)$ ADM tensions corresponding to each of the translation symmetry generators $$\label{eqn:TensionsBB}
\uptau_{i}=-\frac{{r_{+}}^{D-1} v}{16 \pi G \ell^2 L^{i}},$$ where these should be interpreted, in fact, as tensions per unit time, given that a time integration is ommitted. We should note that the same results are obtained by using the holographic stress-energy tensor. Thus, the Smarr relation $$\label{eqn:Smarr2}
M+\sum_{i=1}^{D-2}\uptau_{i}L^i=0$$ can be interpreted as the trace of the conformal stress tensor being zero, as is expected for the dual thermal CFT.
Furthermore, by imposing that the entropy be related to the horizon area via $S=A/4G$ and calculating the temperature with the usual prescription $T=f'({r_{+}})/4\pi$ we obtain, in terms of ${r_{+}}$, $$\label{eqn:EntropyVac}
S=\frac{{r_{+}}^{D-2}v}{4 G}, \qquad T=\frac{D-1}{4}\frac{{r_{+}}}{\pi \ell^2}.$$ It is then straightforward to verify that, if the AdS scale $\ell$ and the compactifying lengths $L^{i}$ are not regarded as thermodynamic variables, the Smarr relation $$\label{eqn:Smarr3}
(D-1)M=(D-2)TS$$ and the first law of thermodynamics $$\label{eqn:ProtoFirstLaw}
dM = TdS$$ are satisfied. However, in the extended thermodynamics, we do allow a variation in $\ell$ and $L^{i}$ to obtain a differential relation similar to (\[eqn:FirstLaw1\]). We calculate the geometric volume of the planar black hole by integrating the volume element over the interior $$\label{eqn:BBVolume}
V=\int_{0}^{{r_{+}}}dr\int_{\mathcal{B}}d\vec{x}\sqrt{-g}=\frac{v{r_{+}}^{D-1}}{D-1},$$ and we identify the spacetime pressure with the cosmological constant $$\label{eqn:Pressure}
P=-\frac{\Lambda}{8\pi G} =\frac{(D-1)(D-2)}{16 \pi G \ell^2}.$$ Let us observe that $$\label{eqn:PVM}
PV=M.$$ Therefore, we can combine (\[eqn:Smarr2\]) and (\[eqn:Smarr3\]) to obtain a Smarr relation that involves all the thermodynamic quantities $$\label{eqn:Smarr4}
(D-3)M=(D-2)TS + \sum_{i=1}^{D-2}\uptau_{i}L^i - 2 P\Theta$$ where the thermodynamic volume $\Theta$ is one half of the geometric volume: $$\label{eqn:theta}
\Theta\equiv V/2.$$ Note that this Smarr relation, unlike (\[eqn:Smarr3\]), can be derived from the usual scaling argument according to the engineering dimensions of the thermodynamic charges. With these quantities, the following first law of planar black hole thermodynamics is satisfied $$\label{eqn:FirstLaw2}
dM=TdS+\sum_{i=1}^{D-2}\uptau_{i}dL^{i}+\Theta dP.$$ It is interesting to note that, although the extended thermodynamics of black branes has been studied for some time, the relation between the geometric and thermodynamic volumes has never been explicitly stated. This relation is true for planar black holes, but in the spherical case the factor of 1/2 is absent. Although this may seem puzzling, the ultimate reason is that when we compactify the transverse dimensions, the thermodynamic variables depend on the dimensionless transverse volume $v$, which is dependent on $\ell$. Varying the AdS scale will then change the value of $v$. In the spherical case, the transverse coordinates are already compactified so this does contribution does not exist.
The thermodynamics of black branes can be recast without making any reference to the geometric volume; we can express the variable conjugate to the pressure solely in terms of $M$ and $\Lambda$. Namely, using (\[eqn:PVM\]), we can write $$\label{eqn:LambdaFirstLaw}
dM=TdS+\sum_{i=1}^{D-2}\uptau_{i}dL^{i}+\frac{M}{2\Lambda}d\Lambda.$$ Armed with this simpler case, we can now proceed to generalize these laws to a spacetime that contains a single scalar field; the generalization to many scalar fields should be straightforward. We will find that the first law as written in (\[eqn:LambdaFirstLaw\]) applies also to the hairy black branes. An equation analogous to (\[eqn:FirstLaw2\]) can also be written using an appropriate generalization for the definition of volume.
In the next section we will review the behaviour of fully back-reacted black brane solutions to the Einstein-Scalar field equations and calculate the asymptotic charges in a very general setting. In section 3 we obtain a Smarr relation akin to (\[eqn:Smarr4\]) by proposing a generalization for the volume term, which turns out to be similar to that conjectured in [@Cvetic:2010jb] and involves a volume integral of the scalar potential between the Poincaré and event horizons. We then prove that the proposed quantities satisfy an extended first law. In section 4 we exhibit various examples from the literature and show explicitly that they satisfy the proposed thermodynamic relations. Section 5 is dedicated to the discussion of our results.
Hairy black brane solutions
===========================
Asymptotic behaviour
--------------------
We seek to describe the thermodynamics of black branes with scalar hair in a very general setup. This means that we need to characterize some universal asymptotic behaviour for spacetimes in which a scalar field $\phi$ couples minimally to gravity in a $D$-dimensional spacetime. This physical situation is described by the bulk action $$\label{eqn:Action}
I_{bulk}=-\int d^{D}x\sqrt{-g}\left(\frac{R}{16\pi G}-\frac{1}{2}\partial_{\mu}\phi\partial^{\mu}\phi-2 \Lambda \mathcal{V}(\phi)\right).$$ As can be read from this expression, the scalar potential is taken to be of the form $V(\phi)=-2\Lambda\mathcal{V}(\phi)$ where $\mathcal{V}(\phi)$ is has no explicit dependence on the parameter $\Lambda$. To solve for the metric, we need to find solutions to the Einstein field equations with a negative cosmological constant in the presence of the scalar field, which gives rise to a non-trivial stress energy tensor.
We are interested in radially symmetric spacetimes, and we will assume that the scalar field respects the same symmetries. With this in mind, the most general metric ansatz in Poincaré coordinates is given by [@Gubser:2008ny] $$\label{eqn:MetricAnsatz}
ds^2=\frac{r^2}{\ell^2}e^{2A(r)}(-h(r)dt^2+\delta_{ij}dx^i dx^j)+\frac{\ell^2}{r^2}\frac{\sigma^2(r)}{h(r)}dr^2, \quad i=2,\ldots, D.$$ The constant $\ell$ in this expression is related to the prefactor $\Lambda$ of the potential, which will be varied in the solutions to establish the extended thermodynamics. For asymptotically AdS spacetimes, the relation is $$\label{eqn:LambdaLength}
\Lambda=-\frac{(D-1)(D-2)}{2 \ell^2};$$ however, this relation may change for different asymptotic behaviours of the solutions. An example of this case are the asymptotically dilaton AdS spacetimes [@Charmousis:2010zz; @Kim:2012pd; @Kastor:2018cqc], where the potential is given by an exponential function of a scalar field with a finite value at infinity. In this work, we will restrict our study to asymptotically AdS spacetimes. In terms of the potential, this means that $\mathcal{V}(\phi_{\infty})=1/16\pi G$, where $\phi_{\infty}$ is the limit of the value of the scalar field as one approaches the conformal boundary.
In terms of the functions $A(r)$, $\sigma(r)$ and $h(r)$, there are three independent gravitational field equations which read $$\label{eqn:eom1}
(D-2) \left(r A'+1\right) \left((D-1) h \left(r A'+1\right)+r h'\right)-8\pi G(r^2 h \phi'^2-2 \ell^2 V(\phi)\sigma^2)=0,$$ $$\label{eqn:eom2}
\left(((D-1) r A'+D) -\frac{r\sigma'}{\sigma}\right) h'+r h''=0,$$ $$\label{eqn:eom3}
r \left((D-2) A''+8 \pi G \phi'^2\right)+(D-2) A'-(D-2)(1+rA')\frac{\sigma'}{\sigma}=0.$$ The condition that the spacetime be asymptotically AdS translates into the limits $A\rightarrow 0 $, $\sigma\rightarrow 1 $ and $h\rightarrow 1$ as $r\rightarrow \infty$. In general, it is no trivial matter to obtain solutions to these coupled equations. The handful of solutions that have been found in the literature use a specific form of the potential that allow for exact solutions (see, for example: [@Acena:2012mr; @Acena:2013jya; @Fan:2015ykb]). Nevertheless, we can make some progress as equation (\[eqn:eom2\]) allows us to express the blackening factor $h$ as the exact integral, independent of the form of the potential. We obtain $$\label{eqn:BlackFactor}
h(r)=(D-1)\mu\int_{r_{+}}^{r}\frac{e^{-(D-1)A(r)}\sigma(r)}{r^D}dr,$$ where the constant $\mu$ is chosen so that $h\rightarrow 1$ as $r\rightarrow \infty$. The radial value $r_{+}$ is a zero of $h$, which is required to exist if we are to have a horizon at $r=r_{+}$. Note that, at large values of $r$, we can expand $$\label{eqn:AsympBlackFactor}
h\sim 1-\frac{\mu}{r^{D-1}}+\ldots$$
The full characterization of the asymptotics of the metric will depend on the particular form of the scalar potential. Without loss of generality, if $\phi\rightarrow 0$ as $r\rightarrow\infty$, the asymptotic behaviour of the scalar field to leading order in $r$ is $$\label{eqn:AsympPhi}
\phi\sim \frac{\phi_{0}}{r^{D-1-\Delta}}+\frac{\phi_{+}}{r^{\Delta}}.$$ Here, the exponent $\Delta$ is fixed by the specific form of the potential and corresponds holographically to the conformal dimension of the dual operator. In the following, this asymptotic behaviour is taken to be fixed.
In the above, we have not made any assumptions about the behaviour of the metric close to $r=0$. In section 3, we will impose a certain condition on the behaviour of the functions so that there is a singularity at $r=0$ and the integral of the potential is finite inside the horizon.
ADM charges
-----------
We face a challenge when calculating the masses of scalar field spacetimes using the ADM method: if we have an asymptotic behaviour of the scalar field characterized by two or more parameters, then it has been shown in [@Lu:2014maa] that the Hamiltonian variation at infinity cannot be expressed as a total derivative, so the mass is not well defined. We will circumvent this problem by taking the pure scalar field spacetime as the background, and calculate the mass contribution of the planar black hole only. Fot this purpose, we will use a version of the Hamiltonian formalism introduced in [@PhysRevD.31.283].
To calculate the ADM mass, first we foliate spacetime with spacelike hypersurfaces $\Sigma$, with unit normal vector $n^{\mu}$. The induced metric in these hypersurface is then $$\label{eqn:InducedMetric}
\gamma_{\mu\nu}=g_{\mu\nu}-n_{\mu}n_{\nu}.$$ The canonical momentum associated to $\gamma_{\mu\nu}$ is $$\label{eqn:ConjugateMomentum}
\pi^{\mu\nu}=-\sqrt{\gamma}\left(K\gamma^{\mu\nu}-K^{\mu\nu}\right)$$ where $K_{\mu\nu}$ is the extrinsic curvature of $\Sigma$. The second ingredient in the calculation is a background metric, which we take to be the pure scalar field spacetime, defined by setting $h(r)=1$ in (\[eqn:MetricAnsatz\]). The induced background metric in the boundary and its conjugate momentum are denoted by $\gamma^{(0)}_{\mu\nu}$ and $\pi^{(0)}_{\mu\nu}$ respectively. The corresponding differences between the quantities on $\Sigma$ and on the background are denoted by $\delta \gamma_{\mu\nu}$ and $\delta \pi_{\mu\nu}$.
The ADM mass is the conserved charge associated to the conserved Noether current correspondent to the time translation symmetry of our spacetime. The background Killing vector that generates this symmetry is $\xi=\partial/\partial t$. If $F$ is the lapse of $\xi$, meaning the component of $\xi$ along the direction of the normal vector field to the foliation $\Sigma$, and $F^{a}$ is the shift, the component parallel to the family of hypersurfaces, then the mass can be expressed as $$\label{eqn:ADMCharge}
M=-\frac{1}{16 \pi G}\int_{\partial\Sigma_{\infty}}da_{c}B^{c}$$ where [@PhysRevD.31.283] $$\label{eqn:Bvector}
B^{c}= F(D^{c}\delta\gamma-D_{b}\delta\gamma^{bc})-\delta\gamma D^{c}F+\delta\gamma^{cb}D_{b}F +\frac{1}{|\gamma|}F^{d}\left( \pi^{(0)ab}\delta\gamma_{ab}\tensor{\gamma}{^{(0)}^c_d}+2\pi^{(0)cb}\delta\gamma_{db}-2\delta\tensor{\pi}{^c_d} \right).$$ Here, $D_{a}$ denotes the covariant derivative corresponding to the background induced metric $s^{(0)}_{ab}$. We note that, in general, there is an additional contribution of matter to the vector $B^{c}$. For our purposes, we will hold the asymptotic behaviour of the scalar field fixed, so that $\delta\phi=0$. We can take a foliation $\Sigma$ such that the normal vector field to the surfaces is parallel to the timelike Killing vector $\xi=\partial/\partial t$. As a consequence, $\xi$ has lapse $F=r/L$ and shift $F^{a}=0$. In our case, the perturbations read $\delta\gamma_{tt}=\delta\gamma_{rr}=\mu/r^{(D-1)}$. Evaluating the mass of the black brane we obtain, in $D$ spacetime dimensions $$\label{eqn:ADMMass}
M=(D-2)\frac{\mu v}{16 \pi G \ell^2}.$$
The tensions can be calculated in a similar way [@Traschen:2001pb]. The only difference is that the foliation of spacetime is now given by timelike hypersurfaces (changing the minus sign in (\[eqn:InducedMetric\])). This foliation is chosen in such a way that the normal vector is parallel to the Killing vector $\xi_{k}=\partial/\partial x^{k}$. For the spatial tension associated to the symmetry in the $k$-th spatial direction we obtain $$\label{eqn:ADMTension}
\uptau_{k}=-\frac{\mu v}{16 \pi G \ell^2 L^{k}},$$ As in the case of the hairless black brane, we find that the mass and the tensions satisfy $$\label{eqn:Traceless}
M+\sum_{i=1}^{D-2}\uptau_{k}L^{k}=0.$$ Having assumed that the spacetime is asymptotically AdS, this is again interpreted as the trace of the stress energy tensor in the dual CFT being zero[^1]. We do not expect this to hold in the case where the asymptotics deviate from anti-de Sitter. Such hyperscaling violating spacetimes and their thermodynamics have been studied, for example, in the case of asymptotically dilaton AdS spacetimes [@Kastor:2018cqc].
It is worthwhile to insist on the fact that the mass calculated above does not include the energy contribution from the scalar field; as such it represents a gravitational energy. To account for the energy contained in the scalar field, one would have to use either the holographic [@Brown:1992br; @Balasubramanian:1999re] or Wald prescriptions [@Wald:1993nt] using appropriate counterterms [@Emparan:1999pm; @Buchel:2013lla; @Papadimitriou:2005ii]. In this paper we do not vary the asymptotic behaviour of the scalar field, so it will not contribute to the first law of thermodynamics. Work in these lines, albeit without varying the cosmological constant, has been carried out in [@Liu:2015tqa; @Gursoy:2018umf].
Extended thermodynamics
=======================
Smarr relation
--------------
We now proceed to derive a Smarr relation for the scalar-black brane solution. Such an equation will relate the mass to the entropy, the tensions and the black brane volume, which is in general different from the vacuum black brane spacetime. Obtaining such an equation tells us what to expect the thermodynamic volume to be equal to, so that we can check the first law of thermodynamics later. There are two ways to go about this: the first is to perform a scaling transformation and use Euler’s theorem, which is straightforward provided that we already know the first law; the second is to use geometric methods, namely a Komar intergration. In this section we will illustrate how to derive the Smarr relation using the scaling argument.
The scaling method was already presented in [@Liu:2015tqa] in the context of a fixed cosmological constant. We utilize this result in order to generalize it to varying compactifying and AdS scales. Under a scaling transformation. Without rescaling $\ell$, there exists a scaling symmetry of the metric $$\label{eqn:Scaling1}
r\rightarrow \lambda r, \qquad t\rightarrow \lambda^{-1}t, \qquad x^{i}\rightarrow x^{i}\lambda^{-1}$$ provided that $\mu\rightarrow \lambda^{D-1}\mu$. The entropy scales as $S\rightarrow \lambda^{D-2}S$, as can be deduced from dimensional analysis. Considering the mass as a function only of the entropy, we observe that $M(S)$ is a homogeneous function such that $$\label{eqn:MassScaling}
M(\lambda^{D-2}S)=\lambda^{D-1}M(S).$$
Euler’s theorem for homogeneous functions states that, if a function $f$ is such that $$f(\lambda^{p}x, \lambda^{q}y,\lambda^{r}z)=\lambda^{s}f(x,y,z),$$ then $$\label{eqn:EulerTheorem}
s f = p\left(\frac{\partial f}{\partial x}\right)+q\left(\frac{\partial f}{\partial y}\right)+r\left(\frac{\partial f}{\partial z}\right).$$ The homogeneity of the mass along with this theorem imply that we must have $$\label{eqn:ProtoSmarr}
(D-1)M=(D-2)\left(\frac{\partial M}{\partial S}\right)S.$$ This relation still holds true once we vary the compactifying and AdS scales. However, it is not useful to define a thermodynamic volume. First of all, one would need to know a priori that $\partial M/\partial S$ is the temperature to obtain a Smarr relation as in equation (\[eqn:Smarr3\]), and for that we need to know that a first law $dM=TdS$ holds true. This is indeed the case, as shown in [@Liu:2015tqa]. Secondly, we need to study the mass dependence on the cosmological constant to obtain a relation akin to (\[eqn:Smarr4\]). We can do this by noticing that the scaling transformation $$\label{eqn:Scaling2}
r\rightarrow \lambda r, \qquad t\rightarrow \lambda t, \qquad x^{i}\rightarrow \lambda x^{i}, \qquad \ell \rightarrow \lambda \ell$$ is a symmetry of the metric. Note that because of the transformation rule for the transversal coordinates $x^{i}$, the compactifying lengths scale as $L^{k}\rightarrow \lambda L^{k}$.
Considering the mass as a function of the $L^{k}$ and the cosmological constant, we see that it is homogeneous in the following way: $$\label{eqn:MassScaling2}
M(\lambda^{D-2}S, \lambda L^{k}, \lambda^{-2}\Lambda)=\lambda^{D-3}M(S,L_{k},\Lambda).$$ Then Euler’s theorem immediately implies that $$\label{eqn:EulerSmarr1}
(D-3)M=(D-2)\left(\frac{\partial M}{\partial S}\right)S+\sum_{k}\left(\frac{\partial M}{\partial L^{k}}\right)L^{k}-2\left(\frac{\partial{M}}{\partial\Lambda}\right)\Lambda.$$ Now, we recall that $M+\sum_{T_{k}}L_{k}=0$ from the fact that the holographic stress energy tensor must be traceless. Then, by subtracting $2M$ on each side of equation (\[eqn:ProtoSmarr\]), it is permissible to write $$\label{eqn:FinalSmarr}
(D-3)M=(D-2)\left(\frac{\partial M}{\partial S}\right)S+\sum_{k}\uptau_{k}L^{k}-2\left(\frac{M}{2\Lambda}\right)\Lambda.$$ From this we could identify, up to some overall coefficients, $\partial M/\partial L^{k}=\uptau^{k}$ and $\partial M/\partial \Lambda =M/2\Lambda$, the latter being proportional to the volume term. This turns out to be the correct identification but an independent derivation of the first law is needed to make it precise.
Komar Integration
-----------------
Before deriving such a first law, we present a geometric derivation of the Smarr relation using Komar integration. This derivation will elucidate a relation between the integral of the potential behind the black brane horizon and the volume term. By tracing out the gravitational field equations, we can express the Ricci tensor as $$\label{eqn:Ricci}
R_{\mu\nu}=8 \pi G \left(\partial_{\mu}\phi\partial_{\nu}\phi + \frac{2}{D-2}g_{\mu\nu}V(\phi)\right);$$ then, we contract this equation with the timelike Killing vector $\xi^{\alpha}$ and, using the Killing identity $\square\xi^{\alpha}=-\tensor{R}{^\alpha_\beta}\xi^{\beta}$, we observe that $$\label{eqn:GaussLaw}
\nabla_{\mu}\nabla^{\mu}\xi^{\nu}+\frac{16 \pi G}{D-2}\xi^{\nu}V(\phi)=0.$$
We have assumed that the scalar field is static, so that $\xi^{\alpha}\partial_{\alpha}\phi=0$. This also implies that the divergence $\nabla_{\alpha}(\xi^{\alpha}V(\phi))$ vanishes. Therefore, there exists a Killing potential, which is a 2-form $\omega^{\alpha\beta}$ such that $$\label{eqn:KillingPotential}
V(\phi)\xi^{\beta}=\nabla_{\alpha}(V(\phi)\omega^{\alpha\beta}).$$ Combining the previous results, we have that $$\label{eqn:GaussLaw2}
\nabla_{\mu}\left(\nabla^{\mu}\xi^{\nu}+\frac{16 \pi G}{D-2}\omega^{\mu\nu}V(\phi)\right)=0.$$ We can then calculate the integral of this vanishing 1-form over a codimension-1 spacelike hypersurface $\Sigma$ and, by virtue of Stokes’ theorem, we may express this result as $$\label{eqn:KomarIntegral}
I=\int_{\partial\Sigma}d\sigma_{\mu\nu}\left(\nabla^{\mu}\xi^{\nu}+\frac{16 \pi G}{D-2}\omega^{\mu\nu}V(\phi)\right)=0.$$ where $\sigma_{\mu\nu}$ is the binormal to the boundary of $\Sigma$, given a choice of orientation. The boundary $\partial\Sigma$ has a component at infinity and a component at the bifurcation surface $H$. Hence, we can cast this as a sum $I_{\infty}-I_{H}=0$.
The assumption of spherical symmetry implies that the only non-vanishing components of $\omega^{\alpha\beta}$ are $\omega^{rt}=-\omega^{tr}$. In fact, we can use equation (\[eqn:KillingPotential\]) to express this component in terms of an integral of the potential $$\label{eqn:KillingPotentialSol}
\omega^{rt}=\frac{1}{\sqrt{-g}V(\phi)}\int_{\rho}^{r} dr' \sqrt{-g}V(\phi).$$ Here $\rho$ is a constant of integration and is therefore arbitrary, and it accounts for the fact that the Killing potential is unique only up to the addition of a closed 2-form. The integral (\[eqn:KomarIntegral\]) can then be rewritten as $$\label{eqn:KomarIntegral2}
I=\int d^{D-2}x\sqrt{-g}\nabla^{r}\xi{^t}+\frac{16 \pi G}{D-2}\int d^{D-2}x\int_{\rho}^{r}dr'\sqrt{-g} V(\phi)=0.$$ We need to evaluate this integral at infinity and at the horizon. Both integrals will depend on our choice of $\rho$, but we expect this dependence to cancel out. Remarkably, there is an exact expression for the integral $I$ that comes from using the equations of motion. First, one must use (\[eqn:eom3\]) to find $\phi'$ in terms of the metric functions $A$ and $\sigma$. This result can be plugged into equation (\[eqn:eom1\]) to solve for the potential. Then the quantity $\sqrt{-g}V(\phi)$ is a total derivative in the radial variable, so the indefinite integral can be readily calculated. As a result we obtain $$\label{eqn:W}
W(r)\equiv\int dr\sqrt{-g}V(\phi)= -\frac{(D-2) r^{D-1} h(r) e^{(D-1) A(r)} \left(r A'(r)+1\right)}{16 \pi G \ell^D \sigma(r)}.$$ There are no integrals over the radial coordinate for the $\nabla^{r}\xi^{t}$ term, so this term can be evaluated without any further complications. The result is $$\nabla^{r}\xi^{t}= \frac{r \left(2 h(r) \left(r A'(r)+1\right)+r h'(r)\right)}{2\ell^2 \sigma^2 (r)}.$$
We can use these results to get a simple expression for the integrand in (\[eqn:KomarIntegral2\]). We obtain $$\label{eqn:IntegralR}
I=\frac{r^D e^{(D-1)A(r)} h'(r)}{2 \ell^2 \sigma(r)} v-\frac{16 \pi G W(\rho)}{D-2}v$$ With this expression at our disposal, we can calculate $I_{\infty}$ at a sphere at infinity and $I_{H}$ at the horizon. Here, we note that $W(r_{+})=0$ provided that $\sigma(r)$, $A(r)$ and its derivative are finite at the horizon, that is, we assume that the coordinate singularity at $r=r_{+}$ comes only from the blackening factor $h(r)$. We also use the fact that $\xi^{\alpha}$ generates the black brane horizon, and hence the integral of $\nabla^{\mu}\xi^{\nu}$ over $H$ is equal to $8\pi G TS$. The asymptotic behaviours of $A$ and $h$ allow us to express $$\label{eqn:IntegralInfty}
I_{\infty}= 8 \pi G \left( \frac{D-1}{D-2}M-\frac{2 W(\rho) v}{D-2}\right),$$ $$\label{eqn:IntegralHorizon}
I_{H}= 8 \pi G \left( T S -\frac{2 W(\rho) v}{D-2}\right).$$ Upon imposing $I_{\infty}=I_{H}$, the $\rho$ dependent terms cancel. We also note that the divergent part of the integral of $\nabla^{r}\xi^{t}$ is always cancelled by the integral of the potential, yielding a term proportional to the mass. We thus obtain the Smarr relation $$\label{eqn:FakeSmarr}
(D-1)M=(D-2)TS,$$ which is the same as (\[eqn:Smarr3\]). Thus, considering $\Lambda$ and $L_{k}$ thermodynamic variables as before and defining $P=-\Lambda/8 \pi G$, we obtain the following Smarr relation analogous to (\[eqn:Smarr4\]): $$\label{eqn:FinalSmarr2}
(D-3)M=(D-2)TS + \sum_{i=1}^{D-2}\uptau_{i}L^i - 2 P\Theta.$$ Here, we have subtracted $2M$ on each side of (\[eqn:FakeSmarr\]) by using (\[eqn:Traceless\]) and by defining the thermodynamic volume $$\Theta\equiv -4\pi G M/\Lambda.$$ This means that the extended thermodynamics of black branes is in fact very simple, and the thermodynamic volume can be readily calculated once we know the mass. Comparing to the purely gravitational case, this makes us wonder what is the relation between this thermodynamic volume and the actual geometric volume enclosed between the singularity and the black brane horizon. This relation was straightforward in the hairless case: all one had to do is to calculate both volumes in terms of $r_{+}$ and compare, then we observed that the thermodynamic volume was one half of the geometric one. In our present case, this relation is a bit more complicated but, as we will see, the integral of the potential over the volume behind the horizon turns out to be proportional to the mass, from which $\Theta$ can be readily calculated. The constant of proportionality depends on the particular form of the solution but has a simple expression when the metric is expressed in Poincaré coordinates. Consider the integral $$\label{eqn:Upsilon}
\Upsilon = \int_{\mathcal{B}}d^{D-2}x \int_{0}^{r_{+}}dr V(\phi)=(W(r_{+})-W(0))v.$$ We know that $W(r_{+})$ is zero, so all that remains to do is to calculate $W(0)$. This may seem complicated because we did not impose a particular behaviour of the metric in the deep infrarred. However, under certain conditions, we can simplify its calculation. First, we need to assume that both $A(r)$ and its derivative are finite as $r$ goes to zero. Secondly, we assume that $\sigma(r)\sim 1/r^{\alpha}$, $\alpha \geq 1-D$ as $r\rightarrow 0$. This last assumption ensures that the integral $\Upsilon$ is finite and that there is a singularity of the metric at $r=0$. We can then calculate the limit using the exact formula for $h(r)$:
$$\begin{aligned}
W(0)&=\lim_{r\rightarrow 0} \left(-\frac{(D-2) r^{D-1} h(r) e^{(D-1) A(r)} \left(r A'(r)+1\right)}{16 \pi G \ell^D \sigma(r)}\right)\nonumber \\
&=-\left(\frac{(D-2) e^{(D-1) A(0)}}{16 \pi G \ell^D}\right)\lim_{r\rightarrow 0}\frac{h(r)}{\sigma(r)/r^{D-1}}. \label{eqn:Wzero1}\end{aligned}$$
We observe that when $\alpha\geq 1-D$, then both $h$ and $\sigma/r^{D-1}$ diverge as $r$ approaches zero. This means that we are allowed to use L’Hopital’s rule to evaluate the above limit.
$$\begin{aligned}
W(0)&=-\left(\frac{(D-2) e^{(D-1) A(0)}}{16 \pi G \ell^D}\right)\lim_{r\rightarrow 0}h'(r)\left(\frac{\sigma'(r)}{r^{D-1}}-(D-1)\frac{\sigma(r)}{r^D}\right)^{-1}\nonumber \\
&=-\left(\frac{(D-2)(D-1)\mu e^{(D-1) A(0)}}{16 \pi G \ell^D}\right)\lim_{r\rightarrow 0}\frac{e^{-(D-1)A(r)}\sigma(r)}{r^{D}}\left(\frac{\sigma'(r)}{r^{D-1}}-(D-1)\frac{\sigma(r)}{r^D}\right)^{-1}\nonumber \\
&=\frac{M}{v}\lim_{r\rightarrow 0}\left(1-\frac{1}{D-1}\frac{r\sigma'(r)}{\sigma(r)}\right)^{-1}=\frac{M}{v}\left(\frac{D-1}{D-1+\alpha}\right).\label{eqn:Wzero2}\end{aligned}$$
Therefore, the integral $\Upsilon$ becomes $$\label{eqn:UpsilonM}
\Upsilon=-\left(\frac{D-1}{D-1+\alpha}\right)M.$$ Plugging this into the thermodynamic volume $\Theta=-4\pi G M/\Lambda$, we conclude that $$\label{eqn:ThetaPotential}
\Theta=8 \pi G \left(1+\frac{\alpha}{D-1}\right)\int_{\mathcal{B}}d^{D-2}x\int_{0}^{r_{+}}dr\mathcal{V}(\phi).$$ This provides evidence for the conjecture that the thermodynamic volume must be proportional to the integral of the potential behind the horizon in the case of a hairy black hole. However, we are not finished yet. We need to prove that we can obtain an extended first law using this proposed definition of thermodynamic volume. This is the purpose of the following subsection.
Extended first law
------------------
The Hamiltonian method that we previously employed to calculate the mass and the tensions can also be used to derive the first law of thermodynamics. The trick is to take the scalar black brane spacetime as the background spacetime, and then calculate the mass of a the spacetime given by a metric with the parameters $\mu$, $\ell$ and $L_{k}$ slightly shifted. This will give us a relation between the perturbed parameters $\delta\mu$, $\delta\ell $ and $\delta L_{k}$, and from here we can obtain the first law of thermodynamics. This procedure is known as the Hamiltonian perturbations formalism [@PhysRevD.31.283; @Kastor:2009wy; @Kastor:2018cqc].
When we have a variable cosmological constant, this procedure needs to be done with more care, as now the Hamiltonian for the scalar fields needs to be taken into account. This is due to the fact that, even though we keep the asymptotic behaviour of the scalar field constant, the potential has an explicit overall factor of $\Lambda$ that needs to be accounted for in the perturbations. The linearized perturbations can be shown to satisfy the Einstein field equations; as a consequence, on a constant time slice $\Sigma$, $$\label{eqn:GaussTherm}
D_{a}\left(\frac{B^{a}}{16 \pi G}+2\delta\Lambda\mathcal{V}(\phi)\omega^{ab}n_{b}\right)=0.$$ Here $B^{a}$ is calculated as in (\[eqn:Bvector\]) and we use the same notation as in section II B. This time, the metric perturbations as $r$ goes to infinity are given by $$\label{eqn:PertTherm1}
\delta\gamma_{rr}=\left(\frac{\ell}{r}\right)^2\frac{\delta\mu}{r^{D-1}},$$ $$\label{eqn:PertTherm2}
\delta\gamma_{ij}=\delta_{ij}\frac{r^2}{\ell^2}\left(2\frac{\delta L_{k}}{L_{k}}-2\frac{\delta\ell}{\ell}\right).$$ To establish the first law, we integrate (\[eqn:GaussTherm\]) over $\Sigma$ and use Stokes’ theorem to express it as a boundary integral. As before, the boundary of $\Sigma$ has disconnected components $\partial\Sigma_{\infty}$ at infinity and $\partial\Sigma_{H}$ at the horizon. We define $$\label{eqn:JInfinity}
J_{\infty}=\int_{\partial\Sigma_{\infty}}d\sigma_{a}\left(\frac{B^{a}}{16 \pi G}+2\delta\Lambda\mathcal{V}(\phi)\omega^{ab}n_{b}\right), \quad \textrm{and}$$ $$\label{eqn:JHorizon}
J_{H}=\int_{\partial\Sigma_H}d\sigma_{a}\left(\frac{B^{a}}{16 \pi G}+2\delta\Lambda\mathcal{V}(\phi)\omega^{ab}n_{b}\right);$$ therefore we have $J_{\infty}-J_{H}=0$. For our purposes, it is easier to express the potential part of the integral as a volume integral. This can be done by virtue of (\[eqn:KillingPotentialSol\]), so that
$$\label{eqn:GaussTherm2}
\int_{\partial\Sigma_{\infty}}d\sigma_{a}\left(\frac{B^{a}}{16 \pi G}\right)+2\delta\Lambda\int d^{D-2}x\int_{r_{+}}^{\infty}dr\sqrt{-g}\mathcal{V}(\phi)=\int_{\partial\Sigma_{H}}d\sigma_{a}\left(\frac{B^{a}}{16 \pi G}\right).$$
This allows us to use the exact expression for the integral of the potential that we calculated previously in (\[eqn:W\]). It only remains to calculate all these integrals and express the variations in terms of $\delta M$. As shown in [@PhysRevD.46.1453], the integral at the horizon evaluates to $-T\delta S$, no matter what gauge fields are present Lagrangian. We proceed to evaluate, for a fixed large $r$ hypersurface $$\label{eqn:HPert1}
\int_{\partial\Sigma_{r}}d\sigma_{a}\left(\frac{B^{a}}{16 \pi G}\right)= -\frac{(D-2)}{8 \pi G} \frac{r^{D-1}}{\ell^{D+1}}\delta\ell+\frac{M}{\ell}\delta \ell-\delta M-\sum_{k=1}^{D-2}\uptau_{k}\delta L^{k},$$ $$\label{eqn:HPert2}
2\delta\Lambda\int d^{D-2}x\int_{r_{+}}^{\infty}dr\sqrt{-g}\mathcal{V}(\phi)=\frac{(D-2)}{8 \pi G} \frac{r^{D-1}}{\ell^{D+1}}\delta\ell-\frac{2M}{\ell}\delta\ell.$$ Putting all of this together and expressing the variations of $\ell$ in terms of $\delta\Lambda$, we get the first law for scalar black brane spacetimes $$\label{eqn:FinalFirstLaw}
dM=TdS+\sum_{k=1}^{D-2}\uptau_{k}L^{k}+\frac{M}{2\Lambda}d\Lambda,$$ which is the same as in the vacuum solution (\[eqn:BBMetric\]). Defining the pressure $P$ as in (\[eqn:Pressure\]) and the thermodynamic volume $\Theta=-4\pi G M/\Lambda$ , we obtain $$\label{eqn:FinalFirstLaw2}
dM=TdS+\sum_{k=1}^{D-2}\uptau_{k}L^{k}+\Theta dP.$$
As we see, the only difference between the hairy and non-hairy cases is that the thermodynamic volume $\Theta$ is related in a different way to the natural generalization of geometric volume $$\label{eqn:GeometricVolume}
V=16\pi G \int_{\mathcal{B}}d^{D-2}x\int_{0}^{r_{+}}dr\mathcal{V}(\phi).$$ From equation (\[eqn:ThetaPotential\]) we have, more precisely, that the relation between these two volumes is $$\label{eqn:GeometricThermVolume}
\Theta=\frac{1}{2}\left(1+\frac{\alpha}{D-1}\right)V,$$ where we recall that $\sigma\sim 1/r^{\alpha}$ close to $r=0$, so it characterizes the infrared behaviour of the metric. This geometric volume[^2] is in general difficult to calculate as, in the explicit solutions, the expressions for the scalar potential are quite complicated. It is remarkable that there is such a simple expression for this integral of the potential in terms of the mass, but we must note that such a relation is not that straightforward in the case of spherical black holes. We leave this analysis for future work.
Application to explicit solutions
=================================
Hypergeometric solution
-----------------------
Having derived a first law and a Smarr relation for hairy black branes, an obvious sanity check is to verify that these hold for some explicit solutions. The Einstein field equations are not easy to solve exactly and moreover, it is only possible to do so when the form of the potential allows for an explicit integration of the metric. When this is possible, the constants of integration give us a family of solutions. In our sample metrics, we will also verify the relation (\[eqn:GeometricThermVolume\]) between the thermodynamic and geometric volumes in some particular cases where the integration of the potential is not too complicated.
The first sample solution that we will study was derived in [@Fan:2015ykb], which we review here. The potential is given by[^3] $$\begin{aligned}
16 \pi G V(\phi)= & 2\Lambda(\cosh\phi)^{\frac{\nu k_{0}^2}{D-2}}\left(1-\eta(\sinh\phi)^{\frac{D-1}{\nu}}{}_{2}F_{1}\left(\frac{D-1}{2\nu},\frac{\nu k_{0}^2}{4(D-2)},\frac{D+2\nu+1}{2\nu},\sinh^2\phi\right)\right)\nonumber \\
& \times\left(1-\frac{\nu^2 k_{0}^2 \tanh^2\phi}{2(D-1)(D-2)}\right)+2\Lambda\eta(\cosh\phi)^{\frac{\nu k_{0}^2}{2(D-2)}}(\sinh\phi)^{\frac{D-1}{\nu}}.\label{eqn:PotentialHG}\end{aligned}$$ Here $\eta$, $k_{0}$ and $\nu$ are numerical constants that parametrize this class of solutions. The constant $\nu$ is required to be in the interval $1/2(D-1)<\nu<(D-1)$ so that the mass of the scalar satisfies the BF stability condition and the potential reaches a local maximum at $\phi=0$. With this potential, it is straightforward to verify that, by plugging in $A(r)=0$ and the following functions in our ansatz (\[eqn:MetricAnsatz\]) $$\sigma(r)=\left(1+\frac{q^{2\nu}}{r^{2\nu}}\right)^{-\frac{\nu k_{0}^2}{4(D-2)}}, \quad h(r)=1-\frac{q^{D-1}\eta}{r^{D-1}}{}_{2}F_{1}\left(\frac{D-1}{2\nu},\frac{\nu k_{0}^2}{4(D-2)},\frac{D+2\nu+1}{2\nu},-\frac{q^{2\nu}}{r^{2\nu}}\right) \label{eqn:MetricHG}$$ and with the scalar field $$\label{eqn:ScalarHG}
\phi=\sinh^{-1}\left(\frac{q^{\nu}}{r^{\nu}}\right)$$ then the Einstein field equations are satisfied. By expanding the function $h(r)$ near infinity, the mass and the tensions and the thermodynamic volume can be read off: $$\label{eqn:ChargesHG}
M=(D-2)\frac{\eta q^{D-1}}{16 \pi G \ell^2}v, \quad \uptau_{k}=-\frac{ \eta q^{D-1}}{16 \pi G L_{k}\ell^2}v, \quad \Theta= \frac{\eta q^{D-1}}{2(D-1)}v$$ The temperature $T$ and the entropy $S$ can be found using the usual methods. We find $$\label{eqn:TempEntropHG}
T=\frac{(D-1)\eta q^{D-1}}{4\pi r_{+}^{D-2}\ell^2}, \quad S=\frac{r_{+}^{D-2}v}{4},$$ so indeed, we verify that $(D-1)M=(D-2)TS$ and, with $P=-\Lambda/(8\pi G)$, we also obtain the extended Smarr relation (\[eqn:FinalSmarr2\]). To derive the first law, we observe that $h$ is a function of $q/r$ so, given that $h(r_{+})=0$, then $q/r_{+}$ must be a constant and the variations are related via $$\label{eqn:SmartIdeaHG}
\delta r_{+}= \frac{r_{+}}{q}\delta q.$$ Using this fact, the extended first law (\[eqn:FinalFirstLaw2\]) can be readily verified.
Now we would like to find the relation between the geometric and thermodynamic volumes. As we can see, the volume integral of the potential is not trivial to carry out. As an example, we use $D=5$, $\nu=2$ and $k_{0}=1$, given that the integral simplifies significantly in this case[^4]. Expressing the potential in terms of the coordinate $r$, we have $$\label{eqn:PotentialHG2}
16\pi G V(\phi(r))=-\frac{\Lambda \left(36 \eta r^{2/3} \left(q^4+r^4\right)^{5/6}-(6 \eta +5) \left(5 q^4+6 r^4\right)\right)}{30 r^{4/3} \left(q^4+r^4\right)^{2/3}}.$$ From this expression, we can calculate the indefinite volume integral of the potential $$\label{eqn:IntegralPotentialHG}
16\pi G \int d^{3}x\int dr \sqrt{-g}V(\phi(r))=-\frac{2}{5\ell^2}\left(\frac{3}{2} (6 \eta +5) r^{10/3} \sqrt[6]{q^4+r^4}-9 \eta r^4\right)v.$$ We would like to integrate this from $r=0$ to $r=r_{+}$. The antiderivative evaluates to zero at $r=0$; to evaluate at $r=r_{+}$, we can find the value of $q^4$ from the equation $h(r_{+})=0$. The hypergeometric function simplifies in this case and we get $$\label{eqn:q4HG}
q^4=\frac{\left(5\times 6^{4/5} \sqrt[5]{\frac{5}{\eta }+6}+\left( 6^{9/5} \sqrt[5]{\frac{5}{\eta }+6}-36\right) \eta \right) {r_{+}}^4}{36 \eta }.$$ Replacing this in the indefinite integral (\[eqn:IntegralPotentialHG\]), and defining $\Upsilon$ as in (\[eqn:Upsilon\]) we obtain $$\label{eqn:UpsilonHG}
\Upsilon=-\frac{18}{5}\frac{q^4\eta}{16\pi G \ell^2}v=-\frac{6}{5}M.$$ The geometric volume defined in (\[eqn:GeometricVolume\]) can also be calculated using this integral. The result is $$\label{eqn:GeomThermHG}
V=\frac{3 q^4\eta}{10}v=\frac{12}{5}\Theta.$$
With the knowledge bestowed to us by equation (\[eqn:GeometricThermVolume\]), we could have arrived to this result without so much trouble. All we need to do is expand the function $\sigma(r)$ around zero. We observe that for small $r$, $\sigma(r)\sim r^{2/3}$. We then identify $\alpha=-2/3$ so, replacing this value in formula (\[eqn:GeometricThermVolume\]) we get $$\label{eqn:GeomThermHG2}
\Theta=5/12 V,$$ as required. In general, it is hard to integrate the hypergeometric function, as it will not always reduce to an algebraic expression like above. However, with our result (\[eqn:GeometricThermVolume\]) we can assert that for general values of the metric parameters $$\label{eqn:GeomThermHG3}
\Theta=\frac{1}{2}\left(1-\frac{\nu^2 k_{0}^2}{2(D-1)(D-2)}\right)V.$$
Double branch solution
----------------------
The second example, which we call the double branch solution, is a somewhat better known solution. It was first proposed in the 5 dimensional case in [@Acena:2012mr] and, more recently, it was generalized to an arbitrary number of spacetime dimensions in [@Acena:2013jya]. The chemistry of spherical black hole solutions analogous to this one has also been recently developed in [@Astefanesei:2019ehu] and [@Rojas:2019hze]. This solution is interesting because it allows to calculate explicitly the beta functions of the relevant deformation couplings in the holographic RG flow, as opposed to the previous example which is dual to a vev-driven flow.
However, for our purposes, this solution is not as convenient as the hypergeometric one. This is due to the fact this solution is not expressed in Poincaré coordinates. We can directly verify the Smarr relation and the first law, whence coordinates do not matter, and then integrate the potential to find that there is indeed a proportionality relation between the thermodynamic and geometric volumes. Under the assumptions we used to derive equation (3), we can locally change coordinates and find the asymptotic behaviour of the metric close to $r=0$. We can then verify that the proportionality constant is given by our prescription, but additional work is needed because we need to recover the Poincaré coordinates in the infrared. We must also note that this solution has two branches in which the radial coordinate is defined over different intervals. The branch chosen depends on the sign of the scalar field, which does not change in this solution.
For simplicity, we will work in the 5 dimensional case. The Smarr relation and first law in the D-dimensional case can be derived in a similar fashion, but integrating the potential becomes easier in 5 dimensions. A dimensionful parameter $\eta$ is introduced to obtain a dimensionless radial coordinate $x$, so the metric takes the form $$\label{eqn:ConformalAnsatz}
ds^2=\Omega(x)\left(-f(x)dt^2+\frac{\eta^2 dx^2}{f(x)}+\delta_{ij}dx^{i}dx^{j}\right).$$ Then, with the potential $$\begin{aligned}
16 \pi G V(\phi)&=2\Lambda\left(\frac{(9\nu^2-5)e^{-\phi l_{\nu}}}{8\nu^2}\right)\left(1-\frac{8\mu}{(\nu^2-25)(9\nu^2-25)}\right)\nonumber \\
&\times \left(\frac{(\nu -1) e^ {-\nu \phi l_{\nu} }}{2 (3 \nu +5)}+\frac{(\nu +1) e^{\nu \phi l_{\nu} }}{2 (3 \nu -5)}+\frac{5 \left(\nu ^2-1\right)}{9 \nu ^2-25}\right)\nonumber\\
&+2\Lambda \mu \frac{e^{3 l_{\nu}\phi/2}}{4 \nu ^3} \left[\frac{5 \left(\nu ^2-1\right)}{\nu ^2-25} \left(\frac{e^{-l_{\nu}\nu\phi/2}}{3 \nu -5}+\frac{e^{l_{\nu}\nu\phi/2}}{3 \nu +5}\right)\right.\nonumber\\
&\left. +\frac{1}{3} \left(\frac{(\nu +1) e^{-3l_{\nu}\nu\phi/2}}{(\nu -5) (3 \nu -5)}+\frac{(\nu -1)e^{3\nu l_{\nu}\phi/2}}{(\nu +5) (3 \nu +5)}\right)\right] \label{eqn:PotentialDB}\end{aligned}$$ the Einstein field equations are solved by the metric functions $$\label{eqn:ConformalFactorDB}
\Omega(x)=\frac{\nu^2 x^{\nu-1}}{\eta^2(x^\nu-1)^2},$$ $$\label{eqn:MetricDB}
f(x)=-\frac{\Lambda}{6}+\Lambda\mu\left(\frac{4}{3(\nu^2-25)(9\nu^2-25)}+\frac{x^{5/2}}{12\nu^3}\left(-\frac{x^{-\frac{\nu }{2}}}{\nu -5}-\frac{x^{\nu /2}}{\nu +5}+\frac{x^{-\frac{3 \nu }{2}}}{3 (3 \nu -5)}+\frac{x^{\frac{3 \nu }{2}}}{3 (3 \nu +5)}\right)\right),$$ and the scalar field $$\label{eqn:ScalarDB}
\phi=l_{\nu}^{-1}\log(x); \qquad l_{\nu}^{-1}=\sqrt{\frac{3(\nu^2-1)}{2}}.$$ The numerical constant $\nu\geq 1$ parametrizes different solutions; the value $\nu=1$ corresponds to the vacuum solution. Only in this case, the change to Poincaré coordinates can be done exactly. The Hamiltonian method can be used to derive the mass and the tensions, and from them this we can identify the thermodynamic volume. We obtain $$\label{eqn:ChargesDB}
M=\left(\frac{1}{16 \pi G}\right)\frac{\mu}{16 \eta^4 \ell^2}v, \qquad \uptau_{k}=-\left(\frac{1}{16 \pi G}\right)\frac{\mu}{48 \eta^4 L^{k} \ell^2}v, \qquad \Theta=\frac{1}{4}\frac{\mu}{96\eta^4}v.$$ Defining $x_{+}$ to be the largest root of $f(x_{+})=0$, the entropy and temperature are found to be $$\label{eqn:TempEntropDB}
T=\frac{\mu\left\lvert x_{+}^\nu -1\right\rvert^3}{48\pi\eta\nu^3 x_{+}^{\frac{3}{2}(\nu-1)}\ell^2}, \qquad S=\frac{\nu^3 x_{+}^{\frac{3}{2}(\nu-1)}}{4 G \eta^3\left\lvert x_{+}^\nu -1\right\rvert^3}v.$$ Both the equality $4M=3TS$ and the full Smarr relation with volume and tension terms can directly be shown to hold with these quantities. Moreover, the extended first law is also satisfied. This must be done with care, because the parameters that are to be varied independently to check the first law are the compactifying lengths, $\Lambda$ and the parameter $\eta$, which will change the scale and hence the position of the horizon when $x_{+}$ is fixed.
In order to find the integral of the potential between the Poincaré and black brane horizons, we note that the radial coordinate ranges between $0$ and $1$, or $1$ and infinity depending on whether the scalar is negative or positive. In both cases, the conformal infinity is located at $x=1$, and the Poincaré horizon is at zero or at infinity respectively. We will work with the second case, which corresponds to stable black branes. The integral is complicated but can be computed exactly for any value of $\nu$. We obtain, following steps similar to the hypergeometric case
$$\label{eqn:IntegralPotentialDB}
16\pi G\int d^3 x \int_{\infty}^{x_{+}}dr\sqrt{-g}V(\phi(x))= -\frac{\nu+1}{3\nu+5}\frac{\mu}{4 \eta^4 \ell^2}v,$$
so the relation between the geometric and thermodynamic volumes is $$\label{eqn:GeomThermDB}
\Theta=\frac{1}{8}\left(\frac{3\nu+5}{\nu+1}\right)V.$$
As an alternative to integrating the lengthy expression for the potential, we may simply use formula (\[eqn:UpsilonM\]). In order to do this, we need find the metric in Poincaré coordinates at least locally close to $x=\infty$, which corresponds to $r=0$. Here, we assume $A(r)$ is finite, so we must have that $\Omega(x(r))\sim r^2$. Under this assumption and from the expression (\[eqn:ConformalFactorDB\]) for $\Omega(x)$, when $x$ goes to infinity we must have $$\label{eqn:PoincareDB}
x\sim r^{-2/(1+\nu)}\quad \textrm{and} \quad dx^2\sim r^{-2(3+\nu)/(1+\nu)}dr^2.$$ Using these, and the form of the ansatz (\[eqn:MetricAnsatz\]), we must have that $$\label{eqn:PoincareDB2}
\sigma\sim r^{-\alpha}, \qquad \alpha=\frac{1-\nu}{1+\nu}.$$ From the value of $\alpha$ and using (\[eqn:GeometricThermVolume\]), we again obtain relation (\[eqn:GeomThermDB\]).
Discussion
==========
In this work we have obtained a very general Smarr relation and First Law of thermodynamics for the gravitational energy of hairy black branes using purely geometric methods. The result for the Smarr relation matches the one obtained from scaling arguments and allows us to define a thermodynamic volume, which can be easily computed from the mass. This volume is the conjugate variable to the pressure in the extended first law, and can be calculated simply from the mass and the cosmological constant. For non-hairy black branes, the thermodynamic and geometric volume are proportional to each other, and this also happens for the general case. However, the constant of proportionality is dependent on the infrared behaviour of the solution, namely $$\Theta=\frac{1}{2}\left(1+\frac{\alpha}{D-1}\right)V,$$ where $\alpha$ characterizes the decay of the $g_{rr}$ component as $r\rightarrow 0$. With this result, we have provided evidence of the conjecture that the integral of the potential behind the horizon is proportional to the thermodynamic volume, if only for this particular horizon topology. To provide a full proof, we must also deal with spherical and hyperbolic horizons. The evidence provided in [@Cvetic:2010jb] suggests that a similar result holds. Proving this more general result is left for future work.
Let us observe that the integral of the potential turns out to be proportional to the mass, where the constant of proportionality is given by the behaviour of the metric close to the Poincaré horizon. Thus, by just knowing the potential inside the horizon, the asymptotic value of the potential and the metric at the infrared, we can know the mass. This means that this information is enough to fix the asymptotics of the blackening factor with a simple prescription. Hopefully this insight will help to find more explicit hairy black brane solutions when prescribing some potential.
A possible extension of this work is to allow for other asymptotic behaviour of the solution. The assumption that the spacetime is asymptotically AdS was relaxed in [@Kastor:2018cqc], where a similar Smarr relation and first law hold. The volume gains some terms that are proportional to the trace of the holographic stress-energy tensor. The example therein holds only for some particular form of the potential, but it is plausible that this result is much more general. With the methods that we have derived, it is likely that such a first law can be generalized. Another possible extension of this work is to include variations of the asymptotic behaviour of scalar field, thus obtaining a full thermodynamics of this spacetime. We would expect that the volume term is modified, as the total energy would include contributions from the scalar field. Deriving such a law could hopefully help us understand more about the holographic interpretation of the volume term, which is still subject to much debate [@Johnson:2014yja; @Dolan:2014cja; @Kastor:2014dra; @Karch:2015rpa; @Couch:2016exn].
Acknowledgements {#acknowledgements .unnumbered}
================
The author would like to thank David Kubizň' ak for the helpful discussions and comments and Robert Myers for inspiring the study of black branes with scalar hair. The work was supported by the Perimeter Institute for Theoretical Physics and by the Natural Sciences and Engineering Research Council of Canada. Research at Perimeter Institute is supported in part by the Government of Canada through the Department of Innovation, Science and Economic Development Canada and by the Province of Ontario through the Ministry of Economic Development, Job Creation and Trade.
[^1]: The dual to a scalar black brane spacetime is a QFT that comes from a deformed CFT. This traceless property refers to the stress energy tensor of the non-deformed theory.
[^2]: The hairless case can be considered to be a solution with a constant scalar field and a potential equal to $2\Lambda/16\pi G$. This is the reason why the prefactor of $16\pi G$ is needed in this definition. Let us observe that when $\alpha=0$, the thermodynamic volume is half of the geometric volume, as expected.
[^3]: This potential and the following one are slightly adapted from the original source so that they are of the form $V(\phi)=2\Lambda \mathcal{V}(\phi)$.
[^4]: i.e. Mathematica can do it.
| {
"pile_set_name": "ArXiv"
} |
-0.6cm PAR–LPTHE 03–11
hep-th/0303165
1.4truecm
Laurent Baulieu, Marc Bellon, Alessandro Tanzini [*Laboratoire de Physique Théorique et Hautes Energies, Universités Pierre et Marie Curie, Paris VI et Denis Diderot, Paris VII*]{}
[**Abstract**]{}: Topological euclidean gravity is built in eight dimensions for manifolds with $Spin(7) \subset SO(8)$ holonomy. In a previous work, we considered the construction of an eight-dimensional topological theory describing the graviton and one graviphoton. Here we solve the question of determining a topological model for the combined system of a metric and a Kalb–Ramond two-form gauge field. We then recover the complete $N=1, D=8$ supergravity theory in a twisted form. We observe that the generalized self-duality conditions of our model correspond to the octonionic string equations.
------------------------------------------------------------------------
Postal address: Laboratoire de Physique Théorique et des Hautes Energies, Unité Mixte de Recherche CNRS 7589, Université Pierre et Marie Curie, boîte postale 126. 4, place Jussieu, F–75252 PARIS Cedex 05
ø
Introduction
============
All types of superstring theories can be formally obtained by suitable anomaly free untwisting of a topological sigma-model [@green]. This suggests the possibility that supergravities, which arise as low energy limits of superstrings, can be understood as topological gravities. In particular, $D=11$ supergravity, which determines all known supergravities in lower dimensions, could be viewed as a topological theory.
In a previous work we have shown that, both in four [@BT1] and in eight dimensions [@BT2], the Einstein action plus the Rarita–Schwinger term (in a twisted form) can be obtained by constructing a topological quantum field theory (TQFT), which implements in a BRST invariant way the gravitational instanton equation.[^1] These constructions only holds for manifolds with special holonomy, [*i.e.*]{}, $SU(2)\subset SO(4)$ in four dimensions and $Spin(7)\subset
SO(8)$ in eight dimensions.
In [@BT2], we left open the delicate point of introducing a sector of the eight-dimensional TQFT involving a two-form gauge field. This case escapes the procedure displayed in [@laroche]. In this paper, the equations for the two-form gauge field appear mixed with the one for other fields. The basic idea is the construction of a TQFT for a general two-tensor field, which is the natural object stemming from the zero slope limit of topological sigma-models. The symmetric part of this tensor describes the metric, while its antisymmetric part gives the two-form field. We use the formalism where the metric is described as a vielbein modulo the Lorentz symmetry. This means adding and subtracting spurious degrees of freedom, a task which is by now familiar in the context of TQFT’s. As in our previous model [@BT2], we find that the topological theory is defined on manifolds with $Spin(7)$ holonomy. In fact, the $Spin(7)$-invariant four-form plays a central rôle in the determination of the topological gauge functions. Moreover, the presence on these manifolds of a covariantly constant spinor allows for the definition of a twist [@BT2] which maps some of the fermionic ghosts and antighosts of the topological model on the spinors of $N=1, D=8$ supergravity. Finally, some ghosts of ghost and ghosts of antighosts can be untwisted into the commuting ghosts of local supersymmetry, explaining the emergence of local supersymmetry in our model.
=ø=Ø=Ł[ [L]{}]{}
Including the two-form in topological gravity
==============================================
In this section we address the question of building a TQFT multiplet for a general tensor $A_{\mu\nu}$ of rank two. We consider an eight-dimensional manifold with holonomy group $Spin(7)\subset SO(8)$. The tensor $A_{\mu\nu}$ can be split into its symmetric and antisymmetric parts. If, as in [@BT2], we only consider the symmetric part, which can be interpreted as a metric, we can construct a TQFT that contains the Einstein action, by using a gravitational octonionic self-duality equation. In the spirit of [@BT1] we can also introduce a coupling to the TQFT for an abelian graviphoton, which has ghost number two. This topological model determines, in a twisted form, a truncation of $N=1,D=8$ supergravity [@BT2]. However, the dilaton, the Kalb–Ramond two-form, one graviphoton and their fermionic superpartners in the $N=1,D=8$ supergravity multiplet [@sase] escape this construction.
The difficulty of determining a TQFT for the antisymmetric part $B_{\mu\nu}$ of $A_{\mu\nu}$ can be appreciated as follows, following the ideas contained in [@laroche]. The two-form $B_{\mu\nu}$ contains 28 components, which give 21 degrees of freedom modulo the gauge invariance $B_{\mu\nu}\sim B_{\mu\nu}+\partial_{[\mu}\Lambda_{\nu]}$. The field strength of $B_{\mu\nu}$ is a three-form $G_3=dB_2$ containing $56=(^8_3)$ components. The analysis in [@laroche] indicates that there is no natural way to chose an holonomy group for the eight-dimensional manifold which allows to write a self-duality equation for $G_3$, which, (i) would count for 21 topological independent equations and, (ii), would solve the relativistic wave equation of a two-form. Moreover, the 28 components of the topological ghost of a two-form cannot be rearranged in eight-dimensional spinor representations. Rather, to determine a TQFT involving the two-form in eight dimensions, we will see that it is necessary to combine the Lorentz invariance and the topological invariance of the two-form, and write self-duality equations that mix the dilaton, the eight-bein and the two-form. That this is not only possible, but a useful complement of the construction of [@BT2] appears immediately from the fact that there were 56 components for the antighost for the vielbein, more than the 28 degrees of freedom it describes.
To start with, we enlarge the question of building a TQFT for $A_{\mu\nu}$ into that of building a TQFT for a vielbein $e^a_\mu$, a Lorentz ghost $\Omega^{ab} $ and a two-form $B_{\mu\nu}$,[^2] $$A_{\mu\nu}= A_{\{\mu\nu\}} \oplus A_{[\mu\nu]} \to (e^a_\mu,
\Omega^{ab},B_{\mu\nu}).$$ The Lorentz symmetry can be used to set to zero the antisymmetric part of the matrix $e^a_\mu$[^3]. The number of degrees of freedom of the system $A_{\mu\nu}=A_{[\mu\nu]} \oplus A_{\{\mu\nu\}}$ is indeed equal to that of the $(e^a_\mu, B_{\mu\nu} ,\Omega^{ab})$ system, when the components are algebraically counted, since the Lorentz ghost $\Omega^{ab}$ counts negatively. Eventually, we will interpret $A_{\{\mu\nu\}}$ as a metric $g_{\mu\nu}$, assuming that $e^a_\mu$ is an invertible matrix.
In contrast with our previous work, graviphotons are not separately introduced. Here, these fields naturally appear as ghosts for the topological ghosts of the two-form $B_{\mu\nu}$. An analogous situation holds for the dilaton, which is a $Spin(7)$-invariant part of a ghost of ghost for the vielbein. Having obtained a truncation of supergravity in [@BT2] is now understood as having consistently retained a part of the topological BRST multiplet.
Eventually, we will recognize that we have a TQFT with an equivariance with respect to the Lorentz$\times$diffeomorphism symmetry, whose gauge fields are the spin connection $\o^{ab}_\mu$ and the vielbein $e^a_\mu$. Moreover, our topological model displays an equivariance with respect to the vector gauge symmetry of the two-form, $B_2\sim B_2
+ d\Lambda_1$. Finally, local supersymmetry will show up as a consequence of the symmetry of the topological ghost of the vielbein, defined modulo reparametrizations. Shortly speaking, the construction of a TQFT for a two-tensor yields all the fields of supergravity. As we will discuss in the next section, the topological gauge functions are given by self-duality equations which mix the symmetric and antisymmetric parts of the two-tensor.
Let us now proceed to the detailed construction of the BRST topological multiplets. Geometry determines the set of ghosts and our first guess for the complete set of fields for an $SO(8)$ invariant TQFT with a vielbein $e^a_\mu$, a two-form $B_{\mu\nu}$ and a spin connection $ \o^{ab}_\mu$ is: =$$\begin{aligned}
\matrix
{ & & e^a_\mu & & \cr
&\Psi^{(1) a}_\mu & & \bar\Psi^{(-1)a }_\mu & \cr
\Phi^{(2)a} & & \Phi^{(0)a} , b^{(0)a }_\mu & & \bar\Phi^{(-2)a}\cr
& \eta^{(1 )a} & & \bar \eta^{(-1 )a} & \cr }
\nonu\end{aligned}$$
$$\begin{aligned}
\matrix
{ & & \o^{ab}_\mu & & \cr
& \tilde\Psi^{(1)ab}_\mu & & \bar{\tilde\Psi}^{(-1)ab}_\mu & \cr
{\tilde \Phi}^{(2)ab } & & {\tilde \Phi}^{(0)ab},{\tilde b}^{(0)ab}_\mu
& & \bar{\tilde\Phi}^{(-2)ab} \cr
&\hw {\tilde\eta}^{(1)ab } \hw& & \hw\bar {\tilde\eta}^{{(-1)ab} } \hw& \cr }
\nonu\end{aligned}$$
$$\begin{aligned}
\matrix
{ & & & B_{\mu \nu} & & & \cr
& & \Psi^{(1)}_{\mu \nu} & & \bar \Psi^{(-1)}_{\mu\nu} & \cr
& A^{(2)}_{\mu } & & A^{(0)}_{\mu } , b^{(0)}_{\mu \nu} &&
A^{(-2)}_{\mu } & \cr
R^{(3)} & & S^{(1)},\Psi^{(1)}_\mu && \bar S^{(-1)},\bar \Psi^{(-1)}_\mu &
& \bar R^{(-3)} \cr
& \hw b_ {S^{(1)}} ^{2},\Phi^{(2)} \hw&&\hw b_ {\bar S^{(-1)}} ^{(0)},\bar
\Phi^{(0)},\Phi^{(0)} \hw& &\hw b_{\bar R^{(-3)}}^{(-2)},\bar \Phi^{(-2)}
\hw& \cr
&& \eta^{(1)} && \bar \eta^{(-1)} & \cr}
\nonu\end{aligned}$$
$$\begin{aligned}
\matrix
{ \xi^{(1) \mu} & & \bar \xi^{(-1) \mu} \cr
& b^{(0) \mu} }
\quad\quad\quad\quad\quad
\matrix
{ \O^ {(1)ab} & & \bar \O^ {(-1)ab} \cr
& b^{(0)ab} & }\end{aligned}$$
For the sake of clarity, we have made explicit (as an upper index) the ghost number of the fields in the “pyramid" that describes the BRST topological multiplets. We could introduce a bigrading that separate the ghost number and antighost number, but this would make heavier the notations. In the above pyramids, the BRST symmetry acts on the south-west direction. The fields which are not on the left edge of each pyramid are topological pairs made of antighosts and their Lagrange multipliers. They satisfy trivial BRST equations[^4]. Actually, each one of the fields that are labeled by a letter $b$ or $\eta$, with various indices, is a bosonic or fermionic Lagrange multiplier field, and is essentially equal to the BRST variation of the antighost that is located at its upper right position, e.g., $s
\Psi^{(-1)}_{\mu\nu}=b^{(0)}_{\mu\nu}+\ldots$, $s\bar S^{(-1)}= b_ {\bar S^{(-1)}} ^{(0)}+\ldots$. As an exception to this notational rule, we find useful to define the fields $\Psi^{(1)}_\mu$ and $\bar \Psi^{(-1)}_\mu$ as the fermionic Lagrange multipliers that stem from the BRST variation of the commuting ghosts of ghosts $A^{(0)}_\mu$ and $A^{(-2)}_\mu$ respectively, [*i.e.*]{} $sA^{(0)}_\mu= \Psi^{(1)}_\mu+\ldots$ and $sA^{(-2)}_\mu= \bar\Psi^{(-1)}_\mu+\ldots$.
The fields which carry the essential geometrical information are on the left edge of the pyramids. Their topological symmetry is defined as: $$\begin{aligned}
se^a_\mu &=& \Psi^{(1)a}_\mu-\O^{ab} e^b_\mu + \L_\xi e^a_\mu ,\nonumber
\\ s\o^{ab}_\mu &=& \t \Psi^{(1)ab}_\mu+
D_\mu\O^{ab} + \L_\xi \o^{ab}_\mu,
\nonumber\\
\nonumber\\
s\Psi^{(1)a}_\mu &=& -\O^{ab} \Psi^{(1)b}_\mu-\L_\Phi
e^a_\mu +\t \Phi^{(2)ab}e^b_\mu +\L_\xi \Psi^a_\mu,
\nonumber\\
s \t \Psi ^{(2)ab}_\mu &=& -\O ^{ac}\t
\Psi^{(2)cb} + D_\mu \t \Phi^{(2)ab} -\L _\Phi \o^{ab}_\mu
+ \L_\xi \t\Psi ^{(2)ab}_\mu,
\nonumber\\
\nonumber\\
s \Phi ^{(2)a} &=& \L_\xi \Phi ^{(2)a } -\O ^{ac} \Phi^{(2)a} ,
\nonumber\\
s \t \Phi ^{(2)ab} &=& -\O ^{ac}\t \Phi^{(2)cb}
+\L_\xi \t\Phi ^{(2)ab},
\nonumber\\
\nonumber\\
sB_{\mu\nu} &=&
\Psi^{(1)}_{\mu\nu}
+\L_\xi B_{\mu\nu},
\nonumber \\
s\Psi^{(1)}_{\mu\nu} &=&\L_\Phi B_{\mu\nu}
+\partial_{[\mu} A^{(2)}_{ \nu]} + \L_\xi \Psi^{(1)}_{\mu\nu},
\nonumber \\
sA^{(2)}_{\mu } &=& \partial_{\mu} R^{(3)} + \L_\xi A^{(2)}_{\mu } ,
\nonumber \\
s R^{(3)} &=& \L_\xi R^{(3)} ,
\nonumber\\
\nonumber\\
s\xi^\mu &=& f_a^\mu \Phi^{(2)a}+\xi^\nu \partial _\nu \xi^\mu,
\nonumber\\
s \O ^{ab} &=& \t \Phi^{(2)ab} -\O ^{ac}\O^{cb} + \L_\xi \O ^{ab} .
\label{brs} \end{aligned}$$ In the variation of $\xi^\mu$, the inverse of the vielbein $e_\mu^a$ appears and we denote it as $f_a^\mu $. We have not introduced ghosts $V_\mu^{(1) }$ and ghost of ghost $
m^{(2)}$ for the gauge invariance of the two-form gauge field, with the standard BRST symmetry $QB_{\mu\nu}=\partial _{[\mu} V^{(1) }_{\nu]}$, $QV^{(1) }_{\mu}=\partial _{ \mu} m^{(2) }$, $Qm^{(2) }=0$, which would yield $s V^{(1)}_\mu=A^{(2)}_\mu$ and $sm^{(2)}= R^{(3)}$. We choose instead to write equivariant BRST transformations with respect to this symmetry. This implies that the square of the BRST transformations on the field $B_{\mu\nu}$ is not zero, but corresponds to a reparametrization with parameter $\Phi^a$ and an abelian transformation $B_{\mu\nu}\to B_{\mu\nu}+\partial
_{[\mu} \Lambda_{\nu]}$, with $\Lambda_\mu=A^{(2) }_\mu$. The BRST operator $s$ is thus nilpotent only modulo gauge transformations for the two-form gauge field. Actually, the topological gauge functions we will use in sect.3 only involve the curvature $G_3=dB_2$, so there is no need to give the details of the gauge symmetry of the two-form. We could further set $\xi^\mu=0$, which would yield a BRST symmetry equivariant with respect to the reparametrization. In this case, the BRST symmetry would be nilpotent also modulo reparametrizations along the vector ghost of ghosts $ e_a^\mu \Phi^{(2)a}$. As for the Lorentz invariance, we will instead carefully keep the Lorentz ghost dependence. Eventually, the corresponding gauge functions will be equivariant with respect to local Lorentz transformations for $Spin(7)\subset SO(8)$.
There is an $U(1)$ invariance for the fields $\Psi^{(1)}_\mu$ and $\bar \Psi^{(-1)}_\mu$. Also for this invariance we prefer to work with an equivariant BRST operator, and we do not write explicitly the related Faddeev–Popov ghosts $sc =\Phi^{(2)}$, $s\bar c =\bar \Phi^{(0)}$ in the BRST transformations $$\begin{aligned}
s A^{(-2)}_\mu &=& \bar \Psi^{(-1)}_\mu +\L_\xi A^{(-2)}_\mu
\ \ \ \ \ \ s A^{(0)}_\mu = \Psi^{(1)}_\mu +\L_\xi A^{(0)}_\mu
\nonumber \\
s\bar \Psi^{(-1)}_\mu &=& \partial _\mu \Phi^{(0)}
+ \L_\xi \bar \Psi^{-1}_\mu
\ \ \ \ \ \ \
s \Psi^{(1)}_\mu = \partial _\mu \Phi^{(2)} +\L_\xi \Psi^{(1)}_\mu
\nonumber \\
s \Phi^{(0)} &=& \L_\xi \Phi^{(0)}
\ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \
\ \ \ s \Phi^{(2)} = \L_\xi \Phi^{(2)}
\nonumber \\
s \bar \Phi^{(0)} &=& \L_\xi \bar \Phi^{(0)} +\eta^{(1)}
\ \ \ \ \ \ \ \ \ \
s\bar \Phi^{(-2)}= \bar \eta^{{(-1)}}+ \L_\xi \bar \Phi^ {{(-2)}}
\nonumber \\
s \eta^{(1)} &=& \L_\xi \eta^{(1)}
\ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \
s \bar \eta^{{(-1)}}= \L_\xi \bar \eta^{{(-1)}} \ .\end{aligned}$$ If we only retain a $Spin(7)\subset SO(8)$ invariance, we can redistribute the degrees of freedom of the antighosts and Lagrange multipliers of the BRST multiplets. In this way, we shall be able to define $Spin(7)\subset SO(8)$ invariant topological gauge functions, which is the key for building the eight-dimensional TQFT. Using the decomposition of a $SO(8)$ valued two-form $M^{ab }=M^{ab^-}+ M^{ab^+} $ in the $Spin(7)$-invariant representations $28=7\oplus21$, we get $$\begin{aligned}
\matrix
{ & & e^a_\mu & & \cr
& \Psi^{(1)a}_\mu & & \bar\Psi^{(-1)ab^-}_\mu,\bar\chi^{(-1)a} & \cr
\Phi^{(2)a} & & \sigma, \Phi^{(0)ab^-}\; b^{(0)ab^-}_\mu, b^{(0) }_\mu
& &\bar\Phi^{(-2)a} \cr
&\chi^{( 1 )}, \eta^{(1 )ab^- } & & \bar \eta^{(-1)a } & \cr
}
\nonu\end{aligned}$$
$$\begin{aligned}
\matrix
{ & & \o^{ab}_\mu & & \cr
& \tilde\Psi^{(1)ab}_\mu & & \bar{\tilde\Psi}^{(-1)ab}_\mu & \cr
{ \tilde\Phi}^{(2)ab ^\pm } & & { \tilde\Phi}^{(0)ab ^\pm },
{\tilde b}^{(0)ab}_\mu & & {\bar{\tilde\Phi}}^{(-2)ab ^\pm } \cr
& { \tilde\eta}^{(1)ab ^\pm } & & {\bar{\tilde\eta}}^{(-1)ab ^\pm } & \cr
}
\nonu\end{aligned}$$
$$\begin{aligned}
\matrix
{ & & & B_{\mu \nu} & & & \cr
& & \Psi^{(1)}_{\mu \nu} & & \bar \Psi^{(-1) }_{\mu\nu^\pm} && \cr
& A^{(2)}_{\mu } && A^{(0)}, A^{(0)}_{\mu \nu^-}, b^{(0) }_{\mu \nu^\pm}
& & A^{(-2)}_{\mu } & \cr
R^{(3)} & & \hw S^{(1)},\Psi^{(1)}_{\mu\nu^-}, \Psi^{(1)}\hw & &
\hw \bar S^{(-1)},\bar \Psi^{(-1)}_{\mu} \hw & & \bar R^{(-3)} \cr
& b_ {S^{(1)}} ^{(2)}, \Phi^{(2)} & & b_ {\bar S^{(-1)}} ^{(0)},
\Phi^{(0)} ,\bar\Phi ^{(0)} & & b_{\bar R^{(-3)}}^{(-2)},\bar \Phi^{(-2)}
& \cr
& & \eta^{(1)} & & \eta^{(-1)}& & \cr
}
\nonu\end{aligned}$$
$$\begin{aligned}
\matrix
{ \xi^ \mu & & \bar \xi^ \mu \cr
& b^\mu }
\quad\quad\quad\quad\quad
\matrix
{
\O^ {ab ^\pm} & & \bar \O^ {ab ^ \pm} \cr
& b^{ab ^\pm} &
}\end{aligned}$$
Let us now clarify the rôle of the various topological fields and their supergravity interpretation. To obtain the relevant set of propagating fields, a certain number of trivial gauge fixing must be done, which give algebraic terms of the type $s(AB)=(sA)B\pm A(sB)=\Psi_A A\pm A\Psi_B$. Such terms eliminate quartets of the form $A,\Psi_A, B,\Psi _B$ in a BRST invariant way by their algebraic equations of motions. After some thinking, one understands that the following $s$-exact terms are needed $$\begin{aligned}
&&s\Big[{\bar{\tilde \Phi}} ^{(-2 )ab^-}
( \t \eta ^{(1 )ab^-}\,\, -\O ^{ ab^-}) \Big]=
{\bar{\tilde \eta}} ^{(-1 )ab^-}
( {\t \eta} ^{(1 )ab^-}\,\, - \O ^{ ab^-})
+ {\bar{\tilde \Phi}} ^{(-2 )ab^-} {\tilde \Phi}^{(2 )ab^-}
\cr
&&s\Big[{\bar{ \Psi}} ^{(-1 ) }_{\mu\nu ^+ }
{ {\tilde \Phi}} ^{(0 )ab^+} e^\mu_ae^\nu_b
\Big]
={\bar{ \Psi}} ^{(-1 ) }_{\mu\nu ^+ }
{ {\tilde \eta}} ^{(1 )ab^+} e^\mu_ae^\nu_b
+{ {b}} ^{(0 ) }_{\mu\nu ^+ }
{ {\tilde \Phi}} ^{(0 )ab^+} e^\mu_ae^\nu_b
\cr
&&s\Big[{\bar{ \Psi}} ^{(-1 ) }_{\mu\nu ^- } { {A}} ^{(0 ) }_{\mu\nu^-} \Big]
= {\bar{ \Psi}} ^{(-1 ) }_{\mu\nu ^- } { {\Psi}} ^{(1 ) }_{\mu\nu^-}
+ {\bar{ b}} ^{(0 ) }_{\mu\nu ^- } { {A}} ^{(0 ) }_{\mu\nu^-}
\nonumber \\
&&s \Big[{\bar S}^{(-1)}A^{(0)} \Big] = \bar S^{(-1)}\Psi^{(1)}
+b _{\bar S^{(-1)}}^{(0)} A^{(0)}\end{aligned}$$ The fields $\Phi^{(2)}, \bar \Phi^{(-2)},\bar \eta^{(-1)}$ are also eliminated, together with the quartets for $A^{(0)}_{\mu\nu}$. The remaining fields are then: $$\begin{aligned}
\matrix
{ & & e^a_\mu & & \cr
& \Psi^{(1)a}_\mu & &\hw \bar\Psi^{(-1)ab^-}_\mu,
\bar\chi^{(-1)a} \hw& \cr
\Phi^{(2)a} & & \sigma, \Phi^{(0)ab^-}\; b^{(0)ab^-}_\mu,b^{(0) }_\mu
& & \bar\Phi^{(-2)a} \cr
&\hw \chi^{( 1) }, \eta^{(1) ab^- } \hw&& \bar \eta^{(-1)a } & \cr
}
\nonu\end{aligned}$$
$$\begin{aligned}
\matrix
{ & & \o^{ab}_\mu & & \cr
& \tilde\Psi^{1ab}_\mu & & \bar{\tilde\Psi}^{-1ab}_\mu & \cr
{ \tilde\Phi}^{2ab ^+ } && \t\Phi^{(0)ab^-}, {\tilde b}^{0ab}_\mu
& & {\bar{\tilde\Phi}}^{-2ab ^+} \cr
& & & { \bar{\tilde\eta}}^{-1ab ^+ } & \cr
}
\nonu\end{aligned}$$
$$\begin{aligned}
\matrix
{ & & & B_{\mu \nu} & & & \cr
& & \Psi^{(1)}_{\mu \nu} & & & \cr
& A^{(2)}_{\mu } & & & & A^{(-2)}_{\mu } & \cr
R^{(3)} & & S^{(1)} & & \bar \Psi^{(-1)}_{\mu}
& & \bar R^{(-3)} \cr
& b_ {S^{(1)}} ^{(2)} & & \Phi^{(0)} ,\bar\Phi ^{(0)}
& & b_{\bar R^{(-3)}}^{(-2)} \cr
& & \eta^{(1)} & & & & & \cr
}
\nonu\end{aligned}$$
$$\begin{aligned}
\matrix
{ \xi^ \mu & & \bar \xi^ \mu \cr
& b^\mu }
\quad\quad\quad\quad\quad
\matrix
{ \O^ {ab } & & \bar \O^ {ab } \cr
& b^{ab } & }
\label{spectrumfinal}\end{aligned}$$
We will shortly see that the component $ \Psi^{( 1) }_{\mu\nu^+}$ of the topological ghost of the two-form $B_{\mu\nu}$ is gauge fixed in an algebraic way, leaving only $ \Psi^{( 1) }_{\mu\nu^-}$ as a propagating field.
Let us analyse the fields that remain after these eliminations. The fields that will a play a rôle as the classical fields of supergravity are: $$\begin{aligned}
&&\mbox{bosons:}\ \ e^a_{\mu }, \sigma, B_{\mu\nu}, A^{(2)}_\mu,
A^{(-2)}_\mu \nn \\
&&\mbox{fermions:}\ \
(\Psi^{(1)a}_\mu, \bar \Psi^{(-1)ab^-}_\mu, \bar
\Psi^{(-1)}_\mu), (\bar
\chi^{(-1)a}, \Psi^{( 1)ab^-},\chi^{( 1)})
\label{classic}\end{aligned}$$ Indeed, these are nothing but the fields of the $N=1,D=8$ supergravity multiplet, up to a twist. The most striking point of our construction is that the [*topological*]{} ghosts of ghosts of the two-form, $A_\mu^{(2)}$ and $A_\mu^{(-2)}$ can be interpreted as the propagating graviphotons of $N=1,D=8$ supergravity. The $Spin(7)$-scalar ghost of ghost $\sigma$, which has ghost number zero, can be interpreted as the dilaton. The corresponding topological ghosts can be recognized as the twisted version of the fermionic part of the spectrum. Modulo some field redefinitions that we will discuss in detail in the following section, the twisted gravitino can be identified with the ghosts $(\Psi^{(1)a}_\mu, \bar \Psi^{(-1)ab^-}_\mu, \bar
\Psi^{(-1)}_\mu)$, and the twisted dilatino with $ (\bar \chi^{(-1)a },
\Psi^{( 1)ab^-},\chi^{( 1)})$. We remark that the BRST variation $sA_\mu^{(-2)}=\bar
\Psi _\mu^{(-1)}$ appears in the twisted gravitino. This is in agreement with the analysis of [@BT1; @BT2], where we found that at least one graviphoton with non zero ghost number is needed both in four and eight dimensions in order to introduce $\bar
\Psi _\mu^{(-1)}$. As we will describe in the following section, the possibility of having $Spin(7)$ decompositions allows one to do all relevant maps of tensors upon spinors.
The fields of the topological multiplets not appearing in eq. (\[classic\]) will be interpreted as the ordinary Faddeev–Popov ghosts and antighosts of the supergravity. The infinitesimal symmetry transformations of supergravity will be deduced from the topological BRST equations. After the eight-dimensional untwisting, one gets from eq. (\[spectrumfinal\]) the propagating fields $(\Phi^{(2)a}, {\Phi}^{(0)ab -},\Phi^{(0)})$, $(\bar \Phi^{(-2)a}, {\t \Phi}^{(0)ab -},\bar \Phi^{(0)})$ and $(\bar \eta^{(-1)a}, \eta^{(1)ab ^-} , \eta^{(1)})$. These can be interpreted respectively as the twisted version of the Faddeev–Popov spinorial ghosts and antighosts for local supersymmetry and the corresponding Lagrange multipliers. The invariance of the untwisted theory therefore displays a variation of the vector-spinor by a derivative of these ghosts: this is presumably sufficient to ensure the full supersymmetry of the untwisted action. This is different from the case of topological Yang-Mills theory, where the complete supersymmetric invariance of the untwisted action is not automatic.
The topological action and its correspondence with supergravity
===============================================================
After having obtained the relevant field spectrum for the TQFT, the task is of finding the topological gauge functions.
A natural extension of the topological gauge function used in [@BT2] in the presence of the two-form field $B_{\mu\nu}$ is given by imposing the octonionic self-duality condition $$\label{octo}
\tilde\o^{ab}-\demi \O^{abcd} \tilde\o^{cd}=0,$$ on the extended connection $$\tilde\o^{ab}\equiv \o^{ab}- G^{ab}_c e^c \ \ ,
\label{otilde}$$ with torsion given by $$\tilde T^a = \tilde D e^a = G^a_{bc}e^b e^c \ \ .
\label{torsion}$$ The condition (\[octo\]) has to be considered together with the gauge function $$d\sigma + 2 *(\O\wedge G_3) =0
\label{g-sigma}$$ involving the dilaton field. We thus write the topological action $$\begin{aligned}
\label{top-act}
\L_{e,B}&=& s
\Big[
\bar \Psi^{(-1)ac^-} (b ^{(0)cb^-} + \o^{cb^-}(e) - G_d^{cb^-} e^d)\V_{ab} \cr
&&\ \ +\bar\chi^{(-1)a} \Big (b^{(0)}_a +\partial_a \sigma
+\Omega_{abcd}G_{bcd} \Big ) \Big],\end{aligned}$$ where we introduced the compact notation for the volume forms ${\cal V}_{a_1\ldots a_i}\equiv {1\over (8-i)!}
\epsilon_{a_1\ldots a_8}e^{a_{(i+1)}}\ldots e^{a_8}$. This topological action, after integration on the Lagrange multipliers $b^{(0)cb^-}$ and $b^{(0)}_a$, gives kinetic terms for the graviton, the two-form and the dilaton of N=1 D=8 supergravity.
First of all, the Einstein Lagrangian, written with the curvature $\tilde R^{ab}$ of the extended connection $\tilde\omega$, is equal to the sum of the ordinary Einstein-Hilbert Lagrangian plus the squared norm of the field-strength $G_3$ $$\tilde {\cal L} = \demi \tilde R^{ab}{\cal V}_{ab}
= \demi R^{ab}{\cal V}_{ab} - {1\over 2} G^{ab}_c G^{ab}_c
{\cal V}.
\label{L-tilde}$$ The Bianchi identity on the field strength $G_3$ and (\[torsion\]) imply that $$\tilde R^{ab}e_b e_a + \tilde T^a\tilde T^a = 0 \ \ .
\label{bi-t}$$ By multiplying (\[bi-t\]) by the self-dual four-form $\Omega$ we get $$\demi \O_{abcd}\tilde R^{cd} {\cal V}^{ab} =
\O_{mnpq}G^a_{mn}G^a_{pq} {\cal V}.
\label{ov}$$ The relation (\[ov\]), together with the $Spin(7)$ decomposition $$\label{deco}
\tilde R^{ab}= \tilde R^{ab^+}+ \tilde R^{ab^-},
\quad \demi \O_{abcd} \tilde R^{ab}= \tilde R^{cd^+}-3 \tilde R^{cd^-} ,$$ allows for the elimination of $\tilde R^{ab^+}$ in the Lagrangian (\[L-tilde\]) $$\begin{aligned}
\label{l-minus}
\tilde {\cal L} &=& \demi \tilde R^{ab}{\cal V}_{ab} =
2\tilde R^{ab^-} {\cal V}_{ab} + {1\over 2}
\O_{abcd}G^{ab}_e G^{cd}_e {\cal V} \\
&=& -4 \tilde\o^{ac^-}\tilde\o^{cb^-}{\cal V}_{ab}
+ 2 \tilde\o^{ab^-}\tilde T^c \V_{abc}
+2d(\tilde\o^{ab^-}\V_{ab})
+{1\over 2}\O_{abcd}G^{ab}_e G^{cd}_e {\cal V}
\ \ ,
\nn\end{aligned}$$ where in the last identity we used that $\tilde R^{ab^-}=\tilde D\tilde\o^{ab^-}
-2\tilde\o^{ac^-}\tilde\o^{cb^-}$ and integrated by parts the term in $\tD\ot^{ab^-}$. By comparing (\[l-minus\]) with (\[L-tilde\]) and using (\[torsion\]) we can finally write the identity $$4 \tilde\o^{ac^-}\tilde\o^{cb^-}{\cal V}_{ab}=
{\cal L}_{EH} - {1\over 2}G^{ab}_c G^{ab}_c\V
+\O_{abcd}G^{ab}_f G^{cd}_f
+4 \o^{ab^-}G_{ab}^c\V_c
+2 d(\ot^{ab^-}\V_{ab}) \ \ ,
\label{omega^2}$$ where we defined the Einstein-Hilbert action as ${\cal L}_{EH}\equiv \demi R^{ab}\V_{ba}$. Let us consider now the square of the gauge function (\[g-sigma\]). By using the identity $\O_{abcd}\O^{afgh}=(6\delta_{bcd}^{[fgh]}
-9\O_{bc}^{[fg}\delta^{h]}_d )$ is easy to find that $$\left(\pa_a\sigma + {1\over 3}\O_{abcd}G^{bcd}\right)^2 =
(\pa_a\sigma)^2 + {2\over 3}G^{ab}_c G^{ab}_c -
\O_{abcd}G^{ab}_f G^{cd}_f +
{2\over 3}\pa^a\left(\sigma\O_{abcd}G^{bcd}\right) \ \ .
\label{sigma^2}$$ By summing (\[omega\^2\]) and (\[sigma\^2\]) (multiplied by the suitable volume form $\V$), we finally get $$\begin{aligned}
&& 4 \tilde\o^{ac^-}\tilde\o^{cb^-}{\cal V}_{ab} +
\left(\pa_a\sigma + {1\over 3}\O_{abcd}G^{bcd}\right)^2 \V
\label{fund-id} \\
&&= {\cal L}_{EH} + \left[(\pa_a \sigma)^2 +
{1\over 6}G^{ab}_c G^{ab}_c\right]\V
+ 4 \o^{ab^-}G_{ab}^c\V_c
+ \mbox{boundary terms} \ \ .
\nn\end{aligned}$$ This identity deserves some attention, since it shows that the sum of the 56 independent terms contained in $ |G _{abc}(B) |^2$, plus those contained in the Einstein action can be obtained as a topological gauge-fixing term, stemming from 64 Lagrange multipliers.
We are now able to compare the topological terms (\[fund-id\]) with the bosonic part of the action of N=1 D=8 supergravity. In the spirit of topological field theory, we can restrict our attention on the kinetic terms for the fields, which simplifies considerably the comparison. In fact, the basic requirement on the gauge-fixing conditions is that they must give a good definition for the propagators of the fields. Interaction terms can be then always added as BRST exact terms in order to get agreement with the complete twisted supergravity action. The first three terms of the topological action (\[fund-id\]) correctly reproduce the kinetic terms for the graviton, the two-form and the dilaton of N=1 D=8 supergravity. Concerning now the mixed term $ \o_c^{ab^-}(e)G_{abc}$ in (\[fund-id\]), it can be reduced in the quadratic approximation to $$\partial _\mu e^a_\mu \Omega_{abcd} G_{bcd} .
\label{gf}$$ This can be done by imposing the $Spin(7)$ invariant Lorentz gauge condition $e^{[a}_\mu V^{b]^-\mu}=0$, where $V^{b\mu}$ is an inverse 8-bein chosen as a reference system. [^5] The expression (\[gf\]) can be absorbed in the gauge-fixing term for the reparametrization invariance $s(\bar \xi^\mu \partial _\nu g_{\mu\nu})$. We thus conclude that the topological action (\[fund-id\]) coincides with the bosonic part of the supergravity action in a given $Spin(7)$-invariant Lorentz gauge. This is not surprising since the condition (\[octo\]) leaves only a residual $Spin(7)$ symmetry group. It is remarkable that the kinetic energies of both the graviton and the two-form stem from the gauge-fixing term (\[octo\]) that only comes from the topological freedom in the vielbein.
We now pass to the fermionic sector. The BRST variation of the first line in (\[top-act\]) gives part of the Rarita–Schwinger action, as in [@BT2]. Notice that the condition $e^{[a}_\mu V^{b]^-\mu}=0$, which must be enforced by the BRST exact term $s (\bar \O ^{[ab]^-}e^{[a}_\mu
V^{b]^-\mu})$, yields the condition $\O ^{[ab]^-}=-\Psi ^{[a}_\mu e ^{b]^-\mu}$, and ensures the $Spin(7)$ invariance of the fermionic part of the contribution of $\L_{e,B}$ to the Rarita–Schwinger action. There is actually a compensation between the BRST variations of $ \o _c^{ab^-}(e)$ and $G ^{ac^-}_c(B) $, which is compulsory to enforce gauge invariance, since $\o ^{ab^-}(e)$ transforms as a connection for Lorentz transformations with self-dual parameter $\O^{ab^-}$.
To determine the complete Rarita–Schwinger action, we must add to $\L_{e,B}$ the following term, as in [@BT2]: $$\begin{aligned}
\L_{\bar \Psi_\mu}=
s \Big[ \partial_{[\mu}A^{(-2)}_{\nu]^-} \Psi^{(1)a}_{[\mu}
e^a_{\nu]}\ \Big]\end{aligned}$$ Looking at the fermionic dependence, since $sA^{(-2)}_\mu=
\bar \Psi^{(-1)}_\mu$, one finds that $ (\Psi^{(1)a}_\mu, \bar \Psi^{(-1) ab^-}_\mu, \bar
\Psi^{(-1)}_\mu)$, can be identified as the twisted gravitino, with 8 chiral and 8 antichiral components, as in [@BT2].
To determine the propagation of $A^{(\pm 2)}_\mu$, we add: $$\begin{aligned}
s\Big[\Psi^{(1)}_{\mu\nu} \pa_{[\mu} A^{(-2)}_{\nu]} \Big]
=\pa_{[\mu} A^{(2)}_{\nu]}\pa_{[\mu} A^{(-2)}_{\nu]}
+\Psi^{(1)}_{\mu\nu}\pa_{[\mu} \Psi^{(-1)}_{\nu]}\end{aligned}$$ This identifies $A^{(\pm 2)}_\mu$ as the two graviphotons of $N=1,D=8$ supergravity. Notice that among the two graviphotons $A_\mu ^{(-2)}$ and $A_\mu ^{(2)}$, only the latter one has a “topological” transformation, since $sA_\mu^{(-2)}=\bar \Psi^{(-1)}_\mu$ while $sA_\mu^{(2)}=\partial _\mu
R^{(3)}$. This is in agreement with the twisted supergravity transformations.
Part of the difficulty of this work was to understand the rôle of the ghost field $\Psi_{\mu\nu}^{(1)}$. On the one hand it is the field that generates by its ghost of ghost symmetry the second graviphoton of the supergravity. On the other hand, in supergravity, the Kalb–Ramond two-form only sees local supersymmetry through the gravitino and the dilatino, and it is challenging to uncover this from topological invariance. This leads to the conclusion that not all the components of $\Psi_{\mu\nu}^{(1)}$ are independent propagating fields. Only $\Psi_{\mu\nu^-}^{(1)}$ survives as an independent field stemming from the topological invariance of the two-form. To enforce the other components of $\Psi_{\mu\nu}^{(1)}$ in a BRST invariant way, we consider the following action, that exhausts the ghost of ghost symmetry in the Lorentz sector: $$\begin{aligned}
\label{el}
\L_{\Psi^{(1)}_{\mu\nu},\tilde{\Phi}^{(2)ab}}&=&s\Big[
e^\mu_c e^\nu_d \ \bar{\tilde\Phi}^{(-2)cd^+}( {
\Psi}^{(1)}_{\mu\nu} -\Psi^{(1)a}_{[\mu}e^a_{\nu]})\Big]
\nn \\
&=&
e^\mu_c e^\nu_d \
\bar {\tilde \eta}^{(-1)cd^+}( { \Psi}_{\mu\nu}^{(1) }
-\Psi^{(1)a}_{[\mu} e^a_{\nu]})
+
\bar {\tilde \Phi}^{ab^+}( \partial _{[a}A^{(2)}_{ b]}
-\tilde \Phi ^{( 2) }_{ab}+\ldots)\nonumber\\\end{aligned}$$ This gauge-fixing allows one to identify $\Psi_{\mu\nu^+}^{(1)}=
\Psi^{(1)a}_{[\mu} e^a_{\nu]^+} $. If we define $$\label{psib}
\Psi_{\mu\nu}^{(1)}= \chi_{\mu\nu^-}^{(1)}
+\Psi^{(1)a}_{[\mu} e^a_{\nu]}$$ then $\chi_{\mu\nu^-}^{(1)}$ and $\chi^{(1)}=s\sigma$ are eight fermionic variables that can be identified with a twisted chiral component of the dilatino. $\bar\chi^{(-1)\mu}$ determines by twist the other chiral component. Moreover, Eq.(\[el\]) accomplishes the elimination of the ghost of ghost dependence in the Lorentz sector, by yielding an algebraic equation of motion for $ {\tilde \Phi}^{( 2)ab^+}$ and $ \bar {\tilde \Phi}^{(-2)ab^+}$, in a $Spin(7)$ invariant way.
Let us summarize the mapping between the fermionic degrees of freedom of the topological action defined by (\[top-act\])–(\[el\]) and that of N=1 supergravity. On a manifold with $Spin(7)$ holonomy there exist a covariantly constant spinor (of norm one) $\varepsilon$ [^6] , which can be used to redefine as in [@BT2] the gravitino $(\lambda, \bar\lambda)$ and the dilatino $(\chi, \bar\chi)$ of $N=1$, $D=8$ supergravity as [^7] $$\begin{aligned}
\lambda &=& \Psi^a \gamma_a \varepsilon \ \ , \nonumber\\
\bar\lambda &=& \bar\Psi \varepsilon +
\bar\Psi^{ab^-}\gamma_{ab}\varepsilon\ ,
\label{twist-1} \\
\chi &=& \bar\chi^a \gamma^a \varepsilon \ \ , \nn\\
\bar\chi &=& \chi\varepsilon +
\chi^{ab^-}\gamma_{ab}\varepsilon \ \ .
\label{twist-2}\end{aligned}$$ On the l.h.s. of (\[twist-1\]), (\[twist-2\]) $(\lambda, \bar\lambda)$ and $(\chi, \bar\chi)$ are spinors of opposite chiralities. The eight-dimensional gamma matrices $\gamma_a$ acts on spinors of definite chirality. Notice that the identification (\[psib\]) implies the appearance in the topological action (\[top-act\]) of the mixed kinetic terms $\bar\Psi^{ac^-}\partial_d\chi^{cb^-} e^d \V_{ab}$, coming from the BRST-variation of the first line in (\[top-act\]), and $\bar\chi^a\O_{abcd}\partial_b\Psi_{cd}$, coming from the BRST-variation of the second line. In order to recover these terms from the twisted supergravity theory we have to impose the field redefinitions $\Psi_b^a\rightarrow \Psi_b^a + \chi^{ab^-}$ and $\bar\Psi_c^{ab^-} \rightarrow \bar\Psi^{ab^-}_c + \delta^{[a}_c
\bar\chi^{b]^-}$. Moreover, since for the bosonic sector the equivalence of the topological and supergravity actions is valid only in a fixed $Spin(7)$-invariant Lorentz gauge, we expect that also some gauge-fixing terms of the same kind are involved in the comparison of the fermionic part. From eq. (\[twist-1\]) and (\[twist-2\]) we see that, modulo the above field redefinitions, the gravitino is mapped to the fields $(\Psi^a, \bar\Psi, \bar\Psi^{ab^-})$ of the topological model, while the dilatino is mapped to the fields $(\bar\chi^a, \chi, \chi^{ab^-})$.
What we have found is interesting. The topological gauge functions are such that the BRST transformation in the effective topological action for $e$ and $B$ is: $$\begin{aligned}
s e^a_\mu &=& \Psi^{(1)a}_\mu +\Omega ^{ab} e^b_\mu
+\ldots
\cr s B_{\mu\nu} &=& \Psi^{(1)a}_{[\mu} e^a_{\nu]}+
{\chi}^{(-1)}_{\mu\nu^-}+
\ldots \end{aligned}$$ Only the symmetrical part of $\Psi ^{(1)a}_\mu$ is involved in $se^a_\mu$, while the Lorentz ghost $\Omega ^{ab}$ allows one to put to zero the antisymmetrical part of $ e^a_\mu$ . This explains how the supersymmetric transformation law of the two-form in the supergravity framework can be interpreted in a topological way, using a suitable gauge function for the topological invariance. What actually happens is that, when one twists the gravitino $\Psi^\alpha_\mu$ into $\Psi^{(1)a}_\mu$, and defines $\Psi^{(1)}_{ \mu\nu}= e_{a\nu} \Psi^{(1)a}_\mu $, then $\Psi^{(1)}_{\{\mu\nu\}}$ and $\Psi^{(1)}_{[\mu\nu]}$ are respectively the topological ghosts of the two-form $B_{\mu\nu}$ and of the metric $g_{\mu\nu}$. Here, the mapping between the spinors is not linear, since it involves various contractions by the vielbein. This a important distinction with the case of the Yang–Mills TQFT, where the mapping is a linear transformation.
Some extra topological functions are needed in order to fix the symmetries of the gauge conditions used so far. These functions exhaust the remaining fields of the BRST topological multiplets. Let us briefly sketch them. To take care of the gauge invariance of $\bar \Psi_\mu ^{(-1)ab^-}$, which follows from the gauge functions, we redefine $b^{(0)ab^-}_\mu \to
b^{(0)ab^-}_\mu +\partial _\mu \t \Phi ^{(0)ab^-}$. Then, to gauge fix the local supersymmetry, which pops up as the gauge invariance of the topological ghosts, we add, as in [@BT1; @BT2] $$\begin{aligned}
\label{gfghost}
\L_{ghosts}=
s\Big[\sqrt{g}(\bar \Phi ^{(-2)a} D_\mu \Psi ^{{(1)a} }_\mu+
{\Phi}^{(0)ab^-} D_\mu \bar \Psi ^{(-1)ab^-}_\mu
+\bar \Phi ^{(0) } \pa_\mu \bar \Psi^{(-1) }_\mu)\Big]\end{aligned}$$ The rôle of this redefinition of $b^{ab^-}$ has been analysed in [@BT1; @BT2]. It ensures the propagation of the field $\t \Phi ^{(0)ab^-}$. The expression of this action identifies $(\Phi^{(2)a},
{\Phi}^{(0)ab -},\Phi^{(0)})$, $(\bar \Phi^{(-2)a}, {\t \Phi}^{(0)ab -},\bar \Phi^{(0)})$ and $(\bar \eta^{(-1)a}, \eta^{(1)ab ^-} , \eta^{(1)})$ as the twisted version of the Faddeev–Popov spinorial ghost and antighosts for local supersymmetry, and their fermionic Lagrange multipliers, respectively.
We also use the topological gauge freedom of the spin connection $\o^{ab }$ to eliminate this field in terms of $e$, by mean of the term $$\begin{aligned}
s\Big [ \bar {\tilde \Psi}^{(-1)ab} e_b \wedge * T^a \Big]\end{aligned}$$ where $*T$ is the Hodge dual of the torsion $T=de+\o\w e $. This gauge fixing, which can be improved by changing $ T^a\to T^a +G^a_{bc} e^be^c$ , also trivially eliminates the dependence of the action upon the topological Lorentz ghosts $ {\tilde \Psi}^{(1) ab}_\mu$ and $\bar {\tilde \Psi}^{(-1)ab}_\mu$, which disappear by their algebraic equations of motion. One must recognize that introducing the Lorentz symmetry is extremely useful, although most of its ingredients are eventually eliminated.
As for the fields $ S^{(1)}$ and $R^{(-3)}$, they are used to fix the ordinary gauge symmetry of $A^{(-2)}$ and $A^{(2)}$ by the action $$\begin{aligned}
s\Big [ S^{ (1)} \partial_\mu A^{(-2)}_\mu +\bar R^{(-3)}\partial_\mu
A^{(2)}_\mu \Big ]\end{aligned}$$ To impose that the vielbein is a symmetrical matrix, and eliminate the $\Omega$ and $\bar \O$ dependence, we just add : $$\begin{aligned}
s \Big[ \bar \Omega ^{ab^+} e^{b}_\mu V ^{\mu a} \Big]
=\bar \Omega ^{ab^+} (\Omega ^{ab} +\ldots)
+b ^{ab^+}
e^{b}_\mu V ^{\mu a},\end{aligned}$$ keeping in mind that we had already used a term $s \Big[ \bar \Omega ^{ab^-} e^{b}_\mu V ^{\mu a} \Big]$. After expansion, and a few field redefinitions, one gets that $ \Omega ^{ab}$ and $ \bar \Omega ^{ab}$ are eliminated by Gaussian integration. As said above, this necessitates the introduction of an inverse vielbein $V^{a\mu}$ as background.
Last of the last, we must add $$\begin{aligned}
s\Big [\bar\xi ^{-1\mu }\partial_\nu g_{\mu\nu} \Big ]\end{aligned}$$ to fix the reparametrization invariance.
The gauge-fixing of the gauge symmetries of topological gauge functions could have been done in a much more refined way, using the technology of equivariant cohomology[^8]. This is a technicality that we have chosen not to present here. It would distract us from our main result, that is, we have finally end up our task of building a TQFT for the Kalb–Ramond field $B_{\mu\nu}$ within the context of topological gravity. The result is that the standard TQFT procedure has lead us to a twisted version of $N=1, D=
8$ supergravity.
It is worth noticing that the topological gauge functions on the extended connection (\[octo\]) and on the dilaton (\[g-sigma\]) are the same appearing in the octonionic superstring equations [@stro]. By coupling our model with a non abelian topological Yang-Mills theory, as it is defined in [@bakasi; @Acharya:1997gp], one could thus obtain a topological theory which effectively describes the transverse properties of the octonionic superstring.
We believe that it will be interesting to study the possible dimensional reductions of the 8-dimensional topological gravity. This idea has already proven to be quite useful in the simpler case of the topological Yang–Mills theory. The reduction to seven dimensions of the octonionic self-duality conditions on the spin connection is of relevance for manifolds with (weak) $G_2$ holonomy [@Bilal]. The study of the dimensional reduction of the generalized self-duality conditions (\[octo\]) and (\[g-sigma\]) could give a more general description of manifolds with $G_2$-structure.
Moreover, one may investigate the possibility of getting a generalization of the Seiberg–Witten theory involving gravity. Generalizing the flat space analysis of [@bakasi], this theory could be derived as an adequate dimensional reduction in four dimensions of the 8-dimensional topological gravity. Further dimensional reductions can give interesting models in 2 and 0 dimensions.
[**Acknowledgments**]{}:
We thank P. de Medeiros and B. Spence for useful discussions. A.T. is supported by a Marie Curie fellowship under contract N$^o$ HPMF-CT-2001-01504.
[15]{}
L. Baulieu, M.B. Green, E. Rabinovici, [ *Superstrings from theories with $N>1$ world-sheet supersymmetry* ]{}, , \[hep-th/9611136\]; [*A Unifying Topological Action for Heterotic and Type II Superstring Theories* ]{}, Phys. Lett. [**B386**]{} (1996) 91, \[hep-th/9606080\].
L. Baulieu and A. Tanzini, [*Topological gravity versus supergravity on manifolds with special holonomy,*]{} JHEP [**0203**]{} (2002) 015, \[hep-th/0201109\].
L. Baulieu, M. Bellon and A. Tanzini, [*Eight-dimensional topological gravity and its correspondence with supergravity,*]{} Phys. Lett. B [**543**]{} (2002) 291, \[hep-th/0207020\].
P. de Medeiros and B. Spence, [*Four-dimensional topological Einstein-Maxwell gravity,*]{} hep-th/0209115.
L. Baulieu and C. Laroche, [ *On Generalized Self-Duality Equations Toward Supersymmetric Quantum Field Theories of Forms,* ]{} Mod. Phys. Lett. [**A13**]{} (1998) 1115, \[hep-th/9801014\].
A. Salam and E. Sezgin, [*d=8 Supergravity: matter couplings, gauging and Minkowski compactification,*]{} Phys. Lett. [**B154**]{}, 37.
L. Baulieu, H. Kanno, I.M. Singer, [ *Special Quantum Field Theories In Eight And Other Dimensions,* ]{} Commun. Math. Phys. [**194**]{} (1998) 149, \[hep-th/9704167\].
B. S. Acharya, M. O’Loughlin and B. Spence, [*Higher-dimensional analogues of Donaldson-Witten theory,*]{} Nucl. Phys. B [**503**]{} (1997) 657 \[hep-th/9705138\].
J. A. Harvey and A. Strominger, [*Octonionic Superstring Solitons,*]{} Phys. Rev. Lett. [**66**]{} (1991) 549.
A. Bilal, J. P. Derendinger and K. Sfetsos, [*(Weak) G(2) holonomy from self-duality, flux and supersymmetry,*]{} Nucl. Phys. [**B628**]{} (2002) 112, \[hep-th/0111274\]; A. Bilal and S. Metzger, [*Compact weak G(2)-manifolds with conical singularities,*]{} hep-th/0302021.
[^1]: In four dimensions, the twist of the complete action including the interaction terms has been discussed in [@mespe].
[^2]: Throughout the paper the Latin indices $a,b,\ldots$ denote flat $SO(8)$ tangent space indices, and $\mu,\nu,\ldots$ are eight-dimensional world indices.
[^3]: Strictly speaking, the two indices of the vielbein represent components in different spaces, so that one can only speak of the antisymmetric part of $e_\mu^a$ once a background vielbein has been chosen. This is not the case for other fields with the same indices as the vielbein, since the vielbein is then available to relate the two spaces.
[^4]: More precisely, all equations for the antighosts appearing in our field spectrum are of the type $ \s \bg = \lambda $, $ \s \lambda =\L_\Phi \bg + \delta_{\t \Phi}\bg $ with $s X= \s X +\L_\xi X+\delta_{{\O} } X$.
[^5]: In the physicist language, this means that one uses the Lorentz invariance to impose that 7 of the components of $e^a_\mu$ vanish in a $Spin(7)$ invariant way. Eventually, the rest of the 21 Lorentz degrees of freedom can be used to enforce that $e^a_\mu$ be a symmetrical $8\times 8$ matrix, that is $e^a_\mu= V^{b\mu } (\delta ^{ab} +
h^{ab})$, where $h^{ab}$ is symmetrical in $a$ and $b$. Then, there is a one-to-one mapping between the metric and this matrix.
[^6]: This spinor can be used to define the self-dual four form as $ \O_{abcd}=\varepsilon^T \gamma_{abcd}\varepsilon$ [@bakasi].
[^7]: Similar results concerning the twist of $N=1$, $D=8$ supergravity have been obtained by P. de Medeiros and B.Spence.
[^8]: As remarked in the previous section, the equivariance is with respect to reparametrizations, $Spin(7)\subset
SO(8)$ Lorentz invariance and two-form gauge symmetry.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Local compositions and structures of alloys have been investigated by Raman and solid-state $^{67}$Zn/$^{25}$Mg nuclear magnetic resonance (NMR) spectroscopies, and by neutron pair-distribution-function (PDF) analyses. The $E_2^\mathrm{low}$ and $E_2^\mathrm{high}$ Raman modes of display Gaussian- and Lorentzian-type profiles, respectively. At higher Mg substitutions, both modes become broader, while their peak positions shift in opposite directions. The evolution of Raman spectra from solid solutions are discussed in terms of lattice deformation associated with the distinct coordination preferences of Zn and Mg. Solid-state magic-angle-spinning (MAS) NMR studies suggest that the local electronic environments of $^{67}$Zn in ZnO are only weakly modified by the 15% substitution of Mg for Zn. $^{25}$Mg MAS spectra of Zn$_{0.85}$Mg$_{0.15}$O show an unusual upfield shift, demonstrating the prominent shielding ability of Zn in the nearby oxidic coordination sphere. Neutron PDF analyses of using a 2$\times$2$\times$1 supercell corresponding to Zn$_7$MgO$_8$ suggest that the mean local geometry of MgO$_4$ fragments concurs with previous density functional theory (DFT)-based structural relaxations of hexagonal wurtzite MgO. MgO$_4$ tetrahedra are markedly compressed along their $c$-axes and are smaller in volume than ZnO$_4$ units by $\approx$6%. Mg atoms in have a shorter bond to the $c$-axial oxygen atom than to the three lateral oxygen atoms, which is distinct from the coordination of Zn. The precise structure, both local and average, of obtained from time-of-flight total neutron scattering supports the view that Mg-substitution in ZnO results in increased total spontaneous polarization.'
author:
- 'Young-Il Kim,$^1$ Sylvian Cadars,$^2$ Ramzy Shayib,$^2$ Thomas Proffen,$^3$ Charles S. Feigerle,$^4$ Bradley F. Chmelka,$^2$ and Ram Seshadri$^{1,5}$'
title: 'Local structures of polar wurtzites studied by Raman and $^{67}$Zn/$^{25}$Mg NMR spectroscopies and by total neutron scattering'
---
Introduction
============
Research in the area of polar semiconductor heterostructures has been growing rapidly, driven in large part by interest in two-dimensional electron gas (2DEG) systems.[@Tsuka; @Koike; @Rajan; @Sasa; @Tampo] 2DEGs are known to form at heterojunction interfaces that bear polarization gradients. They can display extremely high electron mobilities, especially at low temperatures, owing to spatial confinement of carrier motions.[@Davies] Recent reports of 2DEG behaviors in Ga$_{1-x}$Al$_x$N/GaN and /ZnO heterostructures have great significance for the development of novel high-electron-mobility transistors (HEMTs)[@Rajan; @Sasa; @Tampo] and quantum Hall devices.[@Tsuka]
2DEG structures are usually designed by interfacing a polar semiconductor with its less or more polar alloys in an epitaxial manner. Since the quality of the 2DEG depends critically on interface perfection, as well as the polarization gradient at the heterojunction, understanding compositional and structural details of the parent and alloy semiconductors is an important component in 2DEG design and fabrication. The evolution of atomic positions and cell parameters upon alloying can directly affect the magnitude of the polarization gradient and the carrier density at the heterojunction.
/ZnO is one of the more promising heterostructure types for studies of 2DEGs, due to the large polarization of ZnO, the relatively small lattice mismatch, and the large conduction band offsets in the /ZnO heterointerface. Although 2DEG formation in /ZnO heterostructures have been researched for some time, a clear understanding of the alloy structure of is currently lacking. Recently, we have studied composition-dependent changes in the crystal structures of polycrystalline by synchrotron x-ray diffraction and Raman spectroscopy.[@Kim1; @Kim2] For the composition range $0\,\leqslant\,x\,\leqslant\,0.15$, we have shown that Mg-substitution modifies the aspect ratio of the hexagonal lattice through enhanced bond ionicity, and in parallel, decreases static polarization in the crystal, due to decreased internal distortion in the tetrahedral coordinations.
Here, we conduct a detailed and more precise study of the local structure of alloys using Raman and solid-state nuclear magnetic resonance (NMR), in conjunction with neutron diffraction techniques. Raman and NMR spectroscopy are useful probes for addressing the molecular compositions and structures of solid-solution systems. Peak shapes and widths of the Raman spectra reflect compositional fluctuations and both short- and long-range order, whereas NMR is sensitive to the local environments around specific nuclei. Thus, both techniques can provide structural information on that is complementary to diffraction analyses. Here, we examine the details of $E_2^\mathrm{low}$ and $E_2^\mathrm{high}$ Raman modes for ($x=0$, 0.05, 0.10, and 0.15). We have also used spin-echo magic-angle-spinning (MAS) $^{67}$Zn and $^{25}$Mg NMR to study separately zinc and/or magnesium species in ZnO and Zn$_{0.85}$Mg$_{0.15}$O. These measurements have been correlated with average and local crystal structures of and ZnO, as established by Rietveld (diffraction space) and pair-distribution-function (PDF, real-space) analyses of time-of-flight neutron diffraction data. Compared with x-rays, neutron scattering provides much greater sensitivity to oxygen and Mg positions, as well as larger momentum transfer as measured by the larger maximum $Q$ wavevector.[@Proffen] By taking advantage of the increased data quality, we are able to isolate the geometry of MgO$_4$ moieties stabilized in the wurtzite ZnO lattice.
Experimental
============
Polycrystalline powder samples of ZnO and ($x$ = 0.05, 0.10, 0.125, and 0.15) were prepared from oxalate precursors obtained by co-precipitation using Zn(CH$_3$CO$_2$)$_2\cdot$2H$_2$O, Mg(NO$_3$)$_2\cdot$6H$_2$O, and H$_2$C$_2$O$_4$, all of which had purities of 99.999% from Aldrich. The two metal salts were dissolved together in deionized water and added to an oxalic acid solution in the ratio of \[Zn$^{2+}$\]:\[Mg$^{2+}$\]:\[C$_2$O$_4^{2-}$\] = (1$-x$):$x$:1.05. Upon mixing with oxalate, Zn$^{2+}$ and Mg$^{2+}$ immediately coprecipitated as white crystalline oxalate powders, which were washed with deionized water and dried at 60$^{\circ}$C for 4h and subsequently heated at 550$^{\circ}$C for 20h in air to decompose the oxalates to the oxides. Powder x-ray diffraction measurements confirmed the formation of Zn$_{1-x}$Mg$_x$C$_2$O$_4\cdot$2H$_2$O after heating at 60$^{\circ}$C, and after decomposing at 550$^{\circ}$C. For the NMR experiments, samples were also prepared at different conditions of temperature (900$^{\circ}$C) and atmosphere (O$_2$, N$_2$).
Raman spectra for ZnO and ($x=$ 0.05, 0.10, and 0.15) were acquired at room temperature using a Jobin Yvon-Horiba T64000 triple grating (1800gr/mm) spectrometer. The spectra of lightly compressed powders were recorded using micro-Raman sampling in air with 514.5nm and 675.5nm excitations. The spectra reported here were averages of 10 acquisitions of 20 s integrations of the CCD detector with $\approx$10mW of laser power focused onto the samples through a 10$\times$ objective. There was no evidence of degradation of the samples or associated changes in their spectra under these conditions. The spectrometer was calibrated using a 520.7cm$^{-1}$ lattice mode of silicon. The Raman features of wurtzite $E_2^\mathrm{low}$ and $E_2^\mathrm{high}$ phonon modes were analyzed in detail to determine the peak shape, position, and width. Spectral background was removed following Shirley[@Shirley], and the profile fitting was performed using pseudo-Voigt[@Cox] or Breit-Wigner-Fano functions.[@Yoshik]
NMR measurements of the low-gyromagnetic-ratio nuclei $^{67}$Zn and $^{25}$Mg were conducted at high (19.6Tesla) magnetic field strength at the National High Magnetic Field Laboratory in Tallahassee, Florida. These investigations benefited from the enhanced sensitivity and improved resolution for the $^{67}$Zn and $^{25}$Mg nuclei that results from reduced second-order quadrupolar interactions, which scale inversely with the strength of the high applied magnetic field. The experiments were conducted at room temperature at 19.6T ($^1$H resonance frequency of 830MHz), which for $^{67}$Zn ($I=\frac52$, 4.1% natural abundance, *ca.* 1.5 receptivity relative to $^{13}$C) and $^{25}$Mg ($I=-\frac52$, 10.1% natural abundance, *ca.* 0.7 receptivity relative to $^{13}$C) corresponded to Larmor frequencies of 51.88MHz and 50.76MHz, respectively. All of the NMR spectra presented here were recorded on a single-resonance 4-mm probehead under MAS conditions at 10kHz, using a Hahn-echo (*i.e.* $\frac{\piup}{2}$-$\tauup$-$\piup$-$\tauup$-acquisition) with the delays $\tauup$ set to one rotor period $\tauup_R$. The $^{67}$Zn and $^{25}$Mg shifts were referenced to the bulk external secondary standards ZnSe (274ppm relative to 1.0M *aq.* Zn(NO$_3$)$_2$) and MgO (26ppm relative to 3.0M *aq.* MgSO$_4$)[@Dupree], respectively. Pulse lengths of 2$\muup$s and 4$\muup$s were used for the $\frac{\piup}{2}$ and $\piup$ pulses, respectively. A recycle delay of 1s was used in each case, using 16,000 transients for the $^{67}$Zn MAS spectra (*ca.* 5h each), and 160,000 transients for the $^{25}$Mg MAS experiments (*ca.* 46h each). Second-order quadrupolar MAS lineshapes were fitted using the program <span style="font-variant:small-caps;">DMfit</span>.[@Massiot]
Time-of-flight neutron diffraction data for ZnO and were obtained on the neutron powder diffractometer NPDF at the Lujan Neutron Scattering Center at Los Alamos National Laboratory. For ease of structural modeling in the PDF analysis of , we chose a composition of $x=\frac18$ and have used an appropriate 16-atom supercell model of Zn$_7$MgO$_8$. For each sample, $\approx$2g of powder were packed in a vanadium can, and the data were collected for 3h at 25$^{\circ}$C using four detector banks located at 46$^{\circ}$, 90$^{\circ}$, 119$^{\circ}$, and 148$^{\circ}$. Control runs, for the intensity corrections, employed an empty vanadium can, a vanadium piece, and air. The program <span style="font-variant:small-caps;">PDFgetN</span> was used to extract the PDF $G(r)$ from the raw scattering data.[@Peterson] First, the coherent scattering intensity $I(Q)$ was obtained from raw data by the intensity corrections for container, background, and incident beam. Then, the $I(Q)$ was converted to the structure factor $S(Q)$ by the corrections for sample absorption, multiple scattering, and inelasticity. Finally, the PDF $G(r)$ was constructed by the Fourier transform of the reduced structure factor $F(Q)=Q[S(Q)-1]$. In the Fourier transform of $F(Q)$ to $G(r)$, $Q$ data were terminated at 35Å$^{-1}$. PDF refinements were performed using the software program <span style="font-variant:small-caps;">PDFgui</span>.[@Farrow] The average crystal structures of ZnO and were determined by the Rietveld method using the <span style="font-variant:small-caps;">GSAS-EXPGUI</span> software suite.[@Larson; @Toby]
Results and discussion
======================
Raman spectroscopy
------------------
Room temperature Raman spectra for ($x=0$, 0.05, 0.10, and 0.15) are shown in Fig.\[fig:Ramanwide\], along with mode assignments for the observed peaks.[@Calleja; @Damen] The wurtzite lattice, with space group $C_\mathrm{6v}^4$ (Hermann-Mauguin symbol $P6_3mc$), has four Raman-active phonon modes, $A_1+E_1+2E_2$. The two $E_2$ modes are nonpolar, while the $A_1$ and $E_1$ modes are polarized along the $z$-axis and in the $xy$-plane, respectively.[@Calleja; @Damen; @Zhang] The polar modes are further split into longitudinal (LO) and transverse (TO) components due to the macroscopic electric field associated with the LO modes. Raman spectra of the ZnO and compounds are dominated by two intense peaks of $E_2^\mathrm{low}$ ($\approx$98cm$^{-1}$) and $E_2^\mathrm{high}$ ($\approx$436cm$^{-1}$) modes. Raman scattering efficiencies of individual modes in ZnO are known to vary with the excitation energy.[@Calleja] With 514.5nm (2.41eV) excitation, the highest Raman efficiencies are observed from $E_2^\mathrm{high}$ and $E_2^\mathrm{low}$ modes. However, the polar LO modes exhibit a strong resonance effect as the excitation energy approaches the electronic transition energies. In cases when ultraviolet lasers are used for excitation, the Raman spectra of ZnO or are dominated by the signals from LO modes.[@Cheng]
![Raman spectra recorded at room temperature for ($x$ = 0, 0.05, 0.10, and 0.15 from top to bottom). The excitation wavelength used was 514.5nm.[]{data-label="fig:Ramanwide"}](Fig1.jpg){width="8cm"}
![Line profile fittings for (a) $E_2^\mathrm{low}$ and (b) $E_2^\mathrm{high}$ Raman peaks of (observed, gray circles; fit, solid lines; $x$ = 0, 0.05, 0.10, and 0.15 from top to bottom). Simplified lattice vibrations are depicted using large open circles for cations (Zn, Mg) and small gray circles for oxygen.[]{data-label="fig:Ramanfit"}](Fig2.jpg){width="8cm"}
In order to quantify composition-dependent changes in the Raman spectra of , the $E_2^\mathrm{low}$ and $E_2^\mathrm{high}$ line profiles were analyzed by least-squares fitting with standard peak functions. From the fits, peak position ($\omegaup_0$) and the full-width-at-half-maximum (FWHM, $\Gamma$) linewidths were determined. The profile fittings of both $E_2$ modes are shown in Fig.\[fig:Ramanfit\], and the evolution of peak parameters $\omegaup_0$ and $\Gamma$ with Mg-substitution $x$ are plotted in Fig.\[fig:RamanEW\]. Consistent with the previous study,[@Kim2] the $E_2^\mathrm{high}$ peaks of are best represented by Lorentzian Breit-Wigner-Fano (BWF) lineshapes with some asymmetry. However, the $E_2^\mathrm{low}$ line profiles do not fit as well to the Lorentzian-type functions. Instead, the $E_2^\mathrm{low}$ lines of are found to be pseudo-Voigt-type with predominantly Gaussian components of 97%, 83%, 77%, and 81% for $x$ = 0, 0.05, 0.10, and 0.15, respectively.
![(a) Raman shifts $\omegaup_0$ and (b) linewidths $\Gamma$ for the $E_2^\mathrm{low}$ (open circles) and $E_2^\mathrm{high}$ (filled circles) modes of .[]{data-label="fig:RamanEW"}](Fig3.jpg){width="8cm"}
As shown in Fig.\[fig:RamanEW\]a, the two $E_2$ modes of exhibit distinct dependences of phonon energy on the composition. With increased concentration of magnesium, the $E_2^\mathrm{low}$ mode shows a blueshift, while the $E_2^\mathrm{high}$ mode exhibits a redshift. For explaining these opposing trends, the vibrational eigenvectors of the wurtzite $E_2$ modes need to be considered. As shown in Fig.\[fig:Ramanfit\], both $E_2^\mathrm{low}$ and $E_2^\mathrm{high}$ modes are associated with atomic motions in the $ab$-plane. The lower energy branch corresponds mainly ($\approx$85%) to the vibrations of heavier components (cations, in case of ), and conversely the higher energy one corresponds mainly to those of lighter components (oxygen). Consequently, the $E_2^\mathrm{low}$ mode energy of is explicitly affected by the cationic substitution, according to the reduced mass effect. Replacement of Zn with Mg will decrease the reduced mass of the oscillator and in turn increase the phonon energy. By comparison, the change of cation mass should have less influence on the $E_2^\mathrm{high}$ mode energy, and its redshift is adequately attributed to the phonon softening caused by the in-plane lattice expansion. As previously reported, the lattice constant $a$ of increases monotonically with $x$,[@Kim1] which appears to account for the observed $E_2^\mathrm{high}$ mode behavior.
Peak shapes and linewidths of Raman spectra are dictated by the underlying line-broadening mechanisms. Gaussian line broadening is intrinsic to Raman spectra and originates from instrumental resolution. Also, if the sample being examined contains inhomogeneous components with different phonon frequencies, the resulting Raman lineshape would have a Gaussian character, reflecting the statistical nature of the spectrum. On the other hand, Lorentzian line broadening is mostly attributed to finite phonon lifetimes ($t$), according to the energy-time uncertainty relationship $\Gamma/\hbar=1/t$. The phonon lifetime shortening, which will cause linewidth broadening, can occur *via* two independent phonon decay mechanisms; by anharmonic decay into other Brillouin zone phonons or by phonon disruption at crystal defects. The latter, involving lattice imperfections, are quite common for alloy and solid-solution systems. Specifically for , it is expected that Mg-substitution will substantially reduce the size of ordered domains, and thereby block the long-range propagation of zone center phonons.
As noted above, the $E_2^\mathrm{low}$ and $E_2^\mathrm{high}$ lines of ZnO are respectively fitted with Gaussian and Lorentzian functions. It indicates that the $E_2^\mathrm{high}$ mode of ZnO has anharmonic phonon decay channels, whereas the $E_2^\mathrm{low}$ phonon does not. The Lorentzian line broadening behavior of $E_2^\mathrm{high}$ mode, in Fig.\[fig:RamanEW\]b, further manifests that the Mg-substitution increasingly populates the phonon field of with defect centers. The $E_2^\mathrm{low}$ mode also undergoes a gradual line broadening with $x$, and its peak shape changes upon Mg-substitution. While unsubstituted ZnO displays a nearly pure Gaussian peak for the $E_2^\mathrm{low}$ mode, the peaks from ($x>0$) samples are found to contain $\approx$20% of Lorentzian components. Therefore we conclude that, for the system, the same line broadening mechanism applies to both $E_2^\mathrm{low}$ and $E_2^\mathrm{high}$ modes: phonon lifetime shortening by increased crystal defects upon Mg substitution.
Solid-state $^{67}$Zn and $^{25}$Mg NMR spectroscopy
----------------------------------------------------
A different view of the disorder induced by Mg-substitution in ZnO materials can be obtained from solid-state NMR, using the NMR-active isotopes $^{25}$Mg (spin $I=-\frac52$) and $^{67}$Zn ($I=\frac52$). These nuclei are challenging to observe, because their low natural abundances and low gyromagnetic ratios (see Experimental) yield low signal sensitivities, and because second-order quadrupolar interactions associated with nuclear spins $I$ $>$ $\frac12$ can lead to a dramatic broadening of their NMR spectra. Second-order quadrupolar effects, however, scale with the inverse of the static magnetic field $B_0$, and are therefore mitigated at high magnetic fields, which also increase signal sensitivity.
![Solid-state spin-echo $^{67}$Zn MAS NMR spectra acquired at 19.6T, room temperature, and 10kHz MAS: (a) ZnO and (b) Zn$_{0.85}$Mg$_{0.15}$O powders, both prepared at 900$^{\circ}$C in N$_2$. Accompanying the ZnO spectrum in (a) is a fit (dashed line) of the experimental lineshape to a simulated second-order quadrupolar powder pattern.[]{data-label="fig:nmrZn"}](Fig4.jpg){width="8cm"}
Solid-state $^{67}$Zn MAS NMR spectra acquired at 19.6T of ZnO and Zn$_{0.85}$Mg$_{0.15}$O are shown in Fig.\[fig:nmrZn\]. The materials were prepared by heating at 900$^{\circ}$C in N$_2$ and yield similar spectra that clearly exhibit scaled second-order quadrupolar powder MAS lineshape features.[@Kundla] Associated values for the isotropic $^{67}$Zn chemical shift, $\deltaup_\mathrm{iso}$, the quadrupolar coupling constant, $\nuup_\mathrm{Q}$, and the asymmetry parameter, $\etaup_\mathrm{Q}$, can be determined by fitting each resolved second-order quadrupolar MAS lineshape. A simulated powder pattern is shown as the dashed line in Fig.\[fig:nmrZn\]a accompanying the $^{67}$Zn MAS spectrum of ZnO and corresponding to the calculated values $\deltaup_\mathrm{iso}$ = 236ppm, $\nuup_\mathrm{Q}$ = 378kHz, and $\etaup_\mathrm{Q}$ = 0.1. Several discrepancies between the calculated and experimental powder patterns may be due to modest anisotropy of the powder particles and/or to uneven excitation of the different crystallite orientations by the radiofrequency pulses. The powder pattern and fit parameters are nevertheless consistent with a single crystallographically distinct and relatively symmetric $^{67}$Zn site in the wurtzite ZnO structure. The $^{67}$Zn MAS spectrum of Zn$_{0.85}$Mg$_{0.15}$O in Fig. 4b exhibits a similar lineshape and features, which are broadened by the presumably random incorporation of Mg atoms within the Zn-rich wurtzite lattice.
The close similarity between the $^{67}$Zn MAS NMR spectra of ZnO (Fig.\[fig:nmrZn\]a) and Zn$_{0.85}$Mg$_{0.15}$O (Fig.\[fig:nmrZn\]b) indicates that the substitution of 15% of Zn atoms by Mg has little influence on the local electronic environments of a majority of the $^{67}$Zn nuclei. Assuming that Mg atoms are distributed randomly, the probability of having no Mg atom in the second coordination sphere of a given $^{67}$Zn nucleus is only 14% (the first coordination sphere consisting of four O atoms). Significant effects on the $^{67}$Zn spectrum would be expected if the quadrupolar interactions were substantially perturbed by the substitution of Mg atoms for 1, 2, 3, 4 (probabilities of 30%, 29%, 17%, and 7%, respectively) or more of the 12 Zn atoms in the second coordination sphere. However, despite the high probability of modifications of the local chemical environment of the $^{67}$Zn nuclei upon 15% Mg-substitution, dramatic changes of the $^{67}$Zn spectrum are not observed, rather only modest broadening that reflects increased disorder. Under these conditions, the differences induced in ZnO or Zn$_{0.85}$Mg$_{0.15}$O by changing the synthesis atmosphere (*e.g.*, air, O$_2$, or N$_2$) and temperature (over the range 550$^{\circ}$C to 900$^{\circ}$C) were too subtle to be clearly reflected in the corresponding $^{67}$Zn MAS NMR spectra (data not shown), which are similar to those in Fig.\[fig:nmrZn\]a, b. Thus, it appears that $^{67}$Zn NMR measurements of these ZnO-related materials are sensitive principally to substitutions that result in modifications of the first coordination sphere of the zinc nuclei (substitution of the anionic species), or to substitutions resulting in larger lattice distortions or increased electronic disorder than observed here.[@Toberer]
![Spin-echo $^{25}$Mg MAS NMR spectra, recorded at 19.6T, of (a) polycrystalline MgO powder (external reference for $^{25}$Mg NMR shift), (b) Zn$_{0.85}$Mg$_{0.15}$O powder prepared at 550$^{\circ}$C in air, and (c) Zn$_{0.85}$Mg$_{0.15}$O powder prepared at 900$^{\circ}$C in N$_2$.[]{data-label="fig:nmrMg"}](Fig5.jpg){width="8cm"}
Solid-state $^{25}$Mg MAS NMR measurements complement the $^{67}$Zn NMR results by being sensitive to the structural and electronic perturbations experienced directly by Mg atom incorporated into ZnO lattices. The $^{25}$Mg MAS NMR measurements were similarly conducted at 19.6T and the resulting spectra are shown in Fig.\[fig:nmrMg\] for polycrystalline MgO and different Zn$_{0.85}$Mg$_{0.15}$O powders prepared at 550$^{\circ}$C in air and at 900$^{\circ}$C in N$_2$. The $^{25}$Mg MAS spectrum in Fig.\[fig:nmrMg\]a of MgO shows a narrow well-defined $^{25}$Mg peak that reflects the highly symmetric coordination environment of the single type of Mg site in its rock-salt structure.[@MacKenzie1] By comparison, the $^{25}$Mg MAS spectra of the different Zn$_{0.85}$Mg$_{0.15}$O powders contain broad and relatively unstructured lineshapes centered at *ca.* $-$30ppm to $-$40ppm that reflect broad distributions of signal intensity and thus broad distributions of local $^{25}$Mg environments. The first Mg coordination sphere (MgO$_4$) in Zn$_{0.85}$Mg$_{0.15}$O is expected to be significantly distorted from its regular tetrahedral geometry, and the second coordination sphere, composed of Mg(Zn,Mg)$_{12}$, is expected to have an even larger number of different local configurations, bond distances, and/or bond angles. The $^{25}$Mg MAS NMR spectra appear to be more sensitive than the $^{67}$Zn NMR results to local material environments in Zn$_{0.85}$Mg$_{0.15}$O. However, the low signal-to-noise and broad lines observed for natural abundance $^{25}$Mg in Zn$_{0.85}$Mg$_{0.15}$O, even at 19.6T, preclude a detailed and reliable analysis of the isotropic shift distributions and second-order quadrupolar broadening that would be necessary to extract directly quantitative information on nature of the Mg site distributions and/or their disorder. It is noteworthy that the Zn$_{0.85}$Mg$_{0.15}$O powders do not display any spectral feature corresponding to bulk MgO, thereby ensuring the phase purity of ($x\,\leqslant\,0.15$) solid solutions, within the sensitivity limit of the $^{25}$Mg NMR measurements.[@MacKenzie2]
In common for $^{29}$Si, $^{27}$Al, and $^{25}$Mg, an increase of the coordination number for a given cationic center counts for enhanced local shielding, resulting in an upfield displacement of the signal(s) to lower shift values.[@Dupree; @Magi] The spinel MgAl$_2$O$_4$ containing MgO$_4$ fragments exhibits a $^{25}$Mg shift of 52ppm, compared to 26ppm for MgO which is comprised of MgO$_6$ local units.[@Dupree] By comparison, clay minerals in which Mg atoms are coordinated by six (O,OH) ligands yield $^{25}$Mg shifts in the range of 0 to $-$100ppm.[@MacKenzie1] The wurtzite oxide Zn$_{0.85}$Mg$_{0.15}$O has a 4-coordinated geometry of Mg, similar to MgAl$_2$O$_4$, but the former clearly shows a more upfield-shifted $^{25}$Mg resonance at $-30\sim-$40ppm. We infer that $^{25}$Mg MAS NMR signals are considerably influenced by the ligand identity beyond the first coordination shell, and that Zn has a far stronger shielding contribution than Al, Si, or alkaline-earths, thus accounting for the observed resonance values.
Neutron diffraction
-------------------
In previous studies, the crystal structure of ZnO has been reported \[space group $P6_3mc$, Zn at $(\frac13\frac230)$ and O at $(\frac13\frac23u)$\] with lattice constants of $a=$ 3.2427$\sim$3.2501Å and $c=$ 5.1948$\sim$5.2071Å, and atomic position parameter $u=$ 0.381$\sim$0.3826.[@Sawada; @Sabine; @Abrahams; @Harrison; @Albertsson; @Kisi] In our previous synchrotron x-ray study on polycrystalline ZnO, we obtained $a=$ 3.2503Å, $c=$ 5.2072Å, and $u=$ 0.3829.[@Kim1] Here, we reexamine the crystal structures of polycrystalline ZnO and Zn$_{0.875}$Mg$_{0.125}$O utilizing neutron scattering, which has several advantages over x-rays. The coherent neutron scattering length of O (5.804fm) is comparable to those of Zn (5.680fm) and Mg (5.375fm),[@Sears] as distinct from x-ray diffraction, where scattering is strongly weighted by Zn. Moreover, neutron scattering can provide high $Q$ diffraction data with much less attenuation than x-rays, due to there being no fall-off with $Q$ in the form factor.
![Rietveld refinement of time-of-flight powder neutron diffraction profiles of (a) ZnO and (b) . Calculated patterns (gray lines) are superimposed on observed data (open circles), with Bragg positions and difference profiles at the bottom.[]{data-label="fig:Riet"}](Fig6.jpg){width="8cm"}
ZnO Zn$_{0.875}$Mg$_{0.125}$O
----------------------------------------------- ----------------------- ---------------------------
$V$ (Å$^3$) 47.603(1) 47.578(1)
$a$ (Å) 3.24945(1) 3.25058(2)
$c$ (Å) 5.20574(3) 5.19771(3)
$c/a$ 1.60204 1.59940
$u$ 0.38275(5) 0.38214(6)
$U_\mathrm{iso}$(Zn/Mg) (Å$^2$) 0.00648(8) 0.0065(2)
$U_\mathrm{iso}$(O) (Å$^2$) 0.0101(1) 0.0087(2)
$d_{M-\mathrm{O}}$ (Å) (1$\times$) 1.9925(3) (1$\times$) 1.9868(3)
(3$\times$) 1.9729(1) (3$\times$) 1.9743(1)
$\angle_{\mathrm{O}-M-\mathrm{O}}$ ($^\circ$) 108.022(7) 108.081(9)
110.881(7) 110.825(9)
: Structural parameters for ZnO and Zn$_{0.875}$Mg$_{0.125}$O determined from the Rietveld refinement of time-of-flight neutron diffraction pattern.
\[tab:Riet\]
![Average wurtzite parameters $c/a$ and $u$ of ZnO and determined by Rietveld refinements of neutron (diamonds) and synchrotron x-ray (squares, ref.[@Kim1]) diffraction patterns. Note the significantly smaller error bars that emerge from the neutron study. The inset shows the data observed from existing wurtzites (see ref.[@Kim1] for details).[]{data-label="fig:ca_u"}](Fig7.jpg){width="8cm"}
Rietveld refinements of ZnO and Zn$_{0.875}$Mg$_{0.125}$O were carried out on four histograms of data collected at detector locations of 46$^{\circ}$, 90$^{\circ}$, 119$^{\circ}$, and 148$^{\circ}$. The highest-$Q$ data from the 148$^{\circ}$ detector covers the range of $d>0.25\,\AA$ ($Q<25\,\AA^{-1}$). The model used for the refinement in space group $P6_3mc$ has Zn/Mg at the $2b$ Wyckoff position ($\frac13\frac230$) and O also at 2$b$ position ($\frac13\frac23u$). For Zn$_{0.875}$Mg$_{0.125}$O, Zn and Mg were statistically distributed over the common site with fractional occupancies fixed to the respective compositions. The refinement converged with reliability factors of $R_\mathrm{wp}=2.57\,\%, R_\mathrm{p}=1.80\,\%$, and $\chiup^2=2.71$ for ZnO, and $R_\mathrm{wp}=2.89\,\%, R_\mathrm{p}=2.04\,\%$, and $\chiup^2=3.56$ for Zn$_{0.875}$Mg$_{0.125}$O. The Rietveld refinement profiles and the structural parameters are given in Fig.\[fig:Riet\] and Table\[tab:Riet\], respectively. The present refinement results agree well with the findings from our previous synchrotron x-ray study.[@Kim1] Both studies indicate that Mg-substitution expands the $a$-parameter, compresses the $c$-parameter, and decreases the oxygen position parameter $u$, as summed up in Fig.\[fig:ca\_u\]. The two data sets from x-ray and neutron reveal an identical trend in the structural evolution in terms of $a$, $c$, and $u$.
![Comparison of high-$Q$ scattering data from ZnO measured using synchrotron x-ray and pulsed-neutron radiation: (a,b) $F(Q)=Q[S(Q)-1]$, and (c,d) $G(r)$. In the Fourier transform of $F(Q)$ to $G(r)$, $Q$ data were terminated at 28Å$^{-1}$ for x-rays, and 35Å$^{-1}$ for neutrons.[]{data-label="fig:SFG"}](Fig8.jpg){width="8cm"}
For the composition range up to $x=0.15$, the solid solutions remain isostructural with wurtzite ZnO, with both Zn and Mg tetrahedrally coordinated by O atoms. However, since the two cations Zn and Mg have clearly distinct crystal chemistries in oxidic environments, it is conjectured that they may have distinct local geometries within lattices. In order to directly address this question, we have carried out neutron PDF analyses of Zn$_{0.875}$Mg$_{0.125}$O and ZnO. Figure\[fig:SFG\] shows the PDF data for ZnO, comparing the present neutron data with the x-ray data previously measured at Beamline 11-ID-B of the Advanced Photon Source at Argonne National Laboratory. As can be seen from Figs.\[fig:SFG\]a and \[fig:SFG\]b, the neutron study provides scattering information over a significantly higher $Q$ range and also with higher resolution. Correspondingly, the neutron PDF $G(r)$ can be obtained for far wider $r$-ranges than when using x-rays (Fig.\[fig:SFG\]c). Given this, neutrons are expected to enable better description of the alloy structure(s).
![(a) Probability for a cation in to have 0, 1, 2, 3, 4 or 5 Mg neighbors in the 12-membered coordination shell ($r\approx$ 3.25Å), and (b) mean separation between the nearest Mg$-$Mg pairs.[]{data-label="fig:Mgdist"}](Fig9.jpg){width="8cm"}
In order to obtain an appropriate real-space model for the PDF analysis of Zn$_{0.875}$Mg$_{0.125}$O, we estimate as follows, the impact of Mg-substitution on the chemical environment within the ZnO lattice. Cations in the wurtzite lattice form a hexagonal close-packed sublattice, where each cation has 12 nearest neighbors. At the Mg-substitution level of $x=$ 0.125, it is estimated that $\approx$80% of the cations have at least one Mg neighbor in the cation sublattice (Fig.\[fig:Mgdist\]). Each cation (whether Zn or Mg) has as neighbors, 1.5 Mg atoms on average, and the mean Mg$-$Mg distance is as short as 6.5Å. Therefore the coordination geometry of MgO$_4$ will serve a substantial structural factor in the PDF of and needs to be treated with independent parameters in the PDF refinement.
![(Color online) Zn$_7$MgO$_8$ supercell model structure used for neutron PDF analyses of , and coordination environments for three distinct Zn sites therein. In the top part, unlabeled oxygen atoms belong to the O3 type (see text).[]{data-label="fig:supercell"}](Fig10.jpg){width="8cm"}
On the basis of the above considerations, the structural model for was designed as follows. To distinguish the Mg and Zn atoms in the unit cell, while maintaining the composition, a 2$\times$2$\times$1 supercell of the primitive wurtzite structure was selected, with the cell parameters $a=b\approx$ 6.5Å, $c\approx$ 5.205Å, $\alphaup=\betaup=$ 90$^{\circ}$, $\gammaup=$ 120$^{\circ}$, and the unit content Zn$_7$MgO$_8$. In the average structure scheme, in which the atomic positions are set as ($\frac13\frac230$) for cations and ($\frac13\frac23u$) for anions, the external and internal geometries of all the tetrahedra are uniformly defined by the $c/a$ ratio and $u$, respectively. However in the supercell model, we allowed the MgO$_4$ unit to have its own coordination geometry, irrespective of the ZnO$_4$ geometry. As a result, the eight O atoms in the supercell unit are divided into three groups; one O atom bonded to Mg along the $c$-axis (O1), three O atoms bonded to Mg laterally (O2), and the remaining four O atoms (O3). The seven Zn atoms are also grouped according to their proximities to the Mg atom; three Zn atoms having an apical O2 oxygen atom (Zn1), three Zn atoms bonded to O2 atoms laterally (Zn2), and one Zn bonded to only O3 atoms (Zn3). Figure\[fig:supercell\] illustrates the supercell configuration and the three different types of ZnO$_4$ geometries.
The PDF of ZnO was analyzed using a simple wurtzite cell (Zn$_2$O$_2$) obeying the symmetry requirements of the $P6_3mc$ space group. However, for the supercell refinement of , several symmetry constraints were lifted from the average structure description. The $z$-coordinates of the O1, O2, and O3 types of oxygen atoms were independently varied, and the $z$-coordinate of Mg was also refined. We also attempted in-plane displacements of the O2 atoms, but found that these were unstable in the refinement. Therefore, the refinement for included three more position parameters than those for ZnO. For both ZnO and , the refinements of $G(r)$ used the variables of lattice constants, atomic position parameters, isotropic temperature factors, structure scale factor, and peak sharpening coefficients.
![Neutron PDF refinements for (a) ZnO and (b) , parts of the 1.2$-$22Å fittings. Calculated PDFs (gray lines) are superimposed on experimental data (open circles), with the differences shown at the bottom of each.[]{data-label="fig:pdffit"}](Fig11.jpg){width="8cm"}
Figure\[fig:pdffit\] shows the neutron PDF refinements for ZnO and , carried out in the $r$ range of 1.2Å to 22Å. The refinement parameters are summarized in Table\[tab:PDF\]. For both structures, the refinements were achieved with satisfactorily low $R_\mathrm{w}$’s of 7.9% (ZnO) and 8.3% (Zn$_{0.875}$Mg$_{0.125}$O). In comparison with the Rietveld results, the PDF-refined lattice constants are systematically larger by $\approx$0.06%, but good agreements are observed for the $c/a$ ratios. Also, the oxygen atom position in ZnO is reasonably well reproduced from the PDF and Rietveld refinements. An interesting finding emerges from the atomic coordinates of Mg and its adjacent oxygen atoms (O1 and O2) in . As found in Table\[tab:PDF\], the three groups of O atoms have clearly distinct $z$-coordinates from one another. Both the O1 and O2 groups are vertically shifted towards their respective Mg atoms. Mg itself also moves towards the basal plane formed of O2 atoms. The resulting MgO$_4$ unit has a squashed tetrahedral geometry (point group $C_\mathrm{3v}$), with a much shorter height than those of ZnO$_4$ tetrahedra. However the O3 atoms, which are not directly bonded to Mg, have similar $z$-coordinates to those found in ZnO. Therefore, the (Zn3)(O3)$_4$ tetrahedra retain the unperturbed geometry of ZnO. The other types of ZnO$_4$ tetrahedra share corners with MgO$_4$ and are likely to undergo intermediate distortion.
ZnO Zn$_{0.875}$Mg$_{0.125}$O
--------------------------------- --------------------- ---------------------------
$a$ (Å) 3.2514(3) 3.2527(4) $\times$2
$c$ (Å) 5.2089(6) 5.2015(8)
$c/a$ 1.6020 1.5991
$z$(O) 0.3824(4)/0.8824(4) 0.3825(8)/0.8825(8) (O3)
0.386(3) (O2)
0.858(5) (O1)
$z$(Mg) 0.486(8)
$U_\mathrm{iso}$(Zn/Mg) (Å$^2$) 0.0086(8) 0.0063(6)
$U_\mathrm{iso}$(O) (Å$^2$) 0.0070(6) 0.0081(8)
: Structural parameters from the neutron PDF refinements of ZnO and , over the $r$ range of 1.2$-$22Å.
\[tab:PDF\]
![(a) Mg$-$O and Zn$-$O bond distances in , as analyzed by the neutron PDF refinement of a Zn$_7$MgO$_8$ supercell model ($c$-axial bonds, circles; lateral bonds, diamonds), and (b) bond valence sums for each cation site. Horizontal lines indicate the Zn$-$O bond distances (in the upper panel) and BVS for Zn (in the lower panel) in ZnO.[]{data-label="fig:bondl"}](Fig12.jpg){width="8cm"}
To examine the presence of short-range structural characteristics, the PDF refinements were carried out from 1.2Å to different values of $r_\mathrm{max}$ between 10$-$30Å. The lattice constants and position parameters showed insignificant variations with $r_\mathrm{max}$, and the $R_\mathrm{w}$ values had shallow minima at $r_\mathrm{max}\approx22\,\AA$ for both ZnO and . The Mg$-$O and Zn$-$O bond distances plotted in Fig.\[fig:bondl\] represent the averaged results of multiple refinements over 1.2$-r_\mathrm{max}$(Å), where $r_\mathrm{max}$ was increased from 10 to 30Å in 2Å steps. From Fig.\[fig:bondl\]a, we find a meaningful distinction between the MgO$_4$ and ZnO$_4$ bonding structures. The Mg$-$O bonds are shorter than the Zn$-$O bonds, and the former set consists of one short and three longer bonds, in contrast to the latter. These results are consistent with the greater sensitivity of the $^{25}$Mg NMR spectra (Fig.\[fig:nmrMg\]b, c) to Mg substitution into ZnO-rich lattices, compared to the $^{67}$Zn NMR results (Fig.\[fig:nmrZn\]). For Mg and the three Zn atom types in , bond valence sums (BVS)[@Brown1] are evaluated, as shown in Fig.\[fig:bondl\]b. The valence sums for all of the cation and anion sites in are within a reasonable range around the ideal magnitude of 2 valence units (v.u.). The relatively underbonded situation for Zn1 atoms and the slight overbonding of the Mg atom are the results of a rather drastic shift of the bridging O1 atom. The global instability index (GII)[@Brown1] of is calculated as 0.116v.u., which is significantly larger than that of ZnO (0.067v.u.). The GII value is often used as a measure of the residual bond strain, and is known not to exceed 0.2v.u. for ordinary structures in the standard state.[@Brown1; @Brown2] It is worth mentioning, in light of the above GII considerations, that the increased microstrain upon Mg-substitution results in characteristic optical and x-ray line broadening behavior of that we have previously reported.[@Kim3; @Kim4]
The solid solution of has the thermodynamic solubility limit of $x\approx$ 15% on the ZnO-rich side. Obviously the distinct coordination preferences of Mg and Zn prevent the formation of a continuous solid solution across the entire composition range, $x$. While Mg has a point symmetry of $O_\mathrm{h}$ in its binary oxide MgO, wurtzite ZnO possesses only a $C_\mathrm{3v}$ environment. Here, we rationalize the neutron PDF result for the MgO$_4$ tetrahedral geometry, in several ways. Both bond valence[@Brese] and ionic radii[@Shannon] compilations are useful to explain the smaller tetrahedral volume of MgO$_4$, as compared with ZnO$_4$. The bond valence parameter of Mg$-$O ($R_0=0.693\,\AA$) is smaller than that of Zn$-$O ($R_0=0.704\,\AA$), implying that in general Mg$-$O bonds are shorter than Zn$-$O bonds, for the same coordination numbers. Similarly, the 4-coordinate ionic radius of Mg$^{2+}$ (0.57Å) is smaller than that of Zn$^{2+}$ (0.60Å). In order for the MgO$_4$ tetrahedron to have a smaller volume than ZnO$_4$, either the tetrahedral height or the base area should shrink, or an isotropic volume change could occur. From the PDF analysis, the $c$ parameter around Mg showed a clear decrease, whereas the $a$ parameter appears to be nearly unchanged. The $c$-axial parameter contraction of MgO$_4$ unit can be deduced also from the composition-dependent changes of the average structure of .
![$u$-dependent bond valence sums for Mg and Zn in the tetrahedral spaces of their oxides. The dimensions of MgO$_4$ and ZnO$_4$ tetrahedra are from the neutron PDF refinements of and ZnO, respectively. For $u$ near 0.5, a fifth oxygen atom was added to the coordination shell to account for the pseudo-bipyramidal configuration.[]{data-label="fig:calcbvs"}](Fig13.jpg){width="8cm"}
The variable $u$ is the only position parameter in the wurtzite structure. It corresponds to the ratio of the apical bond distance to the $c$ length, $d_\mathrm{apical}/c$, or $d_\mathrm{apical}/2h$, where $h$ is the height of the tetrahedron. For , $u$(Mg) is determined to be 0.395, a value that finds no correspondence to the experimental structures of pure wurtzites. Existing wurtzites have $u$ values within the range of 0.374 to 0.383. Such a high $u$ found for MgO$_4$ is closely linked with its extremely small aspect ratio ($2h/a=1.509$, as compared to the ideal value 1.633). To compare the potential fields within the tetrahedral spaces of MgO$_4$ and ZnO$_4$, the valence sums of Mg and Zn were calculated as functions of $u$ (Fig.\[fig:calcbvs\]). In ZnO, the tetrahedral cavity is slightly oversized for Zn, and the BVS(Zn) has a minimum of $\approx$1.9v.u. at $u\approx$ 0.382. The potential minimum suggested by BVS(Zn) is in a good agreement with the experimental $u$ for ZnO. However, the MgO$_4$ shell in is rather small for Mg, resulting in a global overbonding situation. The BVS(Mg) is minimized at a markedly higher $u$ range than for BVS(Zn), making it clear why the experimental $u$(Mg) is conspicuously larger than $u$(Zn).
Several authors have used first-principles density functional theory (DFT) to predict the relaxed structure of hypothetical wurtzite MgO.[@Limp; @Gopal; @Janotti] An earlier DFT study attempted to relax the energy-minimized wurtzite MgO structure, but only to obtain a hexagonal structure corresponding to the limit of $c$-axis compression ($c/a=1.20$, $u=0.5$).[@Limp] However in later studies, the energy minimizations for wurtzite MgO were achieved within appropriate boundary conditions; $c/a=1.514$, $u=0.398$ by Janotti $et\,al.$[@Janotti], and $c/a=1.520$, $u=0.395$ by Gopal and Spaldin[@Gopal]. Therefore, wurtzite MgO, if it ever occurs, is expected to have an abnormally small $c/a$ ratio and large $u$ value, compared with common wurtzites including ZnO. Interestingly, the above computational approximations of the hexagonal MgO structure agree well with our PDF analyses of , which underscores that the MgO$_4$ fragment has smaller $c/a$ and larger $u$ than ZnO$_4$. Malashevich and Vanderbilt[@Mala1] and Fan $et\,al.$[@Fan] performed DFT computations of supercell models, also deriving consistent conclusions that $c/a$ decreases with Mg-substitution, and that $u$(Mg) $>$ $u$(Zn). The results from the neutron PDF refinement and the above DFT computations are plotted together in Fig.\[fig:MgO4\], in which the distinct internal/external geometries of MgO$_4$ and ZnO$_4$ are contrasted. As a general trend, isolated MgO$_4$ tetrahedra always have larger $u$ parameters than ZnO$_4$ units. The extremely good agreement between the *local* coordination of Mg in determined here from PDF studies and DFT calculations on end-member wurtzite MgO is remarkable. This is in part attributed to our use of the ratio $c/a$ as a comparison parameter. This ratio has fewer systematic errors than do the values of the lattice parameters taken separately.
![Aspect ratio $c/a$ and internal parameter $u$ for MgO$_4$ (filled squares) and ZnO$_4$ (filled circles) tetrahedra in , as analyzed by neutron PDF refinements. DFT results for wurtzite MgO and supercells are compared together: Open squares correspond to the MgO$_4$ geometry in hexagonal MgO (i: ref.[@Janotti] and ii: ref.[@Gopal]), and the dotted open squares and dotted open circles represent the MgO$_4$ and ZnO$_4$ geometries found in supercells[@Mala2], respectively. The gray line defines the border where the axial cation-anion bond along $c$ and three equatorial bonds become equidistant.[]{data-label="fig:MgO4"}](Fig14.jpg){width="8cm"}
The total polarization $P_\mathrm{s}$ of polar oxides can be broken down into electronic, ionic, and piezoelectric contributions, related respectively to the polarizability of component ions, ionic arrangements, and lattice parameters. Mg-substitution decreases the electronic polarization ($P_\mathrm{el}$) of ZnO, because of its greater ionic character, whereas the concomitant $c$-axis compression is expected to increase the piezoelectric polarization ($P_\mathrm{pz}$). Quantitative estimation of the above two contributions ($P_\mathrm{el}$, $P_\mathrm{pz}$) requires sophisticated theoretical analyses and cannot be provided here for ZnO and . Meanwhile, the ionic polarization ($P_\mathrm{ion}$) of can be approximated by a simple electrostatic model using the $c$-axial positions of Zn, Mg, and O. We calculated $P_\mathrm{s}$ of and ZnO, assuming point charges ($q=$+2 for Zn and Mg, and $-$2 for O), and using the atomic coordinates as determined by neutron PDF analysis (see Appendix). Separate reference structures were used for ZnO and , with the corresponding experimental $a$ and $c$ values, but with $u$ = 0.375, so that within isolated (Zn,Mg)O$_4$ moieties the net centers of the cation and anion charges coincide. The calculated $P_\mathrm{s}$ values (= $P_\mathrm{ion}$ here), along $c$, are $-5.18\,\muup\mathrm{C}/\mathrm{cm}^2$ for ZnO and $-5.25\,\muup\mathrm{C}/\mathrm{cm}^2$ for ($\Delta P_\mathrm{ion}=-0.07\,\muup\mathrm{C}/\mathrm{cm}^2$). For comparison, Malashevich and Vanderbilt[@Mala1] have employed the Berry-phase approach, to obtain electric polarization of a Zn$_5$MgO$_6$ ($x=\frac16$) supercell. They find $\Delta P_\mathrm{ion}=-0.22\,\muup\mathrm{C}/\mathrm{cm}^2$, $\Delta P_\mathrm{el}=+0.01\,\muup\mathrm{C}/\mathrm{cm}^2$, and $\Delta P_\mathrm{pz}=-0.81\,\muup\mathrm{C}/\mathrm{cm}^2$, relative to $P_\mathrm{s}(\mathrm{ZnO})=-3.22\,\muup\mathrm{C}/\mathrm{cm}^2$. In fact, the following relations hold true in general for other supercells ($x=\frac14, \frac13, \frac12$) studied by them: $\Delta P_\mathrm{el}>0$, $\Delta P_\mathrm{ion}<0$, $\Delta P_\mathrm{pz}<0$, and $\Delta P_\mathrm{el}$ $\ll$ $\Delta P_\mathrm{ion}$ $<$ $\Delta P_\mathrm{pz}$. We therefore expect that the total polarization of should be larger than that of ZnO, with the increases of piezoelectric and ionic polarizations dominating the decrease of the electronic contribution. We note that $P_\mathrm{s}$(ZnO) from Malashevich and Vanderbilt differs somewhat from the results of other DFT studies[@Gopal; @Corso; @Bernardini; @Noel] which reported $P_\mathrm{s}\mathrm{(ZnO)}=-5.0\sim-5.7\,\muup\mathrm{C}/\mathrm{cm}^2$. But since we are primarily interested in the relative portions of $P_\mathrm{el}$, $P_\mathrm{ion}$, and $P_\mathrm{pz}$ in the $P_\mathrm{s}$, and their changes with respect to the Mg-substitution in ZnO, we ignore the small discrepancies in the absolute values.
The result, an increase of $P_\mathrm{ion}$ upon Mg-substitution, differs from our previous estimates based solely on synchrotron x-ray diffraction analyses.[@Kim1; @Kim2] This inconsistency arises mostly from the improved models for the supercell structures that are permitted by the superior neutron scattering results. Figure\[fig:SFG\] clearly shows the greatly enhanced data quality of the neutron scattering measurements, compared with x-ray experiments. The neutron data lead to better precision (see Table\[tab:Riet\] and ref.[@Kim1]), as well as better accuracy, in the subsequent structure analyses.
It is also noteworthy that the structure modeling for the PDF or Rietveld analyses can confine the range of structural information that is sought. For example, Rietveld refinement of used only one positional parameter, and therefore MgO$_4$ and ZnO$_4$ tetrahedra are characterized by the same $c/a$ and $u$. In the x-ray PDF refinement, two position parameters were refined, so that we could differentiate $u$(Mg) and $u$(Zn), but not their $c/a$ ratios. However, for the neutron PDF refinement, we used four positional parameters as variables, and were able to distinguish the MgO$_4$ and ZnO$_4$ geometries through both $c/a$ and $u$. Again, the choice of a structural model relies on the quality of the diffraction data, for this application on a system with relatively low-atomic-number elements. The neutron data showed sufficient integrity for testing various structure models, which were not possible for the x-ray data alone.
Conclusion
==========
Detailed structural analyses of ($0\,\leqslant\,x\,\leqslant\,0.15$) solid solutions, focusing on the dissimilar local geometries of Mg and Zn, have been performed by using Raman scattering, $^{67}$Zn/$^{25}$Mg NMR spectroscopy, and neutron Rietveld/PDF analyses. Line broadenings of the Raman and $^{67}$Zn/$^{25}$Mg NMR spectra imply that Mg-substitution into a ZnO-rich lattice gradually increases the configurational disorder and crystal defects in the lattice. Rietveld refinement of time-of-flight neutron diffraction patterns and Raman peak profile analyses show that the macroscopic lattice distortions in wurtzites develop in a way such that the $c/a$ ratio decreases with Mg content $x$. A real-space neutron PDF analysis using a supercell model of Zn$_7$MgO$_8$ reveals that the Mg and Zn atoms in have markedly distinct local geometries: MgO$_4$ tetrahedra are smaller in height and volume and have a larger internal parameter $u$. The wurtzite structural parameters $c/a$ and $u$ for MgO$_4$ deviate from their ideal values, in agreement with computational predictions of wurtzite MgO structures. Previous DFT studies on supercells and the present neutron PDF analysis strongly suggest that Mg-substitution will increase the spontaneous polarization of ZnO.
ACKNOWLEDGEMENTS {#acknowledgements .unnumbered}
================
The authors acknowledge support from the National Science Foundation through the MRSEC program (DMR05-20415) and from the Department of Energy, Basic Energy Sciences, Catalysis Science Grant No. DE-FG02-03ER15467. This work has benefited from the use of NPDF at the Lujan Center at Los Alamos Neutron Science Center, funded by DOE Office of Basic Energy Sciences. Los Alamos National Laboratory is operated by Los Alamos National Security LLC under DOE Contract DE-AC52-06NA25396. The authors are grateful to the NSF-supported National High Magnetic Field Laboratory, in Tallahassee, Florida for access to the high-field (19.6T) NMR facilities, and to Zhehong Gan for assistance with the NMR measurements. Andrei Malashevich and David Vanderbilt kindly provided the DFT optimized structural data for supercells. The authors also thank Brent Melot and Daniel Shoemaker for the neutron data collection.
[99]{}
A. Tsukazaki, A. Ohtomo, T. Kita, Y. Ohno, H. Ohno, and M. Kawasaki, Science **315**, 1388 (2007).
K. Koike, K. Hama, I. Nakashima, G. Takada, M. Ozaki, K. Ogata, S. Sasa, M. Inoue, and M. Yano, Jpn. J. Appl. Phys. **43**, L1372 (2004).
S. Rajan, H. Xing, S. DenBaars, U. K. Mishra, and D. Jena, Appl. Phys. Lett. **84**, 1591 (2004).
S. Sasa, M. Ozaki, K. Koike, M. Yano, M. Inoue, Appl. Phys. Lett. **89**, 053502 (2006).
H. Tampo, H. Shibata, K. Matsubara, A. Yamada, P. Fons, S. Niki, M. Yamagata, and H. Kaine, Appl. Phys. Lett. **89**, 132113 (2006).
J. H. Davies, *The Physics of Low-Dimensional Semiconductors: An Introduction (Cambridge University Press, New York, 1998).*
Y.-I. Kim, K. Page, and R. Seshadri, Appl. Phys. Lett. **90**, 101904 (2007).
Y.-I. Kim, K. Page, A. M. Limarga, D. R. Clarke, and R. Seshadri, Phys. Rev. B **76**, 115204 (2007).
Th. Proffen, T. Egami, S. J. L. Billinge, A. K. Cheetham, D. Louca, and J. B. Parise, App. Phys. A **74**, S163 (2002).
D. A. Shirley, Phys. Rev. B **5**, 4709 (1972).
P. Thompson, D. E. Cox, and J. B. Hastings, J. Appl. Cryst. **20**, 79 (1987).
M. Yoshikawa, Mater. Sci. Forum. **52-53**, 365 (1989).
R. Dupree and M. E. Smith, J. Chem. Soc. Chem. Commun. **1988**, 1483.
D. Massiot, F. Fayon, M. Capron, I. King, S. Le Calve, B. Alonso, J.-O. Durand, B. Bujoli, Z. Gan, and G. Hoatson, Mag. Reson. Chem. **40**, 70 (2002).
P. F. Peterson, M. Gutmann, Th. Proffen, and S. J. L. Billinge, J. Appl. Cryst. **32**, 1192 (2000).
C. L. Farrow, P. Juhas, J. W. Liu, D. Bryndin, E. S. Bozin, J. Bloch, Th. Proffen, and S. J. L. Billinge, J. Phys.:Condens. Matter **19**, 335219 (2007).
A. C. Larson and R. B. von Dreele, Los Alamos National Laboratory Report No. LAUR 86-748, 1994.
B. H. Toby, J. Appl. Cryst. **34**, 210 (2001).
J. M. Calleja and M. Cardona, Phys. Rev. B **16**, 3753 (1977)
T. C. Damen, S. P. S. Porto, and B. Tell, Phys. Rev. **142**, 570 (1966).
J. M. Zhang, T. Ruf, M. Cardona, O. Ambacher, M. Stutzmann, J.-M. Wagner, and F. Bechstedt, Phys. Rev. B **56**, 14399 (1997).
H.-M. Cheng, K.-F. Lin, H.-C. Hsu, C.-J. Lin, L.-J. Lin, and W.-F. Hsieh J. Phys. Chem. B **109**, 18385 (2005).
E. Kundla, A. Samoson, and E. Lippmaa, Chem. Phys. Lett. **83**, 229 (1981).
E. S. Toberer, J. D. Epping, B. F. Chmelka, R. Seshadri, Chem. Mater. **18**, 6345 (2006)
K. J. D. MacKenzie and R. H. Meinhold, Am. Miner. **79**, 250 (1994). Mg survey
K. J. D. MacKenzie and R. H. Meinhold, Thermochim. Acta **230**, 331 (1993).
M. Magi, E. Lippmaa, A. Samoson, G. Engelhardt, and A.-R. Grimmer, J. Phys. Chem. **88**, 1518 (1984).
H. Sawada, R. Wang, and A. W. Sleight, J. Solid State Chem. **122**, 148 (1996).
T. M. Sabine and S. Hogg, Acta Crystallogr. B **25**, 2254 (1969).
S. C. Abrahams and J. L. Bernstein, Acta Crystallogr. B **25**, 1233 (1969).
P. Harrison, G. A. Jeffrey, and J. R. Townsend, Acta Crystallogr. **11**, 552 (1958).
J. Albertsson, S. C. Abrahams, and A. Kvick, Acta Crystallogr. B **45**, 34 (1989).
E. H. Kisi and M. M. Elcombe, Acta Crystallogr. C **45**, 1867 (1989).
V. F. Sears, Neutron News **3**, 26 (1992).
I. D. Brown, *The Chemical Bond in Inorganic Chemistry: The Bond Valence Model (Oxford University Press, Oxford, 2002).*
I. D. Brown, Acta Crystallogr. B **48**, 553 (1992).
Y.-I. Kim and R. Seshadri, J. Korean Phys. Soc. in press.
Y.-I. Kim and R. Seshadri, Inorg. Chem. in press.
N. E. Brese and M. O’Keeffe, Acta Crystallogr. B **47**, 192 (1991).
R. D. Shannon, Acta Crystallogr. A **32**, 751 (1976).
S. Limpijumnong and W. R. L. Lambrecht, Phys. Rev. B **63**, 104103 (2001).
A. Janotti, D. Segev, and C. G. Van de Walle, Phys. Rev. B **74**, 045202 (2006).
P. Gopal and N. A. Spaldin, J. Electron. Mater. **35**, 538 (2006).
A. Malashevich and D. Vanderbilt, Phys. Rev. B **75**, 045106 (2007).
X. F. Fan, H. D. Sun, Z. X. Shen, J.-L. Kuo, and Y. M. Lu, J. Phys.:Condens. Matter **20**, 235221 (2008).
A. Malashevich and D. Vanderbilt, private communications.
A. Dal Corso, M. Posternak, R. Resta, and A. Baldereschi, Phys. Rev. B **50**, 10715 (1994).
F. Bernardini, V. Fiorentini, and D. Vanderbilt, Phys. Rev. B **56**, R10024 (1997).
Y. Noel, C. M. Zicovich-Wilson, B. Civalleri, Ph. D’Arco, and R. Dovesi, Phys. Rev. B **65**, 014111 (2001).
APPENDIX {#appendix .unnumbered}
========
Complete list of atomic coordinates for the 16-atom wurtzite supercell Zn$_7$MgO$_8$ ($a=b=6.5054\,\AA$, $c=5.2015\,\AA$, $\alphaup=\betaup=$ 90$^{\circ}$, $\gammaup=$ 120$^{\circ}$), as determined by the neutron PDF analysis.
Atom $x$ $y$ $z$ Type
------ -------- -------- ----------- ------
Zn 0.1667 0.3333 0 Zn1
0.6667 0.3333 0 Zn3
0.1667 0.8333 0 Zn1
0.6667 0.8333 0 Zn1
0.3333 0.1667 0.5 Zn2
0.8333 0.1667 0.5 Zn2
0.8333 0.6667 0.5 Zn2
Mg 0.3333 0.6667 0.486(8) Mg
O 0.1667 0.3333 0.386(3) O2
0.6667 0.3333 0.3825(8) O3
0.1667 0.8333 0.386(3) O2
0.6667 0.8333 0.386(3) O2
0.3333 0.1667 0.8825(8) O3
0.8333 0.1667 0.8825(8) O3
0.3333 0.6667 0.858(5) O1
0.8333 0.6667 0.8825(8) O3
| {
"pile_set_name": "ArXiv"
} |
---
abstract: |
Given a (possibly infinite) connected graph $G$, a measure $\mu$ on the subgraphs of $G$ is said to be $1$*-independent* if events supported on disjoint sets of vertices of $G$ are independent. Let $\mathcal{M}_{1,p}(G)$ denote the collection of $1$-independent measures on $G$ where any given edge is present with probability at least $p$. Let $\mathbb{Z}^2$ denote the square integer lattice. Balister and Bollobás raised the question of determining the critical value $p_{\star}=p_{1,c}(\mathbb{Z}^2)$ such that for all $p>p_{\star}$ and all $\mu \in \mathcal{M}_{1,p}(\mathbb{Z}^2)$, a $\mu$-random subgraph $G$ of $\mathbb{Z}^2$ almost surely contains an infinite component. This can be thought as a $1$-independent analogue of the celebrated Harris–Kesten theorem.
In this paper we investigate both this problem and $1$-independent random graph models more generally. We give two lower bounds on $p_{\star}$ that significantly improve on the previous bounds. Furthermore, motivated by the Russo–Seymour–Welsh lemmas, we define a critical probability for long paths in $1$-independent random graphs and determine its value for the line and ladder lattices. Finally, for finite graphs $G$ we study $f_{1,G}(p)$ (respectively $F_{1,G}(p)$), the infimum (respectively supremum) over all $\mu\in \mathcal{M}_{1,p}(G)$ of the probability of a $\mu$-random subgraph of $G$ being connected. We determine $f_{1,G}(p)$ and $F_{1,G}(p)$ exactly when $G$ is a path, a complete graph and a cycle of length at most $5$.
Many new problems arise from our work, which are discussed in the final section of the paper.
author:
- 'A. Nicholas Day[^1]'
- 'Victor Falgas–Ravry'
- 'Robert Hancock[^2]'
title: 'Long paths and connectivity in $1$-independent random graphs'
---
Introduction {#section: introduction}
============
Bond percolation models, $1$-independence and edge-probability
--------------------------------------------------------------
Let $G$ be a (possibly infinite) connected graph. Subgraphs of $G$ are called *configurations*. In a configuration $H$, an edge is said to be *open* if it belongs to $H$, and *closed* otherwise. A *bond percolation model* on the host graph $G$ is a probability measure $\mu$ on the subgraphs of $G$, i.e. on the space of configurations. Given such a measure, we denote the corresponding random graph model by $\mathbf{G}_{\mu}$, and refer to it as the $\mu$-random graph or $\mu$-random configuration.
In this paper, we study bond percolation models $\mu$ where the states (open or closed) of edges in subsets $F_1,F_2$ of $E$ in a $\mu$-random configuration are independent provided that the edges in $F_1$ and $F_2$ are ‘sufficiently far apart’. To make this more precise, we make use of the following definition.
Two edge sets $F_1, F_2\subseteq E$ are $k$-distant if $F_1\cap F_2 =\emptyset$ and the shortest path of $G$ from an edge in $F_1$ to an edge in $F_2$ contains at least $k$ edges. A bond percolation model $\mu$ on $G$ is $k$-independent if for any pair $(F_1, F_2)$ of $k$-distant edge sets, the intersections $\mathbf{G}_{\mu}\cap F_1$ and $\mathbf{G}_{\mu}\cap F_2$ are independent random variables.
So for example $\mu$ is $0$-independent if each edge of $\mathbf{G}_{\mu}$ is open at random independently of all the others, i.e. $\mu$ can be viewed as a product of Bernoulli measures on the edges of $G$. A well-studied $0$-independent model is the Erd[ő]{}s–Rényi random graph $\mathbf{G}_{n,p}$, where the host graph is $G=K_n$, the complete graph on $n$ vertices, and where $\mu$, known as the $p$*-random measure*, sets each edge to be open with probability $p$, independently of all the others.
In this paper, we focus on the next strongest notion of independence, namely $1$-independence. Measures that are $1$-independent have the property that events supported on disjoint vertex sets are independent. For many $1$-independent models, the randomness can be thought to ‘reside in the vertices’. An important example of a $1$-independent model is that of *site percolation* on the square integer lattice. In this case the host graph is the square integer lattice $\mathbb{Z}^2$ (where two vertices are joined by an edge if they lie at distance $1$ apart), and the measure $\mu=\mu_{\mathrm{site}}(\theta)$ is obtained by switching each vertex of $\mathbb{Z}^2$ *on* at random with probability $\theta$, independently of all the others, and by setting an edge to be open if and only if both of its endpoints are switched on. Site percolation measures may be defined more generally on any host graph in the natural way.
Site percolation is an example of a broader class of 1-independent measures where we independently associate to each vertex $v\in V(G)$ a state $S_v$ at random, and set an edge $uv$ to be open if and only if $f(S_u, S_v)=1$, for some deterministic function $f$ (which may depend on $u$ and $v$). We refer to such measures as *vertex-based measures* (see Section \[section: open problems\] for a formal definition); vertex-based measures are closely related to the notion of [*two-block factors*]{} — see [@LiggettSchonmannStacey97] for further details.
An important point to note is that while all $0$-independent bond percolation models are a product of Bernoulli measures on the edges of $G$ (with varying parameters), it is well-known that a graph $G$ may support many $1$-independent measures which cannot be realised as vertex-based measures or even as measures from the more general class of block factors, see for instance [@HolroydLigget16]. In particular for most graphs $G$, it is not feasible to generate or simulate the collection of $1$-independent measures of $G$.
Given a bond percolation model $\mu$ on a host graph $G$, the (lower)-edge-probability of $\mu$ is $$\mathrm{d}(\mu):=\inf_{e\in E(G)} \mu\{e\textrm{ is open} \}.$$
So for instance a $p$-random measure has edge-probability $p$, while a site percolation measure with parameter $\theta$ has edge-probability $\theta^2$. The collection of $k$-independent bond percolation models $\mu$ on a graph $G$ with edge-probability $d(\mu)\geqslant p$ is denoted by $\mathcal{M}_{k, p}(G)$.
Critical probabilities for percolation and motivation for this paper
--------------------------------------------------------------------
Percolation theory is the study of random subgraphs of infinite graphs. Since its inception in Oxford in the 1950s, it has blossomed into a rich theory and has been the subject of several monographs [@BollobasRiordan06; @Grimmett99; @MeesterRoy96]. The central problem in percolation theory is to determine the relationship between edge-probabilities and the existence of infinite connected components in bond percolation models.
In the most fundamental instance of this problem, consider an infinite, locally finite connected graph $G$, and let $\mu$ be a $0$-independent bond percolation model on $G$. We say that *percolation* occurs in a configuration $H$ on $G$ if $H$ contains an infinite connected component of open edges. By Kolmogorov’s zero–one law, for $G$ and $\mu$ as above, percolation is a tail event whose $\mu$-probability is either zero or one. This allows one to thus define the *Harris critical probability* $p_{0,c}(G)$ for $0$-independent percolation: $$p_{0,c}(G):= \inf\Bigl\{p\in [0,1]: \forall \mu\in \mathcal{M}_{0,p}(G), \ \mu(\{\mathrm{percolation}\})=1\Bigr\}.$$
\[problem: 0-independent percolation\] Given an infinite, locally finite connected graph $G$, determine $p_{0,c}(G)$.
One of the cornerstones of percolation theory — and indeed one of the triumphs of twentieth century probability theory — is the Harris–Kesten theorem, which established the value of $p_{0,c}(\mathbb{Z}^2)$ to be $1/2$.
Let $\mu$ be the $p$-random measure on $\mathbb{Z}^2$. Then $$\mu(\{\mathrm{percolation}\})=\left\{\begin{array}{ll} 0 & \mathrm{if } \quad p\leqslant \frac{1}{2}\\
1 & \mathrm{if }\quad p>\frac{1}{2}. \end{array}\right.$$
In this paper, we focus on the question of what happens to the Harris critical probability in $\mathbb{Z}^2$ if the assumption of $0$-independence is weakened to $k$-independence. In particular, how much can *local* dependencies between the edges postpone the *global* phenomenon of percolation?
Let $G$ be an infinite, locally finite connected graph and let $k\in \mathbb{N}_{0}$. The Harris critical probability for $k$-independent percolation[^3] in $G$ is defined to be: $$p_{k,c}(G):= \inf\Bigl\{p\in [0,1]: \forall \mu\in \mathcal{M}_{k,p}(G), \ \mu(\{\mathrm{percolation}\})=1\Bigr\}.$$
\[problem: Harris critical probabiliy for 1-ipm\] Determine $p_{1, c}(\mathbb{Z}^2)$.
Problem \[problem: Harris critical probabiliy for 1-ipm\] was proposed by Balister and Bollobás [@BalisterBollobas12] in a 2012 paper in which they began a systematic investigation of $1$-independent percolation models. Study of $1$-independent percolation far predates their work, however, due to important applications of $1$-independent percolation models.
A standard technique in percolation is *renormalisation*, which entails reducing a $0$-independent model to a $1$-independent one (possibly on a different host graph), trading in some dependency for a boost in edge-probabilities. Renormalisation arguments feature in many proofs in percolation theory; a powerful and particularly effective version of such arguments was developed by Balister, Bollobás and Walters [@BalisterBollobasWalters05].
Their method, which relies on comparisons with $1$-independent models on $\mathbb{Z}^2$ (in almost all cases) and Monte–Carlo simulations to estimate the probabilities of bounded events, has been applied to give rigorous confidence intervals for critical probabilities/intensities in a wide variety of settings: various models of continuum percolation [@BalisterBollobasWalters05; @BalisterBollobasWalters09; @BalisterBollobas13], hexagonal circle packings [@BenjaminiStauffer13], coverage problems [@BalisterBollobasSarkarKumar07; @HaenggiSarkar13], stable Poisson matchings [@DeijfenHolroydPeres11; @DeijfenHaggstromHolroyd12], the Divide-and-Colour model [@BalintBeffaraTassion13], site and bond percolation on the eleven Archimedean lattices [@RiordanWalters07] and for site and bond percolation in the cubic lattice $\mathbb{Z}^3$ [@Ball14]. The usefulness of comparison with $1$-independent models and the plethora of applications give strong theoretical motivation for the study of $1$-independent percolation.
From a more practical standpoint, many of the real-world structures motivating the study of percolation theory exhibit short-range interactions and local dependencies. For example a subunit within a polymer will interact and affect the state of nearby subunits, but perhaps not of distant ones. Similarly, the position or state of an atom within a crystalline network may have a significant influence on nearby atoms, while long-range interactions may be weaker. Within a social network, we would again expect individuals to exert some influence in aesthetic tastes or political opinions, say, on their circle of acquaintance, and also expect that influence to fade once we move outside that circle. This suggests that $k$-independent bond percolation models for $k\geqslant 1$ are as natural an object of study as the more widely studied $0$-independent ones.
Despite the motivation outlined above, $1$-independent models remain poorly understood. To quote Balister and Bollobás from their 2012 paper: “$1$-independent percolation models have become a key tool in establishing bounds on critical probabilities \[...\]. Given this, it is perhaps surprising that some of the most basic questions about $1$-independent models are open”. There are in fact some natural explanations for this state of affairs. As remarked on in the previous subsection, there are *many* very different $1$-independent models with edge-probability $p$, and they tend to be harder to study than $0$-independent ones due to the extra dependencies between edges. In particular simulations are often of no avail to formulate conjectures or to get an intuition for $1$-independent models in general. Moreover, while the theoretical motivation outlined above is probabilistic in nature, the problem of determining a critical constant like $p_{1,c}(\mathbb{Z}^2)$ is extremal in nature — one has to determine what the worst possible $1$-independent model is with respect to percolation — and calls for tools from the separate area of extremal combinatorics.
In this paper, we continue Balister and Bollobás’s investigation into the many open problems and questions about and on these measures. Before we present our contributions to the topic, we first recall below previous work on $1$-independent percolation.
Previous work on $1$-independent models
---------------------------------------
Some general bounds for stochastic domination of $k$-independent models by $0$-independent ones were given by Liggett, Schonmann and Stacey [@LiggettSchonmannStacey97]. Amongst other things, their results implied $p_{1,c}(\mathbb{Z}^2)<1$. Balister, Bollobás and Walters [@BalisterBollobasWalters05] improved this upper bound via an elegant renormalisation argument and some computations. They showed that in any $1$-independent bond percolation model on $\mathbb{Z}^2$ with edge-probability at least $0.8639$, the origin has a strictly positive chance of belonging to an infinite open component. This remains to this day the best upper bound on $p_{1,c}(\mathbb{Z}^2)$. In a different direction, Balister and Bollobás [@BalisterBollobas12] observed that trivially $p_{1,c}(G)\geqslant \frac{1}{2}$ for any infinite, locally finite connected graph $G$. In the special case of the square integer lattice $\mathbb{Z}^2$, they recalled a simple construction due to Newman which gives $$\begin{aligned}
\label{eq: site percolation bounds on p_{1,c}}
p_{1,c}(\mathbb{Z}^2)\geqslant \left(\theta_{\mathrm{site}}\right)^2+\left(1-\theta_{\mathrm{site}}\right)^2,\end{aligned}$$ where $\theta_{site}$ is the critical value of the $\theta$-parameter for site percolation, i.e. the infimum of $\theta \in [0,1]$ such that switching vertices of $\mathbb{Z}^2$ on independently at random with probability $\theta$ almost surely yields an infinite connected component of on vertices. Plugging in the known rigorous bounds for $0.556\leqslant \theta_{\mathrm{site}}\leqslant 0.679492$ [@VDBErmakov96; @Wierman95] yields $p_{1,c}(\mathbb{Z}^2)\geqslant 0.5062$, while using the non-rigorous estimate $\theta_{\mathrm{site}}\approx 0.592746$ (see for example [@SudingZiff99]) yields the non-rigorous lower-bound $p_{1,c}(\mathbb{Z}^2)\geqslant 0.5172$.
With regards to other lattices, Balister and Bollobás completed a rigorous study of $1$-independent percolation models on infinite trees [@BalisterBollobas12], giving $1$-independent analogues of classical results of Lyons [@Lyons90] for the $0$-independent case. Balister and Bollobás’s results were later generalised to the $k$-independent setting by Mathieu and Temmel [@MathieuTemmel12], who also showed interesting links between this problem and theoretical questions concerning the Lovász local lemma, in particular the work of Scott and Sokal [@ScottSokal05; @ScottSokal06] on hard-core lattice gases, independence polynomials and the local lemma.
Our contributions
-----------------
In this paper, we make a three-fold contribution to the study of Problem \[problem: Harris critical probabiliy for 1-ipm\]. First of all, we improve previous lower bounds on $p_{1, c}(\mathbb{Z}^2)$ with the following theorems.
\[theorem: local construction integer lattice\] For all $d \in \mathbb{N}_{\geqslant 2}$, we have that $$p_{1,c}(\mathbb{Z}^{d}) \geqslant 4 - 2 \sqrt{3} \approx 0.535898\ldots \ .$$
Theorem \[theorem: local construction integer lattice\] strictly improves on the previous best lower bound for $d=2$ given in (\[eq: site percolation bounds on p\_[1,c]{}\]) above; moreover, it is based on a very different idea, which first appeared in the second author’s PhD thesis [@FalgasRavry12]. In addition we give a separate improvement of (\[eq: site percolation bounds on p\_[1,c]{}\]): let $\theta_{\mathrm{site}}$ again denote the critical threshold for site percolation. Then the following holds.
\[theorem: global construction integer lattice\] $$p_{1,c}(\mathbb{Z}^{2}) \geqslant \left(\theta_{\mathrm{site}}(\mathbb{Z}^{2})\right)^{2} + \frac{1}{2}\left(1-\theta_{\mathrm{site}}\left(\mathbb{Z}^{2}\right)\right).$$
Substituting the rigorous bound $\theta_{\mathrm{site}}\geqslant0.556 $ into Theorem \[theorem: global construction integer lattice\] yields the lower bound $p_{1,c}(\mathbb{Z}^{2}) \geqslant 0.531136$, which does slightly worse than Theorem \[theorem: local construction integer lattice\]. However substituting in the widely believed but non-rigorous estimate $\theta_{\mathrm{site}}\approx 0.592746$ yields a significantly stronger lower bound of $p_{1,c}(\mathbb{Z}^{2}) \geqslant 0.554974$.
Secondly, motivated by efforts to improve the upper bounds on $p_{1,c}(\mathbb{Z}^2)$, and in particular to establish some $1$-independent analogues of the Russo–Seymour–Welsh (RSW) lemmas on the probability of crossing rectangles, we investigate the following problems. Let $P_n$ denote the graph on the vertex set $\{1,2,\ldots n\}$ with edges $\{12, 23, \ldots, (n-1)n\}$, i.e. a path on $n$ vertices. Given a connected graph $G$, denote by $P_n\times G$ the Cartesian product of $P_n$ with $G$. A *left-right crossing* of $P_n\times G$ is a path from a vertex in $\{1\}\times V(G)$ to a vertex in $\{n\}\times V(G)$. We define the *crossing critical probability* for $1$-independent percolation on $P_n\times G$ to be $$p_{1, \times}(P_n\times G):=\inf\Bigl\{p\in[0,1]:\ \forall \mu \in \mathcal{M}_{1,p}(P_n\times G), \ \mu(\exists\textrm{ open left-right crossing})>0 \Bigr\},$$ i.e. the least edge-probability guaranteeing that in any $1$-independent model on $P_n \times G$, there is a strictly positive probability of being able to cross $P_n\times G$ from left to right.
\[problem: critical crossing prob\] Given $n\in \mathbb{N}$ and a finite, connected graph $G$, determine $p_{1, \times}(P_n\times G)$.
Problem \[problem: critical crossing prob\] can be thought of as a first step towards the development of $1$-independent analogues of the RSW lemmas; these lemmas play a key role in modern proofs of the Harris–Kesten theorem, and one would expect appropriate $1$-independent analogues to constitute a similarly important ingredient in a solution to Problem \[problem: Harris critical probabiliy for 1-ipm\]. By taking the limit as $n\rightarrow \infty$ in Problem \[problem: critical crossing prob\], one is led to consider another $1$-independent critical probability. Let $G$ be an *infinite*, locally finite connected graph. The *long paths critical probability* for $1$-independent percolation on $G$ is $$p_{1, \ell\mathit{p}}(G):=\inf\Bigl\{p\in[0,1]:\ \forall \mu \in \mathcal{M}_{1,p}(G),\ \forall n\in \mathbb{N}\ \mu(\exists\textrm{ open path of length }n)>0 \Bigr\},$$ i.e. the least edge-probability at which arbitrarily long open paths will appear in all $1$-independent models in $G$.
\[problem: critical long paths prob\] Given an infinite, locally finite, connected graph $G$, determine $p_{1, \ell\mathit{p}}(G)$.
In this paper, we resolve Problem \[problem: critical crossing prob\] in a strong form when $G$ consists of a vertex or an edge (see Theorems \[theorem: connected function Pn\] and \[theorem: ladder construction\]). This allows us to solve Problem \[problem: critical long paths prob\] when $G$ is the integer line lattice $\mathbb{Z}$ and the integer ladder lattice $\mathbb{Z}\times P_2$.
\[theorem: long paths critical prob\] We have that
(i) $p_{1, \ell \mathit{p}}(\mathbb{Z})=\frac{3}{4}$, and
(ii) $p_{1, \ell \mathit{p}}(\mathbb{Z}\times P_2)=\frac{2}{3}$.
Note that part (i) of Theorem \[theorem: long paths critical prob\] above can be read out of earlier work of Liggett, Schonman and Stacey [@LiggettSchonmannStacey97] and Balister and Bollobás [@BalisterBollobas12]. We prove further bounds on both $p_{1, \times}(P_n \times G)$ and $p_{1, \ell\mathit{p}}(\mathbb{Z}\times G)$ for a variety of graphs $G$. We summarise the latter, less technical, set of results below. Let $C_{n}$ and $K_n$ denote the cycle and the complete graph on $n$ vertices respectively.
\[theorem: upper bounds on long paths critical prob\] We have that
(i) $ 0.5359\ldots =4-2\sqrt{3} \leqslant p_{1, \ell\mathit{p}}(\mathbb{Z}\times C_{n})\leqslant p_{1,\ell\mathit{p}}(\mathbb{Z}\times P_n)\leqslant \frac{2}{3} $ for all $n \geqslant 3$;
(ii) $p_{1, \ell\mathit{p}}(\mathbb{Z}\times K_3) \leqslant \frac{1}{16} \Bigl(13 - \frac{55}{\sqrt[3]{128 \sqrt{14} - 251}} + \sqrt[3]{128 \sqrt{14} - 251}\Bigr)= 0.63154\ldots $ ;
(iii) $p_{1, \ell\mathit{p}}(\mathbb{Z}\times C_4) \leqslant (3-\sqrt{3})/2=0.63397\ldots$ ;
(iv) $p_{1, \ell\mathit{p}}(\mathbb{Z}\times C_5) \leqslant 0.63895\ldots$ ;
(v) $0.5359\ldots=4-2\sqrt{3}\leqslant \lim_{n\rightarrow \infty} p_{1, \ell\mathit{p}}(\mathbb{Z}\times K_{n})\leqslant \frac{5}{9}=0.5555\ldots$.
A key ingredient in the proof of Theorems \[theorem: long paths critical prob\] and \[theorem: upper bounds on long paths critical prob\] is a local lemma-type result, Theorem \[theorem: crossing G times Pn\], relating the probability in a $1$-independent model of finding an open left-right crossing of $P_n\times G$ to the probability of a given copy of $G$ being connected in that model. This motivated our third contribution to the study of $1$-independent models in this paper, namely an investigation into the connectivity of $1$-independent random graphs.
Let $G$ be a finite connected graph. For any $p\in [0,1]$, we define the $k$-independent connectivity function of $G$ to be $$f_{k,G}(p):=\inf\Bigl\{\mu(\exists \textrm{ open spanning tree}): \ \mu \in \mathcal{M}_{k,p}(G)\Bigr\}.$$
\[problem: connectivity\] Given a finite connected graph $G$, determine $f_{1,G}(p)$.
We resolve Problem \[problem: connectivity\] exactly when $G$ is a path, a complete graph or a cycle on at most $5$ vertices.
\[theorem: connected function Pn\] Given $n \in \mathbb{N}_{\geqslant 2}$ and $p \in [0,1]$, let $\theta = \theta(p) :=\frac{1 + \sqrt{4p - 3}}{2}$ and $p_{n}:= \frac{1}{4}\left(3 - \tan^{2}\left(\frac{\pi}{n+1} \right) \right)$. We have that $$f_{1,P_{n}}(p) = \begin{cases} \sum_{j = 0}^{n} \theta^{j}(1-\theta)^{n-j} &\text{ for } p \in [p_{n},1] , \\
0 &\text{ for } p \in [0,p_{n}]. \end{cases} \nonumber$$
\[theorem: connected function Kn\] Given $n \in \mathbb{N}_{\geqslant 2}$ and $p \in [0,1]$, let $\theta = \theta(p) := \frac{1 + \sqrt{2p - 1}}{2}$ and $p_{n} := \frac{1}{2}(1 - \tan^{2}(\frac{\pi}{2n}))$. We have that $$f_{1,K_{n}}(p) = \begin{cases} \theta^{n} + (1-\theta)^{n} &\text{ for } p \in [p_{n},1] , \\
0 &\text{ for } p \in [0,p_{n}]. \end{cases} \nonumber$$ In particular, $$f_{1, K_3}(p)=f_{1, C_3}(p)= \begin{cases}
\frac{3p-1}{2} & \text{for } p \in [\frac{1}{3},1], \\
0 & \text{for } p \in [0,\frac{1}{3}].
\end{cases} \nonumber$$
\[theorem: connected function C4\] For $p \in [0,1]$ we have that $$f_{1, C_4}(p)= \begin{cases}
2p-1 & \text{for } p \in [\frac{1}{2},1], \\
0 & \text{for } p \in [0,\frac{1}{2}].
\end{cases} \nonumber$$
\[theorem: connected function C5\] For $p \in [0,1]$ we have that $$f_{1, C_5}(p)= \begin{cases}
\frac{p(3p^2-1)}{3p-1} & \text{for } p \in [\frac{\sqrt{3}}{3},1], \\
0 & \text{for } p \in [0,\frac{\sqrt{3}}{3}].
\end{cases} \nonumber$$
We also consider the opposite problem to Problem \[problem: connectivity\], namely maximising connectivity in $1$-independent random graph models. Let $\mathcal{M}_{k, \leqslant p}(G)$ denote the collection of $1$-independent measures $\mu$ on $G$ such that $\sup_{e\in E(G)}\mu \{e \textrm{ is open}\}\leqslant p$. Set $$F_{k, G}(p):=\sup \{ \mu \left(\exists \textrm{ open spanning tree}\right): \ \mu \in \mathcal{M}_{k,\leqslant p}(G) \}.$$
\[problem: maximising connectivity\] Given a finite connected graph $G$, determine $F_{1,G}(p)$.
We resolve Problem \[problem: maximising connectivity\] exactly when $G$ is a path, a complete graph or a cycle on at most $5$ vertices.
\[theorem: maximising connectivity in Pn\] For all $n\in \mathbb{N}$ with $n\geqslant 2$, $F_{1, P_n}(p)=p^{\lfloor \frac{n}{2}\rfloor}$.
\[theorem: maximising connectivity in Kn\] For all $n\in \mathbb{N}$ with $n\geqslant 2$, $F_{1, K_n}(p)=1-f_{1, K_n}(1-p)$.
\[theorem: maximising connectivity for C4\] $$F_{1, C_4}(p)=\left\{ \begin{array}{ll}
2p-p^2 & \textrm{if }p\in [\frac{2}{3}, 1],\\
2p^2 & \textrm{if }p\in [0, \frac{1}{3}].
\end{array}\right.$$
\[theorem: maximising connectivity for C5\] $$F_{1, C_5}(p)=\left\{ \begin{array}{ll}
\frac{p(2-5p(1-p))}{5-3p} & \textrm{if }p\in [\frac{3}{5}, 1],\\
\frac{5p^2}{3}&\textrm{if }p\in [\frac{1}{2}, \frac{3}{5}],\\
\frac{5p^2(p+1)}{p+4}& \textrm{if }p\in [0, \frac{1}{2}].
\end{array}\right.$$
Together, Theorems \[theorem: connected function Pn\]–\[theorem: connected function C5\] and \[theorem: maximising connectivity in Kn\]–\[theorem: maximising connectivity for C5\] determine the complete connectivity ‘profile’ for $1$-independent measures $\mu$ on $K_n$, $P_n$, $C_4$ and $C_5$ — that is, the range of values $\mu(\{\textrm{connected}\})$ can take if every edge is open with probability $p$. In Figure \[figure: connectivity\], we illustrate these for four of these graphs $G$ with plots of $f_{1,G}(p)$, $F_{1, G}(p)$ and $f_{0, G}(p):=\mu\left(\mathbf{G}_p \textrm{ is connected}\right)$, where $\mathbf{G}_p$ is the $0$-independent model on $G$ obtained by setting each edge of $G$ to be open with probability exactly $p$, independently at random.
; ; ; ; ;
; ; ; ; ;
; ; ; ; ;
; ; ; ; ; ;
Organisation of the paper
-------------------------
Our first set of results, Theorems \[theorem: local construction integer lattice\] and \[theorem: global construction integer lattice\] are proved in Section \[section: lower bound on crit prob for percolation in Z\^2\].
In Section \[section: general upper bound on crossing prob\], we use arguments reminiscent of those used in inductive proofs of the Lovász local lemma to obtain Theorem \[theorem: crossing G times Pn\], which gives a general upper bound for crossing and long paths critical probabilities in $1$-independent percolation models on Cartesian products $\mathbb{Z}\times G$. This result is used in Sections \[section: line lattice\] and \[section: ladder\] to prove Theorem \[theorem: long paths critical prob\] on the long paths critical probability for the line and ladder lattices.
In Sections \[section: line lattice\], \[section: complete graphs\] and \[section: cycles\] and \[section: maximising connectivity\], we prove our results on $f_{1,G}(p)$ and $F_{1,G}(p)$ when $G$ is a path, a complete graph or a short cycle. We apply these results in Section \[section: long paths\] to prove Theorem \[theorem: upper bounds on long paths critical prob\]. Finally we end the paper in Section \[section: open problems\] with a discussion of the many open problems arising from our work.
Notation {#subsection: notation}
--------
We write $\mathbb{N}$ for the set of natural numbers $\{1,2,\ldots\}$, $\mathbb{N}_{0}$ for the set $\mathbb{N} \cup \{0\}$, and $\mathbb{N}_{\geqslant k}$ for the set of natural numbers greater than or equal to $k$.
We set $[n]:=\{1,2, \ldots n\}$. Given a set $A$, we write $A^{(r)}$ for the collection of all subsets of $A$ of size $r$, hereafter referred to as $r$-sets from $A$. We use standard graph theoretic notation. A graph is a pair $G=(V,E)$ where $V=V(G)$ and $E=E(G)\subseteq V(G)^{(2)}$ denote the vertex set and edge set of $G$ respectively. Given a subset $A\subseteq G$, we denote by $G[A]$ the subgraph of $G$ induced by $A$. We also write $N(A)$ for the set of vertices in $G$ adjacent to at least one vertex in $A$.
Throughout this paper, we shall use *$k$-ipm* as a shorthand for ‘$k$-independent percolation model/measure’. In a slight abuse of language, we say that a bond percolation model $\mu$ on an infinite connected graph $G$ percolates if $\mu(\{\textrm{percolation}\})=1$. We refer to a random configuration $\mathbf{G}_{\mu}$ as a $\mu$-random subgraph of $G$. Finally we write $\mathbb{E}_{\mu}$ for the expectation taken with respect to the probability measure $\mu$. For any event $X$, we write $X^{c}$ for the complement event.
Lower bounds on $p_{1,c}(\mathbb{Z}^d)$ {#section: lower bound on crit prob for percolation in Z^2}
=======================================
Let $d\in \mathbb{N}_{\geqslant 2}$. For $k \in \mathbb{N}_{0}$, let $T_{k} := \big\{(x,y) \in \mathbb{Z}^{d}: \max(|x|,|y|) = k\big\}$. Let $q := \sqrt{3} - 1$. For each vertex in $\mathbb{Z}^{d}$, we colour it either Blue or Red, or set it to state $I$, which stands for *Inwards*. The probability that a given vertex will be in each of these states will depend on which of the $T_{k}$ the vertex is in, and we assign these states to each vertex independently of all other vertices.
- If $v$ is a vertex in $T_{k}$, where $k \equiv 0 \!\mod 6$, then we colour $v$ Blue.
- If $v$ is a vertex in $T_{k}$, where $k \equiv 1 \!\mod 6$, then we colour $v$ Red with probability $q/2$ and colour it Blue otherwise.
- If $v$ is a vertex in $T_{k}$, where $k \equiv 2 \!\mod 6$, then we colour $v$ Red with probability $q$ and put it in the Inwards state $I$ otherwise.
- If $v$ is a vertex in $T_{k}$, where $k \equiv 3 \!\mod 6$, then we colour $v$ Red.
- If $v$ is a vertex in $T_{k}$, where $k \equiv 4 \!\mod 6$, then we colour $v$ Blue with probability $q/2$ and colour it Red otherwise.
- If $v$ is a vertex in $T_{k}$, where $k \equiv 5 \!\mod 6$, then we colour $v$ Blue with probability $q$ and put it in the Inwards state $I$ otherwise.
See Figure \[figure: vertex states\] for the possible states of the vertices in $T_{0},T_{1},T_{2}$ and $T_{3}$ when $d=2$.
$\begin{array}{ccccccc}
{\color{red}R} & {\color{red}R} & {\color{red}R} & {\color{red}R} & {\color{red}R} & {\color{red}R} & {\color{red}R}\\
{\color{red}R} & {\color{red}R} I & {\color{red}R} I & {\color{red}R} I & {\color{red}R} I & {\color{red}R} I & {\color{red}R}\\
{\color{red}R} & {\color{red}R} I & {\color{blue}B} {\color{red}R} & {\color{blue}B} {\color{red}R} & {\color{blue}B} {\color{red}R} & {\color{red}R} I & {\color{red}R}\\
{\color{red}R} & {\color{red}R} I & {\color{blue}B} {\color{red}R} & {\color{blue}B} & {\color{blue}B} {\color{red}R} & {\color{red}R} I & {\color{red}R}\\
{\color{red}R} & {\color{red}R} I & {\color{blue}B} {\color{red}R} & {\color{blue}B} {\color{red}R} & {\color{blue}B} {\color{red}R} & {\color{red}R} I & {\color{red}R}\\
{\color{red}R} & {\color{red}R} I & {\color{red}R} I & {\color{red}R} I & {\color{red}R} I & {\color{red}R} I & {\color{red}R}\\
{\color{red}R} & {\color{red}R} & {\color{red}R} & {\color{red}R} & {\color{red}R} & {\color{red}R} & {\color{red}R}\\
\end{array}$
Now suppose that $e = \{v_{1},v_{2}\}$ is an edge in $\mathbb{Z}^{d}$. Firstly we say that the edge $e$ is open if either both $v_{1}$ and $v_{2}$ are Blue or both $v_{1}$ and $v_{2}$ are Red. We also say the edge $e$ is open if, for some $k$, we have that $v_{1} \in T_{k}$, $v_{2} \in T_{k+1}$, and $v_{2}$ is in state $I$. In all other cases we say that the edge $e$ is closed. It is clear that this gives a $1$-independent measure on $\mathbb{Z}^{d}$ as it is vertex-based, and it is also easy to check that every edge is present with probability at least $4 - 2 \sqrt{3}$.
Call this measure $\mu$, and let $G:=\mathbb{Z}^2$. We claim that in $\mathbf{G}_{\mu}$, for all $k \equiv 0 \!\mod 3$, there is no path of open edges from $T_{k}$ to $T_{k+3}$. Suppose this is not the case, and $P$ is some path of open edges from a vertex in $T_{k}$ to $T_{k+3}$. We first note that $P$ cannot include a vertex in state $I$, as such a vertex would be in $T_{k+2}$ and would only be adjacent to a single edge. Thus every vertex of $P$ is either Blue or Red. However, as one end vertex of $P$ is Blue and the other end vertex is Red, and there are no open edges with different coloured end vertices, we have that such a path $P$ cannot exist. As a result, every component of $\mathbf{G}_{\mu}$ is sandwiched between some $T_{k-3}$ and $T_{k+3}$, where $k \equiv 0 \!\mod 3$, and so is of finite size. Thus we have that $p_{1,c}(\mathbb{Z}^{d}) \geqslant 4 - 2 \sqrt{3}$.
The construction in Theorem \[theorem: local construction integer lattice\] can in fact be generalised to certain other graphs and lattices. Given an infinite, connected, locally finite graph $G$, and a vertex set $A \subseteq V(G)$, let $\overline{A}$ be the closure of $A$ under $2$-neighbour bootstrap percolation on $G$. That is, let $\overline{A} := \bigcup_{i \geqslant 0} A_{i}$, where $A_{0} := A$ and for $i \geqslant 1$ $$A_{i} := A_{i-1} \cup \{v \in V(G): v \text{ has } 2 \text{ or more neighbours in } A_{i-1}\}. \nonumber$$ We say that $G$ has the *finite $2$-percolation property* if, for every finite set $A \subseteq V(G)$, we have that $\overline{A}$ is finite.
\[corollary: if 2-percolation property, then lb on p\_[1,c]{}\] If $G$ has the finite $2$-percolation property, then $p_{1,c}(G) \geqslant 4 - 2\sqrt{3}$.
Partition $V(G)$ in the following way: pick any vertex $v$ and set $T_{0} := \{v\}$. For $k \geqslant 1$ let $$T_{k} := \overline{N(T_{k-1})} \setminus \bigcup_{j = 0}^{k-1} T_{j}\nonumber.$$ We have that if $w \in T_{k}$, then $w$ is only adjacent to vertices in $T_{k-1},T_{k}$ and $T_{k+1}$. Moreover, $w$ is adjacent to at most one vertex in $T_{k-1}$ — this is the crucial property needed for our construction. Since $G$ has the finite $2$-percolation property, each $T_{k}$ is finite. Thus we can use the $T_{k}$ to construct a non-percolating $1$-ipm on $G$ in the exact same fashion as done for $\mathbb{Z}^{d}$ in Theorem \[theorem: local construction integer lattice\] (the key being that vertices in state $I$ are still dead ends, being adjacent to a unique edge), which in turn shows that $p_{1,c}(G) \geqslant 4 - 2\sqrt{3}$.
An example of a lattice with the finite $2$-percolation property is the lattice $(3,4,6,4)$, where here we are using the lattice notation of Grünbaum and Shephard [@GrunbaumShephard86]. Riordan and Walters [@RiordanWalters07] showed that the site percolation threshold of this lattice is very likely to lie in the interval $[0.6216,0.6221]$. Thus this estimate, together with Newman’s construction (see equation (\[eq: site percolation bounds on p\_[1,c]{}\])), shows (non-rigorously) that $p_{1,c}\left(\left(3,4,6,4\right)\right) \geqslant 0.52981682$. As this is less than $4-2\sqrt{3}$, we have that our construction gives the (rigorous) improvement of $p_{1,c}\left(\left(3,4,6,4\right)\right) \geqslant 4-2\sqrt{3}$.
Fix $\varepsilon > 0$ sufficiently small so that $q := \theta_{\text{site}}(\mathbb{Z}^{2}) - \varepsilon$ is strictly larger than $1/4$. For each vertex $v \in \mathbb{Z}^{2}$, we assign to it one of three states: On, $L$ or $D$, and we do this independently for every vertex. We assign $v$ to the On state with probability $q$, we assign it to the $L$ state with probability $\frac{1}{2}(1-q)$, and else we assign it to the $D$ state with probability $\frac{1}{2}(1-q)$.
We now describe which edges are open based on the states of the vertices. We first say that the edge $e$ is open if both of its vertices are in the On state. If a vertex is in state $L$, then the edge adjacent and to the left of it is open. Similarly, if a vertex is in state $D$, then the edge adjacent and down from it is open. All other edges are closed. See Figure \[figure: global construction\] for an example of this construction.
![This figure shows the open edges of the construction on a small subset of $\mathbb{Z}^2$. The unlabelled vertices correspond to those in the On state. The black edges are the open edges that are adjacent to two On vertices, while the red edges are the open edges that are either to the left of an $L$ vertex or below a $D$ vertex.[]{data-label="figure: global construction"}](GlobalConstruction)
It is easy to see that this is a $1$-independent measure on $\mathbb{Z}^{2}$ as it is vertex-based, and every edge is present with probability $q^{2} + \frac{1}{2}(1-q)$. Call this measure $\mu$ and let $G:=\mathbb{Z}^2$. We will show that every component of $\mathbf{G}_\mu$ has finite size. We begin by first proving an auxiliary lemma. Let $x \in [0,\frac{1}{2}]$, and let us define another $1$-independent measure on $\mathbb{Z}^{2}$, which we call the *left-down* measure with parameter $x$. In the left-down measure, each vertex of $\mathbb{Z}^{2}$ is assigned to one of three states: Off, $L$ or $D$, and we do this independently for every vertex. For each vertex $v \in \mathbb{Z}^{2}$, we assign it to state $L$ with probability $x$, we assign it to state $D$ with probability $x$, and we assign it to state Off with probability $1-2x$. As above, if a vertex is in state $L$, then the edge adjacent and to the left of it is open, while if a vertex is in state $D$, then the edge adjacent and down from it is open. All other edges are closed. We use $\nu_x$ to denote the left-down measure with parameter $x$.
\[lemma: left-down measure\] If $ 0 \leqslant x \leqslant \frac{3}{8}$, then all components in $\mathbf{G}_{\nu_x}$ are finite.
Let $z := 1 - \sqrt{1-2x}$. As $0 \leqslant x \leqslant \frac{3}{8}$ we have that $0 \leqslant z \leqslant \frac{1}{2}$. We start by taking a random subgraph of $\mathbb{Z}^{2}$ where every edge is open with probability $z$, independently of all other edges. We then further modify it as follows. For each vertex $v = (x,y)$ we look at the state of the edge $e_{1}$ from $v$ to the vertex $(x-1,y)$, and the state of the edge $e_{2}$ from $v$ to the vertex $(x,y-1)$. If at least one of $e_{1}$ or $e_{2}$ is closed, we do not change anything. However, if both $e_{1}$ and $e_{2}$ are open, with probability $\frac{1}{2}$ we close the edge $e_{1}$, and otherwise we close the edge $e_{2}$. We do this independently for every vertex $v$ of $\mathbb{Z}^{2}$. It is easy to see that this is an equivalent formulation of $\nu_x$, the left-down measure with parameter $x$. Thus we have coupled $\nu_x$ to the $0$-independent bond percolation measure $\xi$ on $\mathbb{Z}^{2}$ with edge-probability $z$. In this coupling we have that if an edge $e$ is open in $\mathbf{G}_{\nu_x}$, then it is also open in $\mathbf{G}_{\xi}$. As $z \leqslant 0.5$ we have that all components in $\mathbf{G}_{\xi}$ are finite by the Harris–Kesten theorem, and so we also have that all of the components in $\mathbf{G}_{\nu_x}$ are finite too.
By considering an appropriate branching process it is possible to prove the stronger result that if $ 0 \leqslant x < \frac{1}{2}$, then almost surely all components in $\mathbf{G}_{\nu_x}$ are finite. We make no use of this stronger result in this paper, so we omit its proof. It is also clear that when $x = \frac{1}{2}$, every vertex in $\mathbf{G}_{\nu_x}$ is part of an infinite path consisting solely of steps to the left or steps downwards, and so percolation occurs in $\mathbf{G}_{\nu_x}$ at this point.
Let us return to our original $1$-independent measure $\mu$, where every vertex is in state On, $L$ or $D$. Recall that our aim is to show that all components have finite size in $\mathbf{G}_\mu$. Consider removing all vertices in state $L$ or $D$, and also any edges adjacent to these vertices. What is left will be a collection of components consisting only of edges between vertices in the On state, which we call the *On-sections*. The black edges in Figure \[figure: global construction\] are the edges in the On-sections. As a vertex is On with probability $q < \theta_{\text{site}}(\mathbb{Z}^{2})$, we have that almost surely every On-section is finite. Similarly, consider removing all edges in the On-sections. What is left will be a collection of edges adjacent to vertices in the $L$ or $D$ states. We call these components the $LD$-sections; the red edges in Figure \[figure: global construction\] are the edges in the $LD$-sections. As each vertex is in state $L$ with probability $\frac{1}{2}(1-q) \leqslant \frac{3}{8}$ and in state $D$ with the same probability, Lemma \[lemma: left-down measure\] tells us that almost surely every $LD$-section is finite. For each vertex $v$ in state $L$ orient the open edge to the left of it away from $v$, while for each vertex $v$ in state $D$ orient the open edge below it away from $v$. This gives a partial orientation of the open edges of $\mathbf{G}_\mu$, where every vertex in state $L$ or $D$ has exactly one edge oriented away from it. Moreover, if $v_{1}$ is a vertex in the On state and $v_{2}$ is a vertex in the $L$ or $D$ state, then the edge between them is oriented from $v_{2}$ to $v_{1}$. This implies that every $LD$-section is adjacent to at most one On-section. As such, almost surely every component in $\mathbf{G}_\mu$ consists of at most one On-section, and a finite number of finite $LD$-sections attached to it. Thus almost surely every component in $\mathbf{G}_\mu$ is finite.
A general upper bound for crossing $P_n\times G$ {#section: general upper bound on crossing prob}
================================================
Let $G$ be a finite connected graph. Set $v(G):=\vert V(G)\vert$. Recall that for any $1$-independent bond percolation measure $\mu \in \mathcal{M}_{1,p}(G)$, we have $\mu(\mathbf{G}_{\mu} \textrm{ is connected})\geqslant f_{1,G}(p)$.
\[theorem: crossing G times Pn\] If $p$ satisfies $$\begin{aligned}
\label{equation: bounding equation}
\left(f_{1,G}(p)\right)^2\geqslant \frac{1}{\alpha(1-\alpha)} (1-p)^{v(G)},
\end{aligned}$$ for some $\alpha\in(0,1/2]$, then for every $\ell\in \mathbb{N}$ $$f_{1, P_{\ell}\times G}(p)\geqslant \left((1-\alpha)f_{1,G}(p)\right)^{\ell} .$$
Consider an arbitrary measure $\mu\in \mathcal{M}_{1, p}(\mathbb{Z}\times G)$. For any $n\in \mathbb{N}$, the restriction of $\mu$ to $[n]\times V(G)$ is a measure from $\mathcal{M}_{1,p}(P_n\times G)$, and clearly all such measures can be obtained in this way. Furthermore, for every $n\in \mathbb{N}$, the restriction of $\mu$ to $\{n\}\times V(G)$ is a measure from $\mathcal{M}_{1,p}(G)$, and in particular the subgraph of $\mathbf{(\mathbb{Z}\times G)}_{\mu}$ induced by $\{n\}\times V(G)$ is connected with probability at least $f_{1, G}(p)$.
We consider the $\mu$-random graph $\mathbf{(\mathbb{Z}\times G)}_{\mu}$. For $n\geqslant 1$ let $Y_n$ be the event that $[n]\times V(G)$ induces a connected subgraph. For $n\geqslant 2$, let $X_n$ be the event that $[n-1]\times V(G)$ induces a connected subgraph and at least one vertex in $\{n\}\times V(G)$ is connected to a vertex in $\{n-1\}\times V(G)$. For $n=1$, set $X_1$ to be the trivially satisfied event occurring with probability $1$. Set $$\begin{aligned}
x_n &:= \mu\left(X_n^c \vert \bigcap_{m<n} Y_m\right) & \textrm{ and } && y_n &:=\mu\left(Y_n^c\vert X_n \cap (\bigcap_{m<n} Y_m) \right).
\end{aligned}$$ For $n\geqslant 1$, let $V_n$ be the event that $\{n\}\times V(G)$ induces a connected subgraph, and for $n\geqslant 2$ let $H_n$ be the event that at least one of the edges from $\{n-1\}\times V(G)$ to $\{n\}\times V(G)$ is present. Observe that we have the following easy inclusions:
(a) $(X_{n+1})^c\cap Y_n\subseteq (H_{n+1})^c\cap Y_n$,
(b) $Y_n\cap Y_{n-1}\subseteq (V_n\cap X_n)\cap Y_{n-1}$, and
(c) $(Y_n)^c\cap X_{n}\subseteq (V_n)^c \cap X_n$.
We begin by establishing two inductive relations for the sequences $x_n$ and $y_n$. First of all, using (a) and (b) we have, $$\begin{aligned}
x_{n+1} = \frac{\mu((X_{n+1})^c\cap (\bigcap_{m\leqslant n} Y_m ))} {\mu(\bigcap_{m\leqslant n} Y_m )}% \\
&\leqslant \frac{\mu((H_{n+1})^c \cap (\bigcap_{m\leqslant n} Y_m ))} {\mu(\bigcap_{m\leqslant n} Y_m )}\notag \\
&\leqslant \frac{\mu((H_{n+1})^c ) } {\mu(Y_n \vert (\bigcap_{m<n} Y_m ))} \qquad \textrm{ by $1$-independence}\notag \\
&\leqslant \frac{(1-p)^{v(H)}} {\mu(V_n \vert (\bigcap_{m<n} Y_m )) -\mu((X_n)^c \vert (\bigcap_{m<n} Y_m ))}\notag \\
&\leqslant \frac{(1-p)^{v(H) }}{f_{1,G}(p) -x_n} \qquad \textrm{by $1$-independence}. \label{equation: xn recursive bound}
\end{aligned}$$ Secondly, using (c), $$\begin{aligned}
y_{n} = \frac{\mu((Y_n)^c\cap X_n \cap (\bigcap_{m<n} Y_m ))} {\mu(X_n \cap (\bigcap_{m<n} Y_m ))} &\leqslant \frac{\mu((V_n)^c \cap (\bigcap_{m<n} Y_m ))} {\mu(X_n \cap (\bigcap_{m<n} Y_m ))}\notag \\
&\leqslant \frac{\mu((V_n)^c ) } {\mu(X_n \vert (\bigcap_{m<n} Y_m ))} \qquad \textrm{ by $1$-independence}\notag \\
& \leqslant \frac{(1-f_{1,G}(p))} {1-x_n}. \label{equation: yn recursive bound}
\end{aligned}$$ Now if (\[equation: bounding equation\]) is satisfied, we claim that $x_n\leqslant \alpha f(p)$ for all $n$. Indeed $x_1=0$, and if $x_n< \alpha f_{1,G}(p)$, then by (\[equation: yn recursive bound\]) $$\begin{aligned}
y_n\leqslant\frac{1-f_{1,G}(p)}{1-\alpha f_{1,G}(p)}= 1- \frac{(1-\alpha)f_{1,G}(p)}{1-\alpha f_{1,G}(p)}<1. \nonumber \end{aligned}$$ Furthermore, we have by (\[equation: xn recursive bound\]) and (\[equation: bounding equation\]) that $$\begin{aligned}
x_{n+1}\leqslant \frac{(1-p)^{v(H)}}{(1-\alpha)f_{1,G}(p)}\leqslant \alpha f_{1,G}(p), \nonumber\end{aligned}$$ so our claim follows by induction.
Finally, we have that $$\begin{aligned}
\mu(Y_{\ell})= \prod_{i=1}^{\ell} (1-x_i)(1-y_i)&> \left(\frac{(1-\alpha)f_{1,G}(p)}{1-\alpha f_{1,G}(p)} \right)^{\ell}\left(1-\alpha f_{1,G}(p)\right)^{\ell}= \left( (1-\alpha) f_{1,G}(p)\right)^{\ell}.
\end{aligned}$$
For any finite connected graph $G$, $f_{1,G}(p)$ is a non-decreasing function of $p$ with $f_{1,G}(p)=1$. Thus the function $(f_{1,G}(p))^2$ is also non-decreasing in $p$ and attains a maximum value of $1$ at $p=1$. On the other hand, the function $4(1-p)^{v(G)}$ is strictly decreasing in $p$ and is equal to $4$ at $p=0$. Thus there exists a unique solution $p_{\star}=p_{\star}(G)$ in the interval $[0,1]$ to the equation $$\begin{aligned}
\label{equation: pstar}
(f_{1,G}(p))^2 = 4(1-p)^{v(G)}.\end{aligned}$$ Theorem \[theorem: crossing G times Pn\] thus has the following immediate corollary.
\[corollary: bound on long path constant in products with the line\] Let $G$ be a finite connected graph. Let $p_{\star}=p_{\star}(G)$ be as above. Then $$p_{1, \ell \mathit{p}}(\mathbb{Z}\times G) \leqslant p_{\star}.$$
Apply Theorem \[theorem: crossing G times Pn\] with $\alpha=1/2$.
Imaginary limits of real constructions: a preliminary lemma {#section: preliminary lemma}
===========================================================
In this section we prove a lemma that we shall use in Sections \[section: line lattice\] and \[section: complete graphs\]. The lemma will allow us to use certain vertex-based constructions to create other $1$-ipms that cannot be represented as vertex-based constructions (or would correspond to vertex-based constructions with ‘complex weights’).
\[lemma: preliminary lemma\] Let $G$ be a finite graph, and let $\mathcal{Q} := \{Q_{H}(\theta): H \subseteq G\}$ be a set of polynomials with real coefficients, indexed by subgraphs of $G$. Given $\theta \in \mathbb{C}$, let $\mu_{\theta}$ be the following function from subgraphs of $G$ to $\mathbb{C}$: $$\mu_{\theta}(H) := Q_{H}(\theta). \nonumber$$ Suppose there exists a non-empty interval $I \subseteq \mathbb{R}$ such that, for all $\theta \in I$, the function $\mu_{\theta}$ defines a $1$-ipm on $G$. Suppose further that there exists a set $X \subseteq \mathbb{C}$ such that, for all $\theta \in X$ and all $H \subseteq G$, $\mu_{\theta}(H)$ is a non-negative real number. Then $\mu_{\theta}$ is a $1$-ipm on $G$ for all $\theta \in X$.
We start by proving that $\mu_{\theta}$ is a measure on $G$ for all $\theta \in X$. As $\mu_{\theta}(H)$ is a non-negative real number for all $\theta \in X$ and all $H \subseteq G$, all that is left to prove is that $$\label{equation: 1-ipm check 1}
\left(\sum_{H \subseteq G} Q_{H}(\theta) \right) - 1 = 0.$$ The left hand side of (\[equation: 1-ipm check 1\]) is a polynomial in $\theta$ with real coefficients, and is equal to zero for all $\theta$ in the interval $I$. By the Fundamental Theorem of Algebra, the polynomial is identically zero and so (\[equation: 1-ipm check 1\]) holds for all $\theta$.
We now show that $\mu_{\theta}$ is a $1$-ipm on $G$ for all $\theta \in X$. To do this we must show that the following holds true for all $\theta \in X$, for all $A,B \subseteq V(G)$ such that $A$ and $B$ are disjoint, and all $G_{1}$ and $G_{2}$ such that $G_{1}$ is a subgraph of $G[A]$ while $G_{2}$ is a subgraph of $G[B]$: $$\label{equation: 1-ipm check 2}
%\mu_{\theta}\left(G\left[A\right] \cong G_{1},G\left[B\right] \cong G_{2}\right) = \mu_{\theta}\left(G\left[A\right] \cong G_{1}\right) \mu_{\theta}\left( G\left[B\right] \cong G_{2}\right).
\mu_{\theta}\left(\mathbf{G}_{\mu_{\theta}} \left[A\right] = G_{1},\mathbf{G}_{\mu_{\theta}} \left[B\right] = G_{2}\right) = \mu_{\theta}\left(\mathbf{G}_{\mu_{\theta}}\left[A\right] = G_{1}\right) \mu_{\theta}\left( \mathbf{G}_{\mu_{\theta}} \left[B\right] = G_{2}\right).$$ Both sides of (\[equation: 1-ipm check 2\]) are polynomials in $\theta$ with real coefficients — the left hand side, for example, can be written as $$\begin{aligned}
\sum_{H \subseteq G: \ H[A]=G_1, \ H[B]=G_2} Q_H(\theta).\end{aligned}$$ As $\mu_{\theta}$ is a $1$-ipm on $G$ for all $\theta \in I$, we have that these two polynomials agree on $I$, and so by the Fundamental Theorem of Algebra, they must be the same polynomial. Thus (\[equation: 1-ipm check 2\]) holds as required.
The line lattice $\mathbb{Z}$ {#section: line lattice}
=============================
In this section we prove Theorem \[theorem: connected function Pn\] on the connectivity function of paths. Recall that, given $n \in \mathbb{N}_{\geqslant 2}$ and $p \in [0,1]$, we let $\theta = \theta(p) := \frac{1+\sqrt{4p-3}}{2}$ and $p_{n} := \frac{1}{4}\left(3 - \tan^{2}\left(\frac{\pi}{n+1} \right) \right)$. Let $g_{n}(\theta):= \sum_{j = 0}^{n} \theta^{j}(1-\theta)^{n-j}$.
We begin by constructing a measure $\nu_{p} \in \mathcal{M}_{1,p}(P_{n})$ as follows. Let us start with the case $p \geqslant \frac{3}{4}$. For each vertex of $P_{n}$, we set it to state $0$ with probability $\theta$, and set it to state $1$ otherwise, and we do this independently for every vertex. Recall that for each $j \in [n]$ we write $S_j$ for the state of vertex $j$; in this construction, the states are independent and identically distributed random variables. We set the edge $\{j,j+1\}$ to be open if $S_j \leqslant S_{j+1}$, and closed otherwise. Thus, as $p = \theta + (1-\theta)^{2}$, we have that each edge is open with probability $p$. Moreover $(\mathbf{P}_{n})_{\nu_{p}}$ will be connected if and only if there exists some $j \in [n+1]$ such that $S_{k} = 0$ for all $k < j$, while $S_{k} = 1$ for all $k \geqslant j$. Therefore $(\mathbf{P}_{n})_{\nu_{p}}$ is connected with probability $g_{n}(\theta)$. As this construction is vertex-based, it is clear that it is $1$-independent.
When $p < \frac{3}{4}$ we have that $\theta$ is a complex number, and so the above construction is no longer valid. However, as discussed in Section \[section: preliminary lemma\], we will show that it is possible to extend this construction to all $p \in [p_{n},1]$. For each subgraph $G$ of $P_{n}$, set $Q_{G}(\theta)$ to be the polynomial $\nu_{p}((\mathbf{P}_n)_{\nu_{p}}=G)$ for all $\theta \in [\frac{3}{4},1]$. The following claim, together with Lemma \[lemma: preliminary lemma\], shows that in fact $\nu_{p}$ is a $1$-ipm on $P_{n}$ for all $p \in [p_{n},1]$.
\[claim: Pn book keeping\] For all $p \in [p_{n},\frac{3}{4})$ and all $G \subseteq P_{n}$ we have that $Q_{G}\left(\theta\left(p\right)\right)$ is non-negative real number.
We proceed by induction on $n$. When $n = 2$ we have that there are only two possible subgraphs of $P_{2}$, which are $P_{2}$ itself and its complement $\overline{P_{2}}$. We have that $Q_{P_{2}}(\theta(p)) = p$ and $Q_{\overline{P_{2}}}(\theta(p)) = 1-p$, so the claim holds as required for $n=2$.
Let us now assume that $n > 2$ and that the claim is true for all cases from $2$ up to $n-1$. We split into two further subcases. We first deal with the case that $G = P_{n}$. We have that $Q_{P_{n}}\left(\theta\left(p\right)\right) = g_{n}(\theta)$. For $p < \frac{3}{4}$ we can write $$\label{equation: g_n(theta)}
g_{n}(\theta) = \frac{\theta^{n+1} - (1-\theta)^{n+1}}{2\theta -1}.$$ When $p < \frac{3}{4}$ we have that $\theta$ and $1-\theta$ are complex conjugates, and also that $2\theta - 1$ is a pure imaginary number. Thus both the numerator and denominator of the above fraction are pure imaginary, and so $g_{n}\left(\theta\left(p\right)\right)$ is a real number for all $p < \frac{3}{4}$. By writing $\theta = re^{i \phi}$, where $r := \sqrt{1-p}$ and $\phi := \arctan \left(\sqrt{3-4p}\right)$, we can rewrite (\[equation: g\_n(theta)\]) as $$g_{n}\left(\theta\left(p\right)\right) = \frac{2 r^{n+1}}{\sqrt{3-4p}}\sin \left(\left(n+1\right) \phi \right). \nonumber$$ From this equation we can see that $g_{n}\left(\theta\left( p \right)\right) = 0$ if and only if $\left(n+1\right) \phi = m \pi$ for some $m \in \mathbb{Z}$, which occurs if and only if $p = \frac{1}{4}\left(3- \tan^{2} \left( \frac{m \pi}{n+1} \right) \right)$ for some $m \in \mathbb{Z}$. Therefore the largest value of $p$ in $[0,\frac{3}{4})$ for which $g_{n}\left(\theta\left( p \right)\right) = 0$ is $p = p_{n}$, and this occurs when $m \equiv \pm 1 \!\mod (n+1)$. Let $p_{\star} := \frac{1}{4}\left(3 - \tan^{2}\left(\frac{\pi}{2(n+1)} \right) \right)$ and $\phi_{\star} := \arctan \left(\sqrt{3-4p_{\star}}\right)$. We have that $p_{n} < p_{\star} < \frac{3}{4}$ and $\sin \left(\left(n+1\right) \phi_{\star}\right) = 1$, and so $g_{n}\left(\theta\left( p_{\star} \right)\right) > 0$. Thus $g_{n}\left(\theta\left( p \right)\right)$ is non-negative for all $p$ in the interval $[p_{n},\frac{3}{4})$ as required.
We now deal with the case that $G \neq P_{n}$. Let us consider the vertex-based construction from which $Q_{G}(\theta)$ was defined. As not every edge is present in $G$ we have that there exists some $j \in [n-1]$ such $\{j,j+1\}$ is not an edge, and so $S_{j} = 1$ while $S_{j+1} = 0$. Note that if $j \geqslant 2$, then the edge $\{j-1,j\}$ is present in $G$ regardless of the state of vertex $j-1$. Similarly, if $j \leqslant n-2$, then the edge $\{j+1,j+2\}$ is present in $G$ regardless of the state of vertex $j+2$. If we write $G_{1} := G\left[\{1,\ldots,j-1\}\right]$ and $G_{2} := G\left[\{j+2,\ldots,n\}\right]$, then we have that $$\label{equation: polynomials}
Q_{G}(\theta) = \theta(1-\theta)Q_{G_{1}}(\theta)Q_{G_{2}}(\theta).$$ Now, by induction, we have that $Q_{G_{1}}\left( \theta\left( p \right)\right)$ and $Q_{G_{2}}\left( \theta \left( p \right) \right)$ are positive real numbers for all $p \in [p_{n},\frac{3}{4})$; to make this inductive step work we are using the fact that $(p_{n})_{n \geqslant 2}$ forms an increasing sequence, and so $p \geqslant p_{n}$ implies that $p \geqslant p_{s}$ for all $s \leqslant n$. As $\theta\left(p\right)\left(1-\theta\left(p\right)\right) = 1-p$, we have that (\[equation: polynomials\]) is a positive real for all $p \in [p_{n},\frac{3}{4})$, and so we have proven the claim.
Note that as this proof shows that $g_{n}\left(\theta \left(p_{n}\right)\right) = 0$, we have that the probability $(\mathbf{P}_{n})_{\nu_{p_{n}}}$ is connected is equal to $0$. As $\nu_{p_{n}} \in \mathcal{M}_{1,p}(P_{n})$ for all $p \leqslant p_{n}$, we have that $f_{1,P_{n}}(p) = 0$ for all $p \leqslant p_{n}$.
We now prove that this construction is optimal with respect to the connectivity function. Note that the following proof involves essentially following the proof of Theorem \[theorem: crossing G times Pn\] when $G$ consists of a single point and checking that the above construction is tight at every stage of this proof.
The above construction discussed shows that $$f_{1,P_{n}}(p) \leqslant \begin{cases} g_{n}(\theta) &\text{ for } p \in [p_{n},1] , \\
0 &\text{ for } p \in [0,p_{n}]. \end{cases} \nonumber$$ It is clear that $f_{1,P_{n}}(p) \geqslant 0$ for all $p$, and so all that remains to show is that $f_{1,P_{n}}(p) \geqslant g_{n}\left(\theta\left(p\right)\right)$ for all $p \in [p_{n},1]$.
Let $\mu \in \mathcal{M}_{1,p}(P_{n})$. For $k \in [n]$, let $X_{k}$ be the event that the subgraph of $(\mathbf{P}_{n})_{\nu_{p}}$ induced by the vertex set $[k]$ is connected, and let $H_{k}$ be the event that the edge $\{k-1,k\}$ is not present in $(\mathbf{P}_{n})_{\nu_{p}}$. Let $q_{2}^{\mu} := 1-p$, and for $k > 2$ let $q_{k}^{\mu} := \mu\left(\left(X_{k}\right)^{c}| X_{k-1}\right)$. We have that $$\begin{aligned}
q_{k}^{\mu} &=& \frac{\mu(H_{k} \cap X_{k-1})}{\mu(X_{k-1})}\nonumber \\
& \leqslant & \frac{\mu(H_{k} \cap X_{k-2})}{\mu(X_{k-1})} \label{equation: qn 1}\\
&=& \frac{\mu(H_{k})\mu(X_{k-2})}{\mu(X_{k-1})} \qquad \textrm{ by $1$-independence} \nonumber \\
&\leqslant & \frac{1-p}{1-q_{k-1}^{\mu}}.\label{equation: qn 2}\end{aligned}$$ Note that $\mu((\mathbf{P}_{n})_\mu = P_{n}) = \prod_{j = 2}^{n}(1-q_{n}^{\mu})$. Thus to show that the previous construction is optimal with respect to the connectivity function it is enough to show that equality holds for inequalities (\[equation: qn 1\]) and (\[equation: qn 2\]) when $\mu = \nu_{p}$. In the measure $\nu_{p}$, we have that every edge is present with probability exactly $p$, thus $\nu_{p}(H_{k}) = 1-p$ and so equality holds in (\[equation: qn 2\]). To prove that equality holds in (\[equation: qn 1\]), it is sufficient so show that $$\label{equation: qn 3}
\mu(H_{k} \cap X_{k-1}) = \mu(H_{k} \cap X_{k-2}).$$ Both the left and right hand sides of (\[equation: qn 3\]) can be expressed as polynomials in $\theta(p)$, and so it is sufficient to show that equality holds for $p \geqslant \frac{3}{4}$, as that will show they are the same polynomial (and so equality holds for all $p \in [p_{n},1])$. Suppose that the event $(H_{k} \cap X_{k-2})$ occurs. As $H_{k}$ has occurred we have that $S_{k-1} = 1$ while $S_{k} = 0$. As $S_{k-1} = 1$, we have that edge $\{k-2,k-1\}$ is open, regardless of $S_{k-2}$. Thus, as $ X_{k-2}$ has occurred we also have that $X_{k-1}$ has occurred. Therefore $(H_{k} \cap X_{k-1})$ has also occurred, and so we are done.
We remark in similar fashion to the above proof, we have that the following holds for any $\mu \in \mathcal{M}_{1,p}(P_{n})$: $$\begin{aligned}
\mu(X_{n}) &\geqslant& \mu(X_{n-1}) - \mu(X_{n-1} \cap H_{n}) \nonumber \\
&\geqslant& \mu(X_{n-1}) - \mu(X_{n-2} \cap H_{n}) \nonumber \\
&=& \mu(X_{n-1}) - \mu(X_{n-2})\mu(H_{n}^{c}) \qquad \textrm{ by $1$-independence}\nonumber \\
&\geqslant& \mu(X_{n-1}) - (1-p)\mu(X_{n-2}). \nonumber\end{aligned}$$ Moreover, by once again considering what states of vertices can lead to the various events, we have that equality holds for all of the above inequalities when $\mu = \nu_{p}$. This leads us to another way to define $g_{n}\left(\theta \left(p\right)\right)$: let $g_{2}\left(\theta \left(p\right)\right) := p$, $g_{3}\left(\theta \left(p\right)\right) := 2p-1$, and for all $n \geqslant 4$ we have that $$g_{n}\left(\theta \left(p\right)\right) = g_{n-1}\left(\theta \left(p\right)\right) - (1-p)g_{n-2}\left(\theta \left(p\right)\right). \nonumber$$
The ladder lattice $\mathbb{Z}\times K_2$ {#section: ladder}
=========================================
In this section we construct a family of $1$-ipms on segments of the ladder $\mathbb{Z}\times P_2$ with edge-probability close to $2/3$ for which with probability $1$ there are no open left-right crossings. The idea of this construction is due to Walters and the second author [@FalgasRavry12] (though the technical work involved in rigorously showing the construction works is new).
Let $p$ be fixed with $1/2<p<2/3$. We define sequences $\left(p_n\right)_{n \in \mathbb{N}}$, $\left(r_n\right)_{n\in \mathbb{N}}$ and $\left(s_n\right)_{n \in \mathbb{N}}$ iteratively as follows. We begin by setting $p_1:=r_1:=1$ and $s_1:=0$. Then for $n \in \mathbb{N}$, we let $$\begin{aligned}
p_{n+1}&:=\left\{\begin{array}{ll}
1-\frac{1-p}{r_n+s_n} & \textrm{if }r_n+s_n\geqslant 1-p,\\
0 & \textrm{otherwise;}
\end{array}\right. \\
r_{n+1}&:=\left\{\begin{array}{ll}
1-\frac{1-p}{p_n} & \textrm{if }p_n\geqslant 1-p,\\
0 & \textrm{otherwise;}
\end{array}\right. \\
s_{n+1}&:=\left\{\begin{array}{ll}
\max\left\{1-2r_{n+1}+\frac{r_{n+1}-(1-p)}{p_{n+1}}, 0 \right\} & \textrm{if }p_{n+1}>0,\\
0 & \textrm{otherwise.}
\end{array}\right.\end{aligned}$$
\[lemma: sequences work\]The following hold for all $n\in \mathbb{N}$:
(i) $p_n, r_n\in [0,1]$,
(ii) $s_n \in [0, 1-r_n]$,
(iii) $p_{n+1}\leqslant p_n$,
(iv) $r_{n+1}\leqslant r_n$,
(v) $r_{n+1}+s_{n+1}\leqslant r_n+s_n$.
We prove the lemma by induction on $n$. By definition of our sequences, $p_1=r_1=1\geqslant p=p_2=r_2$, $s_1=0$, and $0<s_2=\frac{(2p-1)(1-p)}{p}< 1-p=r_1+s_1-r_2$, and thus (i)–(v) all hold in the base case $n=1$.
Suppose now (i)–(v) hold for all $n \leqslant N$, for some $N \geqslant 1$. Since $p_N$ and $r_N+s_N$ both lie in $[0,1]$, the definition of $p_{N+1}$ and $r_{N+1}$ implies these also both lie in $[0,1]$. This establishes (i) for $n=N+1$. By construction, $s_{N+1}\geqslant 0$, and by the inductive hypotheses (ii) and (v), we have $$\begin{aligned}
s_{N+1}\leqslant r_{N}+s_N-r_{N+1}\leqslant 1-r_{N+1},
\end{aligned}$$ whence (ii) holds for $n=N+1$.
If $p_{N+2}=0$, then $p_{N+2}\leqslant p_{N+1}$ trivially holds (since $p_{N+1}\geqslant 0$ by (i)). On the other hand, suppose $p_{N+2}=1-\frac{1-p}{r_{N+1}+s_{N+1}}>0$. Then we have $r_{N+1}+s_{N+1}> 1-p$, which by our inductive hypothesis (v) implies $r_N+s_N\geqslant r_{N+1}+s_{N+1}> 1-p$. The definition of $p_{N+1}$ then implies $$\begin{aligned}
p_{N+2}=1-\frac{1-p}{r_{N+1}+s_{N+1}} \leqslant 1-\frac{1-p}{r_{N}+s_{N}}=p_{N+1},
\end{aligned}$$ as desired, establishing that (iii) holds for $n=N+1$. Arguing in exactly the same way (using the inductive hypothesis (iii) instead of (v)), we obtain that $r_{N+2}\leqslant r_{N+1}$. Hence (iv) holds for $n=N+1$.
Finally we consider (v) for $n=N+1$, which is the most delicate part of the induction. For convenience, set $t_n:=r_n+s_n$. If $p_{N+2}=0$, then by (iv) which we have just established and the definition of $s_{N+2}$, we have that $t_{N+2}=r_{N+2}\leqslant r_{N+1}\leqslant t_{N+1}$, and so (v) holds in this case. Also if $p_{N+2}>0$ and $r_{N+2}=0$, then by construction $p_{N+1}<1-p$, and thus (iii) that we have established above then tells us that $p_{N+2}\leqslant p_{N+1}\leqslant 1-p$. The definition of $s_{N+2}$ then gives $$\begin{aligned}
t_{N+2}=0+s_{N+2}=\max\{1- \frac{1-p}{p_{N+2}}, 0\}=0\leqslant t_{N+1},
\end{aligned}$$ and (v) holds in this case also. So finally, let us assume that $p_{N+2}$ and $r_{N+2}$ are both strictly positive.
\[claim: intermediate result — s\_i >0 for all smaller i\] If $p_{N+2}$ and $r_{N+2}$ are both strictly positive, then for all $i\in \{2,\ldots,N+2\}$, we have that $s_i>0$.
Fix $i\in \{2,\ldots,N+2\}$. By Lemma \[lemma: sequences work\](iii)–(iv) and since $i\geqslant 2$, we have $0 < p_{N+2}\leqslant p_i\leqslant p_2=p$ and $0 < r_{N+2} \leqslant r_i\leqslant r_2=p$. Since $r_{i+1}>0$, we in fact have $p_{i}>1-p$. We also have that $$\begin{aligned}
1-2r_i+\frac{r_i-(1-p)}{p_i}=\frac{1}{p_i}\left(p_i +r_i -2r_ip_i -(1-p)\right)=:\frac{1}{p_i} f(p_i, r_i).
\end{aligned}$$ Now for fixed $y\in [1/2, 1]$, the function $x\mapsto f(x,y)$ is a non-increasing function of $x$. Thus if $r_i\geqslant 1/2$, we have $f(p_i, r_i)\geqslant f(p, r_i)= (2p-1)(1-r_i)>0$. On the other hand for fixed $y\in (0,1/2)$, the function $x\mapsto f(x,y)$ is strictly increasing in $x$. Therefore if $r_i<1/2$, we have $f(p_i, r_i)> f(1-p, r_i)= r_i(2p-1)>0$. In either case, $f(p_i, r_i)>0$, and thus $s_i=\max\left(\frac{1}{p_i}f(p_i, r_i), 0 \right)>0$.
We thus have, for all $i \in [N+1]$ that $$\begin{aligned}
\label{eq: expressions for pi, ri, si}
p_{i+1}=1-\frac{1-p}{r_{i}+s_{i}},&&r_{i+1}=1-\frac{1-p}{p_{i}},
&&s_{i+1}&=1-2r_{i+1}+\frac{r_{i+1}-(1-p)}{p_{i+1}}.
\end{aligned}$$ Combining these equations we obtain for $i\in \{2,\ldots,N+1\}$ that: $$\begin{aligned}
\label{eq: p_{i+1} as a function of p_i, p_{i-1}}
p_{i+1}=1-\frac{p_{i}p_{i-1}(1-p)}{p(p_{i-1}-p_{i}+1)+p_{i}-1}.
\end{aligned}$$
\[claim: intermediate result — p\_[i+1]{} as function of p\_i for all smaller i\] For all integers $i\in [N+1]$, we have $$\begin{aligned}
p_{i+1}=\frac{p_{i}-(1-p)}{(2-p)p_{i}-(1-p)}.
\end{aligned}$$
Since $p_1=1$ and $p_2=p$, our claim holds for $i=1$. Suppose it holds for some $i\leqslant N$. Then by rearranging terms, we have $$\begin{aligned}
p_{i}= \frac{(1-p)(1-p_{i+1})}{1- (2-p)p_{i+1}}.\end{aligned}$$ Substituting this into the formula for $p_{i+2}$ given by (\[eq: p\_[i+1]{} as a function of p\_i, p\_[i-1]{}\]), we see our claim holds for $i+1$ as well.
It follows from Claim \[claim: intermediate result — p\_[i+1]{} as function of p\_i for all smaller i\] and (\[eq: expressions for pi, ri, si\]) that for all $i\in [N+1]$, we can write $t_{i+1}$ as a function $p_{i}$: $$\begin{aligned}
\label{eq: t_{i+1 formula}}
t_{i+1}&=& 1-r_{i+1}+\frac{r_{i+1}-(1-p)}{p_{i+1}} \nonumber \\
&=&\frac{p_{i}p(2-p)-(1-p)}{p_{i}-(1-p)}\nonumber \\
&=&p(2-p)-\frac{(1-p)^3}{p_i-(1-p)}.\end{aligned}$$ For $\frac{1}{2}<p<\frac{2}{3}$ fixed and $p_{i}>(1-p)$ (which we recall holds since $r_{i+1}>0$), the expression above is an increasing function of $p_{i}$. By our inductive hypothesis (iii) that $p_{N+1} \leqslant p_{N}$ it follows that $t_{N+2} \leqslant t_{N+1}$ and we have verified that (v) holds for $n=N+1$.
\[lemma: ladder construction is 2-inaccessible if suff far\] Suppose $p= \frac{2}{3}-\varepsilon$, for some $\varepsilon \in (0, \frac{1}{6})$. We have that $p_n=r_n=s_n=0$ for all $n\geqslant N_{\varepsilon}$, where $N_{\varepsilon}:= \lceil 2\varepsilon^{-1}\rceil$.
Recall that, as in the proof of Lemma \[lemma: sequences work\], we set $t_n:=s_n+r_n$. Suppose first that there exists $m\in [N_{\varepsilon}-2]$ such that $r_{m}=0$. Then $p_{m+1}=0$ and $s_{m+1}=0$ by construction and $r_{m+1}=0$ by Lemma \[lemma: sequences work\](iv). Lemma \[lemma: sequences work\](iii)-(v) then implies $p_{n}=r_{n}=s_{n}=0$ for all $n\in \mathbb{N}_{\geqslant m+1}$, as required.
Suppose instead that $r_{n} > 0$ for all $n \in [N_{\varepsilon}-2]$ and there exists some $m\in [N_{\varepsilon}-2]$ such that $p_{m}\leqslant 1-p$. Then $r_{m+1}=0$, and thus by the argument above, we have that $p_{n}=r_{n}=s_{n}=0$ for all $n\in \mathbb{N}_{\geqslant m+2}$, as required.
Finally, suppose $p_n>1-p$ and $r_n>0$ both hold for all $n \in [N_{\varepsilon}-2]$. By Claim \[claim: intermediate result — s\_i >0 for all smaller i\], we have $s_n>0$ for all $n\in \{2,\ldots,N_{\varepsilon}-2\}$. This allows us in turn to apply Claim \[claim: intermediate result — p\_[i+1]{} as function of p\_i for all smaller i\] to all $n$ in this interval and to deduce that $$\begin{aligned}
\label{ineq: p_n -p_{n-1} less than -eps}
p_{n-1}- p_{n}&=&p_{n-1}-\frac{p_{n-1}-(1-p)}{(2-p)p_{n-1}-(1-p)} \nonumber\\
&=&\frac{1}{(2-p)p_{n-1}-(1-p)}\left((2-p) \left(p_{n-1}-\frac{1}{2}\right)^2+\frac{2-3p}{4} \right) \nonumber \\
&\geqslant& \frac{3\varepsilon}{4}.
\end{aligned}$$ Recall that $p_1=1$. As such, it follows from inequality (\[ineq: p\_n -p\_[n-1]{} less than -eps\]) that $p_{n} \leqslant 1 - (n-1)\frac{3\varepsilon}{4}$ for all $n\in [N_{\varepsilon} -2 ]$. As $N_{\varepsilon}=\lceil 2 \varepsilon^{-1}\rceil >3 +\frac{8}{9}\varepsilon^{-1}$ and $p < \frac{2}{3}$, we have that there exists some $n\in [N_{\varepsilon}-2]$ such that $p_{n} < 1-p$, which is a contradiction.
\[theorem: ladder construction\] Fix $p\in (\frac{1}{2}, \frac{2}{3})$. Then there exists $N\in \mathbb{N}$ such that for all $n \geqslant N$, $$p_{1, \times}(P_n\times P_2)\geqslant p.$$
Given $p=2/3-\varepsilon$, with $\varepsilon \in (0, \frac{1}{6})$ fixed, we construct the sequences $(p_n)_{n\in \mathbb{N}}, (r_n)_{n\in \mathbb{N}}, (s_n)_{n\in \mathbb{N}}$ as above and let $N=N_{\varepsilon}$ be the integer constant whose existence is given by Lemma \[lemma: ladder construction is 2-inaccessible if suff far\]. We construct a $1$-ipm $\mu \in \mathcal{M}_{1,p}\left(P_{N}\times P_2\right)$ as follows. We write the vertex set $V(P_{N} \times P_{2})$ as $[N] \times [2]$. As in the case of the line lattice, for each vertex $(n,y)$ we assign it a state $S_{(n,y)}$, which we can think of as an independent random variable, as follows. If $n+y$ is even, then $$\begin{aligned}
S_{(n,y)}:=\begin{cases}
2 & \textrm { with probability }p_n,\\
0 & \textrm { with probability }1-p_n.
\end{cases}
\end{aligned}$$ If instead $n+y$ is odd, then $$\begin{aligned}
S_{(n,y)}:=\begin{cases}
2 & \textrm { with probability }r_n,\\
1 & \textrm{ with probability }s_n,\\
0 & \textrm { with probability }1-r_n-s_n.
\end{cases}
\end{aligned}$$ We observe here that by Lemma \[lemma: sequences work\](i)–(ii), these are well-defined probability distributions. Now we give the conditions for which edges are open:
- for each $n\in [N-1]$ and $y\in [2]$, the horizontal edge $\{(n,y),(n+1, y)\}$ is open if and only if $S_{(n,y)}\leqslant S_{(n+1, y)}$,
- for each $n\in [N]$, the vertical edge $\{(n,1), (n,2)\}$ is open if and only if $\left(S_{(n,1)}-S_{(n,2)}\right)(1-S_{(n,1)})(1-S_{(n,2)})=0$.
Note the condition for a vertical edge $\{(n,1), (n,2)\}$ to be open can be rephrased as if and only if either $S_{(n,1)}=S_{(n,2)}$ or one of $S_{(n,1)}, S_{(n,2)}$ is equal to $1$. So intuitively, the value of the $S_{(n,y)}$ must increase from left to right along horizontal edges, and it must stay constant along vertical edges unless one of the endpoints is in the special state $1$ which allows free passage up or down. Clearly $\mu$ is a $1$-independent bond percolation model on $G:=P_N\times P_2$ as it is vertex-based.
\[claim: ladder measure has right edge probability\] We have that $d(\mu)\geqslant p$.
For $(n,y)\in [N-1]\times [2]$, consider the horizontal edge $\{(n,y), (n+1,y)\}$, . If $n+y$ is even, then by definition of $r_{n+1}$, $$\begin{aligned}
\mu\left(\{(n,y), (n+1,y)\}\in \mathbf{G}_\mu \right)&=\mu \left(S_{(n,y)}\leqslant S_{(n+1,y)}\right)=r_{n+1}+(1-r_{n+1})(1-p_n)\geqslant p.
\end{aligned}$$ Similarly if $n+y$ is odd, then by definition of $p_{n+1}$, $$\begin{aligned}
\mu \left(\{(n,y), (n+1,y)\}\in \mathbf{G}_\mu \right)&=\mu \left(S_{(n,y)}\leqslant S_{(n+1,y)}\right)=p_{n+1}+(1-p_{n+1})(1-r_n-s_n)\geqslant p.
\end{aligned}$$ Finally, for a vertical edge $\{(n,1), (n,2)\}$, $n\in [N]$, we have $$\begin{aligned}
\mu \left(\{(n,1), (n,2)\}\in \mathbf{G}_\mu \right)&=\mu \left(S_{(n,1)}=S_{(n,2)} \textrm{ or } 1\in \{S_{(n,1)}, S_{(n,2)} \}\right)\\
&=s_n + p_{n}r_{n}+ (1-p_{n})(1-r_{n}-s_{n}).
\end{aligned}$$ Now, if $p_n=0$, then $r_{n-1}\leqslant 1-p$ by definition of $p_n$, whence $r_n\leqslant 1-p$ by Lemma \[lemma: sequences work\](iv), and so the expression above equals $1-r_n \geqslant p$. On the other hand if $p_n\neq 0$, then by definition of $s_n$ the expression above is at least $p$. Thus each horizontal edge and each vertical edge is open in $\mathbf{G}_\mu$ with probability at least $p$, and $d(\mu)\geqslant p$ as claimed.
There is no open path in $\mathbf{G}_\mu$ from $\{1\}\times [2]$ to $\{N\}\times [2]$.
By construction, $p_1=r_1=1$, whence $S_{(1,1)}=S_{(1,2)}=2$. Furthermore, by Lemma \[lemma: ladder construction is 2-inaccessible if suff far\] and our choice of $N$, $p_N=r_N=s_N=0$, whence $S_{(N,1)}=S_{(N,2)}=0$.
Let $N'$ be the largest $n\in [N]$ for which there exists an open path in $\mathbf{G}_\mu$ from $\{1\}\times [2]$ to $\{n\}\times [2]$. Let $\mathcal{P}$ be such a path, and let $v_0\in \{1\}\times [2]$, $v_1\in \{2\}\times[2]$, $v_2$, ... , $v_{\ell} \in \{n\}\times [2]$ be the vertices of $\mathcal{P}$ traversed from left to right. Observe that in this ordering of the vertices of $\mathcal{P}$, every horizontal edge $\{(n,y),(n+1,y)\}$ of $\mathcal{P}$ is traversed from left to right.
We claim that for all $i\in [\ell]$, we have $S_i\in\{1,2\}$. Indeed, by construction $S_{v_0}=2$. Suppose there exists some $1\leqslant i<\ell$ such that $S_{v_j}\in\{1,2\}$ for all $j < i$. If $S_{v_i}=2$, then the edge $v_iv_{i+1}$ can be open in $\mathbf{G}_\mu$ only if $S_{v_{i+1}}\in \{1,2\}$. What is more, $S_{v_{i+1}}$ can be equal to $1$ if and only if $v_iv_{i+1}$ is a vertical edge. On the other hand, suppose $S_{v_i}=1$. Then $v_{i-1}v_i$ was a vertical edge (since there is no edge both of whose endpoints are in state $1$ and since horizontal edges are traversed from left to right by $\mathcal{P}$), and so $v_{i+1}=v_i+(1,0)$. But then $v_iv_{i+1}$ open in $\mathbf{G}_\mu$ implies $S_{v_{i+1}}=2$. Thus for every vertex $v_i$ of $\mathcal{P}$, we have that $S_{v_i}$ is indeed in state $1$ or $2$.
This implies in particular that $v_{\ell}\notin \{N\}\times [2]$ (since as we remarked above $S_{(N,1)}=S_{(N,2)}=0$). Thus there is no open path in $\mathbf{G}_\mu$ from $\{1\}\times [2]$ to $[N]\times [2]$.
Thus $\mu$ is an element of $\mathcal{M}_{1,p}(P_N\times [2])$ for which $$\mu(\exists\textrm{ open left-right crossing})=0.$$ Given $n\geqslant N$, we may extend $\mu$ to an element $\mu'\in \mathcal{M}_{1,p}(P_n\times P_2)$ by letting every edge in $P_n\times P_2\setminus P_N\times P_2$ be open independently at random with probability $p$. In this way we obtain a $1$-independent bond percolation measure $\mu'$ on $P_n\times P_2$ with edge-probability $p$ for which there almost surely are no open left–right crossings of $P_n\times P_2$, giving the required lower bound on $p_{1, \times}(P_n\times P_2)$.
Trivially, the $1$-independent connectivity function of the path on $2$ vertices $P_2$ (i.e. the graph consisting of a single edge) is $f_{1,P_2}(p)=p$. Thus the constant $p_{\star}(P_2)$ defined by equation (\[equation: pstar\]) is the unique solution in $[0,1]$ to the equation $x^2=4(1-x)^2$, namely $p_{\star}(P_2)=\frac{2}{3}$. By Corollary \[corollary: bound on long path constant in products with the line\], this implies $p_{1, \ell \mathit{p}}(\mathbb{Z}\times P_2)\leqslant \frac{2}{3}$.
For the lower bound, fix $p \in (\frac{1}{2},\frac{2}{3})$. In the proof of Theorem \[theorem: ladder construction\], we showed there exist some integer $N\in \mathbb{N}$ and $\mu \in \mathcal{M}_{1,p}(P_N\times P_2)$ such that
(i) $\mu(\exists\textrm{ open left-right crossing})=0$;
(ii) $\mu\left(\{(1,1), (1,2)\} \textrm{ and } \{(N,1), (N,2)\} \textrm{ are open}\right)=1$.
We use this measure to create a measure $\nu \in \mathcal{M}_{1,p}(\mathbb{Z} \times P_2)$. Let $G:=\mathbb{Z} \times P_2$. For each $i\in \mathbb{Z}$, we let the subgraphs $\mathbf{G}_{\nu}[ (i(N-1)+[N])\times [2]]$ on horizontal shifts of the ladder $P_N\times P_2$ by $i(N-1)$ be independent identically distributed random variables with distribution given by $\mu$. Thanks to property (ii) recorded above, the random subgraphs agree on the vertical rungs $\{1+i(N-1)\}\times P_2$ of the ladder, and this gives rise to a bona fides $1$-independent model $\nu$ on $\mathbb{Z} \times P_2$ with edge-probability $p$. Furthermore, property (i) implies all connected components in $\mathbf{G}_{\nu}$ have size at most $4(N-1)-2=4N-6$. In particular, $p_{1, \ell}(\mathbb{Z}\times P_2)\geqslant p$. Since $p<\frac{2}{3}$ was chosen arbitrarily, this gives the required lower bound $p_{1, \ell}(\mathbb{Z}\times P_2)\geqslant \frac{2}{3}$.
Complete graphs {#section: complete graphs}
===============
In this section we will prove Theorem \[theorem: connected function Kn\]. Recall that, given $ n \in \mathbb{N}_{\geqslant 2}$ and $p \in [0,1]$, we let $\theta = \theta(p) := \frac{1 + \sqrt{2p - 1}}{2}$ and $p_{n} := \frac{1}{2}(1 - \tan^{2}(\frac{\pi}{2n}))$. Let $g_{n}(\theta) := \theta^{n} + (1-\theta)^{n}$.
An upper bound for $f_{1, K_n}(p)$
----------------------------------
Before proving Theorem \[theorem: connected function Kn\], let us give a simple vertex-based construction of a measure $\nu_{p} \in \mathcal{M}_{1,p}(K_{n})$ that shows $f_{1,K_{n}}(p) \leqslant g_{n}(\theta)$ for $p \geqslant \frac{1}{2}$. We call this measure the *Red-Blue construction*. We think of $K_{n}$ as the complete graph on vertex set $[n]$, and we colour each vertex Red with probability $\theta$ and colour it Blue otherwise, and we do this independently for all vertices. The edge $\{i,j\} \in [n]^{(2)}$ is open if and only if $i$ and $j$ have the same colour. As $p = \theta^{2} + (1- \theta)^{2}$, we have that each edge is present in $(\mathbf{K}_{n})_{\nu_{p}}$ with probability $p$. Note that $(\mathbf{K}_{n})_{\nu_{p}}$ will either be either a disjoint union of two cliques, in which case it is disconnected, or the complete graph $K_{n}$, in which case it is connected. This latter case occurs if and only if every vertex receives the same colour, and so the probability that $(\mathbf{K}_{n})_{\nu_{p}}$ is connected is equal to $g_{n}(\theta)$. As this construction is vertex-based, it is clear that it is $1$-independent.
If $p < \frac{1}{2}$ then $\theta$ is a complex number, and so the Red-Blue construction is no longer valid. However, as discussed in Section \[section: preliminary lemma\], we will show that it is possible to extend this construction to all $p \in [p_{n},1]$. Given $j \in \{0,1,\ldots,n\}$, let $$g_{n,j}(\theta) := \theta^{j}(1-\theta)^{n-j} + \theta^{n-j}(1-\theta)^{j}. \nonumber$$ When $j = 0$ or $j = n$ we have that $g_{n,0}(\theta)$ and $g_{n,n}(\theta)$ are each equal to $g_{n}(\theta)$, and so we just write the latter instead. Given some $A \subseteq [n]$, let $H_{A}$ be the disjoint union of a clique on $A$ with a clique on $[n] \setminus A$. Note that when $A = \emptyset$ or $[n]$ we have that $H_{A}$ is equal to $K_{[n]}$, and more generally that $H_{A} = H_{[n]\setminus A}$. For $p \in [0,1]$, let $\mu_{p}$ be the following function on subgraphs $G$ of $K_{n}$: $$\mu_{p}\left( G \right) := \begin{cases} g_{n,|A|}\left( \theta\left(p \right)\right) &\text{ if } G = H_{A} \text{ for some } A \subseteq [n] , \\
0 &\text{ else}. \end{cases} \nonumber$$ For $p \in [\frac{1}{2},1]$ this function matches the Red-Blue construction given above, and so by defining $\nu_{p}((\mathbf{K}_n)_{\nu_p}=G):=\mu_p(G)$ for all subgraphs $G \subseteq K_n$, we obtain a measure $\nu_p$ which is a $1$-ipm defined without making reference to states of vertices. The following claim, together with Lemma \[lemma: preliminary lemma\], shows that in fact $\nu_{p}$ is a $1$-ipm on $K_{n}$ for all $p \in [p_{n},1]$.
\[claim: Kn book keeping\] For all $p \in [p_{n},\frac{1}{2}]$ and all $j \in \{0,\ldots,n\}$ we have that $g_{n,j}\left(\theta\left(p\right)\right)$ is non-negative real number.
Let us begin with the case $j = n$. As $p \leqslant \frac{1}{2}$, we have that $\theta$ and $1-\theta$ are complex conjugates, and so $g_{n}\big(\theta(p)\big)$ is a real number for all $p$ in this range. By writing $\theta = re^{i \phi}$, where $r := \sqrt{\frac{1-p}{2}}$ and $\phi := \arctan \left(\sqrt{1-2p}\right)$, we can write $$g_{n}\left(\theta\left(p\right)\right) = 2 r^{n} \cos \left(n \phi \right). \nonumber$$ From this equation we can see that $g_{n}\left(\theta\left(p\right)\right) = 0$ if and only if $n \phi = \frac{\pi}{2} + m \pi$ for some $m \in \mathbb{Z}$, which occurs if and only if $p = \frac{1}{2}\left(1-\tan^{2} \left(\frac{\pi}{2n} + \frac{m \pi}{n} \right)\right)$ for some $m \in \mathbb{Z}$. Therefore the largest value of $p$ for which $g_{n}\left(\theta\left(p\right)\right) = 0$ is $p = p_{n}$, and this occurs when $m \equiv -1 \text{ or } 0 \!\mod n$. As $g_{n}\left(\theta\left(\frac{1}{2}\right)\right) > 0$, we have that $g_{n}\left(\theta\left(p\right)\right)$ is a non-negative real number for all $p \in [p_{n},\frac{1}{2}]$, which proves the claim when $j = n$. For general $j \in \{0,\ldots,n\}$, we have that $$\label{equation: Kn book keeping}
g_{n,j}\left(\theta\left(p \right)\right) = \begin{cases}
\left(\frac{1-p}{2}\right)^{j} g_{n-2j} \left(\theta \left(p \right)\right)&\text{ if } n \geqslant 2j,\\
\left(\frac{1-p}{2}\right)^{n-j} g_{2j-n} \left(\theta \left(p \right)\right)&\text{ if } n \leqslant 2j. \end{cases} \nonumber$$ Therefore the previous case of the claim shows that $g_{n,j}\left(\theta\left(p \right)\right) \in [0,1]$ for all $p \in [p_{n},\frac{1}{2}]$; at this stage we are using the fact that $(p_{n})_{n \geqslant 2}$ forms an increasing sequence, and so $p \geqslant p_{n}$ implies that $p \geqslant p_{s}$ for all $s \leqslant n$.
Note that as this proof shows that $g_{n}\left(\theta \left(p_{n}\right)\right) = 0$, we have that the probability $(\mathbf{K}_{n})_{\nu_{p_{n}}}$ is connected is equal to $0$. As $\nu_{p_{n}} \in \mathcal{M}_{1,p}(K_{n})$ for all $p \leqslant p_{n}$, we have that $f_{1,K_{n}}(p) = 0$ for all $p \leqslant p_{n}$. We now prove that this construction is optimal with respect to the connectivity function.
A lower bound on $f_{1,K_n}(p)$
-------------------------------
The previous constructions discussed show that $$f_{1,K_{n}}(p) \leqslant \begin{cases} g_{n}(\theta) &\text{ for } p \in [p_{n},1] , \\
0 &\text{ for } p \in [0,p_{n}]. \end{cases} \nonumber$$ It is clear that $f_{1,K_{n}}(p) \geqslant 0$ for all $p$, and so all that remains to show is that $f_{1,K_{n}}(p) \geqslant g_{n}(\theta)$ for $p \in [p_{n},1]$. We will prove this result by induction on $n$. The inequality is trivially true when $n = 2$, so let us assume that $n > 2$ and that the inequality is true for all cases from $2$ up to $n-1$. First, we note that $g_{n}(\theta) = g_{j}(\theta)g_{n-j}(\theta) - g_{n,j}(\theta)$ for all $j \in \{0,1,\ldots,n\}$. Thus, if we multiply both sides of this equation by $\binom{n}{j}$ and sum over all $j \in \{0,1,\ldots,n\}$, we have that $$\label{equation: f-equation}
2^{n}g_{n}(\theta) = \Bigg( \sum_{j = 0}^{n}\binom{n}{j}g_{j}(\theta)g_{n-j}(\theta)\Bigg) - 2,$$ Let $\mu \in \mathcal{M}_{1,p}(K_{n})$ and let $C$ be the event that $(\mathbf{K}_n)_{\mu}$ is connected. Given $A \subseteq [n]$, let $X_{A}$ be the event that $(\mathbf{K}_n)_{\mu}[A]$ and $(\mathbf{K}_n)_{\mu}[A^{c}]$ are each connected, where $A^{c} = [n] \setminus A$. Moreover, let $Y_{A}$ be the event that $(\mathbf{K}_n)_{\mu}[A]$ and $(\mathbf{K}_n)_{\mu}[A^{c}]$ are each connected, and there are no edges between $A$ and $A^{c}$ in $(\mathbf{K}_n)_{\mu}$. For all $A \subseteq [n]$, we have that $$\label{inequality: f-inequality2}
\mu(C) \geqslant \mu(X_{A}) - \mu(Y_{A}).$$ Note that when $A = \emptyset$ or $A = [n]$, the above equation is trivially true due to the fact that $C, X_{\emptyset},X_{[n]},Y_{\emptyset}$ and $Y_{[n]}$ are all the same event. As $\mu$ is $1$-independent we have that if $A$ is a non-empty proper subset of $[n]$, then, by induction on $n$, we have $$\label{inequality: f-inequality1}
\mu(X_{A}) \geqslant g_{|A|}(\theta)g_{n - |A|}(\theta).$$ Note that here we are using the fact that $(p_{n})_{n \geqslant 2}$ forms an increasing sequence, and so $p \geqslant p_{n}$ implies that $p \geqslant p_{s}$ for all $s \leqslant n$. We are also using the fact that $g_{1}(\theta) = 1$ for all $\theta \in [0,1]$. We proceed by summing (\[inequality: f-inequality2\]) over all non-empty proper subsets of $[n]$, and then applying (\[inequality: f-inequality1\]) to obtain $$\begin{aligned}
\label{inequality: f-inequality3}
(2^{n} - 2)\mu(C) \geqslant \Bigg(\sum_{A \subseteq [n]} g_{|A|}(\theta)g_{n-|A|}(\theta)\Bigg) - 2g_{0}(\theta)g_{n}(\theta) \nonumber \\
- \Bigg(\sum_{A \subseteq [n]} \mu(Y_{A})\Bigg) + \mu(Y_{\emptyset})+ \mu(Y_{[n]}).\end{aligned}$$ We apply (\[equation: f-equation\]) and the fact that the events $C,Y_{\emptyset}$ and $Y_{[n]}$ are all the same event to (\[inequality: f-inequality3\]) to get $$\begin{aligned}
\label{inequality: f-inequality4}
(2^{n} - 4)\mu(C) \geqslant (2^{n}-4)g_{n}(\theta) +2 - \Bigg(\sum_{A \subseteq [n]} \mu(Y_{A})\Bigg).\end{aligned}$$ Note that for all $A \subseteq [n]$, the events $Y_{A}$ and $Y_{A^{c}}$ are the same event, and so $\sum_{A \subseteq [n]} \mu(Y_{A}) = 2\sum_{1 \in A \subseteq [n]} \mu(Y_{A})$. Moreover, the set $\{ Y_{A} : 1 \in A \subseteq [n]\}$ consists of pairwise disjoint events, and so $\sum_{1 \in A \subseteq [n]} \mu(Y_{A}) \leqslant 1$. Thus $$\label{inequality: Y-inequality}
\sum_{A \subseteq [n]}\mu(Y_{A}) \leqslant 2.$$
We apply (\[inequality: Y-inequality\]) to (\[inequality: f-inequality4\]) to obtain $(2^{n}-4)\mu(C) \geqslant (2^{n}-4)g_{n}(\theta)$. As $n >2$, we have that $\mu(C) \geqslant g_{n}(\theta)$ and so we are done.
A remark on $f_{k, K_n}(p)$ for $k\geqslant 2$
----------------------------------------------
Clearly we can define $f_{k,G}(p)$ analogously to $f_{1,G}(p)$ for $k \in \mathbb{N}_0$. For $k=0$, $f_{0, K_n}(p)$ is exactly the probability that an instance of the Erd[ő]{}s–Rényi random graph $\mathbf{G}_{n,p}$ contains a spanning tree. As far as we know, there is no nice closed form expression for this function.
In this section, we have computed $f_{1, K_n}(p)$ exactly, which is the other interesting case, as for $k\geqslant 2$ the connectivity problem is trivial.
\[prop: f\_[k, K\_n]{}(p) for k at least 2\] For all $k,n \in \mathbb{N}_{\geqslant 2}$, we have that $$f_{k, K_n}(p)=\begin{cases}
0 & \textrm{if }p\leqslant 1 -\frac{2}{n},\\
1-\frac{n(1-p)}{2}& \textrm{otherwise.}
\end{cases}$$
For the lower bound, consider $\mu\in \mathcal{M}_{k,p}(K_n)$. Since any subgraph of $K_n$ with at least $\binom{n}{2}-(n-1)$ edges is connected, we can apply Markov’s inequality to show that $$\begin{aligned}
1-\mu(\{\textrm{connected}\})&\leqslant \mu (\{\exists \ \geqslant(n-1) \textrm{ closed edges}\})
\leqslant \frac{1}{n-1}\mathbb{E}_{\mu} \{\# \textrm{ closed edges}\}=\frac{n(1-p)}{2}.\end{aligned}$$ For the upper bound, consider the random graph $\mathbf{G}$ obtained as follows. Let $x:=\frac{1-p}{2}$. With probability $\min(nx, 1)$, select a vertex $i\in [n]=V(K_n)$ uniformly at random, and let $\mathbf{G}$ be the subgraph of $K_n$ obtained by removing all edges incident with $i$. Otherwise, let $\mathbf{G}$ be the complete graph $K_n$. It is easy to check that $\mathbf{G}$ is a $2$-independent model with edge-probability $p$ and that $\mathbf{G}$ is connected if and only if $\mathbf{G}=K_n$, an event which occurs with probability $1-\min(1,nx)=\max\left(0, 1-n(1-p)/2\right)$.
Cycles {#section: cycles}
======
Linear programming for calculating $f_{1,G}(p)$ {#subsection: linear programming}
-----------------------------------------------
In this subsection we describe how we can represent the problem of finding $f_{1,G}(p)$, for any graph $G$, as a (possibly non-linear) programme.
Given a graph $G$ on vertex set $[n]$, let $\mathcal{H}=\mathcal{H}(G)$ be the set of all labelled subgraphs of $G$. Throughout this section we treat these subgraphs as subsets of $E(G)$, and always imagine them to be on the full vertex set $[n]$. For each labelled subgraph of $G$ we write $$\begin{aligned}
\mu (S) := \mu(S \subseteq \mathbf{G}_{\mu}) && \text{and} && \mu (\hat{S}) := \mu(\mathbf{G}_{\mu} \cong S). \end{aligned}$$ Recall that for a function $\mu:\mathcal{H}\rightarrow \mathbb{R}_{\geqslant 0}$, we have $\mu \in \mathcal{M}_{1,p}(G)$ if and only if the following three conditions all hold:
1. [$\mu$ is a probability measure on labelled subgraphs of $G$,]{}
2. [Every edge of $G$ is open with probability at least $p$,]{}
3. [Given non-empty $S,T \in \mathcal{H}$ such that $S$ and $T$ are supported on disjoint subsets of $[n]$, $\mu(S) \cdot \mu(T)= \mu(S \cup T)$.]{}
Note that as we are interested in determining $f_{1,G}(p)$, for our purposes it is sufficient to assume that in fact every edge of $G$ is open with probability $p$. Bearing this in mind, we can we can rewrite these conditions in the following way:
1. [ $\sum_{H \in \mathcal{H}} \mu(\hat{H}) = 1$]{},
2. [ For all edges $e \in E(G)$, we have that$\sum_{H \in \mathcal{H}} \mathbbm{1}(e\in H)\mu(\hat{H}) \geqslant p$]{},
3. For all non-empty $S,T \in \mathcal{H}$, such that $S$ and $T$ are supported on disjoint subsets of $[n]$, we have that $$\label{equation: third bullet product}
\sum_{H \in \mathcal{H}} \mu(\hat{H}) \Big( \mathbbm{1}\big( (S \cup T) \subseteq H \big) - \mathbbm{1}\big(S \subseteq H\big) \mu(T) \Big) = 0.$$
Let $A=A(G)$ be a matrix which has columns indexed by $\mathcal{H}$, and a row for each piece of information given by one of the above conditions. That is:
1. We have a row for the empty set such that $A_{\emptyset,H}:= 1$.
2. [We have a row for $e \in G$; the entry $A_{e,H}:=\mathbbm{1}(e \in H)$;]{}
3. [We have a row for each pair $S,T \in \mathcal{H}\setminus \{\emptyset\}$ supported on disjoint subsets of $[n]$; the entry $A_{\{S,T\},H}:=\mathbbm{1}((S \cup T) \subseteq H) - \mu(T) \cdot \mathbbm{1}(S \subseteq H)$.]{}
Let $\mathbf{q}=\mathbf{q}(G)$ be a vector with indexing the same as the rows of $A$; let $q_\emptyset:=1$, $q_e:=p$ for $e \in G$, and $q_{\{S,T\}}:=0$ for each pair $S,T \in \mathcal{H} \setminus \{\emptyset\}$ supported on disjoint subsets of $[n]$. Then a vector $\mathbf{w}$, whose entries are indexed by $\mathcal{H}$, which satisfies $w_{H} \geqslant 0$ for all $H \in \mathcal{H}$, and also $A \mathbf{w} = \mathbf{q}$ corresponds precisely to a measure $\mu \in \mathcal{M}_{1,p}(G)$.
Let $\mathbf{c}$ be a vector indexed by $\mathcal{H}$ defined by $c_{H} := \mathbbm{1}(H \text{ is connected})$. Just to make it clear, we say that $H \in \mathcal{H}$ is connected if it contains a spanning tree of $[n]$. Then for a given value of $p$ the vector $\mathbf{w}(p)$ satisfying $A \mathbf{w}(p) = \mathbf{q}$ corresponds to a measure $\mu \in \mathcal{M}_{1,p}(G)$ such that $\mu (H \text{ is connected}) = f_{1,G}(p)$.
Observe that for any graph with five vertices or fewer, any partition of the graph into two parts has that one part must have at most two vertices in it. In particular, if $G$ is a graph on $[5]$, and $S$ and $T$ are non-empty subgraphs of $G$ supported on disjoint subsets of $[5]$, then one of $S$ and $T$ must consist of precisely one edge of $G$. By choosing $T$ to be this subgraph, we can always choose $S$ and $T$ for (\[equation: third bullet product\]) so that $\mu(T)=p$. Thus for any choice of $p$, we can turn the problem of finding $f_{1,G}(p)$ into the following linear programme: $$\label{equation: lp minimum}
a^{*}=\min_{\mathbf{w}} \mathbf{c}^{T}\mathbf{w} \quad \text{subject to } A\mathbf{w}=\mathbf{q}, \mathbf{w} \geqslant 0.$$ (Note that for graphs with six or more vertices, one may find $S$ and $T$ such that $\mu(T)$ (in (\[equation: third bullet product\])) is an unknown function of $p$, and thus the programme is not linear; for example, this indeed is the case for $C_6$.)
The duality theorem states that the asymmetric dual problem has the same optimal solution $a^{*}$: $$\label{equation: lp maximum}
a^{*}=\max_{\mathbf{x}} \mathbf{q}^{T}\mathbf{x} \quad \text{subject to } A^{T}\mathbf{x} \leqslant \mathbf{c}.$$ One can easily solve the linear programmes above for a specific value of $p$, for example using the software Maple, and the *LPSolve* function it contains. However we of course wish to find solutions for all values of $p \in [0,1]$.
By writing $A=(a_{ij})$, $\mathbf{w}=(w_j)$, $\mathbf{c}=(c_j)$, $\mathbf{q}=(q_i)$ and $\mathbf{x}=(x_i)$ any solutions $\mathbf{w}$ and $\mathbf{x}$ must satisfy $\sum_j a_{ij} w_j = q_i$, $\sum_i a_{ij} x_i \leqslant c_j$ and $w_i \geqslant 0$. Thus we have $$\sum_{i} q_i x_i = \sum_i \left( \sum_j a_{ij} w_j \right) x_i = \sum_j \left( \sum_i a_{ij} x_i \right) w_j \leqslant \sum_j c_j w_j.$$ In particular for optimal solutions we have $\sum_{i} q_i x_i = \sum_{j} c_j w_j$ and so the inequality must be an equality, that is $$\left(\sum_i a_{ij} x_i \right) w_j = c_j w_j, \quad \text{ for all $j$.}$$ Consequently for each $j$ we either have $w_j=0$ or $\sum_i a_{ij} x_i=c_j$. Thus in our attempt to obtain a function for all $p$, it seems reasonable to look at an optimal solution for one value of $p$ and see which $w_j$ have been set to zero; assume for these indices that we always have $w_j=0$ and attempt to directly solve the equations that result from this. This motivates the following method:
- [Solve (\[equation: lp minimum\]) with a specific value of $p$ to obtain a solution $\mathbf{w}(p)$ and a set $J:=\{ j \in [|\mathbf{w}|] : w_j(p)=0\}$.]{}
- [Solve the set of equations $\{(A\mathbf{w})_i=q_i,w_j=0 : i \in [|\mathbf{w}|],j \in J\}$ to obtain functions of $p$ for all $w_k$, $k \in [|\mathbf{w}|]$, which we write as $w'_k(p)$.]{}
- [Solve the set of equations $\{(A^T\mathbf{x})_i=c_i: i \in [|\mathbf{w}|] \setminus J\}$ to obtain functions of $p$ for all $x_k$, $k \in [|\mathbf{w}|]$, which we write as $x'_k(p)$.]{}
- [Write $w^*(p):= \mathbf{c}^{T}\mathbf{w'}(p)$ and $x^*(p):=\mathbf{q}^T\mathbf{x'}(p)$.]{}
- [For a certain interval $P \subseteq [0,1]$ of values of $p$, check that $(A^T\mathbf{x}')_i(p) \leqslant c_i$ and $w'_i(p) \geqslant 0$, for all $i \in [|\mathbf{w}|]$.]{}
For the given interval $P$ which works above, the conditions above ensure that the $\mathbf{w}'(p)$ and $\mathbf{x}'(p)$ obtained in this way are feasible solutions to (\[equation: lp minimum\]) and (\[equation: lp maximum\]) respectively. Thus if $w^*(p)=x^*(p)$, then by the duality theorem we have $f_{1,G}(p)=w^*(p)$. Furthermore, a measure $\mu$ on the subgraphs of $G$ which is extremal is given directly by $\mathbf{w}'(p)$. In the following subsection we give, as examples, two results which are proved using the above method.
The connectivity function of small cycles
-----------------------------------------
In this subsection we prove Theorems \[theorem: connected function C4\] and \[theorem: connected function C5\] using the above method. Furthermore, the method gives us an extremal example in each case.
For $C_4$ and $p \in [\frac{1}{2},1]$ an extremal construction is given by the measure $\mu$, defined by $$\mu(\hat{H}) = \begin{cases}
2p-1 & \text{if } H=C_4;\\
\frac{p(1-p)}{2} & \text{if $H$ is contains precisely two edges, which are adjacent;}\\
(1-p)^2 & \text{if $H$ is contains precisely two edges, which are not adjacent;}\\
0 & \text{otherwise.}\\
\end{cases}
\nonumber$$ For $C_4$ and $p \in [0,\frac{1}{2}]$ an extremal construction is given by the measure $\mu$, defined by $$\mu(\hat{H}) = \begin{cases}
1-2p & \text{if $H$ is the empty graph;}\\
\frac{p(1-p)}{2} & \text{if $H$ is contains precisely two edges, which are adjacent;}\\
p^2 & \text{if $H$ is contains precisely two edges, which are not adjacent;}\\
0 & \text{otherwise.}\\
\end{cases}
\nonumber$$
We can in fact give a direct combinatorial proof of the lower bound in Theorem \[theorem: connected function C4\]: for any $\mu \in \mathcal{M}_{1,p}(C_4)$, we have by $1$-independence that $$\begin{aligned}
\mu(\{\textrm{connected}\})&\geqslant \mu(\{12, 34 \textrm{ are open}\})-\mu(\{23, 14 \textrm{ are closed}\})\geqslant p^2-(1-p)^2=2p-1.\end{aligned}$$ Together with the first of the constructions of measures $\mu$ above (which can be found by analysing how the bound in the inequality above can be tight), this gives a second and perhaps more insightful proof of Theorem \[theorem: connected function C4\] than the one obtained from applying the linear optimisation method. However for the next result, on $f_{1,C_5}(p)$, we do not have a combinatorial proof, and our result relies solely on linear optimisation.
For $C_5$ and $p \in [\frac{\sqrt{3}}{3},1]$ an extremal construction is given by the measure $\mu$, defined by $$\mu(\hat{H}) = \begin{cases}
\frac{p(3p^2-1)}{3p-1} & \text{if } H=C_5;\\
\frac{p(1-p)(2p-1)}{5(3p-1)} & \text{if $H$ is missing precisely two edges, which are adjacent;}\\
\frac{p(1-p)^2}{5(3p-1)} & \text{if $H$ is missing precisely two edges, which are not adjacent;}\\
\frac{(2p-1)(1-p)^2}{3p-1} & \text{if $H$ is the empty graph;}\\
0 & \text{otherwise.}\\
\end{cases}
\nonumber$$ For $C_5$ and $p \in [0,\frac{\sqrt{3}}{3}]$ an extremal construction is given by the measure $\mu$, defined by $$\mu(\hat{H}) = \begin{cases}
\frac{5p^3-5p^2-2p+2}{3p+2} & \text{if $H$ is the empty graph;}\\
\frac{p(1-3p^2)}{3p+2} & \text{if $H$ consists of precisely two edges, which are adjacent;}\\
\frac{p^3}{3p+2} & \text{if $H$ is missing precisely two edges, which are adjacent;}\\
\frac{p^2(1+p)}{3p+2} & \text{if $H$ is missing precisely two edges, which are not adjacent;}\\
0 & \text{otherwise.}\\
\end{cases}
\nonumber$$
General bounds for cycles of length at least [$6$]{}
----------------------------------------------------
We can use Markov’s inequality to derive the following simple lower bound on $f_{1,C_n}(p)$ for $n\geqslant 6$.
\[proposition: Markov for Ck\] For $n \in \mathbb{N}$, with $n \geqslant 6$, and $p \in [0,1]$, we have $f_{1,C_n}(p) \geqslant \frac{np-(n-2)}{2}$.
A small adjustment to this argument gives the following improvement for $n=6$.
\[proposition: improved Markov for C6\] For $p \in [0,1]$ we have that $f_{1,C_6}(p) \geqslant -p^3+3p^2-1$.
Let $\mu \in \mathcal{M}_{1,p}(C_k)$. Note that $\mathbf{G}_{\mu}$ is connected if and only if it has at most one closed edge. Thus by Markov’s inequality, we have $$\begin{aligned}
f_{1,C_k}(p) & = & 1-\mu (\exists \geqslant 2 \text{ closed edges in $C_k$}) \nonumber \\
& \geqslant & 1- \frac{\mathbb{E}_{\mu}(\#\text{ closed edges in $C_k$})}{2} \nonumber \\
& = & 1 - \frac{k(1-p)}{2} = \frac{kp-(k-2)}{2}. \nonumber\end{aligned}$$
Let $X$ be the number of closed edges in $\mathbf{G}_{\mu}$. We have $$\begin{aligned}
f_{1,C_6}(p) & = 1- \mu(X \geqslant 2) = 1 - \frac{\mathbb{E}_{\mu}(X)}{2}+ \\
& \left(\frac{\mu(X=1)+\mu(X=3)+2\mu(X=4)+3\mu(X=5)+4\mu(X=6)}{2}\right) \\
& \geqslant 1-\frac{6(1-p)}{2}+(1-p)^3=-p^3+3p^2-1.\end{aligned}$$ Proving the inequality involves considering the weights given to each labelled subgraph of $C_6$. The simple idea is that the sum of weights of subgraphs of the graph consisting of $3$ disjoint edges must sum to $(1-p)^3$ by $1$-independence, and this weight is smaller than the weight of all subgraphs; we omit the full details since it is a tedious exercise of notation and simple algebra.
Maximising connectivity {#section: maximising connectivity}
=======================
In this section, we derive our results for maximising connectivity in $1$-independent modes. First of all Theorem \[theorem: connected function Kn\] allow us to easily determine the value of $F_{1, K_n(p)}$ and hence prove Theorem \[theorem: maximising connectivity in Kn\].
Given a $1$-independent model $\mathbf{G}$ on $K_n$ with edge-probability at least $1-p$, observe that the complement $\mathbf{G}^c$ of $\mathbf{G}$ in $K_n$ is a $1$-independent model in which every edge is open with probability at most $p$. Furthermore, $\mathbf{G}^c$ is connected whenever $\mathbf{G}$ fails to be connected. This immediately implies $$\begin{aligned}
\label{eq: complement of slow connectivity}
1-f_{1, K_n}(1-p)\leqslant F_{1, K_n}(p).
\end{aligned}$$ Furthermore, observe that the Red-Blue measure $\nu_p$ we constructed to obtain the upper bound on $f_{1, K_n}(p)$ in the proof of Theorem \[theorem: connected function Kn\] has the property that a $\nu_{p}$-random graph is connected if and only if its complement fails to be connected. This immediately implies that we have equality in (\[eq: complement of slow connectivity\]).
For paths, a simple construction achieves the obvious upper bound for $F_{1,P_n}(p)$.
For any measure $\mu\in \mathcal{M}_{1, \leqslant p}(P_n)$, we have by $1$-independence that $$\begin{aligned}
\mu\left(\{\textrm{connected}\}\right)=\mu(\{P_n\})\leqslant \mu\left(\bigcap_{1\leqslant i \leqslant \lfloor \frac{n}{2}\rfloor}\{\textrm{the edge $\{2i-1, 2i\}$ is open}\}\right)\leqslant p^{\lfloor \frac{n}{2}\rfloor},
\end{aligned}$$ which implies $F_{1,P_n}(p)\leqslant p^{\lfloor \frac{n}{2}\rfloor}$. For the lower bound, we construct a $1$-ipm as follows. For each integer $i$: $1\leqslant i \leqslant n/2$, we assign a state On to the vertex $2i$ with probability $p$, and a state Off otherwise, independently at random. Then set an edge of $P_n$ to be open if one of its endpoints is in state On, and closed otherwise. This is easily seen to yield a $1$-ipm $\mu$ on $P_n$ in which every edge is open with probability $p$, and for which $$\begin{aligned}
\mu\left(\{\textrm{connected}\}\right)=\mu\left(\bigcap_{1\leqslant i \leqslant \lfloor \frac{n}{2}\rfloor}\{\textrm{the vertex $2i$ is in state On}\}\right)= p^{\lfloor \frac{n}{2}\rfloor}.
\end{aligned}$$ Thus $F_{1,P_n}(p)\geqslant p^{\lfloor \frac{n}{2}\rfloor}$, as claimed.
The case of cycles $C_n$ appears to be slightly more subtle. For the $4$-cycle, as in the previous section, we can give two proofs, one combinatorial and the other via linear optimisation.
The theorem immediately follows from an application of the linear optimisation techniques from Section \[section: cycles\]. Alternatively, we can obtain the upper bound by a direct argument. For any measure $\mu\in \mathcal{M}_{1, \leqslant p}(C_4)$, we have by $1$-independence that $$\begin{aligned}
1-\mu(\{\textrm{connected} \})&\geqslant \mu\left(\{\textrm{both $12$ and $34$ are closed}\}\right)\geqslant (1-p)^2,
\end{aligned}$$ and, by a simple union bound and $1$-independence, $$\begin{aligned}
\mu(\{\textrm{connected} \})&\leqslant \mu\left(\{\textrm{both $12$ and $34$ are open}\}\cup \{\textrm{both $23$ and $14$ are open}\}\right) \leqslant 2p^2.
\end{aligned}$$ Combining these two inequalities and using $1-(1-p)^2=2p-p^2$, we obtain $$\begin{aligned}
\mu(\{\textrm{connected} \})&\leqslant \min\left(2p^2, 2p-p^2\right),
\end{aligned}$$ which gives the claimed upper bound on $F_{1, C_4}(p)$.
For the lower bound, we give two different constructions, depending on the value of $p$. For $p \in [\frac{2}{3},1]$ consider the measure $\mu$ defined by $$\mu(\hat{H}) = \begin{cases}
p(3p-2) & \text{if } H=C_4;\\
p(1-p) & \text{if $H$ contains precisely three edges;}\\
1-p(2-p) & \text{if $H$ is the empty graph with no edges;}\\
0 & \text{otherwise.}\\
\end{cases}
\nonumber$$ It is easily checked that $\mu \in \mathcal{M}_{1,\leqslant p}(C_4)$ and that $\mu(\{\textrm{connected}\})=1-(1-p(2-p))=2p-p^2$, which is maximal for $p$ in that range.
For $p \in [0,\frac{2}{3}]$, consider the measure $\mu$ defined by $$\mu(\hat{H}) = \begin{cases}
\frac{p^2}{2} & \text{if $H$ contains precisely three edges;}\\
\frac{p(2-3p)}{2} & \text{if $H$ contains precisely one edge;}\\
1-4p(1-p) & \text{if $H$ is the empty graph with no edges;}\\
0 & \text{otherwise.}\\
\end{cases}
\nonumber$$ Again, it is easily checked that $\mu \in \mathcal{M}_{1,\leqslant p}(C_4)$ and that $\mu(\{\textrm{connected}\})=\mu\left(\{\geqslant 3\textrm{ edges open}\}\right)=2p^2$, which is maximal for $p$ in that range.
We simply apply the linear optimisation method from Section \[section: cycles\] — here again we do not have a combinatorial proof. In addition to establishing the theorem, this gives us constructions of extremal $1$-independent measures maximising connectivity.
For $p \in [\frac{3}{5},1]$ an extremal construction is given by the measure $\mu$, defined by $$\mu(\hat{H}) = \begin{cases}
\frac{p(5p-3)}{5-3p} & \text{if $H=C_5$;}\\
\frac{p(1-p^2)}{5-3p} & \text{if $H$ contains precisely four edges;}\\
\frac{3p^3-7p^2+5p-1}{5-3p} & \text{if $H$ contains precisely two edges, which are adjacent;}\\
\frac{2(1-p)^3}{5-3p} & \text{if $H$ contains precisely one edge;}\\
0 & \text{otherwise.}\\
\end{cases}
\nonumber$$ For $p \in [\frac{1}{2}, \frac{3}{5}]$ an extremal construction is given by the measure $\mu$, defined by $$\mu(\hat{H}) = \begin{cases}
\frac{p^2}{3} & \text{if $H$ contains precisely four edges;}\\
\frac{p(2-3p)}{3} & \text{if $H$ contains precisely two edges, which are adjacent;}\\
\frac{p(2p-1)}{3} & \text{if $H$ contains precisely one edge;}\\
\frac{3-5p}{3} & \text{if $H$ is the empty graph with no edges;}\\
0 & \text{otherwise.}\\
\end{cases}
\nonumber$$ For $p \in [0,\frac{1}{2}]$ an extremal construction is given by the measure $\mu$, defined by $$\mu(\hat{H}) = \begin{cases}
\frac{p^2(p+1)}{p+4} & \text{if $H$ contains precisely four edges;}\\
\frac{p^2(1-2p)}{p+4} & \text{if $H$ is missing precisely two edges, which are adjacent;}\\
\frac{p(p^2-3p+2)}{p+4} & \text{if $H$ contains precisely two edges, which are adjacent;}\\
\frac{5p^2-9p+4}{p+4} & \text{if $H$ is the empty graph with no edges;}\\
0 & \text{otherwise.}\\
\end{cases}
\nonumber$$
Proof of Theorem \[theorem: upper bounds on long paths critical prob\] {#section: long paths}
======================================================================
Combining Corollary \[corollary: bound on long path constant in products with the line\] with our results on $1$-independent connectivity, much of Theorem \[theorem: upper bounds on long paths critical prob\] is immediate.
For the lower bound in part (i), we note that $\mathbb{Z}\times C_n$ has the finite 2-percolation property. Thus, as described after the proof of Theorem \[theorem: local construction integer lattice\], we have that $p_{1, \ell\mathit{p}}(\mathbb{Z}\times C_n)\geqslant 4-2 \sqrt{3}$. For the upper bound in part (i), since the long paths critical probability is non-decreasing under the addition of edges, we have $$p_{1, \ell\mathit{p}}(\mathbb{Z}\times C_n)\leqslant p_{1, \ell\mathit{p}}(\mathbb{Z}\times P_n) \leqslant p_{1, \ell\mathit{p}}(\mathbb{Z}\times P_2),$$ which is at most $2/3$ by Theorem \[theorem: long paths critical prob\](ii).
For the upper bounds (ii)–(iv) Theorem \[theorem: upper bounds on long paths critical prob\] follow directly from our results on $1$-independent connectivity functions. For $G=K_3, C_4, C_5$, we plug in the value of $f_{1, G}(p)$ in equation (\[equation: pstar\]), solve for $p_{\star}(G)$ and apply Corollary \[corollary: bound on long path constant in products with the line\].
In part (v), we begin by noting that as we are considering an increasing nested sequence of graphs, the sequence $\left(p_{1, \ell \mathit{p}}(\mathbb{Z}\times K_n)\right)_{n\in \mathbb{N}}$ is non-increasing in $[0,1]$ and hence tends to a limit as $n\rightarrow \infty$. For the lower bound in (v), observe that for any $n\in \mathbb{N}$ the graph $\mathbb{Z}\times K_n$ has the finite 2-percolation property – indeed for any finite $k$, the closure of a copy of $P_k\times K_n$ under $2$-neighbour bootstrap percolation in $\mathbb{Z}\times K_n$ is equal to itself. We construct a $1$-ipm $\mu$ on $\mathbb{Z}\times K_n$ as in Corollary \[corollary: if 2-percolation property, then lb on p\_[1,c]{}\] but with starting set $T_0=\{0\}\times V(K_n)$ and hence $T_k=(\{k\}\times V(K_n))\cup (\{-k\}\times V(K_n))$. It is easily checked that $\mu$-almost surely, all components (and hence all paths) in a $\mu$-random graph have length at most $5n$. Since by construction $d(\mu)=4-2\sqrt{3}$, this proves $$p_{1, \ell \mathit{p}}\left(\mathbb{Z}\times K_n\right)\geqslant 4-2\sqrt{3}$$ for all $n \in \mathbb{N}$. For the upper bound, we perform some simple analysis. By solving a quadratic equation, we see that $$\left(\frac{1+\sqrt{2p-1}}{2}\right)^2 > (1-p)$$ for all fixed $p\in (\frac{5}{9}, 1)$. Then by Theorem \[theorem: connected function Kn\], for any such fixed $p$ and all $n$ sufficiently large, we have that $$\begin{aligned}
\left(f_{1, K_n}(p)\right)^2&> \left(\frac{1+\sqrt{2p-1}}{2}\right)^{2n}>4(1-p)^n= 4(1-p)^{v(K_n)}.
\end{aligned}$$ Thus $p_{\star}(K_n)<p$ for all $n$ sufficiently large, which by Corollary \[corollary: bound on long path constant in products with the line\] implies $p_{1, \ell \mathit{p}}\left(\mathbb{Z}\times K_n\right)<p$.
Open problems {#section: open problems}
=============
More tractable subclasses of $1$-independent measures {#subsection: problems on simpler classes of 1-ipms}
-----------------------------------------------------
The most obvious open problem about $1$-independent percolation is of course whether the known lower and upper bounds on $p_{1, c}(\mathbb{Z}^2)$ can be improved. This problem is, we suspect, very hard in general. However, it may prove more tractable if we restrict our attention to a smaller family of measures.
Let $G$ be a graph. A $G$-partition is a partitioned set $\sqcup_{v\in V(G)}\Omega_v$, with non-empty parts indexed by the vertices of $G$. A $G$-partite graph is a graph $H$ on a $G$-partition $V(H)=\sqcup_{v\in V(G)}\Omega_v$ whose edges are a subset of the union of the complete bipartite graphs $\sqcup_{uv\in E(G)}\{\omega_u\omega_v: \ \omega_u \in \Omega_u, \omega_v \in\Omega_v\}$ corresponding to the edges of $G$.
Given a $G$-partite graph $H$ on a $G$-partition $\sqcup_{v\in V(G)}\Omega_v$, we have a natural way of constructing $1$-independent bond percolation models: given a family $\mathbf{X}=\left(S_v\right)_{v\in V(G)}$ of independent random variables with $S_v$ taking values in $\Omega_v$, the *$(H, \mathbf{X})$-random subgraph* of $G$, denoted by $H[\mathbf{X}]$, is the random configuration on $E(G)$ obtained by setting $uv$ to be open if and only if $S_uS_v\in E(H)$.
\[def: state based measure\] Let $G$ be a graph. A measure $\mu \in \mathcal{M}_{1,p}(G)$ is said to be vertex-based if there exist
- a $G$-partition $\sqcup_{v\in V(G)}\Omega_v$,
- an associated $G$-partite graph $H$, and
- a collection of independent random variables $(S_v)_{v\in V(G)}$ with $S_v$ taking values in $\Omega_v$,
such that the $(H, \mathbf{X})$-random subgraph $H[\mathbf{X}]$ has the same distribution as the $\mu$-random graph $\mathbf{G}_{\mu}$.
Let $\mathcal{M}_{\mathrm{vb}, p}(G)$ denote the collection of all vertex-based measures on $G$ with edge-probability $p$.
\[problem: state-based measures for percolation\] Determine $\inf\Bigl\{p\in [0,1]: \ \forall \mu \in \mathcal{M}_{\mathrm{vb}, p}(\mathbb{Z}^2), \ \mu(\{\mathrm{percolation}\})=1 \Bigr\}$.
Vertex-based measures arise naturally in renormalising arguments, and are thus a natural class of examples to consider. A special case of Problems \[problem: Harris critical probabiliy for 1-ipm\] and \[problem: state-based measures for percolation\] is obtained by further restricting our attention to the case where the $\Omega_v$ have bounded size.
A vertex-based measure $\mu$ on a graph $G$ is $N$-uniformly bounded if it as in Definition \[def: state based measure\] above and in addition for each $v\in V(G)$, $\vert \Omega_v\vert \leqslant N$. Furthermore, a vertex-based measure $\mu$ on a graph $G$ is uniformly bounded if it is $N$-uniformly bounded for some $N\in \mathbb{N}$.
Let $\mathcal{M}_{N-\mathrm{ubvb}, p}(G)$ and $\mathcal{M}_{\mathrm{ubvb}, p}(G)$ denote the collection of all vertex-based measures on $G$ with edge-probability $p$ that are $N$-uniformly bounded and uniformly bounded respectively.
\[problem: ub state-based measures for percolation\]
(i) For $N\in \mathbb{N}$, determine $$\inf\Bigl\{p\in [0,1]: \ \forall \mu \in \mathcal{M}_{N-\mathrm{ubvb}, p}(\mathbb{Z}^2), \ \mu(\{\mathrm{percolation}\})=1 \Bigr\}.$$
(ii) Determine $$\inf\Bigl\{p\in [0,1]: \ \forall \mu \in \mathcal{M}_{\mathrm{ubvb}, p}(\mathbb{Z}^2), \ \mu(\{\mathrm{percolation}\})=1 \Bigr\}.$$
Finally, let us note that the second most obvious problem arising from our work, besides that of improving the bounds on $p_{1, c}(\mathbb{Z}^2)$, is arguably that of giving bounds on $p_{1, \ell \mathit{p}}(\mathbb{Z}^2)$ and closely related variants. Such problems, which correspond to new questions in extremal graph theory, are discussed in the subsections below. For these problems too we believe restrictions to the class of uniformly bounded vertex-based $1$-ipms could be both fruitful and interesting in their own right.
Harris critical probability for $k$-ipms and/or other graphs
------------------------------------------------------------
Since any $k$-independent model is also a $(k+1)$-independent model, for any infinite, locally finite connected graph $G$, the sequence of Harris critical probabilities $\left(p_{k,c}(G)\right)_{k\in \mathbb{Z}_\geqslant 0}$ is non-decreasing in $[0,1]$ and hence tends to a limit as $k\rightarrow \infty$. It is natural to ask for the value of this limit when $G=\mathbb{Z}^2$.
\[problem: limit of Harris as k tends to infinity\] Determine the value of $$p_{\infty, c}(\mathbb{Z}^2):=\lim_{k\rightarrow \infty} p_{k,c}(\mathbb{Z}^2).$$
In particular, is it the case that $p_{\infty, c}(\mathbb{Z}^2)<1$?
One may also ask about bounds on $p_{1, c}(G)$ for some of the other commonly studied lattices in percolation theory.
\[problem: 1-indep crit prob for other lattices\] Give good bounds on the value of $p_{1,c}(G)$ when $G$ is one of the eleven Archimedean lattices in the plane or the $d$-dimensional integer lattice $\mathbb{Z}^d$.
This problem is particularly interesting when $G$ is the triangular lattice or the honeycomb lattice (two lattices for which the $0$-independent Harris critical probability is known exactly), or the cubic integer lattice $\mathbb{Z}^3$ (which is important in applications). A challenge in all cases is finding constructions of non-percolating $1$-independent measures with high edge-probability — indeed, our arsenal of constructions for $1$-independent percolation problems is so sparse that any new construction could be of independent interest.
In a different direction, we can observe that $\mathbb{Z}^{d+1}$ contains a copy of $\mathbb{Z}^d$, whence the sequence $\left(p_{1, c}(\mathbb{Z}^d)\right)_{d\in \mathbb{N}}$ is non-increasing in $[0,1]$ and converges to a limit. Balister and Bollobás asked for its value:
\[problem: limit of 1-ip Harris crit prob for Zd\]\[Balister, Bollobás [@BalisterBollobas12]\] Determine $\lim_{d\rightarrow \infty} p_{1,c}(\mathbb{Z}^d)$.
Other notions of $1$-independent critical probabilities {#subsection: different critical probabilities}
-------------------------------------------------------
Let $G$ be an infinite, locally finite connected graph, and $v_0$ a fixed vertex of $G$. Given a bond percolation model $\mu$ on $G$, we let $\mathbf{C}_{v_0}$ denote the connected component of $\mathbf{G}_{\mu}$ containing $v_0$.
If $\mu$ is $0$-independent, then $\mu(\{\mathrm{percolation}\})=1$ if and only if $\mu(\{\vert \mathbf{C}_0\vert =\infty\})>0$. However this need not be true for a $1$-independent measure. Indeed, consider the $1$-ipm on $\mathbb{Z}^2$ obtained by taking the measure constructed in the proof of Theorem \[theorem: local construction integer lattice\] to determine the state of the edges in the $\ell_{\infty}$ ball of radius $3$ around the origin and setting every other edge to be open independently at random with probability $4-2\sqrt{3}$. Then in this model percolation occurs almost surely, but the origin is contained inside a component of order at most $28$.
Thus in principle there are different edge-probability thresholds in $1$-independent percolation on a graph $G$ for percolation to occur *somewhere* with probability $1$ and for it to occur *anywhere* with strictly positive probability. Indeed, if $p_{1,c}(\mathbb{Z}^2)$ were strictly less than $3/4$, then one could obtain examples of such a graph $G$ by attaching a long path to the origin in $\mathbb{Z}^2$.
Another critical edge-probability of interest is the *Temperley critical probability*, which in $0$-independent percolation is the threshold $p_T$ at which $\mathbb{E}\vert \mathbf{C}_v\vert =\infty$ for any vertex $v$ (and every $0$-independent measure with edge-probability $>p_T$). In general this threshold is different from the Harris critical probability. Again for $1$-independent percolation we have that the threshold for *some* vertex $v\in V(G)$ to satisfy $\mathbb{E}\vert \mathbf{C}_v\vert =\infty$ and for the threshold for *all* vertices of $G$ to satisfy this are different.
\[problem: different critical prob\] Given an infinite, locally finite connected graph $G$, determine the following four critical probabilities: $$\begin{aligned}
p_{1,T_1}(G)&:=\inf\Bigl\{p\in [0,1]: \ \forall \mu \in \mathcal{M}_{1, p}(G), \exists v\in V(G): \ \mathbb{E}_{\mu} \vert \mathbf{C}_v\vert =\infty \Bigr\},\\
p_{1,T_2}(G)&:=\inf\Bigl\{p\in [0,1]: \ \forall \mu \in \mathcal{M}_{1, p}(G), \forall v\in V(G): \ \mathbb{E}_{\mu} \vert \mathbf{C}_v\vert =\infty \Bigr\},\\
p_{1,H_1}(G)&:=\inf\Bigl\{p\in [0,1]: \ \forall \mu \in \mathcal{M}_{1, p}(G), \exists v\in V(G): \ \mu\left(\vert \mathbf{C}_v\vert =\infty\right)>0 \Bigr\},\\
p_{1,H_2}(G)&:=\inf\Bigl\{p\in [0,1]: \ \forall \mu \in \mathcal{M}_{1, p}(G), \forall v\in V(G): \ \mu\left(\vert \mathbf{C}_v\vert =\infty\right)>0 \Bigr\}.
\end{aligned}$$
It follows from their definition that these four critical probabilities satisfy $$\begin{aligned}
\label{eq: ineq between four crit prob}
p_{1,T_1}(G)\leqslant p_{1,T_2}(G) \leqslant p_{1, H_2}(G) \quad \textrm{ and } \quad p_{1,T_1}(G)\leqslant p_{1,H_1}(G) \leqslant p_{1, H_2}(G).\end{aligned}$$ Given the possibility of joining graphs with very different properties by long lines, we conjecture that all these critical probabilities are different in general.
\[conjeture: optimal inequalities between four crit prob\]
(i) None of the inequalities in (\[eq: ineq between four crit prob\]) may be replaced by an equality.
(ii) Neither $p_{1, T_2}(G)\leqslant p_{1, H_1}(G)$ nor the reverse inequality are true in general.
Observe that $p_{1, H_1}(G)$ is the $1$-independent Harris critical probability $p_{1,c}(G)$ studied in this paper; given Conjecture \[conjeture: optimal inequalities between four crit prob\], we more precisely should call it the *first* Harris critical probability for $1$-independent percolation. Our construction for the proof of Theorem \[theorem: local construction integer lattice\] and the argument of Balister, Bollobás and Walters from [@BalisterBollobasWalters05] give the following bounds when $G=\mathbb{Z}^2$: $$4-2\sqrt{3} \leqslant p_{1,T_1}(\mathbb{Z}^2) \leqslant p_{1,H_2}(\mathbb{Z}^2)\leqslant 0.8639.$$
\[question: equality of the critical prob in Z\^2\] Are any of the four critical probabilities from Problem \[problem: different critical prob\] equal when $G=\mathbb{Z}^2$?
Finally, note that Problem \[problem: different critical prob\] asks, in essence, how much we can delay percolation phenomena relative to the $0$-independent case by exploiting the local dependencies between the edges allowed by $1$-independence. While perhaps less useful in applications, it is an equally natural and appealing extremal problem to ask how much we can use these local dependencies to instead hasten the emergence of an infinite connected component. Balister and Bollobás were the first to consider this problem in [@BalisterBollobas12], which it would be remiss not to mention here.
\[def: fast percolation\] Let $G$ be an infinite, locally finite connected graph, and let $\mathcal{M}_{k, \leqslant p}(G)$ be as before the collection of $k$-ipms $\mu$ on $G$ satisfying $\sup_{e\in E(G)}\mu\{e\textrm{ is open}\}\leqslant p$. The critical threshold for fast $k$-independent percolation on $G$ is $$p_{k,F}(G) :=\inf\Bigl\{p\in[0,1]: \ \exists \mu \in \mathcal{M}_{k,\leqslant p}(G): \ \mu(\{\mathrm{percolation}\})=1\Bigr\}.$$
Balister and Bollobás determined $p_{1,F}(G)$ when $G$ is an infinite, locally finite tree, and also gave the simple general bounds $$\begin{aligned}
\label{eq: Balister Bollobas bounds on fast percolation}
\frac{1}
{\left(\mu_{\mathrm{conn.}}(G)\right)^2 } \leqslant p_{1,F}(G)\leqslant \left(\theta_{\mathrm{site}}(G)\right)^2,\end{aligned}$$ where $\mu_{\mathrm{conn.}}(G)$ is the connective constant of $G$ and $\theta_{\mathrm{site}}(G)$ the critical value of the $\theta$-parameter for site percolation on $G$. For the square integer lattice, this gives a lower bound on $p_{1,F}(\mathbb{Z}^2)$ of $0.1393$ from known upper bounds on $\mu_{\mathrm{conn.}}(\mathbb{Z}^2)$. In the other direction, we get a rigorous upper bound of $p_{1,F}(\mathbb{Z}^2)$ of $0.4618$ and non-rigorous upper bound of $0.3515$ from bounds and estimates for $\theta_{\mathrm{site}}(\mathbb{Z}^2)$. This obviously leaves a big gap, which Balister and Bollobás asked to reduce.
\[question: fast per threshold\]\[Balister and Bollobás [@BalisterBollobas12]\] What is $p_{1, F}(\mathbb{Z}^2)$?
Long paths critical probability {#subsection: problems on long paths}
-------------------------------
An obvious problem is to tighten the bounds in Theorem \[theorem: upper bounds on long paths critical prob\](v), which are not too far apart (compared to many of the other bounds on critical probabilities for $1$-independent model).
\[problem: improving lp crit constant bounds for Kn times long paths\] Determine $\lim_{n\rightarrow \infty} p_{1, \ell \mathit{p}}\left(\mathbb{Z}\times K_n\right)$ (which must be an element of $[4-2\sqrt{3}, \frac{5}{9}]$).
In a similar vein, the sequence $p_{1, \ell \mathit{p}}(\mathbb{Z}\times P_n)$ is a non-increasing function of $n$ (since $\mathbb{Z}\times P_{n+1}$ contains $\mathbb{Z}\times P_n$ as a subgraph). In this paper, we have given constructions showing that for all integers $n\geqslant 3$, $$4-2\sqrt{3} \leqslant p_{1, \ell\mathit{p}}(\mathbb{Z}\times C_n)\leqslant p_{1, \ell\mathit{p}}(\mathbb{Z}\times P_n)
%,$$ = p\_[1, ]{}(P\_2). Thus the sequence $\left(p_{1, \ell \mathit{p}}(\mathbb{Z}\times P_n)\right)_{n \in \mathbb{N}}$ tends to a limit in the interval $[4-2\sqrt{3}, \frac{2}{3}]$ as $n\rightarrow \infty$.
\[problem: limit of longs paths prob on Z times Pn as n tends to infinity\] Determine $$p_{1, \ell \mathit{p}} \left(\mathbb{Z}\times P_{\infty}\right):=\lim_{n\rightarrow \infty} p_{1, \ell \mathit{p}}(\mathbb{Z}\times P_{n}).$$
An in principle different but related problem is determining the value of the long paths critical in probability in $\mathbb{Z}^2$ (which need not be equal to the quantity $p_{1, \ell \mathit{p}} \left(\mathbb{Z}\times P_{\infty}\right)$ defined above).
\[problem: limit of longs paths prob on Z times Z\] Determine $p_{1, \ell \mathit{p}} \left(\mathbb{Z}^2\right)$.
We can also ask for $k$-independent versions of the long paths critical probability. Defining $p_{k, \ell \mathit{p}}\left(G\right)$ mutatis mutandis, we have for the same reason as in the previous subsection that the sequence $\left(p_{k, \ell \mathit{p}}(\mathbb{Z}\times P_n)\right)_{k \in \mathbb{N}}$ is monotone non-decreasing in $k$ and hence tends to a limit as $k\rightarrow \infty$.
It is straightforward to adapt our arguments and constructions from Section \[section: line lattice\] to show the following result, which follows directly from the work of Ligget, Schonman and Stacey [@LiggettSchonmannStacey97] on stochastic domination of $0$-independent measures on $\mathbb{Z}$ by $k$-independent ones.
\[theorem: k-independent long paths crit prob\]\[Liggett, Schonman and Stacey [@LiggettSchonmannStacey97]\] For any $k\in \mathbb{N}_{0}$, we have $$p_{k, \ell \mathit{p}}(\mathbb{Z})=1-\frac{k^k}{(k+1)^{k+1}},$$ with the convention that $0^0=1$.
Now, for any fixed $n$, given a $k$-independent percolation measure $\mu$ on $\mathbb{Z}$ such that almost surely there is no open path of length more than $\ell$, we can construct a $2(n+k)$-independent percolation measure $\mu'$ on $\mathbb{Z}\times P_n$ with the same edge-probability such that almost surely there is no open path of length more than $\ell n$. Indeed, use $\mu$ to determine the state of the horizontal edge in $\mathbb{Z}\times \{1\}$, set every vertical edge to be open with probability $1$, and set all horizontal edges $e\times \{i\}$, $i\in [n]\setminus\{1\}$ to have the same state as the bottom edge $e\times \{1\}$ to obtain the desired measure $\mu'$. In particular, this construction together with Theorem \[theorem: k-independent long paths crit prob\] implies the following.
\[corollary: k-limit of long paths critical probability is 1\] For any fixed $n\in \mathbb{N}$, $$\lim_{k \rightarrow \infty} p_{k, \ell \mathit{p}} (\mathbb{Z}\times P_n)=1. \qquad \qed$$
This leads to two natural questions:
\[question: k-limit of long paths crit prob for Ztimes Pf(k) for sufficiently fast growing f\] Is there a non-decreasing function $f: \mathbb{N}\rightarrow \mathbb{N}$ such that $$\liminf_{k\rightarrow \infty}p_{k, \ell \mathit{p}}(\mathbb{Z}\times P_{f(k)})<1 \qquad ?$$
If the answer to this question is affirmative, one could further ask whether a linearly growing function $f(k)=O(k)$ might suffice. In a different direction, one can ask what happens in $\mathbb{Z}^2$.
\[question: k-limit of long paths crit prob on Zsquared\] What is the value of $\lim_{k \rightarrow \infty}p_{k, \ell \mathit{p}}(\mathbb{Z}^2)$ ?
Finally, as in Section \[subsection: different critical probabilities\], we should observe that the almost sure existence of arbitrarily long open paths in a $1$-independent model on $G$ does not imply that for every $\ell\in \mathbb{N}$ every vertex of $G$ has a strictly positive probability of being part of a path of length at least $\ell$. Thus we may actually define a second long paths critical probability, $$p_{1, \ell\mathit{p}_2}(G):=\inf\Bigl\{p\in [0,1]: \ \forall \mu \in \mathcal{M}_{1,p}(G), \forall v\in V(G), \forall \ell\in \mathbb{N},\ \mu(\exists \textrm{ open path from $v$ of length }\ell )>0 \Bigr\}.$$
Determine $p_{1, \ell\mathit{p}_2}(\mathbb{Z}^2)$.
Our construction in the proof of Theorem \[theorem: local construction integer lattice\] shows that $p_{1, \ell\mathit{p}_2}(\mathbb{Z}^2)\geqslant 4-2\sqrt{3}$, and we know it is upper-bounded by $p_{1, H_2} (\mathbb{Z}^2)\leqslant 0.8639$. As in Section \[subsection: problems on simpler classes of 1-ipms\], it may be fruitful to study the long paths critical constant when one restricts one’s attention to a smaller class of $1$-ipms. In particular, by considering the class of uniformly bounded vertex-based measures, one is led to the following intriguing problem in graph theory.
Given an $n$-uniformly bounded $\mathbb{Z}^2$-partite graph $H$ with partition $\sqcup_{v\in \mathbb{Z}^2} \Omega_v$. A *transversal subgraph* of $H$ is a subgraph of $H$ induced by a set of distinct representatives $S$ for the parts of $H$, i.e. a set of vertices of $H$ such that $\vert S\cap \Omega_v\vert=1$ for all $v\in \mathbb{Z}^2$. The *$G$-partite density* of $H$ is $$d_G(H):=\inf \Bigl\{ \frac{ e(H[\Omega_u \sqcup \Omega_v])}{\vert \Omega_u\vert \cdot \vert \Omega_v \vert}: \ uv\in E(\mathbb{Z}^2) \Bigr\} .$$
\[question: graph theory\] Suppose $H$ is an $n$-uniformly bounded $\mathbb{Z}^2$-partite graph in which in every transversal subgraph the connected component containing the origin is...
(a) ... of size at most $C$, for some constant $C\in \mathbb{N}$.
(b) ... finite.
How large can $d_G(H)$ be?
This question can be viewed as a problem from extremal multipartite graph theory. Plausibly some tools from that area, in particular the work of Bondy, Shen, Thomassé and Thomassen [@BondyShenThomasseThomassen06] and Pfender [@Pfender12], could be brought to bear on it.
Connectivity function
---------------------
We determined in Sections \[section: cycles\] and \[section: maximising connectivity\] the connectivity function $f_{1,C_n}(p)$ for cycles $C_n$ of length at most $5$. It is natural to ask what happens for longer cycles.
\[problem: connectivity of cycles\] Determine $f_{1,C_n}(p)$ for $n \in \mathbb{N}_{\geqslant 6}$.
As mentioned in Section \[section: cycles\], the problem of finding $f_{1,C_6}(p)$ is non-linear. Nevertheless, one can use software, such as Maple and its contained NLPSolve function, to try to estimate the answer. This suggests the following:
- [The threshold at which $f_{1,C_6}(p)$ becomes nonzero is approximately $p=0.59733$;]{}
- [For $p$ just above this threshold, the best ‘asymmetric’ (see the next subsection for a definition) measure is *better* than the best ‘symmetric’ measure; e.g. at $p=0.62$ we have $f_{1,C_6}(0.62)$ is approximately $0.007$, but is as high as $0.11$ when restricted to ‘symmetric’ measures.]{}
More generally, one can ask what happens in cycles if we have higher dependency or if we try to maximise connectivity rather than minimise.
\[problem: minimising connectivity for cyles\] Determine $f_{k, C_n}(p)$ for all $p\in [0,1]$, $k\in \mathbb{N}$ and integers $n\geqslant k+2$.
\[problem: maximising connectivity for cyles\] Determine $F_{k, C_n}(p)$ for all $p\in [0,1]$, $k\in \mathbb{N}$ and integers $n\geqslant k+2$.
Beyond paths, cycles and complete graphs, the $1$-independent connectivity problem is perhaps most natural to study in the hypercube graph $Q_n$ and in the $n\times n$ toroidal grid $C_n\times C_n$. Progress on either of these would likely lead to progress on other problems in $1$-independent percolation as well.
\[problem: connectivity of hypercubes\] Determine $f_{1,Q_n}(p)$ for all $n\geqslant 3$.
\[problem: connectivity of toroidal grids\] Determine $f_{1,C_n\times C_n}(p)$ for all $n \geqslant 3$.
In a different direction, we can ask whether the extremal measures attaining $f_{1, G}(p)$ can be required to have ‘nice’ properties. For $C_4$ and $p \in [0,1/2]$ another extremal construction for $f_{1,C_4}(p)$ is given by the measure $\mu$, defined by $$\mu(\hat{H}) = \begin{cases}
1-2p & \text{if $H$ is the empty graph;}\\
p(1-p) & \text{if $H$ is $\{12,14\}$ or $\{23,34\}$};\\
p^2 & \text{if $H$ is contains precisely two edges, which are not adjacent;}\\
0 & \text{otherwise.}\\
\end{cases}
\nonumber$$ Motivated by the above, we call a measure $\mu \in \mathcal{M}_{1,p}(G)$ *symmetric* if for any pair of labelled subgraphs $S$ and $T$ of $G$ such that there exists an automorphism of $G$ mapping $S$ to $T$, then $\mu(\hat{S})=\mu(\hat{T})$. Note that the above measure is an example of a non-symmetric extremal construction for $f_{1,C_4}(p)$, whereas the measure given at the end of Section \[subsection: linear programming\] is symmetric. This leads to the following question.
\[question: symmetry\] For any $G$ and any $p \in [0,1]$, does there always exist a symmetric measure $\mu \in \mathcal{M}_{1,p}(G)$ which achieves $f_{1,G}(p)$?
It is not too hard to show that the answer to this question is affirmative for all graphs $G$ for which the program for solving $f_{1,G}(p)$ attained via our method in Section \[subsection: linear programming\] is linear (see the appendix for a proof of this fact). Another natural question is when the extremal connectivity can be attained by vertex-based measures.
\[question: state-based\] For which $G$ and which $p$ does there exist a vertex-based measure $\mu \in \mathcal{M}_{1,p}(G)$ which achieves $f_{1,G}(p)$?
Acknowledgements {#acknowledgements .unnumbered}
================
First and foremost, the second author wants to express his debt of gratitude to Mark Walters, who introduced him to $1$-independent percolation during his PhD thesis, and with whom he constructed the $1$-independent measure on the ladder given in Section \[section: ladder\].
Secondly the authors would like to collectively thank Paul Balister, Alexander Holroyd, Oliver Riordan and Amites Sarkar for stimulating discussions, encouragement and useful references and advice.
The work in this paper was undertaken during the Spring semester 2018, while the third author was a guest researcher in the Ume[å]{} under the auspices of the Ume[å]{}–Birmingham Erasmus exchange agreement. The Erasmus support that made this visit possible is gratefully acknowledged. Research of the first and second author was funded by a grant from the Swedish Research Council, whose support is also gratefully acknowledged.
Appendix {#appendix: symmetric proof .unnumbered}
========
Let $S_n$ be the symmetric group on $n$ elements. For a graph $G$ let $\Gamma(G)$ be the automorphism group of $G$, $\Gamma(G):=\{ \sigma \in S_n: \sigma(i) \sigma(j) \in E(G) \text{ if and only if } ij \in E(G), \text{ for all }i,j \in [n]\}$. Enumerate the elements of $\Gamma(G)$ as $\sigma_i$, $i \in [|\Gamma(G)|]$, and for $H \subseteq G$ write $H_{\sigma_i}$ for its image under $\sigma_i$. Recall $\mu$ is *symmetric* if for all $H \subseteq G$ and for all $i,j$ we have $\mu(\hat{H}_{\sigma_i})=\mu(\hat{H}_{\sigma_j})$.
Let $G$ be a graph with vertex set $[n]$ such that the optimisation problem for $f_{1,G}(p)$ is linear. Suppose that $\mu$ is a non-symmetric measure which achieves $f_{1,G}(p)$ for some value of $p$. Then for this same value of $p$ there exists another $\mu'$ which is symmetric and also achieves $f_{1,G}(p)$.
Let $\mu$ be the non-symmetric measure which achieves $f_{1,G}(p)$. For all $H \subseteq G$ define $$\mu'(\hat{H}):=\frac{1}{|\Gamma(G)|} \sum_{j=1}^{|\Gamma(G)|} \mu(\hat{H}_{\sigma_j}).
\nonumber$$ First note that we have the following: $$\begin{aligned}
\label{mu1}
\mu'(S_{\sigma_j}) & =& \sum_{H \subseteq G} \mathbbm{1}(S_{\sigma_j} \subseteq H) \mu'(\hat{H}) \nonumber \\
&=& \frac{1}{|\Gamma(G)|} \sum_{H \subseteq G} \sum_{i=1}^{|\Gamma(G)|} \mathbbm{1}(S_{\sigma_j} \subseteq H_{\sigma_i}) \mu'(\hat{H}) \nonumber \\
& =& \frac{1}{|\Gamma(G)|} \sum_{H \subseteq G} \sum_{i=1}^{|\Gamma(G)|} \mathbbm{1}(S_{\sigma_i} \subseteq H) \mu'(\hat{H}) \nonumber \\
&= & \frac{1}{|\Gamma(G)|} \sum_{i=1}^{|\Gamma(G)|} \mu(S_{\sigma_i}).\end{aligned}$$ The first and final equalities follow by definition. The second equality follows by summing through each automorphism of $H$ and the fact that $\mu'(\hat{H}_{\sigma_i})=\mu'(\hat{H}_{\sigma_j})$ for all $i,j$. The third equality follows by swapping automorphisms of $H$ to automorphisms of $S$, which again works since $\mu'(\hat{H}_{\sigma_i})=\mu'(\hat{H}_{\sigma_j})$.
Now note that if $S$ is the empty graph or a single edge, then $\mu(S_{\sigma_i})=\mu(S_{\sigma_j})$ for all $i,j$ and thus we obtain $\mu'(S_{\sigma_i})=\mu(S_{\sigma_i})$ for all $i$. It easily follows that $\mu'$ is a measure with edge-probability $p$. We must show $\mu'(S) \cdot \mu'(T) = \mu'( S \cup T)$ for all $S,T$ which are labelled non-empty subgraphs of $G$ supported on disjoint subsets of vertices. If the optimisation problem is linear, then without loss of generality we have $\mu'(T)=p$, and so this follows by linearity and (\[mu1\]). It remains to show that $\mu'$ also achieves $f_{G}(p)$. Again this follows easily since $$\begin{aligned}
\mu (\text{A $\mu'$-random graph is connected}) &=& \sum_{H \subseteq G} \mathbbm{1}(\text{$H$ is connected}) \mu'(H) \nonumber \\
&= & \frac{1}{|\Gamma(G)|} \sum_{H \subseteq G} \sum_{i=1}^{|\Gamma(G)|} \mathbbm{1}(\text{$H_{\sigma_i}$ is connected}) \mu'(H_{\sigma_i}) \nonumber\\
&= & \mu (\text{A $\mu$-random graph is connected}),\nonumber\end{aligned}$$ where the second equality follows by summing through each automorphism of $H$, and the third since $H_{\sigma_i}$ is connected if and only if $H_{\sigma_j}$ is connected, for all $i,j$.
[^1]: Institutionen för Matematik och Matematisk Statistik, Ume[å]{} Universitet, 901 87 Ume[å]{}, Sweden. Emails: [email protected] and [email protected]. Research supported by Swedish Research Council grant 2016-03488.
[^2]: Faculty of Informatics, Masaryk University, Botanická 68A, 602 00 Brno, Czech Republic. Email: [email protected] This work has received funding from the European Research Council (ERC) under the European Union’s Horizon 2020 research and innovation programme (grant agreement No 648509) and from the MUNI Award in Science and Humanities of the Grant Agency of Masaryk University. This publication reflects only its authors’ view; the European Research Council Executive Agency is not responsible for any use that may be made of the information it contains.
[^3]: As observed in [@FalgasRavry12], a simple $k$-independent variant of Kolmogorov’s zero–one law shows that percolation remains a tail event when we consider $k$-independent models.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Cooperation between terminals has been proposed to improve the reliability and throughput of wireless communication. While recent work has shown that relay cooperation provides increased diversity, increased multiplexing gain over that offered by direct link has largely been unexplored. In this work we show that cooperative multiplexing gain can be achieved by using a half duplex relay. We capture relative distances between terminals in the high SNR diversity multiplexing tradeoff (DMT) framework. The DMT performance is then characterized for a network having a single antenna half-duplex relay between a single-antenna source and two-antenna destination. Our results show that the achievable multiplexing gain using cooperation can be greater than that of the direct link and is a function of the relative distance between source and relay compared to the destination. Moreover, for multiplexing gains less than 1, a simple scheme of the relay listening 1/3 of the time and transmitting 2/3 of the time can achieve the 2 by 2 MIMO DMT.'
author:
-
bibliography:
- 'IEEEabrv.bib'
- 'isit09.bib'
title: Cooperative Multiplexing in the Multiple Antenna Half Duplex Relay Channel
---
Introduction
============
There is a growing interest in the design of cooperative schemes that provide diversity and multiplexing gain for communication via wireless relays. Cooperative diversity refers to the additional diversity gain (compared to direct link) offered by cooperation. Similarly if a relay provides additional degrees of freedom (compared to direct link) it is said to provide a cooperative multiplexing gain [@4698542].
Diversity multiplexing tradeoff (DMT) [@1197843] has been widely used to analyze and compare the performance of cooperative schemes. DMT for the half-duplex single relay network has been studied extensively in literature [@1362898][@4305423][@sameer]. For the case with single antennas at all terminals the $2\times 1$ MISO DMT bound has recently been shown [@sameer] to be achievable. In this paper we study the DMT for the multiple antenna half-duplex relay channel having $m,n$ and $k$ antennas at source, destination and relay respectively. This was studied in [@4305423] but results were shown only for the special case $m=n=1$. We calculate the maximum achievable DMT for the $m=1,n=2,k=1$ configuration. In the process we also demonstrate techniques that enable results for general $m,n$ and $k$. $m=1,n=2,k=1$ is the simplest configuration where relay cooperation provides additional multiplexing gain compared to direct link. We show that if source and relay are relatively close to each other, cooperative multiplexing gain is achievable even with half duplex relaying. Note that the full-duplex case has been studied in [@4698542]. Moreover, for multiplexing gains less than $1$, if source-relay SNR (measured in dB) is at least two times the source-destination SNR a simple scheme with the relay listening $\frac{1}{3}$ of the time and transmitting $\frac{2}{3}$ of the time can achieve the $2\times2$ MIMO DMT.
These results lend fresh insight into the fundamental limits of cooperative multiplexing in the half-duplex relay channel. We demonstrate the use of two key techniques that enable our results.
1. *Distance between terminals:* In most results it is seen that relative distances between source, relay and destination do not affect DMT performance of the relay channel. Since DMT is calculated at high SNR the path loss and therefore distances are not easily captured in results. We overcome this apparent limitation by scaling the average SNR’s of the various links differently.
Our approach enriches the DMT framework by adding insights about network geometry.
2. *MIMO with half duplex antenna:* The min-cut capacity bound has been used in [@4305423] to calculate an upper bound for DMT performance. Notice in [ \[fig:model\]]{} that the $\{S,R\},\{D\}$ cut corresponds to a $2\times 2$ MIMO system with one source antenna that remains active only for a fraction of total communication time ($R$ is half duplex). It was noted [@4305423] that an upper bound for mutual information across such a cut is hard to compute. Due to this, DMT bounds have only been reported for the special case of $m=n=1$.
In [Sec \[calc\]]{} we demonstrate a simple channel decomposition that allows us to compute the cut-set DMT bound for the $m=1,n=2,k=1$ configuration. The technique can be applied towards computing DMT bounds for general $m,n$ and $k$. Recent results in [@4595031] [@Avestimehr:EECS-2008-128] show that a simple relaying scheme called “quantize-map” can achieve a rate within constant gap of the cut-set capacity. In [Sec \[achieve\]]{} we discuss this scheme and show that it achieves the cut-set DMT bound.
![Relay channel with $2$ antennas at destination and $S\to R$ proximity gain ${\eta}$. \[fig:model\]](system.eps)
System Model {#model}
============
Consider the system in [ \[fig:model\]]{} with source $S$, relay $R$ and destination $D$ having $1,1$ and $2$ antennas respectively. Let $D_{j}, j\in\{1,2\}$ denote the $j$th antenna at $D$. The channel gain for $S \to R$ is $h_{sr}$, gains for $S \to D_{j}$ are $h_{sj}$ and $R \to D_{j}$ are $h_{dj}$. All the channel gains are assumed to be flat fading having i.i.d. ${\mathcal{CN}}(0,1)$ distribution. We assume quasi-static fading, i.e. once realized, channel gains remain unchanged for the duration of the codeword and change independently between codewords. Noise at all receivers is additive i.i.d. ${\mathcal{CN}}(0,1)$ and independent of all other variables in the system. Transmit power at $S$ and $R$ is limited by an average power constraint. Since noise power at receiver is normalized to $1$, the transmit power constraint is specified by the average Signal to Noise Ratio (SNR). $R$ is assumed to operate under a half-duplex constraint. For simplicity it is assumed that transmissions at $S$ and $R$ are synchronous at symbol level.
We assume an asymmetrical network geometry. $S$ and $R$ are modeled to be close to each other as compared to $\{S,R\}$ and $D$. $S\to D$ and $R \to D$ are assumed to have the same average SNR denoted by $\rho$. $S\to R$ on the other hand is modeled to have SNR higher than $\rho$ by a factor ${\eta}$ on dB scale, i.e. the $S\to R$ average SNR is $\rho^{{\eta}}$. The $S\to R$ channel (cooperation link) thus has ${\eta}-1$ more degrees of freedom than other channels in the network. We call ${\eta}$ the proximity gain and assume ${\eta}\geq1$.
No channel state information (CSI) is available at $S$ i.e. only average channel statistics $\rho,{\eta}$ are known. However, at $D$ all channel realizations $h_{sr}$, $h_{sj},h_{dj}$ are completely known.
We identify three models for relaying strategy.
- *Global:* The relay uses knowledge of all instantaneous channel realizations to optimize its strategy.
- *Local:* The relay can measure $h_{sr}$ and uses only this (local) information.
- *Blind:* The relay only uses average channel statistics.
The *global* strategy is discussed in [Sec \[results\]]{} while *local* and *blind* are discussed in [Sec \[dynamic\]]{}.
Diversity-Multiplexing Tradeoff {#results}
===============================
![$d(r)$ for various values of proximity gain ${\eta}$.\[fig:1\]](fig1.eps)
\[thm:ga\] The maximum achievable DMT for network described in [Sec \[model\]]{} is given by, $$\label{eq:ga}
d(r)=\left\lbrace\begin{array}{cc}
\min\{{\eta}+2,4\}-3r & 0\leq r\leq1,{\eta}\geq1 \\
(2{\eta}-{\eta}r-1)/({\eta}-r) & 1\leq r\leq2-\frac{1}{{\eta}}, {\eta}\geq 2 \\
{\eta}-\frac{1}{2-r} & 1\leq r\leq 2-\frac{1}{{\eta}}, 1\leq {\eta}\leq2
\end{array}\right.$$
For system model described in [Sec \[model\]]{} the maximum achievable multiplexing gain $r^{*}=\inf_{r\geq0} \{r|d(r)=0\}$ is, $$\label{eq:rmax}
r^{*}=2-\frac{1}{{\eta}}$$
For a symmetrical geometry with all channels having the same degrees of freedom (${\eta}=1$) we get $r^{*}=1$ i.e. cooperation doesn’t provide additional maximum multiplexing gain. To enable higher multiplexing gain the $S\to R$ channel (cooperation link) needs to have more degrees of freedom than the $S\to D$ channel (communication link).
Let $d_{2\times 2}(r)$ represent the DMT of the $2\times2$ MIMO channel. For finite ${\eta}$ it can be seen that $d(r)\leq d_{2\times2}(r)$ with strict inequality over a non-empty region of $r$. This suggests that for distributed antennas the finite capacity of the cooperation channel $(S\to R)$ poses a fundamental limitation on the achievable DMT performance. It can easily be verified that, $$\lim_{{\eta}\to \infty} d(r) \to d_{2\times2}(r)$$ [ \[fig:1\]]{} shows $d(r)$ for several values of ${\eta}$.
We prove [Theorem \[thm:ga\]]{} in two steps. In [Sec \[calc\]]{} we show that the cut-set DMT upper bound for network in [Sec \[model\]]{} is given by . In [Sec \[achieve\]]{} we show that this bound is achievable.
Cut-Set DMT upper bound {#calc}
-----------------------
Let $f (0\leq f \leq 1)$ denote a listen-transmit schedule for the half duplex relay. $R$ listens for a fraction $f$ (listening phase) of total communication time and transmits for fraction $(1-f)$ (cooperation phase). The two cuts of the network $\mathcal{C}_{D}=\{S,R\},\{D\}$ and $\mathcal{C}_{S}=\{S\}\{R,D\}$ are shown in [ \[fig:cuts\]]{} for these two phases. In the listening phase let $\mathbf{X}_{S}^{1}$ denote the sequence of symbols transmitted by $S$ while $\mathbf{Y}_{R}$ and $\mathbf{Y}_{D}^{1}$ denote received signals at $R$ and $D$ respectively. Similarly for the cooperation phase $\mathbf{X}_{S}^{2}$ and $\mathbf{X}_{R}$ are the symbol sequences transmitted from $S$ and $R$ while $\mathbf{Y}_{D}^{2}$ is received at $D$. The instantaneous mutual information across the two cuts can be written as,
![Two cuts of network during listen $(f)$ and cooperation phase.$(1-f)$.\[fig:cuts\]](cuts.eps)
[$$\begin{aligned}
I_{\mathcal{C}_{S}}&=&fI(\mathbf{X}_{S}^{1};\mathbf{Y}_{R},\mathbf{Y}_{D}^{1}|\mathbf{X}_{R})+(1-f)I(\mathbf{X}_{S}^{2};\mathbf{Y}_{D}^{2}|\mathbf{X}_{R}) \label{eq:icut1}\\
I_{\mathcal{C}_{D}}&=&fI(\mathbf{X}_{S}^{1};\mathbf{Y}_{D}^{1})+(1-f)I(\mathbf{X}_{S}^{2},\mathbf{X}_{R};\mathbf{Y}_{D}^{2}) \label{eq:icut2}\end{aligned}$$ ]{}
To maximize these mutual information expressions we need to choose zero-mean complex Gaussian distributions for $\mathbf{X}_{S}^{1},\mathbf{X}_{S}^{2}$ and $\mathbf{X}_{R}$ that have covariance matrices which satisfy their respective average power constraints. Using these distributions we can write mutual information upper bounds ${I_{\mathcal{C}_{S}}^{'}}$ and ${I_{\mathcal{C}_{D}}^{'}}$ for $I_{\mathcal{C}_{S}}$ and $I_{\mathcal{C}_{D}}$ respectively. [ $$\begin{aligned}
I_{\mathcal{C}_{S}}\leq I_{\mathcal{C}_{S}}^{'}&=&f\log(1+\rho^{{\eta}}|h_{sr}|^{2}+\rho||{\mathbf{h}}_{s}||^{2}) \\
&&+(1-f)\log(1+\rho||{\mathbf{h}}_{s}||^{2}) \\
&\approx& f\max\{\log(1+\rho^{{\eta}}|h_{sr}|^{2}),\log(1+\rho||{\mathbf{h}}_{s}||^{2})\} \\
&& + (1-f)\log(1+\rho||{\mathbf{h}}_{s}||^{2}) \\
I_{\mathcal{C}_{D}}\leq I_{\mathcal{C}_{D}}^{'}&=&f\log(1+\rho||{\mathbf{h}}_{s}||^{2}) \\
&&+(1-f)\log\det({\mathbf{I}}+\rho{\mathbf{H}}{\mathbf{H}}^{\dagger}) \end{aligned}$$ ]{} where ${\mathbf{h}}_{s}=\left[\begin{array}{c} h_{s1} \\ h_{s2}\end{array}\right], {\mathbf{h}}_{r}=\left[\begin{array}{c} h_{r1} \\ h_{r2}\end{array}\right]$ and ${\mathbf{H}}=[\begin{array}{cc}{\mathbf{h}}_{s} & {\mathbf{h}}_{r}\end{array}].$ It can be verified that the approximation is tight within one bit.
Note that the expression for ${I_{\mathcal{C}_{D}}^{'}}$ is a linear combination of the capacities of two Raleigh fading Gaussian channels having correlated channel matrices ${\mathbf{H}}_{1}=[\begin{array}{cc}{\mathbf{h}}_{s} & 0 \end{array}]$ and ${\mathbf{H}}=[\begin{array}{cc}{\mathbf{h}}_{s} & {\mathbf{h}}_{r}\end{array}]$. Outage probability for the MIMO channel was calculated in [@1197843] using eigenvalue decomposition of the channel matrix. Following the same technique there would require computing the joint eigenvalue distributions for the two correlated hermitian matrices, ${\mathbf{H}}_{1}{\mathbf{H}}_{1}^{\dagger}$ and ${\mathbf{H}}{\mathbf{H}}^{\dagger}$. It was noted in [@4305423] that this is hard to compute. We propose an easier decomposition to solve this problem. The second term in ${I_{\mathcal{C}_{D}}^{'}}$ is the capacity of a $2\times2$ MIMO channel which can be represented by two parallel Gaussian channels having gains $g_{1}$ and $g_{2}$ shown in [ \[fig:parchan\]]{}.
![Parallel Channel Model for $2\times 2$ MIMO.\[fig:parchan\]](parchan.eps)
The channel can be written as, [ $$\left[\begin{array}{c}y_{1} \\ y_{2}\end{array}\right]=\sqrt{\rho}\left[\begin{array}{cc}g_{1}& 0 \\ 0 & g_{2}\end{array}\right]\left[\begin{array}{c}x_{1} \\ x_{2}\end{array}\right]+\left[\begin{array}{c}w_{1} \\ w_{2}\end{array}\right]$$ ]{} where $E[|x_{i}|^{2}]=1$, $w_{1},w_{2}\sim{\mathcal{CN}}(0,1)$. The capacity for $x_{i}\to y_{i}$ is given by $\log(1+\rho g_{i}^{2})$. It was shown in [@679134][@1197843] that a D-BLAST transmission scheme with a MMSE successive interference cancellation receiver achieves the mutual information of the MIMO channel. For this scheme $g_{1}^{2}$ and $g_{2}^{2}$ can be calculated to be, [ $$\begin{aligned}
g_{2}^{2}&=&||\mathbf{h}_{r\perp s}||^{2}+\frac{||\mathbf{h}_{r\parallel s}||^{2}}{1+\rho||\mathbf{h}_{s}||^{2}}\\
g_{1}^{2}&=&||\mathbf{h}_{s}||^{2}\end{aligned}$$ ]{} where $\mathbf{h}_{r\perp s}$ and $\mathbf{h}_{r\parallel s}$ respectively denote the perpendicular and parallel components of $\mathbf{h}_{r}$ with respect to $\mathbf{h}_{s}$.
Note that while $g_{1}^{2}$ and $g_{2}^{2}$ are correlated, $\mathbf{h}_{s},\mathbf{h}_{r\perp s}$ and $\mathbf{h}_{r\parallel s}$ are mutually independent. The correlation between $g_{1}^{2}$ and $g_{2}^{2}$ can therefore be explicitly calculated. The destination decodes $\mathbf{X}_{R}$ in the presence of interference from $\mathbf{X}_{S}^{2}$. It then cancels $\mathbf{X}_{R}$ from its received signal before decoding $\mathbf{X}_{S}^{2}$. Therefore $S$ effectively sees an interference free channel (with gain $g_{1}$) to $D$ during both listen and cooperation phases. [ $$\begin{aligned}
{I_{\mathcal{C}_{D}}^{'}}&=&f\log(1+\rho g_{1}^{2}) \\
&&+(1-f)[\log(1+\rho g_{1}^{2})+\log(1+\rho g_{2}^{2})] \\
&=& \log(1+\rho g_{1}^{2})+(1-f)\log(1+\rho g_{2}^{2})\end{aligned}$$ ]{}
Let $\alpha_{sr},\alpha_{1}$ and $\alpha_{2}$ represent channel realizations via the following variable transformations. [ $$\begin{aligned}
\alpha_{sr}&=&\lim_{\rho\to\infty}\frac{\log(1+\rho^{{\eta}} |h_{sr}|^{2})}{\log{\rho}} \\
\alpha_{1}&=&\lim_{\rho\to\infty}\frac{\log(1+\rho g_{1}^{2})}{\log{\rho}} \\
\alpha_{2}&=&\lim_{\rho\to\infty}\frac{\log(1+\rho g_{2}^{2})}{\log{\rho}}\end{aligned}$$ ]{} This gives us simplified expressions for mutual information upper bounds. [ $$\begin{aligned}
\frac{{I_{\mathcal{C}_{S}}^{'}}}{\log \rho}&=&\alpha_{1}+f(\alpha_{sr}-\alpha_{1})^{+} \label{eq:cone}\\
\frac{{I_{\mathcal{C}_{D}}^{'}}}{\log \rho}&=&\alpha_{1}+(1-f)\alpha_{2} \label{eq:ctwo}\end{aligned}$$ ]{}
To achieve desired multiplexing gain $r$ at high SNR ($\rho \to \infty$) the network must achieve a rate ${\mathcal{R}}=r\log\rho$. The network is in outage if, $\min\{{I_{\mathcal{C}_{S}}^{'}},{I_{\mathcal{C}_{D}}^{'}}\}\leq r\log\rho$. For a given $r$ and schedule $f$ we can define the outage region $\mathcal{O}(r,f)$ over channel realizations $\vec{\alpha}=(\alpha_{1},\alpha_{2},\alpha_{sr})$. [ $$\mathcal{O}(r,f)=\{\vec{\alpha}|\frac{\min\{{I_{\mathcal{C}_{S}}^{'}},{I_{\mathcal{C}_{D}}^{'}}\}}{\log\rho}\leq r\}$$ ]{}
The outage probability $\mathcal{P}_{out}$ is, [ $$\mathcal{P}_{out}=\int_{\vec{\alpha}\in\mathcal{O}(r,f)} f_{\vec{\alpha}}(\alpha_{1},\alpha_{2},\alpha_{sr})$$ ]{} where $f_{\vec{\alpha}_{1}}(\alpha_{1},\alpha_{2},\alpha_{sr})$ is the joint distribution of $(\alpha_{1},\alpha_{2},\alpha_{sr})$.
[Proof see [Appendix \[appa\]]{}]{}[ $$f_{\vec{\alpha}}(\alpha_{1},\alpha_{2},\alpha_{sr})\doteq \rho^{-s(\vec{\alpha})}$$ ]{} where $0\leq \alpha_{1},\alpha_{2}\leq 1$, $0 \leq \alpha_{sr}\leq {\eta}$ and $$s(\vec{\alpha})=\left\lbrace\begin{array}{cc}{\eta}+4-3\alpha_{1}-2\alpha_{2}-\alpha_{sr} & \alpha_{1}+\alpha_{2}\leq 1 \\ {\eta}+3-2\alpha_{1}-\alpha_{2}-\alpha_{sr} & \alpha_{1}+\alpha_{2}>1\end{array}\right.$$ \[lem1\]
For a given listen-transmit schedule $f$ the cut-set DMT bound is therefore given by, $$d(r,f)=\min_{\vec{\alpha}\in \mathcal{O}(r,f)} s(\vec{\alpha}) \label{eq:drt}$$ To get the DMT upper bound we can optimize over all listen-transmit schedules, $$d(r)=\min_{\vec{\alpha}\in \mathcal{O}(r)} \max_{f} s(\vec{\alpha})$$ Note that this optimization is performed on a per realization basis, i.e. the optimal $f$ depends on all channel realizations $\alpha_{sr},\alpha_{1}$ and $\alpha_{2}$. Therefore this corresponds to the *global* strategy discussed in [Sec \[model\]]{}.
It is easy to see that the globally optimal schedule $f_{glob}$ is one which sets ${I_{\mathcal{C}_{S}}^{'}}={I_{\mathcal{C}_{D}}^{'}}$. $$f_{glob}=\frac{\alpha_{2}}{(\alpha_{sr}-\alpha_{1})^{+}+\alpha_{2}}$$ This leads to the solution for $d(r)$ given in .
Achievability: Relaying Scheme {#achieve}
==============================
The “quantize-map" relaying scheme proposed in [@4595031] and [@Avestimehr:EECS-2008-128] has been shown to be DMT optimal for the single antenna relay channel [@sameer]. We show that “quantize-map” adapts naturally to the network described in [Sec \[model\]]{} and with some modification achieves the cut-set DMT bound. For the sake of completeness we include a short description of the scheme.
Description of scheme {#scheme}
---------------------
$S$ has a sequence of messages $w_n
\in\{1,2,\ldots,2^{T{\mathcal{R}}}\}$, $n=1,2,\ldots$ to be transmitted. At both source $S$ and relay we create random Gaussian codebooks. $S$ randomly maps each message to one of its Gaussian codewords and transmits it using $T$ symbol times giving an overall transmission rate of ${\mathcal{R}}$. Due to the half-duplex nature of the relay, it must operate using listen-transmit cycles. Relay listens to the first $fT$ time symbols of each block i.e. $\mathbf{X}_{S}^{1}$. It quantizes $\mathbf{Y}_{R}$ to $\hat{\mathbf{Y}}_R$ and then randomly maps it into a Gaussian codeword ${\mbox{${\bf X }$} }_R$ using a random mapping function $f_R(\hat{ {\mbox{${\bf Y }$} }}_R)$. It transmits this codeword during the next $(1-f)T$ symbol times. Given the knowledge of all the encoding functions and signals received, $D$ attempts to decode the message sent by $S$.
DMT of Quantize-Map
-------------------
By Theorem 7.4.1 in [@Avestimehr:EECS-2008-128], for any fixed listen-transmit schedule $f$, the quantize-map relaying scheme, uniformly over all channel realizations achieves a rate within a constant gap to the cut-set upper bound $\min\{{I_{\mathcal{C}_{S}}^{'}},{I_{\mathcal{C}_{D}}^{'}}\}$ for that particular $f$. The random Gaussian code-book generated at source is independent of $f$. Also the code-book generated at relay depends on $f$ only to determine the length of each codeword $(1-f)T$.
The relay can generate a larger code-book with each codeword of length $T$. If the relay now chooses a listen-transmit schedule $f$, it can use the first $(1-f)T$ symbols of the codeword to compose ${\mbox{${\bf X }$} }_R$. The destination always knows the schedule $f$ and hence can adapt its decoder accordingly. This construction allows us to claim that “quantize-map" achieves a rate within a constant gap of $\min\{{I_{\mathcal{C}_{S}}^{'}},{I_{\mathcal{C}_{D}}^{'}}\}$ uniformly for each dynamic choice of $f$ i.e. [ $$\label{eq:QMRate}
\min\{{I_{\mathcal{C}_{S}}^{'}},{I_{\mathcal{C}_{D}}^{'}}\} -\kappa \leq
R_{\text{quantize-map}}(h_{sr},{\mathbf{h}}_{s},{\mathbf{h}}_{r},\rho,f)$$]{}
The constant $\kappa$ in the above equation does not depend on the channel gains and $\mbox{SNR}$. At the order of DMT which assumes high SNR $(\rho\to \infty)$ the effect of $\kappa$ becomes negligible and hence we have the following theorem for achievability.
\[thm:achieve\] For dynamic listen-transmit schedules, the modified quantize-map relaying scheme as described above achieves the diversity multiplexing tradeoff of $\min\{{I_{\mathcal{C}_{S}}^{'}},{I_{\mathcal{C}_{D}}^{'}}\}$, where ${I_{\mathcal{C}_{S}}^{'}}$ and ${I_{\mathcal{C}_{D}}^{'}}$ are given by .
Achievability: Listen-Transmit Schedule {#dynamic}
=======================================
In [Sec \[calc\]]{} the cut-set DMT upper bound was calculated for the *globally* optimal listen-transmit schedule $f_{glob}$. However in a practical communication scenario global knowledge of instantaneous channel realizations may not be available at the relay. To account for this we defined the *local* and *blind* relaying strategies in [Sec \[model\]]{}. In this section we refine [Theorem \[thm:ga\]]{} to calculate DMT bounds for *local* and *blind* schedules.
Blind Scheduling
----------------
\[thm:lowrate\] For the low rate region i.e. $r\leq1$, the *blind* scheduling strategy is DMT optimal. Additionally for ${\eta}\geq2$ the *blind* strategy achieves the $2\times2$ MIMO DMT bound for $r\leq1$. The optimal blind schedule for this region is $f_{blind}=\frac{1}{3}$ $$\begin{array}{cc}
d_{blind}(r)=\min\{{\eta}+2,4\}-3r & r\leq 1
\end{array}$$
![$d(r)$ $d_{local}(r)$ and $d_{blind}(r)$ comparison for $r\geq 1$.\[fig:2\]](fig3.eps)
From the DMT bound for blind scheduling can be written as, $$d_{blind}(r)= \max_{f} \min_{\vec{\alpha}\in \mathcal{O}(r,f)} s(\vec{\alpha}) \label{eq:blind}$$ i.e. $f$ is optimized without knowledge of channel realizations $\vec{\alpha}$. Solving this optimization for $r\leq1$ yields [Theorem \[thm:lowrate\]]{}. This suggests that as long as cooperative multiplexing is not necessary i.e. desired rate ${\mathcal{R}}=r\log(\rho)$ is such that $r\leq1$, static scheduling at relay is sufficient to achieve the DMT upper bound. $f_{blind}=\frac{1}{3}$ turns out to be the optimal listen-transmit schedule for this region.
For the high rate region $(r>1)$, the analytical solution for is tedious to obtain. The optimization is convex and can be solved numerically. [ \[fig:2\]]{} shows a comparison between $d(r)$ and $d_{blind}(r)$ for $r\geq1$. It can be seen that for cooperative multiplexing $(r\geq1)$ static scheduling is insufficient to achieve DMT upper bound.
Local Scheduling
----------------
Similarly, the DMT bound for local scheduling can be expressed as an optimization problem from . $$d_{local}(r)=\min_{\alpha_{sr}} \max_{f} \min_{\alpha_{1},\alpha_{2}\in\mathcal{O}(r,f)} s(\vec{\alpha})
\label{eq:local}$$ $f_{local}$ can be optimized using knowledge of $\alpha_{sr}$ only. The DMT performance of local scheduling must be at-least as good as blind scheduling, therefore by [Theorem \[thm:lowrate\]]{} for $r\leq 1$ $d_{local}(r)=d_{blind}(r)=d(r)$.
Numerical solution to for the high rate $r\geq1$ region is shown in [ \[fig:2\]]{}. It can be seen that local scheduling performs better than blind, but for higher ${\eta}$ this advantage diminishes.
Acknowledgements
================
The authors wish to acknowledge the contributions of the students, faculty and sponsors of the Berkeley Wireless Research Center and the National Science Foundation Infrastructure Grant No. 0403427.
Proof of [Lemma \[lem1\]]{} {#appa}
===========================
Marginal Distribution of $\alpha_{sr}$
--------------------------------------
$f_{\alpha_{sr}}(\alpha)$ is calculated as, $$\begin{aligned}
\mathcal{P}[\alpha_{sr}<\alpha]&=&\lim_{\rho\to\infty}\mathcal{P}[|h_{sr}|^{2}<\rho^{\alpha_{sr}-{\eta}}] \\
f_{\alpha_{sr}}(\alpha)&\doteq&\rho^{{\eta}-\alpha} (0 \leq \alpha_{sr}\leq {\eta})\end{aligned}$$
Joint Distribution of $\alpha_{1}$ and $\alpha_{2}$
---------------------------------------------------
Note that $g_{1}^{2}$ has a $\chi^{2}_{4}$ distribution, the marginal distribution of $\alpha_{1}$ is given by, $$\begin{aligned}
f_{\alpha_{1}}(\alpha_{1})&=f_{g_{1}^{2}}(\rho^{-(1-\alpha_{1})}) \frac{dg_{1}^{2}}{d\alpha_{1}} \\
&\doteq \rho^{-2(1-\alpha_{1})} (0\leq\alpha_{1}\leq1)\end{aligned}$$ Now for $(0\leq\alpha_{1},\alpha_{2}\leq1)$ their joint CDF can be written as, $$\begin{aligned}
F_{\alpha_{1},\alpha_{2}}(\alpha_{1},\alpha_{2})&=&\mathcal{P}[g_{1}^{2}\leq \rho^{\alpha_{1}-1},g_{2}^{2}\leq \rho^{\alpha_{2}-1}] \\
&=&\int_{x=0}^{\alpha_{1}} \mathcal{P}[g_{1}^{2}= \rho^{x-1},g_{2}^{2}\leq \rho^{\alpha_{2}-1}] dx \\
&\doteq& \int_{x = 0}^{\alpha_1}\rho^{-2(1-x)} \\ && \mathcal{P}(|h_{s\perp r}|^2 +
\frac{|h_{s\parallel r} |^2}{1 + \rho^{x}} \leq \rho^{\alpha_2 - 1})dx \\
&\doteq& \int_{x = 0}^{\alpha_1}\rho^{-2(1-x)} \\ && \int_{y=0}^{\alpha_2
-1}\rho^{y}\mathcal{P}(|h_{s\parallel r} |^2\dot\leq \rho^{x+\alpha_2 - 1})\ dy\ dx\\\end{aligned}$$
### Case $\alpha_{1}+\alpha_{2}\leq 1$
$$\begin{aligned}
F_{{\alpha_1,\alpha_2}}(\alpha_1,\alpha_2) &\doteq& \int_{x =
0}^{\alpha_1}\rho^{-2(1-x)}\int_{y=0}^{\alpha_2
-1}\rho^{x+y+\alpha_2 - 1}\ dy\ dx \\
&\doteq& \int_{x = 0}^{\alpha_1}\rho^{3x+2\alpha_2 - 4}\ dx\\
&\doteq& \rho^{3\alpha_1+2\alpha_2 - 4}\\
f_{\alpha_1,\alpha_2}(\alpha_{1},\alpha_{2})
&\doteq& \rho^{3\alpha_1+2\alpha_2 - 4}\end{aligned}$$
### Case $\alpha_{1}+\alpha_{2} > 1$
$$\begin{aligned}
F_{{\alpha_1,\alpha_2}}(\alpha_1,\alpha_2) &\doteq& \int_{x = 0}^{1-\alpha_2}\rho^{3x+2\alpha_2 - 4}\ dx
\\&& +\int_{x =1-\alpha_2}^{\alpha_1}\rho^{2x + \alpha_2 -3}\ dx\\
&\doteq& \rho^{2\alpha_2 + \alpha_1 -3} \\
f_{\alpha_1,\alpha_2}
&\doteq& \rho^{2\alpha_1 + \alpha_2 -3}\end{aligned}$$
Since $\alpha_{sr}$ is independent of $\alpha_{1},\alpha_{2}$ we get [Lemma \[lem1\]]{}.
| {
"pile_set_name": "ArXiv"
} |
[**Every bordered Riemann surface is a complete proper curve in a ball**]{}
[**Antonio Alarcón and Franc Forstnerič**]{}
> [**Abstract**]{} We prove that every bordered Riemann surface admits a complete proper holomorphic immersion into a ball of $\c^2$, and a complete proper holomorphic embedding into a ball of $\c^3$.
>
> [**Keywords**]{} Riemann surfaces, complex curves, complete holomorphic immersions.
>
> [**MSC (2010):**]{} 32B15, 32H02, 14H50, 53C42.
Introduction {#sec:intro}
============
The question of whether there exist complete bounded complex submanifolds of a Euclidean space $\c^n$ for $n>1$ was raised by Yang [@Yang1; @Yang2]. (An immersed submanifold $f\colon M\to \c^n$ is said to be [*complete*]{} if the induced Riemannian metric $f^* ds^2$ on $M$, obtained by pulling back the Euclidean metric $ds^2$ on $\c^n$ by the immersion, is a complete metric on $M$. A submanifold is bounded if it is contained in a relatively compact subset of the ambient space.) The first important result in this direction was obtained by Jones [@Jones] who constructed a holomorphic immersion of the unit complex disc $\d$ into $\c^2$, and an embedding into $\c^3$, with bounded image and complete induced metric. His method is strongly complex analytic and is based on the BMO duality theorem.
The closely related question on the existence of complete bounded minimal surfaces in $\r^3$ was a classical problem in the theory of minimal surfaces, known as the [*Calabi-Yau problem*]{}. The first affirmative answer was given by Nadirashvili [@Nad]; his method uses Runge’s theorem for holomorphic functions on planar labyrinths of compact sets.
No new examples of complete bounded complex submanifolds in $\c^n$ were discovered until recently when Martín et al. [@MUY] extended Jones’ result to complete bounded complex curves in $\c^2$ with arbitrary finite genus and finitely many ends. Their technique is completely different from the one of Jones and is much more geometric; they used the existence of a simply connected complete bounded null holomorphic curve in $\c^3$ [@AL] and modified a technique developed by F. J. López [@Lopez] for constructing complete minimal surfaces in $\r^3$.
Recently Alarcón and López [@AL] constructed complete bounded holomorphic curves in $\c^2$ and null curves in $\c^3$ of arbitrary topological type. Their method also comes from the theory of minimal surfaces in $\r^3$ and null curves in $\c^3$, and it strongly relies on Runge’s and Mergelyan’s approximation theorems. Their examples have the extra feature that they can be made proper in any convex domain of $\c^2$; in particular, in the unit ball $\b=\{z\in\c^2\colon |z|<1\}$. In the same spirit, Ferrer et al. [@FMM] constructed complete minimal surfaces with arbitrary topology and properly immersed in any convex domain of $\r^3$.
Despite their flexibility, none of these methods enables one to control the complex structure on the curve, except of course in the simply connected case when any such curve is biholomorphic to the unit disc.
The aim of this paper is to construct complete bounded complex curves that are normalized by any given bordered Riemann surface. Our main result is the following.
\[th:intro\] Every bordered Riemann surface admits a complete proper holomorphic immersion to the unit ball of $\c^2$, and a complete proper holomorphic embedding to the unit ball of $\c^3$.
Theorem \[th:intro\] is an immediate corollary of the following more precise result.
\[th:intro2\] Let $\Rcal$ be a bordered Riemann surface. Every holomorphic map $f\colon \overline \Rcal\to\c^n$ $(n>1)$ can be approximated, uniformly on compacta in $\Rcal$, by complete proper holomorphic immersions (embeddings if $n>2$) into any open ball in $\c^n$ containing the image $f(\overline \Rcal)$.
By a minor modification of our proof, one can replace balls by arbitrary convex domains. Our method, coupled with the techniques from [@DF], also yields complete proper holomorphic immersions of bordered Riemann surfaces into any Stein manifold of dimension $>1$; see Theorem \[th:Stein\] in Sec. \[sec:Stein\].
Our construction is inspired by that of Alarcón and López [@AL], but we use additional complex analytic tools. The examples in [@AL] appear after two different deformation procedures, starting from a given holomorphic immersion $f\colon \overline{\Rcal} \to\c^2$; the first one using Mergelyan’s theorem, and the second one Runge’s theorem. Since the latter provides no information on the placement in $\c^2$ of some parts of the curve, one must shrink the curve to guarantee its boundedness, thereby losing control of its complex structure. In the present paper we replace Runge’s theorem by the solution of certain Riemann-Hilbert boundary value problems. This gives sufficient control of the position of the complex curve in the ambient space to avoid shrinking. The use of the Riemann-Hilbert problem in the construction of proper holomorphic maps has a long history; see the 1992 paper of Forstnerič and Globevnik [@FG] and the introduction and references in [@DF].
Another important tool in our construction is the method of Forstnerič and Wold [@FW] for exposing boundary points of a complex curve in $\c^n$. This method is used in the inductive step to insure that any curve terminating near a specific finite set of boundary points of $\Rcal$ becomes sufficiently long. On the boundary arcs separated by these points we apply the Riemann-Hilbert method to insure that any curve ending on such an arc is sufficiently long. In a certain precise sense, the first method enables us to localize the length estimates furnished by the second method, thereby preventing the appearance of any shortcuts in the new complex curve.
These additions allow us to simplify the construction in [@AL] and, what is the main point, to avoid changes of the complex structure on the Riemann surface.
It is classical that any open Riemann surface immerses properly holomorphically into $\c^2$ [@Bi; @Nar] (see also [@AL2]), but it is an open question whether it embeds into $\c^2$ [@BN]. Forstnerič and Wold [@FW] gave an affirmative answer for bordered Riemann surfaces whose closures embed holomorphically into $\c^2$, and for all circular domains in $\c$ without punctures [@FW2]. (See also [@Majcen].) Since the immersions into $\c^2$ constructed in the present paper may have self-intersections, the following problem remains open.
Does there exist a complete (properly) embedded complex curve in a ball of $\c^2$?
As we have already mentioned, there exist complete bounded immersed minimal surfaces and null curves with arbitrary topology [@AL; @FMM]. The first examples of complete properly [*embedded null curves*]{} with arbitrary topology in a ball of $\c^3$ have been constructed recently in [@AF Corollary 6.2]. However, the methods in those papers do not enable one to control the complex structure on these surfaces. In the light of Theorem \[th:intro\], the following is therefore a natural question.
Does every bordered Riemann surface admit a complete conformal minimal immersion into a ball of $\r^3$? Even more, does every such surface admit a complete null holomorphic immersion (or embedding) into a ball of $\c^3$?
By using the theorem of Jones [@Jones] one also finds examples of bounded complete complex submanifolds of arbitrary dimension in $\c^N$ for large $N$. In particular, we have the following corollary to Theorem \[th:intro\]; we wish to thank J. E. Fornæss for this observation. Let $\mathbb{B}^k$ denote the unit ball of $\c^k$ and $(\mathbb{B}^k)^l \subset \c^{kl}$ the Cartesian product of $l$ copies of $\mathbb{B}^k$.
\[cor:higher\] Every relatively compact, strongly pseudoconvex domain in a Stein manifold of dimension $n$ admits a complete proper holomorphic immersion into $(\mathbb{B}^2)^{2n} \subset \c^{4n}$, and a complete proper holomorphic embedding into $(\mathbb{B}^3)^{2n+1} \subset \c^{6n+3}$.
Denote by $\mathbb{D}^k$ the unit polydisc in $\c^k$. According to [@DF2010], any relatively compact strongly pseudoconvex domain $D$ in a Stein manifold of dimension $n$ admits a proper holomorphic immersion $g\colon D\to \mathbb{D}^{2n}$ and a proper holomorphic embedding $h\colon D\hookrightarrow \mathbb{D}^{2n+1}$. (For similar results see also the papers [@Dor1995; @Low2; @Stensones1989].) Theorem \[th:intro\] furnishes a proper complete holomorphic immersion $\xi\colon \mathbb{D}\to \mathbb{B}^2$ of the disc into the ball of $\c^2$, and a proper complete holomorphic embedding $\eta \colon \mathbb{D}\to \mathbb{B}^3$ into the ball of $\c^3$. For any integer $k\in\mathbb{N}$ let $\xi^k\colon \mathbb{D}^k \to \c^{2k}$ be defined by $$\xi^k(z_1,z_2,\ldots,z_k)=(\xi(z_1),\xi(z_2),\ldots, \xi(z_k)),\quad (z_1,\ldots,z_k)\in \mathbb{D}^k.$$ Clearly $\xi^k$ immerses $\mathbb{D}^k$ properly into $(\mathbb{B}^2)^k \subset \c^{2k}$. Similarly we define the map $\eta^k\colon \mathbb{D}^k \to \c^{3k}$ which embeds $\mathbb{D}^k$ properly into $(\mathbb{B}^3)^k$. It is now easily verified that the maps $\xi^{2n} \circ g \colon D \to (\mathbb{B}^2)^{2n} \subset \c^{4n}$ and $\eta^{2n+1}\circ h\colon D \to (\mathbb{B}^3)^{2n+1} \subset \c^{6n+3}$ satisfy the conclusion of the corollary.
The dimensions in Corollary \[cor:higher\] are probably far from optimal. It would be interesting to answer the following question.
Let $n>1$. Does the ball $\mathbb{B}^n\subset \c^n$ admit a proper complete holomorphic immersion (or embedding) into the ball $\mathbb{B}^N$ for some $N>n$? What is the answer if $n=2$ and $N=3,4,\ldots$?
The paper is organized as follows. In Sec. \[sec:preli\] we collect the technical tools. The central part of the paper is Sec. \[sec:mainlemma\] where we state and prove the main technical result, Lemma \[lem:main\]. Theorem \[th:intro2\] is proved in Sec. \[sec:proof\] by a recursive application of Lemma \[lem:main\]. In Sec. \[sec:Stein\] we outline the construction of proper complete holomorphic immersions of bordered Riemann surfaces to an arbitrary Stein manifold of dimension $>1$.
Preliminaries {#sec:preli}
=============
We denote by $\langle\cdot,\cdot\rangle$, $|\cdot|$, $\dist(\cdot,\cdot)$, and $\length(\cdot)$, respectively, the hermitian inner product, norm, distance, and length on $\c^n$. Hence $\Re \langle\cdotp,\cdotp\rangle$ is the Eulidean inner product on $\r^{2n}\cong\c^n$.
Given $u\in\c^n$, set $\langle u\rangle^\bot=\{v\in\c^n\colon \langle u,v\rangle=0\}$ and $\span\{u\}=\{\zeta u\colon \zeta\in\c\}$. If $u\neq 0\in\c^n$, then $\langle u\rangle^\bot$ is a complex hyperplane in $\c^n$.
For any compact topological space $K$ and continuous map $f\colon K\to\c^n$, denote by $\|f\|_{0,K}=\sup_{x\in K} |f(x)|$ the maximum norm of $f$ on $K$. If $K$ is a subset of a smooth manifold $M$ and $f$ is of class $\Ccal^r$ for some $r\in\z_+=\n\cup\{0\}$, we denote by $\|f\|_{r,K}$ the $\Ccal^r$-maximum norm of $f$ on $K$, measured with respect the expression of $f$ in a system of local coordinates in some fixed finite open cover of $K$. Similarly we define these norms for maps $M\to X$ to a smooth manifold $X$ by using a fixed cover by coordinate patches on both manifolds.
Given a topological surface $M$ with boundary, we denote by $bM$ the $1$-dimensional topological manifold determined by the boundary points of $M$. By a [*domain*]{} in $M$ we mean an open connected subset of $M\setminus bM$. A surface is said to be [*open*]{} if it is non-compact and does not contain any boundary points. A [*Riemann surface*]{} is an oriented surface together with the choice of a complex structure.
An open connected Riemann surface, $\Rcal$, is said to be a [*bordered Riemann surface*]{} if it is the interior of a compact one dimensional complex manifold, $\overline{\Rcal}$, with smooth boundary $b\overline{\Rcal}$ consisting of finitely many closed Jordan curves.
A domain in a bordered Riemann surface $\Rcal$ is said to be a [*bordered domain*]{} if it is a bordered Riemann surface, hence with smooth boundary. It is classical that every bordered Riemann surface is biholomorphic to a relatively compact bordered domain in a larger Riemann surface $\widehat \Rcal$ which can be chosen either open or compact.
Denote by $\Bscr(\Rcal)$ the family of bordered domains $\Mcal\Subset\Rcal$ such that $\Rcal$ is a tubular neighborhood of $\overline{\Mcal}$; that is to say, $\overline{\Mcal}$ is a Runge subset of $\Rcal$ and $\Rcal\setminus \overline{\Mcal}$ consists of finitely many open annuli.
Let $\Rcal$ be a bordered Riemann surface and $X$ be a complex manifold. Given a number $r\ge 0$, we denote by $\Ascr^r(\Rcal,X)$ the set of all maps $f\colon \overline \Rcal\to X$ of class $\Ccal^r$ that are holomorphic on $\Rcal$. (This space is defined also for non-integer values of $r$ by Hölder continuity.) For $r=0$ we write $\Ascr^0(\d,X)=\Ascr(\d,X)$. When $X=\c$, we write $\Ascr^r(\Rcal,\c) = \Ascr^r(\Rcal)$ and $\Ascr^0(\Rcal) = \Ascr(\Rcal)$. Note that $\Ascr^r(\Rcal,\c^n) = \Ascr^r(\Rcal)^n$ is a complex Banach space. For any complex manifold $X$ and any number $r\ge 0$, the space $\Ascr^r(\Rcal,X)$ carries a natural structure of a complex Banach manifold [@F2007].
Let $\Iscr(\Rcal,\c^n)$ denote the set of all $\Ccal^1$ immersions $\overline{\Rcal}\to\c^n$ that are holomorphic in the interior. We shall also write $\Iscr(\Rcal)$ when the target is clear from the context.
Given an immersion $f\colon \Rcal\to\c^n$, we denote by $\sigma_f^2$ the Riemannian metric in $\Rcal$ induced by the Euclidean metric of $\c^n$ via $f$; that is, $$\sigma_f^2(p,v) = \langle df_p(v), df_p(v)\rangle, \quad p\in \Rcal,\ v\in T_p\Rcal.$$ If $f$ is holomorphic, then $\sigma_f^2$ is a conformal metric on $\Rcal$. We denote by $\dist_{(\Rcal,f)}(\cdot,\cdot)$ the distance function in the Riemannian surface $(\Rcal,\sigma_f^2)$.
A curve $\gamma\colon [0,1)\to \Rcal$ is said to be [*divergent*]{} if the map $\gamma$ is proper; that is, if the point $\gamma(t)$ leaves any compact subset of $\Rcal$ when $t\to 1$.
\[def:complete\_imm\] An immersion $f\colon \Rcal\to\c^n$ is said to be [*complete*]{} if $(\Rcal,\sigma_f^2)$ is complete as a Riemannian surface; that is to say, if the image curve $f\circ \gamma$ in $\c^n$ has infinite length for any connected divergent curve $\gamma$ in $\Rcal$.
A Riemann-Hilbert problem {#sec:RH}
-------------------------
We shall need approximate solutions of certain Riemann-Hilbert problems over the unit disc $\d=\{\zeta\in\c\colon |\zeta|<1\}$. Such results have been used by several authors; the precise version of the following lemma is taken from the papers [@DF; @DF2].
\[lem:Hilbert\] Fix an integer $n\in\n$. Let $f\in\Ascr(\d)^n$, and let $g\colon b\overline{\d}\times\overline{\d}\to\c^n$ be a continuous map such that for every fixed $\zeta\in b\overline{\d}$ we have $g(\zeta,\cdot)\in \Ascr(\d)^n$ and $g(\zeta,0)=f(\zeta)$. Given numbers $\epsilon>0$ and $0<r<1$, there exist a number $r'\in [r,1)$ and a map $h\in\Ascr(\d)^n$ satisfying the following conditions:
- $\dist(h(\zeta),g(\zeta,b\overline{\d}))<\epsilon$ for all $\zeta\in b\overline{\d}$,
- $\dist(h(\rho\zeta),g(\zeta,\overline{\d}))<\epsilon$ for all $\zeta\in b\overline{\d}$ and $\rho\in [r',1)$, and
- $|h(\zeta)-f(\zeta)|<\epsilon$ when $|\zeta|\leq r'$.
In particular, if $J$ is a compact arc in the circle $ b\overline{\d}$ and $g(\zeta,\cdot)=f(\zeta)$ is the constant disc for all points $\zeta\in b\overline{\d}\setminus J$, then for any open neighborhood $C$ of $J$ in $\overline{\d}$ one can choose $h$ such that $\|h-f\|_{0,\overline{\d}\setminus C}<\epsilon$.
A characteristic feature of the map $h$ in the above lemma is that its boundary values must spin very fast in a small neighborhood of the torus in $\c^n$ that is formed by the boundaries of the analytic discs $g(\zeta,\cdotp)$ for $|\zeta|=1$. In the model situation in $\c^2$, with $f(\zeta)=(\zeta,0)$ and $g(\zeta,z)=(\zeta,z)$ for $|\zeta|=1$ and $|z|\le 1$, a suitable (exact!) solution is the map $h(\zeta)=(\zeta,\zeta^N)$ for large values of $N\in \n$. Indeed, the proof of the lemma amounts to a reduction to this model situation.
A gluing lemma {#sec:gluing}
--------------
An important technical step in our construction is to glue pairs of holomorphic mappings on special geometric configurations in a Riemann surface. We first explain this gluing for maps to $\c^n$; in the next subsection we address the correspoding problem for maps to an arbitrary complex manifold. (The general case is only used in Sec. \[sec:Stein\].)
\[def:Cartan\] A [*Cartan decomposition*]{} of a bordered Riemann surface $\overline \Rcal$ is a pair $(A,B)$ of compact domains with smooth boundaries in $\overline\Rcal$ such that $C:=A\cap B$ also has smooth boundary, and we have $$A\cup B =\overline \Rcal,\qquad \overline{A\setminus B}\cap \overline{B\setminus A} =\emptyset.$$
Compare with the notion of a [*Cartan pair*]{} (Chapter 5 in [@F2011]). In applications in this paper, one of the two sets $A,B$ will be a disjoint union of finitely many discs.
The following well known lemma is a solution of the Cousin-I problem with bounds; c.f. Lemma 5.8.2 in [@F2011]. The operators $\Acal$ and $\Bcal$ are obtained by using a smooth cut-off function for the pair $(A,B)$ and a bounded solution operator for the $\overline\partial$-problem on $\overline \Rcal$.
\[linear-splitting\] Let $A\cup B =\overline \Rcal$ be a Cartan decomposition of a bordered Riemann surface $\overline \Rcal$. Set $C:=A\cap B$. For every integer $r\in\z_+$ there exist bounded linear operators $\Acal \colon \Ascr^r(C) \to \Ascr^r(A)$, $\Bcal \colon \Ascr^r(C) \to \Ascr^r(B)$, satisfying the condition $$c = \Acal c - \Bcal c \qquad \forall c\in \Ascr^r(C).$$
\[linear-gluing\] [(Assumptions as in Lemma \[linear-splitting\].)]{} For every pair of integers $n\in \n$ and $r\in\z_+$ there is a constant $M=M_{n,r}>0$ with the following property. Given a pair of mappings $f\in \Ascr^r(A)^n$, $g\in \Ascr^r(B)^n$, there exists $F\in \Ascr^r(\overline \Rcal)^n$ such that $$||F-f||_{r,A} \le M ||f-g||_{r,C},\qquad ||F-g||_{r,B} \le M ||f-g||_{r,C}.$$
The corollary follows immediately from Lemma \[linear-splitting\] by setting $$c=f|_C-g|_C \in \Ascr^r(C)^n, \quad a=\Acal c\in \Ascr^r(A)^n,\quad b=\Bcal c\in \Ascr^r(B)^n,$$ and noting that as a consequence we have $f-a=g-b$ on $C$. Hence the maps $f-a\colon A\to\c^n$ and $g-b\colon B\to\c^n$ amalgamate into a map $F\colon \overline \Rcal\to\c^n$ with the stated properties. The constant $M$ depends on the norms of the operators $\Acal,\Bcal$ in Lemma \[linear-splitting\].
Gluing holomorphic sprays {#sec:sprays}
-------------------------
The simple gluing method in the previous subsection no longer works in the absence of a linear structure on the target manifold. It can be replaced by the method of gluing pairs of holomorphic sprays of maps that was developed by Drinovec Drnovšek and Forstnerič [@DF; @DF2; @F2007]; this method applies to maps with values in an arbitrary complex manifold. We recall the main results and refer for further details to the cited papers, or to Sect. 5.9 in [@F2011].
\[def:spray\] Let $r\in\z_+$. A [*spray of maps*]{} $\overline \Rcal\to X$ of class $\Ascr^r(\Rcal,X)$ is a $\Ccal^r$ map $F\colon P \times \overline{\Rcal}\to X$ that is holomorphic on $P\times\Rcal$, where $P$ is a domain of a complex Euclidean space $\c^m$ containing the origin $0\in\c^m$. The spray is said to be [*dominating*]{} if the partial differential $\frac{\partial}{\partial t} F(t,z) \colon T_t\c^m\cong \c^m \to T_{F(t,z)}X$ is surjective for all $(t,z)\in P\times \overline \Rcal$. The map $f=F(0,\cdot)\colon \Rcal\to X$ is the [*core map*]{}, and $P$ is the [*parameter set*]{} of the spray $F$.
A spray of maps in the above definition can be viewed as a holomorphic map from the domain $P\subset \c^m$ to the Banach manifold $\Ascr^r(\Rcal,X)$.
[(Existence of sprays, [@DF].)]{} \[lem:Espray\] Let $r\in\z_+$. For every map $f\in \Ascr^r(\Rcal,X)$ there exists a dominating spray of maps $F\colon P\times\overline{\Rcal}\to X$ of class $\Ccal^r$ such that $F(0,\cdot)=f$.
When $X=\c^n$, we can simply take $F(t,z)=f(z)+t$ for $t\in\c^n$ to get a dominating spray of maps $\c^n\times \Rcal \to \c^n$ with the core map $f$.
[(Gluing sprays of maps, [@DF].)]{} \[lem:Gspray\] Let $r\in\z_+$. Let $(A,B)$ be a Cartan decomposition of a bordered Riemann surface $\overline \Rcal$ (Def. \[def:Cartan\]), and set $C=A\cap B$. Given a domain $0\in P_0\subset\c^m$ and a dominating spray of maps $F_A \colon P_0\times A \to X$ of class $\Ccal^r$, there exists a domain $P\Subset P_0\subset\c^m$ containing $0\in\c^m$ such that the following holds.
For every spray of maps $F_B\colon P_0\times B \to X$ of class $\Ccal^r$ that is sufficiently $\Ccal^r$-close to $F_A$ on $P_0\times C$ there exists a spray of maps $F\colon P\times \overline \Rcal \to X$ of class $\Ccal^r$ such that:
- the restriction $F \colon P\times A\to X$ is close to $F_A$ in the $\Ccal^r$-topology (depending on the $\Ccal^r$-distance between $F_A$ and $F_B$ on $P_0\times C$), and
- $F(t,z)\in \{F_1(s,z)\colon s\in P_0\}$ for all $t\in P$ and $z\in B$.
\[rem:r=0\] For applications in this paper it suffices to use these gluing methods in the basic case $r=0$, i.e., with continuity up to the boundary. This is because every map $f\in \Ascr(\Rcal,X)$ can be approximated, uniformly on $\overline \Rcal$, by maps holomorphic in a neighborhood of $\overline \Rcal$ in a larger Riemann surface [@DF].
The Main Lemma {#sec:mainlemma}
==============
In this section we prove an approximation result, Lemma \[lem:main\], which is the core of this paper; Theorem \[th:intro\] will follow by a standard recursive application of it. We focus on maps to $\c^2$, although the same method works for any $\c^n$, $n>1$.
Let $\b(s)=\{u\in\c^2\colon |u|<s\}$ denote the open ball of radius $s>0$ in $\c^2$, centered at the origin, and let $\overline{\b}(s)=\{u\in\c^2\colon |u|\leq s\}$ be the corresponding closed ball. Recall from Sec. \[sec:preli\] that $\Bscr(\Rcal)$ denotes the set of all bordered Runge domains $\Mcal\Subset\Rcal$, and $\Iscr(\Rcal)$ is the set of all immersions $\overline \Rcal\to \c^2$ that are holomorphic in $\Rcal$.
\[lem:main\] Let $\Rcal$ be a bordered Riemann surface, let $\Mcal\in\Bscr(\Rcal)$, let $z_0$ be a point in $\Mcal$, let $f\in \Iscr(\Rcal)$, and let $\epsilon$, $\rho$, and $s>\epsilon$ be positive constants. Assume that
1. $f(\overline{\Rcal} \setminus \Mcal)\subset \b(s) \setminus \overline{\b}(s-\epsilon)$, and
2. $\dist_{(\Rcal,f)}(z_0,b\overline{\Rcal})>\rho$.
Then, for any $\hat{\epsilon}>0$ and $\delta>0$, there exists an immersion $\hat{f}\in \Iscr(\Rcal)$ satisfying the following conditions:
1. $\|\hat{f}-f\|_{1,\overline{\Mcal}}<\hat{\epsilon}$,
2. $\hat{f}(b\overline{\Rcal})\subset \b(\hat{s}) \setminus \overline{\b}(\hat{s}-\hat{\epsilon})$, where $\hat{s} =\sqrt{s^2+\delta^2}$,
3. $\hat{f}(\overline{\Rcal}\setminus \Mcal)\cap\overline{\b}(s-\epsilon)=\emptyset$, and
4. $\dist_{(\Rcal,\hat{f})}(z_0,b\overline{\Rcal})>\hat{\rho} :=\rho+\delta$.
The lemma asserts that any holomorphic immersion $f\colon \overline{\Rcal}\to\c^2$ such that $f(b\overline{\Rcal})$ lies near a given sphere in $\c^2$ (see [(i)]{}) can be deformed into another holomorphic immersion $\hat{f}\colon \overline{\Rcal}\to\c^2$ mapping the boundary $b\overline{\Rcal}$ as close as desired to a larger sphere in $\c^2$ (see [(L2)]{}). The deformation is strong near the boundary of $\Rcal$ and is arbitrarily small on a given compact subset of $\Rcal$. In this process both the extrinsic and the intrinsic diameters of the surface grow, but their respective growths are related in a Pythagoras’ way; compare the bounds $\hat{s}$ in [(L2)]{} and $\hat{\rho}$ in [(L4)]{}. This link, which is the key to obtain completeness while preserving boundedness, follows the spirit of Nadirashvili’s original construction of complete bounded minimal surfaces in $\r^3$ [@Nad]. The novelty of Lemma \[lem:main\] is that do not have to modify the complex struct ure on $\Rcal$ in order to keep the extrinsic diameter of the curve suitably bounded. This represents a clear difference with respect to previous results, and is the main improvement obtained in this work.
The proof of Lemma \[lem:main\] consists of three main steps which we now describe.
In the first step (Subsec. \[sec:step1\]) we split the boundary $b\overline{\Rcal}$ into finitely many compact Jordan arcs $\alpha_{i,j}$, with endpoints $p_{i,j-1}$ and $p_{i,j}$, so that deformations of $f$ near $\alpha_{i,j}$, preserving the complex direction $f(p_{i,j})\in\c^2$, keep the image of $\alpha_{i,j}$ disjoint from the ball $\overline{\b}(s-\epsilon)$; at the same time, the length of any segment in $\c^2$, orthogonal to $f(p_{i,j})$ and connecting $f(\alpha_{i,j})$ to the boundary of $\b(\hat{s})$, is bigger than $\delta$.
In the second step (Subsec. \[sec:step2\]) we deform $f$ into another holomorphic immersion $f_0\colon \overline{\Rcal}\to\c^2$. The deformation is large near the points $p_{i,j}$ and is small elsewhere. The new immersion $f_0$ maps a neighborhood of each point $p_{i,j}$ close to the boundary of the ball $\b(\hat{s})$ (see [(d6)]{}), and the length of the image of any curve in $\Rcal$ connecting $z_0$ to $p_{i,j}$ is larger than $\hat{\rho}$ (see [(d9)]{}). Roughly speaking, $f_0$ satisfies items [(L1)]{}, [(L3)]{}, and [(L4)]{} in Lemma \[lem:main\], but it meets [(L2)]{} only near the points $p_{i,j}$. This step is inspired by the method of [*exposing boundary points*]{} of a complex curve in $\c^2$ that was developed by Forstnerič and Wold [@FW] (see also Sec. 8.9 in [@F2011]).
In the third step (Subsecs. \[sec:step3\] and \[sec:step4\]) we work on the part $\beta_{i,j}^2$ of the arc $\alpha_{i,j}$ where $f_0$ does not meet the condition [(L2)]{}; that is, outside a neighborhood of the endpoints $p_{i,j-1}$ and $p_{i,j}$. We use Lemma \[lem:Hilbert\] to solve a suitable Riemann-Hilbert problem over a closed disc $\overline{D}_{i,j}$ containing the arc $\beta_{i,j}^2$ in its boundary. This gives a holomorphic map $h_{i,j}\colon \overline{D}_{i,j}\to\c^2$ which satisfies [(L2)]{} over $\beta_{i,j}^2$ (see [(f1)]{}), is close to $f_0$ outside a small neighborhood of $\beta_{i,j}^2$ in $\overline{D}_{i,j}$ (see [(f3)]{}), and whose orthogonal projection onto the complex line spanned by $f_0(p_{i,j})$ is close to the projection of $f_0$ (see [(f2)]{}). We then glue $f_0$ and the $h_{i,j}$’s into a new immersion $\hat{f}\in \Iscr(\Rcal)$.
Finally, in Subsec. \[sec:step5\] we verify that the map $\hat{f}$ obtained in this way satisfies Lemma \[lem:main\].
**Proof of Lemma \[lem:main\]. Replacing $\Mcal$ by a larger bordered domain in $\Bscr(\Rcal)$ if necessary, we may assume that $$\label{eq:lemma}
\dist_{(\Rcal,f)}(z_0,b\overline{\Mcal})>\rho;$$ see the strict inequality [(ii)]{} and recall that $b\overline{\Rcal}$ is compact. Moreover, we can realize $\Rcal$ as a bordered domain in an open Riemann surface $\widehat{\Rcal}$ such that $\Rcal \in \Bscr(\widehat{\Rcal})$. Then, by Mergelyan’s theorem, we may assume that $f\in\Iscr(\Rcal)$ extends to a holomorphic immersion $f\in\Iscr(\widehat{\Rcal})$.**
Splitting $b\overline{\Rcal}$ {#sec:step1}
-----------------------------
Note that for any $u\in\c^2\setminus \{0\}$, $u+\langle u\rangle^\bot$ is the affine complex line passing through $u$ and orthogonal to $u$. From [(i)]{}, the definition of $\hat{s}$ in [(L2)]{}, and Pythagoras’ theorem, one has $$\dist\big( u\,,\, b{\b}(\hat{s}) \cap (u+\langle u\rangle^\bot)\big)>\delta \quad \forall u\in f(b\overline{\Rcal}).$$ Then, by continuity and up to decreasing $\hat{\epsilon}>0$ if necessary, any point $u\in f(b\overline{\Rcal})$ admits an open neighborhood $U_u$ in $\c^2$ such that $$\label{eq:Pytha}
\dist\big( v\,,\, (\b(\hat{s}) \setminus \overline{\b}(\hat{s}-\hat{\epsilon}))\cap (w+\langle z\rangle^\bot)\big)>\delta\quad \forall v,w,z\in U_u.$$ Condition [(i)]{} also implies that $$(u+\langle u\rangle^\bot)\cap\overline{\b}(s-\epsilon)=\emptyset\quad \forall u\in f(b\overline{\Rcal}).$$ Hence any point $u\in f(b\overline{\Rcal})$ admits an open neighborhood $V_u\subset \c^2$ such that $$\label{eq:far}
(v+\langle w\rangle^\bot)\cap\overline{\b}(s-\epsilon)=\emptyset\quad \forall v,w\in V_u;$$ we are taking into account the compactness of $\overline{\b}(s-\epsilon)$. Set $$\Uscr_u=U_u\cap V_u,\; u\in f(b\overline{\Rcal}), \qquad \Uscr=\{\Uscr_u\colon u\in f(b\overline{\Rcal})\}.$$
Denote by $\alpha_1,\ldots,\alpha_\igot$ the connected boundary curves of $\overline{\Rcal}$, so $b\overline{\Rcal}=\cup_{i=1}^\igot \alpha_i$. Since $\Rcal\in\Bscr(\widehat{\Rcal})$, the $\alpha_i$’s are pairwise disjoint smooth closed Jordan curves in $\widehat{\Rcal}$. As $\Uscr$ is an open covering of the compact set $f(b\overline{\Rcal})\subset\c^2$, there exist an integer $\jgot \geq 3$ and compact connected subarcs $\{\alpha_{i,j}\colon (i,j) \in \{1,\ldots,\igot\}\times\z_\jgot\}$, where $\z_\jgot=\{0,\ldots,\jgot-1\}$ denotes the additive cyclic group of integers modulus $\jgot$, satisfying the following conditions:
1. $\cup_{j=1}^\jgot\alpha_{i,j}=\alpha_i$,
2. $\alpha_{i,j}$ and $\alpha_{i,j+1}$ have the common endpoint $p_{i,j}$ and are otherwise disjoint,
3. there exist points $a_{i,j}\in \b(\hat{s})\setminus \overline{\b}(\hat{s}-\hat{\epsilon})$ such that $$(a_{i,j}+\langle f(p_{i,k})\rangle^\bot)\cap \overline{\b}(\hat{s}-\hat{\epsilon})=\emptyset \quad \forall k\in\{j,j+1\}, \quad \text{and}$$
4. for every $(i,j) \in \{1,\ldots,\igot\}\times\z_\jgot$ there exists a set $\Uscr_{i,j}\in\Uscr$ containing the curve $f(\alpha_{i,j})$. In particular $f(p_{i,j})\in \Uscr_{i,j}\cap\Uscr_{i,j+1}$.
A splitting with these properties is found by choosing the arcs $\alpha_{i,j}$ such that their images $f(\alpha_{i,j}) \subset \c^2$ have small enough diameter.
Stretching from the points $p_{i,j}$ {#sec:step2}
------------------------------------
We adapt to our needs the method from [@FW] that was used in that paper for exposing boundary points. (See also Sections 8.8 and 8.9 in [@F2011].) Our goal here is a different one: we wish to modify the immersion so that the images of a certain finite collection of arcs in $\Rcal$, terminating at points of $b\overline \Rcal$, become very long in $\c^2$.
Recall that $\overline \Rcal$ is a compact bordered domain in $\widehat \Rcal$. For every $(i,j)\in\{1,\ldots,\igot\}\times\z_\jgot$ we choose an embedded real analytic arc $\gamma_{i,j}\subset\widehat{\Rcal}$ that is attached to $\overline{\Rcal}$ at the endpoint $p_{i,j}$ and intersects $b\overline \Rcal$ transversely there, and is otherwise disjoint from $\overline{\Rcal}$. We insure that the arcs $\gamma_{i,j}$ are pairwise disjoint. Let $q_{i,j}$ denote the other endpoint of $\gamma_{i,j}$. We split $\gamma_{i,j}$ into compact subarcs $\gamma_{i,j}^1$ and $\gamma_{i,j}^2$, with a common endpoint, so that $p_{i,j}\in\gamma_{i,j}^1$ and $q_{i,j}\in\gamma_{i,j}^2$.
Next we choose compact smooth embedded arcs $\lambda_{i,j} \subset \c^2$ satisfying the following properties:
1. $\lambda_{i,j}\subset \b(\hat{s})\setminus \overline{\b}(s-\epsilon)$,
2. $\lambda_{i,j}$ is split into compact subarcs $\lambda_{i,j}^1$ and $\lambda_{i,j}^2$ with a common endpoint,
3. $\lambda_{i,j}^1$ agrees with the arc $f(\gamma_{i,j})$ near the endpoint $f(p_{i,j})$; recall that $f\in\Iscr(\widehat{\Rcal})$,
4. $\lambda_{i,j}^1\subset \Uscr_{i,j}\cap\Uscr_{i,j+1}$; see properties [(b3)]{} and [(a4)]{},
5. if $J\subset\lambda_{i,j}^1$ is a Borel measurable subset, then $$\begin{aligned}
\min\{\length(\pi_{i,j}(J)),\length(\pi_{i,j+1}(J))\}&+&\\
\min\{\length(\pi_{i,j}(\lambda_{i,j}^1\setminus J)),\length(\pi_{i,j+1}(\lambda_{i,j}^1\setminus J))\}&>&\delta,\end{aligned}$$ where $\pi_{i,k}\colon \c^2\to\span\{f(p_{i,k})\}\subset\c^2$ denotes the orthogonal projection (observe that [(i)]{} implies $f(p_{i,k})\neq 0$),
6. $(\lambda_{i,j}^2+\langle f(p_{i,k})\rangle^\bot)\cap \overline{\b}(s-\epsilon)=\emptyset$ for $k\in \{j,j+1\}$; see [(a4)]{}, [(b4)]{}, and (\[eq:far\]), and
7. the endpoint $v_{i,j}$ of $\lambda_{i,j}$ contained in the subarc $\lambda_{i,j}^2$ lies in the shell $\b(\hat{s})\setminus \overline{\b}(\hat{s}-\hat{\epsilon})$ and satisfies $(v_{i,j}+\langle f(p_{i,k})\rangle^\bot)\cap \overline{\b}(\hat{s}-\hat{\epsilon})=\emptyset$ for $k\in\{j,j+1\}$; see [(a3)]{}.
The existence of such arcs $\lambda_{i,j}$ is an easy exercise. Property [(b1)]{} is compatible with the rest thanks to conditions [(a3)]{} and [(a4)]{}. To enjoy [(b4)]{} and [(b5)]{}, the curve $\lambda_{i,j}^1$ must be highly oscillating and with small diameter in $\c^2$. On the other hand, to satisfy [(b6)]{} and [(b7)]{}, one can simply take $\lambda_{i,j}^2$ to be a straight line segment in $\b(\hat{s})\setminus \overline{\b}(s-\epsilon)$, connecting $\Uscr_{i,j}\cap\Uscr_{i,j+1}$ (see [(b4)]{}) to a point $v_{i,j}$ as those given by [(a3)]{}.
By property [(b3)]{} we can find a smooth map $f_{\rm e}\colon \widehat \Rcal \to\c^2$ which agrees with $f$ in an open neighborhood of $\overline \Rcal$, and which maps the arcs $\gamma_{i,j}^1$ and $\gamma_{i,j}^2$ diffeomorphically onto the corresponding arcs $\lambda_{i,j}^1$ and $\lambda_{i,j}^2$ for all $(i,j)$.
The compact set $K:=\overline{\Rcal}\cup(\cup_{i,j}\gamma_{i,j})\subset \widehat{\Rcal}$ clearly admits a basis of open neighborhoods that are Runge in $\widehat \Rcal$. By Mergelyan’s theorem one can therefore approximate $f_{\rm e}$, uniformly on an open neighborhood of $\overline{\Rcal}$ and in the $\Ccal^1$-topology on each of the arcs $\gamma_{i,j}$, by a holomorphic immersion $\tilde{f}\colon \widehat \Rcal \to\c^2$.
We shall now apply [@FW Theorem 2.3] (see also [@F2011 Theorem 8.8.1]). Choose a small open neighborhood $V\subset \widehat \Rcal$ of $K$. Further, for every $(i,j)$ we choose a pair of small neighborhoods $W'_{i,j} \Subset W_{i,j} \Subset \widehat \Rcal\setminus \overline \Mcal$ of the point $p_{i,j}$ and a neighborhood $V_{i,j} \Subset \widehat \Rcal \setminus \overline \Mcal$ of the arc $\gamma_{i,j}$. The cited theorem furnishes a smooth diffeomorphism $\phi\colon \overline {\Rcal}\to\phi (\overline{\Rcal})\subset V$ satisfying the following properties (see Fig. \[fig:Phi\]):
- $\phi \colon \Rcal \to \phi(\Rcal)$ is biholomorphic,
- $\phi$ is as close as desired to the identity in the $\Ccal^1$-topology on $\overline \Rcal \setminus \cup_{i,j} W'_{i,j}$, and
- $\phi(p_{i,j}) = q_{i,j}$ and $\phi(\overline \Rcal \cap W'_{i,j}) \subset W_{i,j} \cup V_{i,j}$ for all $(i,j)\in\{1,\ldots,\igot\}\times\z_\jgot$.
Intuitively speaking, $\phi$ hardly changes $\overline \Rcal$ outside small neighborhoods of the chosen boundary points $p_{i,j}$, while at each of these points it creates a narrow spike reaching up to the opposite endpoint $q_{i,j}$ of the arc $\gamma_{i,j}$. (See Fig. 8.1 in [@F2011 p. 367].) Although it is not claimed in the cited sources that the arc $\gamma_{i,j}$ actually belongs to the image $\phi(\overline \Rcal)$, an inspection of the construction shows that we can move it very slightly to a nearby arc $\gamma'_{i,j}\subset \widehat\Rcal$ (keeping its endpoints fixed) to obtain this property. If the new arc $\gamma'_{i,j}$ is close enough to $\gamma_{i,j}$ (which can be achieved by a suitable choice of $\phi$), we can also replace the arc $\lambda_{i,j} \subset \c^2$ by the image $\tilde f (\gamma'_{i,j})$ without disturbing any of the properties [(b1)]{}–[(b7)]{}. In the sequel we drop the primes and assume that, in addition to the above, we have
- $\gamma_{i,j}\setminus\{q_{i,j}\}\subset\phi(\Rcal \setminus \overline{\Mcal})$ for all $(i,j)$.
Set $$\label{eq:sigma}
\sigma_{i,j}^k=\phi^{-1}(\gamma_{i,j}^k)\; \text{for $k\in\{1,2\}$}, \qquad \sigma_{i,j}=\phi^{-1}(\gamma_{i,j})=\sigma_{i,j}^1\cup \sigma_{i,j}^2.$$ (See Fig. \[fig:Phi\].) If the approximations desribed above are close enough, then the composition $$\label{eq:f0}
f_0=\tilde{f}\circ\phi\colon \overline{\Rcal}\to\c^2$$ is a holomorphic immersion satisfying the following properties:
1. $f_0(p_{i,j})\in \b(\hat{s})\setminus \overline{\b}(\hat{s}-\hat{\epsilon})$ and $(f_0(p_{i,j})+\langle f(p_{i,k})\rangle^\bot)\cap \overline{\b}(\hat{s}-\hat{\epsilon})=\emptyset$ for $k\in\{j,j+1\}$ and $(i,j)\in\{1,\ldots,\igot\}\times\z_\jgot$; recall that $f_0(p_{i,j})=\tilde{f}(q_{i,j})\approx f_{\rm e}(q_{i,j})=v_{i,j}$ and see [(b7)]{},
2. $\|f_0-f\|_{1,\overline{\Mcal}}<\hat{\epsilon}/2$,
3. $f_0(\overline{\Rcal}\setminus \Mcal)\subset \b(\hat{s})\setminus \overline{\b}(s-\epsilon)$; see the open conditions [(i)]{} and [(b1)]{} and recall that $\lambda_{i,j}$ is compact, and
4. $\dist_{(\Rcal,f_0)}(z_0,b\overline{\Mcal})>\rho$; see the strict inequality (\[eq:lemma\]).
Furthermore, if the above approximations are close enough and the neighborhood $V$ of $K$ (containing the image $\phi(\overline \Rcal)$) is small enough, then taking into account properties [(b1)]{}–[(b7)]{} and [(c1)]{}–[(c4)]{}, one can easily find simply connected neighborhoods $U_{i,j}^3\Subset U_{i,j}^2\Subset U_{i,j}^1$ of the point $p_{i,j}$ in $\overline{\Rcal}\setminus \overline{\Mcal}$ for any $(i,j)\in\{1,\ldots,\igot\}\times\z_\jgot$ satisfying the following properties (see Figs. \[fig:Phi\] and \[fig:R\]):
1. $\overline{U}_{i,j}^1\cap \overline{U}_{i,k}^1=\emptyset$ if $k\neq j$,
2. $\overline{U}_{i,j}^1\cap \alpha_{i,k}=\emptyset$ if $k\notin\{j,j+1\}$,
3. $\overline{U}_{i,j}^1\cap \alpha_{i,k}$ is a connected compact Jordan arc for $k\in\{j,j+1\}$,
4. $\beta_{i,j}^k :=\overline{\alpha_{i,j}\setminus (U_{i,j}^{k+1}\cup U_{i,j-1}^{k+1})}$ are connected compact Jordan arcs for $k\in\{1,2\}$, and $f_0(\beta_{i,j}^1)\subset\Uscr_{i,j}$; see [(a4)]{},
5. $\sigma_{i,j}^1\subset \overline{U_{i,j}^1\setminus U_{i,j}^2}$, $\sigma_{i,j}^2\subset\overline{U}_{i,j}^2$ (see (\[eq:sigma\])), and $\phi^{-1}(p_{i,j})\in (b\overline{U}_{i,j}^1)\cap\Rcal$,
6. $f_0(\overline{U}_{i,j}^3)\subset\b(\hat{s})\setminus \overline{\b}(\hat{s}-\hat{\epsilon})$ and $(f_0(\overline{U}_{i,j}^3)+\langle f(p_{i,k})\rangle^\bot)\cap \overline{\b}(\hat{s}-\hat{\epsilon})=\emptyset$ for $k\in\{j,j+1\}$; see [(c1)]{},
7. $\left( f_0(\overline{U}_{i,j}^2) + \langle f(p_{i,k})\rangle^\bot \right) \cap \overline{\b}(s-\epsilon)=\emptyset$ for all $k\in \{j,j+1\}$; see [(b6)]{},
8. $f_0(\overline{U_{i,j}^1\setminus U_{i,j}^2})\subset \Uscr_{i,j}\cap\Uscr_{i,j+1}$; see [(b4)]{}, and
9. if $\gamma\subset \overline{U}_{i,j}^1$ is an arc connecting $\Rcal\setminus U_{i,j}^1$ and $\overline{U}_{i,j}^2$, and $J\subset \gamma$ is a Borel measurable subset, then (see [(b5)]{}) we have $$\begin{aligned}
\min\{\length(\pi_{i,j}(f_0(J))),\length(\pi_{i,j+1}(f_0(J)))\}&+&\\
\min\{\length(\pi_{i,j}(f_0(\gamma \setminus J))),\length(\pi_{i,j+1}(f_0(\gamma\setminus J)))\}&>&\delta.\end{aligned}$$
In fact, the above properties hold if $U_{i,j}^3$ is chosen sufficiently small around the point $p_{i,j}$, $U_{i,j}^2$ is sufficiently small around the arc $\sigma_{i,j}^2$, and $U_{i,j}^1$ is sufficiently small around the arc $\sigma_{i,j}$ (see (\[eq:sigma\])).
Stretching from the arcs $\alpha_{i,j}$ {#sec:step3}
---------------------------------------
We shall now stretch the images of the central parts of the arcs $\alpha_{i,j}$ (away from the two endpoints) close to the sphere of radius $\hat s$ in $\c^2$ in order to fulfill the condition [(L2)]{}.
For each $(i,j)\in\{1,\ldots,\igot\}\times\z_\jgot$ we choose a smoothly bounded closed disc $\overline{D}_{i,j} \subset\overline{\Rcal}\setminus \overline{\Mcal}$ such that the following hold (see Fig. \[fig:R\]):
1. $\overline{D}_{i,j}\cap\overline{D}_{i,k}=\emptyset$ if $j\neq k$,
2. $\overline{D}_{i,j}\cap \alpha_{i,j}$ is a compact connected Jordan arc that contains $\beta_{i,j}^2$ in its relative interior, and $\overline{D}_{i,j}\cap \alpha_{i,k}=\emptyset$ for all $k\neq j$,
3. $\overline{D}_{i,j}\cap \sigma_{i,k}=\emptyset$ for all $k\in \z_\jgot$, and
4. $f_0\big(\overline{D_{i,j}\setminus (U_{i,j-1}^2\cup U_{i,j}^2)}\big)\subset \Uscr_{i,j}$. (Here $f_0$ is the map (\[eq:f0\]) constructed in the previous subsec.)
The existence of such discs $D_{i,j}$ is trivially implied by properties [(d4)]{} and [(d8)]{}.
At this point we use an approximate solution of a Riemann-Hilbert problem. By property [(e2)]{} and the second part of [(d6)]{} we can easily find a continuous map $g_{i,j}\colon b\overline{D}_{i,j}\times\overline{\d}\to\c^2$ satisfying the following requirements:
1. $g_{i,j}(p,\cdot)\in \Ascr(\d)^2$ and $g_{i,j}(p,0)= f_0(p)$ for all $p\in b\overline{D}_{i,j}$,
2. $g_{i,j}(p,\overline{\d})\subset f_0(p)+\langle f(p_{i,j})\rangle^\bot$ for all $p\in b\overline{D}_{i,j}$,
3. $g_{i,j}(p,b\overline{\d})\subset \b(\hat{s})\setminus \overline{\b}(\hat{s}-\hat{\epsilon})$ for all $p\in b\overline{D}_{i,j}\cap \alpha_{i,j}$, and
4. $g_{i,j}(p,\cdotp) \equiv f_0(p)$ is the constant disc for all $p\in \overline{b {D}_{i,j}\setminus \beta_{i,j}^2}$.
By using a conformal diffeomorphism $\overline{D}_{i,j} \stackrel{\cong}{\longrightarrow} \d$ onto the unit disc, Lemma \[lem:Hilbert\] gives a map $h_{i,j}\in\Ascr(D_{i,j})^2$ satisfying the following conditions:
1. $h_{i,j}(p)\in \b(\hat{s})\setminus \overline{\b}(\hat{s}-\hat{\epsilon})$ for all $p\in \overline{D}_{i,j}\cap \alpha_{i,j}= b\overline{D}_{i,j}\cap \alpha_{i,j}$; see $iii)$,
2. $\pi_{i,j}\circ h_{i,j}$ is close to $\pi_{i,j}\circ f_0$ on $\overline{D}_{i,j}$; see $ii)$, and
3. $h_{i,j}$ is close to $f_0$ outside a small open neighborhood $C_{i,j}$ of $\beta_{i,j}^2$ in $\overline{D}_{i,j}$ with $b\overline{C}_{i,j} \cap b \overline{D}_{i,j}\subset\alpha_{i,j}$; see $iv)$.
The sets $C_{i,j}\subset D_{i,j}$ and the arcs $\beta^2_{i,j}\subset \alpha_{i,j}$ are illustrated on Fig. \[fig:R\].
Gluing $f_0$ and $h_{i,j}$ {#sec:step4}
--------------------------
Consider the domains $D_0,D_1$ in $\Rcal$ defined by $$D_0=\Rcal\setminus \cup_{(i,j)\in\{1,\ldots,\igot\}\times\z_\jgot} \overline{C}_{i,j}
\quad\text{and}\quad
D_1=\cup_{(i,j)\in\{1,\ldots,\igot\}\times\z_\jgot} D_{i,j}.$$ We may assume that the pair $A=\overline D_0$ and $B=\overline D_1$ is a smooth Cartan decomposition of $\overline \Rcal$ (see Def. \[def:Cartan\]). (Smoothness of $bA$ is easily insured by a suitable choice of the neighborhoods $C_{i,j}$ of $\beta_{i,j}^2$, and the separation property follows from the definition of the sets $C_{i,j}$ and $D_{i,j}$.)
If the approximation of $f_0$ by $h_{i,j}$ on $\overline D_0 \cap \overline{D}_{i,j}$ is close enough (which is guaranteed by [(f3)]{}), one can apply Corollary \[linear-gluing\] to obtain a holomorphic map $\hat{f}\colon \overline{\Rcal}\to\c^2$ which is as close as desired to $f_0$ on $\overline D_0$, and to $h_{i,j}$ on $\overline{D}_{i,j}$, for all $(i,j)$. (The cited corollary furnishes a map in $\Ascr(\Rcal)^2$, but we can approximate it by a holomorphic map.)
\[gluing-by-sprays\] In the general case, with $\c^2$ replaced by an arbitrary complex manifold $X$ (see Sec. \[sec:Stein\]), we can instead apply Lemma \[lem:Gspray\] to obtain a holomorphic map $\hat{f} \colon \overline{\Rcal}\to X$ which is close to the map $f_0$ on $A$, and is close to $h_{i,j}$ on $\overline{C}_{i,j}$ for all $(i,j)$. This can be done provided that we have a suitable procedure for approximating maps $A\to X$ of class $\Ascr(A,X)$, uniformly on $A \cap B$ and uniformly with respect to a parameter $t\in P\subset\c^m$, by maps $B\to X$ of class $\Ascr(B,X)$. We proceed as follows. By Lemma \[lem:Espray\] we can embed $f_0\colon A \to X$ as the core map $f_0=F_A(0,\cdot)$ in a dominating spray of maps $F_A\colon P \times A \to X$, where $P$ is an open set in $\c^m$ containing the origin. By the approximation procedure we find a spray of maps $F_B \colon P \times B \to X$ such that (up to a shrinking of $P$) the map $F_B(t,\cdot)$ is as close as desired to $F_A(t,\cdotp)$ on $A \cap B$ for every $t\in P$. Lemma \[lem:Gspray\] furnishes a new spray of maps $P\times\overline{\Rcal}\to X$ whose core map $\hat{f} \colon \overline{\Rcal}\to X$ is uniformly close to $f_0$ on $A$, and is close to $h_{i,j}$ on $\overline{C}_{i,j}$ for all $(i,j)$.
Checking the properties of $\hat{f}$ {#sec:step5}
------------------------------------
By approximation and general position argument we may assume that $\hat f$ is a holomorphic immersion of a neighborhood of $\overline \Rcal$ in $\widehat \Rcal$ to $\c^2$. Furthermore, if all the approximations in our construction (namely,
1. of $f_0$ by $h_{i,j}$ on $\overline{D}_{i,j}\setminus C_{i,j}$,
2. of $f_0$ by $\hat{f}$ on $\overline D_0$,
3. of $h_{i,j}$ by $\hat{f}$ on $\overline{C}_{i,j}$, and
4. of $\pi_{i,j}\circ f_0$ by $\pi_{i,j}\circ h_{i,j}$ over $\overline{D}_{i,j}$, for all $(i,j)$;
see the above subsections and [(f2)]{}) are sufficiently close, then $\hat{f}$ satisfies the following properties which we verify item by item:
- $\|\hat{f}-f\|_{1,\overline{\Mcal}}<\hat{\epsilon}$.
Indeed, just notice that $\|\hat{f}-f_0\|_{1,\overline{\Mcal}}\approx 0$ (see [(A2)]{} and observe that $\overline{\Mcal}\subset \overline D_0$) and take into account [(c2)]{}.
- $\hat{f}(b\overline{\Rcal})\subset \b(\hat{s})\setminus \overline{\b}(\hat{s}-\hat{\epsilon})$.
Indeed, let $p\in b\overline{\Rcal}$. If $p\in C_{i,j}$ for some $(i,j)$, then [(A3)]{} and [(f1)]{} give that $\hat{f}(p) \approx h_{i,j}(p)\in \b(\hat{s})\setminus \overline{\b}(\hat{s}-\hat{\epsilon})$. If on the other hand $p\in \alpha_{i,j}\setminus C_{i,j}\subset \overline{U}_{i,j}^3\cup\overline{U}_{i,j+1}^3$ (see Fig. \[fig:R\]), then properties [(A2)]{} and [(d6)]{} insure that $\hat{f}(p) \approx f_0(p)\in \b(\hat{s})\setminus \overline{\b}(\hat{s}-\hat{\epsilon})$ as well.
- $\hat{f}(\overline{\Rcal}\setminus \Mcal)\cap\overline{\b}(s-\epsilon)=\emptyset$.
Indeed, let $p\in \overline{\Rcal}\setminus \Mcal$. If $p\notin \overline{D}_{i,j}$, then [(A2)]{} and [(c3)]{} imply that $\hat{f}(p) \approx f_0(p)\notin \overline{\b}(s-\epsilon)$. Assume now that $p\in \overline{D}_{i,j}$. Then either $p\in C_{i,j}$ and [(A3)]{} gives $\hat{f}(p) \approx h_{i,j}(p)$, or $p\in \overline{D}_{i,j}\setminus C_{i,j}$ and [(A2)]{} insures that $\hat{f}(p) \approx f_0(p)$. In either case, property [(A4)]{} guarantees that $\pi_{i,j}(\hat{f}(p))\approx \pi_{i,j}(f_0(p))$ for $p\in \overline{D}_{i,j}$. Assuming that the approximation is close enough (and using compactness of $\overline{\Rcal}\setminus \Mcal$), it thus suffices to verify that $\pi_{i,j}(f_0(p))$ does not belong to the disc $\overline{\d}(s-\epsilon)$ (the $\pi_{i,j}$-projection of the ball $\overline{\b}(s-\epsilon)$). If $p\in \overline{D}_{i,j} \cap\overline{U}_{i,k}^2$, $k\in\{j-1,j\}$, then $\pi_{i,j}(f_0(p)) \in \pi_{i,j}(f_0(\overline{U}_{i,k}^2))$ which is disjoint from $\overline{\d}(s-\epsilon)$ by [(d7)]{}. Finally, if $p\in \overline{D}_{i,j}\setminus (\overline{U}_{i,j-1}^2\cup\overline{U}_{i,j}^2)$, then [(e4)]{} gives that $f_0(p)\in \Uscr_{i,j}$ and so $\pi_{i,j}(f_0(p))\notin \overline{\d}(s-\epsilon)$ as well; see (\[eq:far\]) and [(a3)]{}.
- $\dist_{(\Rcal,\hat{f})}(z_0,b\overline{\Rcal})>\hat{\rho}=\rho+\delta$.
By [(A2)]{} and [(c4)]{} it suffices to check that $\dist_{(\Rcal,\hat{f})}(b\overline{\Mcal},b\overline{\Rcal})>\delta$. Let $\gamma$ be any curve in $\overline \Rcal\setminus \Mcal$ connecting $b\overline{\Mcal}$ and $b\overline{\Rcal}$, and let us show that $\length(\hat{f}(\gamma))>\delta$.
Assume first that $\gamma\cap \overline{U}_{i,j}^2\neq\emptyset$ for some $(i,j)$. Then there exists a subarc $\hat{\gamma}\subset \overline{U}_{i,j}^1$ connecting $\Rcal\setminus U_{i,j}^1$ and $\overline{U}_{i,j}^2$; see Fig. \[fig:R\]. By [(A2)]{} and [(A3)]{} one has $$\begin{aligned}
\length(\hat{f}(\hat{\gamma}))& \approx & \length (f_0(\hat{\gamma}\cap \overline{D}_0))
\\
& & +\ \length (h_{i,j}(\hat{\gamma}\cap \overline{C}_{i,j})) + \length (h_{i,j+1}(\hat{\gamma}\cap \overline{C}_{i,j+1}))\\
& \geq & \length (f_0(\hat{\gamma}\cap \overline{D}_0)) +\ \length (\pi_{i,j}(h_{i,j}(\hat{\gamma}\cap \overline{C}_{i,j}))) \\
& & +\ \length (\pi_{i,j+1}(h_{i,j+1}(\hat{\gamma}\cap \overline{C}_{i,j+1}))),\end{aligned}$$ and by [(A4)]{} and [(d9)]{}, the above is $$\begin{aligned}
& \approx & \length (f_0(\hat{\gamma}\cap \overline{D}_0))
\\
& & +\ \length (\pi_{i,j}(f_0(\hat{\gamma}\cap \overline{C}_{i,j}))) + \length (\pi_{i,j+1}(f_0(\hat{\gamma}\cap \overline{C}_{i,j+1}))) > \delta.\end{aligned}$$ Therefore, $\length(\hat{f}(\gamma)) \ge \length(\hat{f}(\hat{\gamma}))>\delta$ as claimed.
Assume now that $\gamma\cap\overline{U}_{i,j}^2=\emptyset$ for all $(i,j)\in\{1,\ldots,\igot\}\times\z_\jgot$. Then there exist $(i,j)$ and a subarc $\hat{\gamma}\subset \overline{D_{i,j}\setminus (U_{i,j-1}^2\cup U_{i,j}^2)}$ connecting a point $p_0\in \overline{D_{i,j}\setminus (U_{i,j-1}^2\cup U_{i,j}^2\cup C_{i,j})}$ to a point $p_1\in \beta_{i,j}^1$; see Fig. \[fig:R\]. In this case, [(A2)]{} and [(A3)]{} imply that $\hat{f}(p_0)\approx f_0(p_0)$ and $\hat{f}(p_1)\approx h_{i,j}(p_1)$. By [(e4)]{} one has $\{f_0(p_0),f_0(p_1)\}\subset\Uscr_{i,j}$, by [(f1)]{} one has $h_{i,j}(p_1)\in \b(\hat{s})\setminus \overline{\b}(\hat{s}-\hat{\epsilon})$, and by [(A4)]{} one has $\pi_{i,j}(h_{i,j}(p_1))\approx \pi_{i,j}(f_0(p_1))$; hence inequality (\[eq:Pytha\]) gives that $\dist(f_0(p_0),h_{i,j}(p_1))>\delta$. Assuming as we may that $\hat{f}(p_0)$ and $\hat{f}(p_1)$ are close enough to $f_0(p_0)$ and $h_{i,j}(p_1)$, respectively, we thus get $\length(\hat{f}(\gamma)) \ge \length(\hat{f}(\hat{\gamma}))>\delta$.
Thus $\hat{f}$ satisfies the conclusion of Lemma \[lem:main\] and the proof is complete.
Complete proper immersions to Euclidean balls {#sec:proof}
=============================================
We now show how Lemma \[lem:main\] implies Theorem \[th:intro2\].
Let $\overline{\Rcal}$ be a bordered Riemann surface and $h \colon \overline{\Rcal}\to\c^m$ $(m>1)$ be a holomorphic map whose image $h(\overline{\Rcal})$ is contained in an open ball $B \subset \c^m$. Given a compact set $K\subset \Rcal$ and a number $\eta>0$, we must find a complete proper holomorphic immersion $\hat h\colon \Rcal \to B$ (embedding if $m>2$) such that $\|\hat{h}-h\|_{1,K}<\eta$.
We focus on immersions; the necessary modifications to find embeddings in $\c^m$ for $m>2$ are indicated at the end.
By a translation and a dilation of coordinates we may assume that $B=\b$ is the unit ball. By general position we can replace $h$ by an immersion (embedding if $m>2$) such that $h(b\overline{\Rcal})$ does not contain the origin of $\c^m$. Pick numbers $0< \xi < r<1$ so that $h(b\overline{\Rcal}) \subset \b(r)\setminus \overline{\b}(r-\xi)$. Choose a bordered domain $\Ncal\in\Bscr(\Rcal)$ in $\Rcal$ such that $$K\subset \Ncal \quad \hbox{and} \quad
h(\overline{\Rcal} \setminus \Ncal)\subset \b(r)\setminus \overline{\b}(r-\xi).$$ Set $c :=\sqrt{6(1-r^2)}/\pi>0$. Fix a point $\zeta_0\in \Ncal$ and define sequences $\rho_n,r_n>0$ $(n\in \z_+)$ recursively as follows: $$\label{eq:r-rho}
\rho_0=\dist_{(\Rcal,h)}(\zeta_0,b\overline{\Ncal}),\enskip
\rho_n=\rho_{n-1}+\frac{c}{n},\enskip r_0=r, \enskip r_n=\sqrt{r_{n-1}^2+\frac{c^2}{n^2}}.$$ It is immediate that $$\label{eq:diverges}
\lim_{n\to\infty}\rho_n=+\infty$$ and $$\label{eq:converges}
\lim_{n\to\infty}r_n=1.$$ To get (\[eq:converges\]), observe that $r_n^2=r_{n-1}^2+\frac{c^2}{n^2}$ and hence $$\lim_{n\to\infty}r_n^2 = r^2 + \sum_{n=1}^\infty \frac{c^2}{n^2} =
r^2+ c^2 \frac{\,\,\pi^2}{6} = r^2 + \frac{6(1-r^2)}{\pi^2} \frac{\,\,\pi^2}{6} = 1.$$
Set $(\Ncal_0,h_0,\xi_0)=(\Ncal,h,\xi)$. We shall inductively construct sequences of bordered domains $\Ncal_n\in\Bscr(\Rcal)$, holomorphic immersions $h_n\in\Iscr(\Rcal)$, and constants $\xi_n>0$ $(n\in\n)$ satisfying the following conditions:
1. $\Ncal_{n-1}\Subset \Ncal_n$,
2. $\displaystyle \xi_n <\min\left\{ \xi_{n-1}\,,\, \frac{\eta}{2^{n}} \;,\; \frac{1}{2^{n}}\min\left\{ \min_{\Ncal_k} |dh_k| \colon k=1,\ldots,n-1 \right\}\right\}$; this minimum is positive since $dh_k\ne 0$ on $\overline \Rcal$ for all $k\in\{1,\ldots,n-1\}$,
3. $\|h_n-h_{n-1}\|_{1,\overline{\Ncal}_{n-1}}<\xi_n$,
4. $h_n(\overline{\Rcal}\setminus \Ncal_n)\subset \b(r_n)\setminus \overline{\b}(r_n-\xi_n)$,
5. $h_n(\overline{\Rcal}\setminus \Ncal_{n-1})\cap \overline{\b}(r_{n-1}-\xi_{n-1})=\emptyset$, and
6. $\dist_{(\Rcal,h_n)}(\zeta_0,b\overline{\Ncal}_n)>\rho_n$.
We will additionally insure that $$\label{eq:cupN}
\Rcal=\cup_{n\in\n}\Ncal_n.$$
We proceed by induction. To begin, pick a number $\xi_1>0$ so that [(b$_1$)]{} holds. Now Lemma \[lem:main\] can be applied to the data $$(\Mcal \,,\, z_0 \,,\, f \,,\, \epsilon \,,\, \rho \,,\, s \,,\, \hat{\epsilon} \,,\, \delta)= (\Ncal_0 \,,\, \zeta_0 \,,\, h_0 \,,\, \xi_0 \,,\, \rho_0 \,,\,r_0 \,,\, \xi_1 \,,\,c),$$ and define $h_1\in\Iscr(\Rcal)$ as the immersion $\hat{f}$ furnished by Lemma \[lem:main\]. Properties [(c$_1$)]{}, [(e$_1$)]{} correspond to [(L1)]{}, [(L3)]{} in Lemma \[lem:main\], respectively; take into account (\[eq:r-rho\]). Moreover, [(L2)]{} and [(L4)]{} give that $h_1(b\overline{\Rcal})\subset \b(r_1)\setminus \overline{\b}(r_1-\xi_1)$ and $\dist_{(\Rcal,h_1)}(\zeta_0,b\overline{\Rcal})>\rho_1$ (see (\[eq:r-rho\]) again). Since $b\overline{\Rcal}$ is compact, it follows that [(d$_1$)]{} holds for any sufficiently large domain $\Ncal_1\in\Bscr(\Rcal)$ satisfying [(a$_1$)]{}.
For the inductive step, assume that for some $n\ge 2$ we already have $(\Ncal_{j},h_{j},\xi_{j})$ satisfying properties [(a$_j$)]{}–[(f$_j$)]{} for all $j\in\{1,\ldots,n-1\}$. Choose any positive number $\xi_n>0$ satisfying [(b$_n$)]{}. Property [(d$_{n-1}$)]{} allows to apply Lemma \[lem:main\] to the data $$(\Mcal \,,\, z_0 \,,\, f \,,\, \epsilon \,,\, \rho \,,\, s \,,\, \hat{\epsilon} \,,\, \delta)= (\Ncal_{n-1} \,,\, \zeta_0 \,,\, h_{n-1} \,,\, \xi_{n-1} \,,\, \rho_{n-1} \,,\,r_{n-1} \,,\, \xi_n \,,\,\frac{c}{n}).$$ Let $h_n$ denote the immersion $\hat f$ furnished by Lemma \[lem:main\]. By choosing a sufficiently large domain $\Ncal_n \in \Bscr(\Rcal)$, the triple $(\Ncal_n,h_n,\xi_n)$ meets all requirements [(a$_n$)]{}–[(f$_n$)]{}.
Finally, in order to guarantee (\[eq:cupN\]), one simply chooses the bordered domain $\Ncal_n$ large enough in each step of the inductive process. This concludes the construction of the sequence $\{(\Ncal_n,h_n,\xi_n)\}_{n\in\n}$.
From [(c$_n$)]{}, [(b$_n$)]{}, and (\[eq:cupN\]) we infer that the sequence $\{h_n\colon \overline {\Rcal}\to\c^2\}_{n\in\n}$ converges uniformly on compacta in $\Rcal$ to a holomorphic map $\hat{h} \colon \Rcal\to\c^2$. Let us show that $\hat{h}$ satisfies the desired properties.
- $\hat{h}\colon \Rcal\to\c^2$ is an immersion.
Indeed, let $p\in \Rcal$. Pick $n_0\in\n$ so that $p\in\Ncal_{n_0}$. From [(c$_n$)]{} and [(b$_n$)]{} one has $$\begin{aligned}
|d\hat{h}(p)| & \geq & |dh_{n_0}(p)|-\sum_{n>n_0} |dh_{n}(p)-dh_{n-1}(p)|
\\
& \geq & |dh_{n_0}(p)|-\sum_{n>n_0} \|h_{n}-h_{n-1}\|_{1,\Ncal_{n_0}}
\\
& \geq & |dh_{n_0}(p)|-\sum_{n>n_0} \xi_n
\\
& \geq & |dh_{n_0}(p)|-\sum_{n>n_0} \frac{1}{2^n} |dh_{n_0}(p)|> \frac12 \, |dh_{n_0}(p)|>0;\end{aligned}$$ recall that $h_{n_0}\in\Iscr(\Rcal)$. This shows that $\hat{h}$ is an immersion as claimed.
- $\hat{h}\colon \Rcal\to\c^2$ is complete.
Indeed, arguing as above, one infers that $$\dist_{(\Rcal,\hat{h})}(\zeta_0,b\overline{\Ncal}_n)>\frac12\,\dist_{(\Rcal,h_n)}(\zeta_0,b\overline{\Ncal}_n)> \frac{\rho_n}2\quad\forall n\in\n;$$ see [(f$_n$)]{}. Taking limits in the above inequality as $n$ goes to infinity, one gets the completeness of $\hat{h}$ from (\[eq:diverges\]).
- $\|\hat{h}-h\|_{1,\overline{\Ncal}}<\eta$.
This follows trivially from [(b$_n$)]{} and [(c$_n$)]{}.
- $\hat{h}(\Rcal)\subset\b$ and $\hat{h}\colon \Rcal\to\b$ is proper.
Indeed, let $p\in\Rcal$. From [(d$_n$)]{} and the Maximum Principle, $|h_n(p)|<r_n$ for every $n\in\n$. By (\[eq:converges\]), taking limits as $n$ goes to infinity, one has $|\hat{h}(p)|\leq1$ and, again by the Maximum Principle, the first assertion holds.
For the properness it suffices that $\hat{h}^{-1}(\overline{\b}(t))$ is a compact subset of $\Rcal$ for any $0<t<1$. Observe first that [(b$_n$)]{} and [(c$_n$)]{} imply $$\label{eq:hath-h}
\|\hat{h}-h_n\|_{0,\Ncal_n}<\eta/2^n\quad \forall n\in\n.$$ Since $r_n\to 1$ and $\xi_n\to 0$ as $n\to\infty$, we can take $n_0\in\n$ large enough so that $$\label{eq:r-t}
t+\xi_{n-1}+\eta/2^n<r_{n-1}\quad \forall n\geq n_0.$$ Combining [(e$_n$)]{} and (\[eq:hath-h\]) one infers that $$\hat{h}(\overline{\Ncal}_n\setminus \Ncal_{n-1})\cap \overline{\b}(r_{n-1}-\xi_{n-1}-\eta/2^n)=\emptyset\quad \forall n\geq n_0.$$ Therefore, (\[eq:r-t\]) gives that $$(\overline{\Ncal}_n\setminus \Ncal_{n-1})\cap \hat{h}^{-1}(\overline{\b}(t))=\emptyset\quad \forall n\geq n_0,$$ hence $\hat{h}^{-1}(\overline{\b}(t))\subset \Ncal_{n_0}$ is compact in $\Rcal$ and we are done.
This completes the proof of Theorem \[th:intro2\] for the case of immersions. If $m>2$, we construct the sequence $h_n\colon \overline \Rcal \to \b\subset \c^m$ so that, in addition, $h_n$ is an embedding on the bordered domain $\overline \Ncal_n \subset \Rcal$ for every $n$; this is possible by applying the general position argument at each step. Furthermore, if the approximation of $h_n$ by $h_{n+1}$ is sufficiently close on $\overline \Ncal_n$ (which is insured by choosing the sequence $\xi_n>0$ to converge to zero fast enough), then the limit map $\hat h=\lim_{n\to\infty} h_n$ is also an embedding. This can be done, for instance, taking in addition to [(b$_n$)]{}, $$\xi_n<\frac1{2n^2} \inf\left\{ |h_{n-1}(p)-h_{n-1}(q)|\colon p,q\in \Ncal_{n-1},\,
{\sf d}(p,q)>\frac1{n}\right\},$$ where ${\sf d}(\cdot,\cdot)$ is any fixed Riemannian metric in $\Rcal$; see the proof of Theorem 4.5 in [@AL3] or [@DF].
Complete proper immersions to Stein manifolds {#sec:Stein}
=============================================
By combining our methods with those in [@DF] we also find complete proper holomorphic immersions of any bordered Riemann surface to an arbitrary Stein manifold of dimension $>1$.
\[th:Stein\] Let $X$ be a Stein manifold of dimension $>1$ endowed with a hermitian metric $ds^2_X$, and let $\overline \Rcal$ be a bordered Riemann surface. Every holomorphic map $f\colon \overline \Rcal\to X$ can be approximated, uniformly on compacta in $\Rcal$, by proper complete holomorphic immersions $\hat f\colon \Rcal\to X$. If $\dim X\ge 3$ then $\hat f$ can be chosen an embedding.
We outline the necessary modifications to the proof of Theorem \[th:intro2\].
Consider first the case when $X$ is a relatively compact, smoothly bounded, strongly pseudoconvex domain in another Stein manifold $Y$, and the hermitian metric $ds^2_X$ extends to a metric $ds^2_Y$ on $Y$. Choose a smooth strongly plurisubharmonic function $\rho$ in a neighborhood of $\overline X$ in $Y$ such that $X=\{\rho<0\}$ and $d\rho\ne 0$ on $bX=\{\rho<0\}$. Pick a number $c_0>0$ such that $\rho$ has no critical values in $[-c_0,+c_0]$.
Every point $p\in bX$ admits a pair of open neighborhoods $U'_p\Subset U_p\Subset Y$ and a holomorphic coordinate map $\theta_p\colon U_p \stackrel{\cong}{\longrightarrow} \b\subset \c^n$ onto a ball $\b$ in $\c^n$, centered at $0=\theta_p(p)$, such that $\theta(U'_p)$ is a smaller ball $\b' \subset \b$ also centered at $0$, and the function $\rho_p := \rho\circ \theta_p^{-1}\colon \b \to\r$ is strongly convex. In particular, $\Sigma_p := \theta_p(bX\cap U_p)$ is a strongly convex hypersurface in $\b$ that may be chosen $\Ccal^2$-close to a spherical cap. Furthermore, by shrinking $U_p$ we insure that the metric $ds^2_Y$ pulls back by $\theta_p$ to a hermitian metric on $\b$ that is comparable to the standard Euclidean metric. By choosing the neighborhood $U'_p$ small enough compared to $U_p$ we can also arrange that, for any point $q\in U'_p$, the intersection of the tangent plane at the point $q':=\theta_p(q) \in\b'$ to the strongly convex hypersurface $\{\rho_p=\rho_p(q')\} \subset \b$ with the convex domain $D_p:=\theta_p(D\cap U_p) \subset\b$ is a relatively compact subset of $\b$.
By compactness of $bX$ there are finitely many holomorphic coordinate charts $\Ucal=\{(U_j,\theta_j)\}_{j=1,\ldots,m}$ as above, and the corresponding subsets $U'_j \Subset U_j$, such that $bX \subset U':=\cup_{j=1}^m U'_j$.
By decreasing the constant $c_0>0$ if necessary we can insure that $U'$ contains the collar $\{z \in Y \colon -c_0\le \rho(z) \le +c_0\}$.
Fix a map $f\colon \overline R\to X$ of class $\Ascr(\Rcal,X)$. Pick a constant $c_1>0$ such that $f(\overline \Rcal) \subset \{z \in X\colon \rho(z)<-c_1\}$. Now choose a constant $c$ with $0<c<\min\{c_0,c_1\}$, and a number $\epsilon$ with $0<\epsilon <c$. By [@DF] we can approximate the map $f$, uniformly on a given compact set in $\Rcal$, by a map $f_0\in \Ascr(\Rcal,X)$ satisfying $-c<\rho(f_0(x)) < -c+\epsilon$ for all $x\in b\Rcal$. The proof in [@DF] uses the tools described in Sec. \[sec:preli\] above; in particular, the Riemann-Hilbert problem and the method of gluing sprays. By general position we may assume that $f_0$ is an immersion. Now replace $f$ by $f_0$ and assume that $f$ satisfies these properties.
We now follow the construction in Sec. \[sec:mainlemma\] to find a new immersion $\hat f\colon \overline \Rcal \to X$ which satisfies an analogue of Lemma \[lem:main\] in this setting. We begin by subdividing the boundary $b\Rcal$ into subarcs $\alpha_{i,j}$ as in Subsec. \[sec:step1\] such that each arc $f(\alpha_{i,j})\subset D$ is contained in one of the sets $U'_k$, and it satisfies the relevant conditions stated in Subsec. \[sec:step1\] with respect to the local holomorphic coordinates on $U'_k$. We then perform the same construction as in the proof of Lemma \[lem:main\] within the local chart $(U_k,\theta_k)$. The geometric conditions described above enable the use of stretching, first from the two endpoints of $\alpha_{i,j}$, and then from the middle segment, towards the boundary $bX\cap U_k$ so that the induced boundary distance in $\Rcal$ increases by a specific amount. Since the geometry in each local chart is essentially the same as in the model case ( recall that $\theta_k(bX\cap U_k)$ is close to a spherical cap and the hermitian metric from $X$ is comparable with the Euclidean metric), the relevant estimates in Sec. \[sec:mainlemma\] remain valid up to a uniformly bounded numerical factor whose presence does not present any problem in the proof. Each of the local modifications obtained in this way is glued with the existing immersion on the rest of the domain by the method of gluing sprays; see Lemma \[lem:Gspray\] and Remark \[gluing-by-sprays\].
In this way we approximate $f$, uniformly on a given compact in $\Rcal$, by a new immersion $\hat f\colon \overline \Rcal \to X$ so that the boundary moves closer to $bX$ by a controlled amount, and the boundary distance in the immersed curve also increases by a presribed amount, the two numbers being related in the Pythagoras’ way; compare with Lemma \[lem:main\]. The details are very similar to those given above and will be left out. The proof is finished by an induction just as in the proof of Theorem \[th:intro2\].
This settles the case when $X$ is a bounded strongly pseudoconvex domain.
The general case is obtained as follows. Choose a smooth strongly plurisubharmonic exhaustion function $\rho\colon X\to\r$ with Morse critical points. Let $f\colon \overline \Rcal\to X$ be a map of class $\Ascr(\Rcal,X)$. Pick an increasing sequence $c_0<c_1<c_2\cdots$, with $\lim_{j\to\infty} c_j=+\infty$, such that every $c_j$ is regular value of $\rho$ and $f(\overline \Rcal)\subset \{\rho<c_0\}$. The sets $X_j=\{\rho<c_j\}$ for $j=0,1,\ldots$ are smoothly bounded strongly pseudoconvex domains exhausting $X$. Fix a point $p_0\in \Rcal$. Choose an increasing sequence $0< M_0< M_1 < M_2<\cdots$ with $\lim M_j=+\infty$. By the special case explained above we can approximate $f$, uniformly on a given compact set $K\subset \Rcal$, by a holomorpic immersion $f_0\colon \overline \Rcal \to X_0$ such that $f_0(b\Rcal)$ is very close to $bX_0$, and for any curve $\gamma\subset \Rcal$ connecting $p_0$ to the boundary $b\Rcal$, the image curve $f_0(\gamma) \subset X_0$ has length a t least $M_0+1$ with respect to $ds^2_X$. We can find a relatively compact subdomain $\Rcal_0\Subset \Rcal$ very close to $\Rcal$ such that the distance from $p_0$ to $\Rcal\setminus \Rcal_0$ with respect to the metric $f_0^* (ds^2_X)$ is $>M_0$, and such that $f_0(\overline \Rcal \setminus \Rcal_0)$ lies in a small neighborhood of $bX_0$. Applying again the special case we approximate the map $f_0$, uniformly on $\overline \Rcal_0$, by a holomorphic immersion $f_1\colon \overline \Rcal \to X_1$ such that $f_1(b\Rcal)$ is contained in a small neighborhood of $bX_1$ and the distance from $p_0$ to $b\Rcal$ with respect to the metric $f_1^* (ds^2_X)$ is $>M_1+1$. Next choose a domain $\Rcal_1\Subset \Rcal$ containing $\overline \Rcal_0$ such that the distance from $p_0$ to $\Rcal\setminus \Rcal_1$ in the metric $f_1^* (ds^2_X)$ is $>M_1$ and $f_1(\overline \Rcal\setminus \Rcal_1)$ is very close to $bX_1$.
Continuing inductively, we obtain a sequence of holomorphic immersions $f_j\colon \overline\Rcal \to X_j$ that converges uniformly on compacta in $\Rcal$ to a proper complete holomorphic immersion $f=\lim_{j\to\infty} f_j\colon \Rcal\to X$. When $\dim X\ge 3$, the maps $f_j$ in the above sequence, and also the limit map $f$, can be chosen to be embeddings.
**Acknowledgements. A. Alarcón is supported by Vicerrectorado de Política Científica e Investigación de la Universidad de Granada, and is partially supported by MCYT-FEDER grants MTM2007-61775 and MTM2011-22547, Junta de Andalucía Grant P09-FQM-5088, and the grant PYR-2012-3 CEI BioTIC GENIL (CEB09-0010) of the MICINN CEI Program.**
F. Forstnerič is supported by the research program P1-0291 from ARRS, Republic of Slovenia.
[12]{}
Alarcón, A.; Forstnerič, F.: Null curves and directed immersions of open Riemann surfaces. `arXiv:1210.5617`
Alarcón, A.; López, F.J.: Minimal surfaces in $\r^3$ properly projecting into $\r^2$. J. Diff. Geom. **90**, (2012) 351–382
Alarcón, A.; López, F.J.: Null curves in $\c^3$ and Calabi-Yau conjectures. Math. Ann. **355**, (2013) 429–-455
Alarcón, A.; López, F.J.: Proper holomorphic embeddings of Riemann surfaces with arbitrary topology into $\c^2$. J. Geom. Anal., in press. `arXiv:1104.1893`
Bell, S.R.; Narasimhan, R.: Proper holomorphic mappings of complex spaces. Several complex variables, VI, pp. 1–38. Encyclopaedia Math. Sci., 69, Springer-Verlag, Berlin (1990)
Bishop, E.: Mappings of partially analytic spaces. Amer. J. Math. **83**, (1961) 209–242
Dor, A.: Immersions and embeddings in domains of holomorphy. Trans. Am. Math. Soc. **347**, (1995) 2813–2849
Drinovec Drnovšek, B.; Forstnerič, F.: Holomorphic curves in complex spaces. Duke Math. J. **139**, (2007) 203–254
Drinovec Drnovšek, B.; Forstnerič, F.: Strongly pseudoconvex Stein domains as subvarieties of complex manifolds. Am. J. Math. **132**, (2010) 331–360
Drinovec Drnovšek, B.; Forstnerič, F.: The Poletsky-Rosay theorem on singular complex spaces. Indiana Univ. Math. J., in press. `arXiv:1104.3968`
Ferrer, L.; Martín, F.; Meeks, W.H. III: Existence of proper minimal surfaces of arbitrary topological type. Adv. Math. **231**, (2012) 378–413
Forstnerič, F.: Manifolds of holomorphic mappings from strongly pseudoconvex domains. Asian J. Math. **11**, (2007) 113–126
Forstnerič, F.: Stein Manifolds and Holomorphic Mappings (The Homotopy Principle in Complex Analysis). Ergebnisse der Mathematik und ihrer Grenzgebiete, 3. Folge, 56, Springer-Verlag, Berlin-Heidelberg (2011)
Forstnerič, F.; Globevnik, J.: Discs in pseudoconvex domains. Comment. Math. Helv. **67**, (1992) 129–145
Forstnerič, F.; Wold, E.F.: Bordered Riemann surfaces in $\c^2$. J. Math. Pures Appl. **91**, 100–114 (2009)
Forstnerič, F.; Wold, E.F.: Embeddings of infinitely connected planar domains into $\mathbb C^2$. Analysis & PDE, in press. `arXiv:1110.5354`
Jones, P.W.: A complete bounded complex submanifold of $\c^3$. Proc. Amer. Math. Soc. **76**, (1979) 305–306
López, F.J.: Hyperbolic complete minimal surfaces with arbitrary topology. Trans. Amer. Math. Soc. **350**, (1998) 1977–1990
Løw, E.: Embeddings and proper holomorphic maps of strictly pseudoconvex domains into polydiscs and balls. Math. Z. **190**, (1985) 401–410
Majcen, I.: Embedding certain infinitely connected subsets of bordered Riemann surfaces properly into $\c^2$. J. Geom. Anal. **19**, (2009) 695–707
Martín, F.; Umehara, M.; Yamada, K.: Complete bounded holomorphic curves immersed in $\c^2$ with arbitrary genus. Proc. Amer. Math. Soc. **137**, (2009) 3437–3450
Nadirashvili, N.: Hadamard’s and Calabi-Yau’s conjectures on negatively curved and minimal surfaces. Invent. Math. **126**, (1996) 457–465
Narasimhan, R.: Imbedding of holomorphically complete complex spaces. Amer. J. Math. **82**, (1960) 917–934
Stensønes, B.: Proper holomorphic mappings from strongly pseudoconvex domains in $\c^2$ to the unit polydisc in $\c^3$. Math. Scand. **65**, (1989) 129–139
Yang, P.: Curvatures of complex submanifolds of $\c^n$. J. Diff. Geom. **12**, (1977) 499–511
Yang, P.: Curvature of complex submanifolds of $\c^n$. Proc. Sympos. Pure Math., vol. 30, part 2, pp. 135–137. Amer. Math. Soc., Providence (1977)
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'This is the second paper in a series of three papers aiming to study cohomology of group theoretic Dehn fillings. In the present paper, we derive a spectral sequence for Cohen-Lyndon triples which can be thought of as a refined version of the classical Lyndon-Hochschild-Serre spectral sequence in the settings of group theoretic Dehn fillings. In the next paper [@sun2019cohomologyiii], we will apply this spectral sequence to study cohomological finiteness properties of Dehn fillings of acylindrically hyperbolic groups.'
author:
- Bin Sun
bibliography:
- 'bin\_refs.bib'
title: 'Cohomology of group theoretic Dehn fillings II: A spectral sequence'
---
Introduction
============
#### 1.1 Dehn surgery in $3$-manifolds.
In $3$-dimensional topology, *Dehn surgery* is a process of cutting off a solid torus from a $3$-manifold and then gluing the torus back in a different way. The Lickorish-Wallace theorem reveals an interesting and useful aspect of Dehn surgery: every closed connected orientable $3$-manifold can be obtained from the $3$-sphere by applying finitely many Dehn surgeries.
The second step of the surgery, called *Dehn filling*, begins with a $3$-manifold $M$ with toral boundary and glues a solid torus to $M$ by identifying their boundaries, resulting in a new manifold. Topologically distinct ways of gluing a solid torus are parametrized by free homotopy classes of essential simple closed curves on $\partial M$, called *slopes*. Dehn filling interacts nicely with hyperbolicity:
\[thm. thurson hyperbolic Dehn filling\] Let $M$ be a compact orientable $3$-manifold with boundary a torus, such that $M\setminus\partial M$ admits a complete finite volume hyperbolic structure. Then for all but finitely many slopes on $\partial M$, the corresponding Dehn filling results in a hyperbolic $3$-manifold.
#### 1.2 Group theoretic Dehn filling.
There is also a group theoretic version of Dehn fillings. Let $G$ be a group with a subgroup $H$ and let $N$ be a normal subgroup of $H$. The *group theoretic Dehn filling* corresponding to the triple $(G,H,N)$ is the quotient group $G/{\langle\hspace{-.7mm}\langle }N \rr$, where ${\langle\hspace{-.7mm}\langle }N \rr$ is the normal closure of $N$ in $G$.
Relations of these two kinds of Dehn fillings can be seen via the following example: Under the assumptions of Theorem \[thm. thurson hyperbolic Dehn filling\], the natural map $\pi_1(\partial M)\rightarrow \pi_1(M)$ is injective and thus $\pi_1(\partial M)$ can be naturally thought of as a subgroup of $\pi_1(M)$. Let $G=\pi_1(M)$ and $H=\pi_1(\partial M)$. Every slope $s$ on $\partial M$ generates a normal subgroup $N_s=\langle s \rangle \lhd H$. Let $M_s$ be the Dehn filling corresponding to a slope $s$, then $\pi_1(M_s)=G/{\langle\hspace{-.7mm}\langle }N_s \rr$ by the Seifert-van Kampen theorem.
Dehn filling is a useful tool in group theory. The solution of the virtually Haken conjecutre makes use of Dehn fillings in word-hyperbolic groups [@agol2013virtual]. For certain relatively hyperbolic groups, Dehn fillings are used to study the Farrel-Jones conjecture and isomorphism problem [@antolin2017farrell; @dahmani2018recognizing]. For mapping class groups of surfaces, [@dahmani2017hyperbolically] constructs purely pseudo-Anosov normal subgroups by applying Dehn fillings. Other applications of Dehn fillings can be found in [@agol2016alternate; @groves2016boundaries].
Analogs of Theorem \[thm. thurson hyperbolic Dehn filling\] can be proved for groups satisfying certain negative curvature conditions. The first such result is due to Osin [@osin2007peripheral] and independently to Groves-Manning [@groves2008dehn] in the settings of relatively hyperbolic groups. Later, a generalization of relative hyperbolicity based on the notion of a hyperbolically embedded subgroup was proposed and a generalization of the results of [@groves2008dehn; @osin2007peripheral] was obtained [@dahmani2017hyperbolically]. We refer to [@dahmani2017hyperbolically; @osin2017groups] for the corresponding definitions and results.
#### 1.3 Motivation: a question on cohomology.
If a manifold $M$ satisfies the assumptions of Theorem \[thm. thurson hyperbolic Dehn filling\] and $M_s$ is the Dehn filling corresponding to a slope $s$ on $\partial M$, then Theorem \[thm. thurson hyperbolic Dehn filling\] can be applied to compute the group cohomology $\operatorname{H}^{\ast}(\pi_1(M_s);\cdot)$. Indeed, Theorem \[thm. thurson hyperbolic Dehn filling\] asserts that in most cases, $M_s$ is a hyperbolic $3$-manifold and thus the universal cover of $M_s$, namely $\mathbb{H}^3$, is contractible. It follows that $M_s$ is a model of the classifying space of $\pi_1(M_s)$. However, in the more general settings of hyperbolically embedded subgroups, no geometric construction is involved and it is unclear how the group cohomology of the Dehn filling quotient $\operatorname{H}^{\ast}(G/{\langle\hspace{-.7mm}\langle }N \rr;\cdot)$ can be computed. Therefore, it is natural to ask the following.
\[question. main question\] Let $G$ be a group, let $H$ be a hyperbolically embedded subgroup of $G$ (denoted as $H\hookrightarrow_h G$), and let $N$ be a normal subgroup of $H$. What can be said about the group cohomology $\operatorname{H}^{\ast}(G/{\langle\hspace{-.7mm}\langle }N \rr; \cdot)$?
In this series of three papers, we answer this question and obtain several applications. The first paper [@sun2018cohomologyi] proves a Cohen-Lyndon type theorem for Dehn filling kernels. In the present paper, we derive a spectral sequence for Cohen-Lyndon triples, which can be regarded a refined version of the classical Lyndon-Hochschild-Serre spectral sequence, to compute cohomology of the corresponding Dehn filling quotients. With the aid of this spectral sequence, the third paper [@sun2019cohomologyiii] will study cohomological properties of Dehn fillings and obtain applications on acylindrically hyperbolic groups.
#### Acknowledgement.
I would like to thank my supervisor, Professor Denis Osin, for the valuable discussions. This paper would not have been written without his help. I would also like to thank Professor Anna Marie Bohmann, who gave me many useful suggestions on an early version of this paper.
Main results {#sec. main results}
============
#### 2.1 Cohen-Lyndon type theorems.
Recall that in the first paper [@sun2018cohomologyi] of this series, we obtained a free product structure, called the Cohen-Lyndon property, for sufficiently deep Dehn fillings of hyperbolically embedded subgroups.
Let $G$ be a group with a subgroup $H$. We say that a property $P$ holds *for all sufficiently deep* normal subgroups $N\lhd H$ if there exists a finite set $\mathcal{F}\subset H\setminus\{1\}$ such that $P$ holds for all normal subgroups $N\lhd H$ with $N\cap\mathcal{F}=\emptyset$.
Let $G$ be a group, let $H$ be a subgroup of $G$, and let $N$ be a normal subgroup of $H$. $(G,H,N)$ is called a *Cohen-Lyndon triple* if there exists a left transversal $T$ of $H{\langle\hspace{-.7mm}\langle }N \rr$ in $G$ such that ${\langle\hspace{-.7mm}\langle }N \rr$ is the free product of the groups $tNt^{-1},t\in T$, denoted as $${\langle\hspace{-.7mm}\langle }N \rr=\prod^{\ast}_{t\in T}tNt^{-1}.$$
\[thm. simple CL-property\] Suppose that $G$ is a group with a subgroup $H\hookrightarrow_h G$. Then for all sufficiently deep normal subgroups $N\lhd H$, $(G,H,N)$ is a Cohen-Lyndon triple.
Cohen-Lyndon type theorems were first considered by Cohen-Lyndon [@cohen1963free], hence the name “Cohen-Lyndon triple”. The result of [@cohen1963free] was later generalized by [@edjvet1987cohen Theorem 1.1] and [@groves2016boundaries Theorem 4.8].
The result of [@sun2018cohomologyi] is more general than Theorem \[thm. simple CL-property\] as it deals with the more general setting of a family of weakly hyperbolically embedded subgroups. Applications of the general result can be found in graphs of groups, e.g., amalgated free products and HNN extensions (see [@sun2018cohomologyi Corollary 6.8]).
#### 2.2 A spectral sequence for Dehn fillings.
Given a Cohen-Lyndon triple $(G,H,N)$, we introduce the following notation $${\overline{G}}=G/{\ll\mathcal{N}\rr},~~~~{\overline{H}}=H/N.$$
The main result of this paper is the following.
\[thm. simple spectral sequence\] If $(G,H,N)$ is a Cohen-Lyndon triple, then for every $\mathbb{Z}{\overline{G}}$-module $A$, there exists a spectral sequence of cohomological type $$\label{eq. simple spectral sequence}
E^{p,q}_2 =
\begin{cases}
\operatorname{H}^p({\overline{H}};\operatorname{H}^q(N;A))&, \text{ if }q\neq 0\\
\operatorname{H}^p({\overline{G}};A)&, \text{ if }q=0
\end{cases}
\Rightarrow \operatorname{H}^{p+q}(G;A).$$ Here, the action of $G$ on $A$ factors through ${\overline{G}}$. In particular, the action of $N$ on $A$ fixes every element.
Combining Theorems \[thm. simple CL-property\] and \[thm. simple spectral sequence\], we obtain:
Suppose that $G$ is a group with a subgroup $H\hookrightarrow_h G$. Then for all sufficiently deep normal subgroups $N\lhd H$ and every $\mathbb{Z}{\overline{G}}$-module $A$, there exists a spectral sequence of cohomological type .
Historically, spectral sequences were introduced by Leray [@Leray1946L'anneau] in his attempt to compute cohomology of sheafs. In the proof of Theorem \[thm. simple spectral sequence\], we make use of the Lyndon-Hochschild-Serre spectral sequence, which was discovered by Lyndon [@lyndon1948cohomology] and then put into its current form by Hochschild-Serre [@hochschild1953cohomology].
In next paper [@sun2019cohomologyiii], we will run spectral sequence backwards, compute $\operatorname{H}^{\ast}({\overline{G}};A)$ from information about $\operatorname{H}^{\ast}(G;A)$ and $\operatorname{H}^{\ast}({\overline{H}};\operatorname{H}^{\ast}(N;A))$, and answer Question \[question. main question\]. To enhance our answer, we supplement Theorem \[thm. simple spectral sequence\] by relating the differentials of to the differentials of the standard Lyndon-Hochschild-Serre spectral sequence of the extension $1\rightarrow N \rightarrow H \rightarrow{\overline{H}}\rightarrow 1$ (see Remark \[rem. description of the differentials\]).
#### 2.3 Applications and remarks.
Recall that the *cohomological dimension* of a group $G$ is $$\operatorname{cd}(G)=\sup\{n\in\mathbb{N}\mid\operatorname{H}^n(G;A)=\{0\}\text{ for some }\mathbb{Z}G\text{-module }A\}.$$ It follows immediately from Theorem \[thm. simple spectral sequence\] that if $(G,H,N)$ is a Cohen-Lyndon triple, then $\operatorname{cd}({\overline{G}})\leqslant \max\{\operatorname{cd}(G),\operatorname{cd}({\overline{H}})+\operatorname{cd}(N)+1\}$. In the next paper [@sun2019cohomologyiii], we will obtain a finer estimate of $\operatorname{cd}({\overline{G}})$. Moreover, we will provide conditions that guarantee ${\overline{G}}$ to be of type $FP_k$ for $1\leqslant k \leqslant \infty$. As a further application, we will construct useful quotients of acylindrically hyperbolic groups that inherit certain cohomologcial finiteness properties of the mother groups.
In fact, we deal with a general version of Cohen-Lyndon triples which is defined for a family of subgroups and normal subgroups. The corresponding generalized version of Theorem \[thm. simple spectral sequence\] turns out to be useful in the next paper [@sun2019cohomologyiii] when we construct particular quotients of acylindrically hyperbolic groups.
If $G$ is a group hyperbolic relative to its subgroup $H$ (in particular, $H$ is a hyperbolically embedded subgroup of $G$ [@dahmani2017hyperbolically Proposition 2.4]) and both $G$ and $H$ are of type $FP_{\infty}$, then for sufficiently deep Dehn fillings, [@wang2018spectral Theorem 1.1] provides a spectral sequence of homological type which computes the relative group cohomology $\operatorname{H}^{\ast}({\overline{G}},{\overline{H}};\mathbb{Z}{\overline{G}})$ from certain combination of homology and cohomology. The spectral sequence should not be confused with the spectral sequence of [@wang2018spectral], as there is no homology involved in . It is worth noting that can also be applied to compute relative cohomology, as we will see in the next paper [@sun2019cohomologyiii].
#### 2.4 Outline of the proof.
The main idea of the proof of Theorem \[thm. simple spectral sequence\] is the following. The Lyndon-Hochschild-Serre spectral sequence for a $\mathbb{Z}{\overline{G}}$-module $A$ and the group extension $$1\rightarrow {\langle\hspace{-.7mm}\langle }N \rr\rightarrow G \rightarrow {\overline{G}}\rightarrow 1$$ takes the form $$\label{eq. simple LHS spectral sequence}
E^{p,q}_2= \operatorname{H}^p({\overline{G}};\operatorname{H}^q({\langle\hspace{-.7mm}\langle }N\rr;A))\Rightarrow \operatorname{H}^{p+q}(G;A).$$ Theorem \[thm. simple spectral sequence\] follows from a computation of $\operatorname{H}^p({\overline{G}};\operatorname{H}^q({\langle\hspace{-.7mm}\langle }N\rr;A))$, which relies on the proposition below. To state it, we first note that if $(G,H,N)$ is a Cohen-Lyndon triple, then the natural map ${\overline{H}}\rightarrow {\overline{G}}$ is injective [@sun2018cohomologyi Lemma 6.4], identifying ${\overline{H}}$ with a subgroup of ${\overline{G}}$. Therefore, it makes sense to talk about $\operatorname{CoInd}^{\mathbb{Z}{\overline{G}}}_{\mathbb{Z}{\overline{H}}}$, the co-induction from $\mathbb{Z}{\overline{H}}$-modules to $\mathbb{Z}{\overline{G}}$-modules.
\[prop. simple coinduce\] If $(G,H,N)$ is a Cohen-Lyndon triple, then for all $q\in \mathbb{Z}\setminus\{0\}$ and every $\mathbb{Z}{\overline{G}}$-module $A$, there is a $\mathbb{Z}{\overline{G}}$-module isomorphism $$\label{simple coinduce}
\operatorname{H}^q({\langle\hspace{-.7mm}\langle }N \rr; A)\cong \operatorname{CoInd}^{\mathbb{Z}{\overline{G}}}_{\mathbb{Z}{\overline{H}}}\hspace{1mm} \operatorname{H}^q(N;A).$$
Thus, Shapiro’s lemma implies
\[prop. simple iso of iterative cohomology\] If $(G,H,N)$ is a Cohen-Lyndon triple, then for all $q\in \mathbb{Z}\setminus\{0\}$ and every $\mathbb{Z}{\overline{G}}$-module $A$, there is an abelian group isomorphism $$\label{eq. simple isomorphism of iterative cohomology}
\operatorname{H}^{\ast}({\overline{G}};\operatorname{H}^q({\langle\hspace{-.7mm}\langle }N\rr;A))\cong \operatorname{H}^{\ast}({\overline{H}};\operatorname{H}^q(N;A)).$$
Notice that for $q=0$, we have $$\label{eq. when q=0}
E^{\ast,0}_2= \operatorname{H}^{\ast}({\overline{G}};\operatorname{H}^0({\langle\hspace{-.7mm}\langle }N \rr;{A}))\cong \operatorname{H}^{\ast}({\overline{G}};{A}^{{\langle\hspace{-.7mm}\langle }N \rr})\cong \operatorname{H}^{\ast}({\overline{G}};{A}),$$ where ${A}^{{\langle\hspace{-.7mm}\langle }N \rr}$ is the ${\langle\hspace{-.7mm}\langle }N \rr$-fixed-points of ${A}$. As ${A}$ is a $\mathbb{Z}{\overline{G}}$-module, the ${\langle\hspace{-.7mm}\langle }N \rr$-action on ${A}$ fixes every point and thus ${A}^{{\langle\hspace{-.7mm}\langle }N \rr}={A}$.
The spectral sequence is obtained by substituting terms of for the terms on the right-hand sides of and .
A natural way to prove Propostion \[prop. simple coinduce\] is to decompose $\operatorname{H}^q({\langle\hspace{-.7mm}\langle }N \rr; A)$ into a direct product $\prod_{t\in T}\operatorname{H}^q(tNt^{-1};A)$, which can be achieved by starting with a model $X$ of the classifying space of $N$ and then taking wedge sum of copies of $X$ to obtain a model of the classifying space of ${\langle\hspace{-.7mm}\langle }N \rr$. The problem with this approach is that one loses the information about the action ${\overline{G}}\curvearrowright \operatorname{H}^q({\langle\hspace{-.7mm}\langle }N \rr;A)$ and deriving Proposition \[prop. simple coinduce\] becomes impossible. Therefore, we must take another approach. We first prove that $\operatorname{CoInd}^{\mathbb{Z}{\overline{G}}}_{\mathbb{Z}{\overline{H}}}\hspace{1mm} \operatorname{H}^{\ast}(N;A)\cong \operatorname{Ext}^{\ast}_{\mathbb{Z}{\langle\hspace{-.7mm}\langle }N \rr}(\mathbb{Z}[G/H],A)$ as $\mathbb{Z}{\overline{G}}$-modules. And then we show that there exist projective resolutions $P\rightarrow \mathbb{Z}[G/H]$ and $R\rightarrow\mathbb{Z}$ over $\mathbb{Z}{\langle\hspace{-.7mm}\langle }N \rr$ such that $P$ and $R$ coincide at dimension $1$ and beyond. It follows that for all $q\neq 0$, we have $$\operatorname{H}^q({\langle\hspace{-.7mm}\langle }N \rr; A)\cong \operatorname{Ext}^q_{\mathbb{Z}{\langle\hspace{-.7mm}\langle }N \rr}(\mathbb{Z}[G/H],A)\cong \operatorname{CoInd}^{\mathbb{Z}{\overline{G}}}_{\mathbb{Z}{\overline{H}}}\hspace{1mm} \operatorname{H}^q(N;A)$$ as $\mathbb{Z}{\overline{G}}$-modules.
This paper is organized as follows. We begin by recalling necessary definitions in Section \[sec. preliminaries\], where we define Cohen-Lyndon triples and introduce useful notations. We prove Theorem \[thm. simple spectral sequence\] in Section \[sec. proof\], where we first assume Proposition \[prop. simple coinduce\] and show Corollary \[prop. simple iso of iterative cohomology\] and Theorem \[thm. simple spectral sequence\], and then we prove Proposition \[prop. simple coinduce\] in Section \[sec. proof of prop coind\].
Preliminaries {#sec. preliminaries}
=============
We begin by introducing conventions and notations. Throughout this paper, all modules (resp. group actions) are left modules (resp. actions). We write $\otimes$ for $\otimes_{\mathbb{Z}}$, the tensor product over $\mathbb{Z}$. Given a group $G$, we write $\operatorname{Hom}_G,\operatorname{Ext}_G$ for $\operatorname{Hom}_{\mathbb{Z}G},\operatorname{Ext}_{\mathbb{Z}G}$, respectively. If $G$ is the free product of its subgroups $G_{\lambda},\lambda\in\Lambda$, then we write $G=\prod^{\ast}_{\lambda\in\Lambda}G_{\lambda}$. If $S$ is a subset of $G$, then ${\langle\hspace{-.7mm}\langle }S \rr$ is the normal closure of $S$ in $G$. If $H$ is a subgroup of $G$, then $LT(H,G)$ is the set of left transversals of $H$ in $G$, and we write $\operatorname{CoInd}^G_H$ for the co-induction from $\mathbb{Z}H$-modules to $\mathbb{Z}G$-modules, i.e., $$\operatorname{CoInd}^G_H A=\operatorname{Hom}_H(\mathbb{Z}G,A)\text{ for all }\mathbb{Z}H\text{-module }A.$$
We briefly recall several actions and refer to [@brown1982cohomology] for details. For every $\mathbb{Z}H$-module $A$, there is a $G$-action on $\operatorname{CoInd}^G_H A$ given by $$(g\circ f)(x)=f(x\cdot g)\text{ for all }g\in G,x\in\mathbb{Z}G,f\in\operatorname{CoInd}^G_H A.$$
Suppose that $H$ is a normal subgroup of $G$ and $B,C$ are $\mathbb{Z}G$-modules. Then there is a $G/H$-action on $\operatorname{Hom}_H(B,C)$ induced by the following $G$-action $$\label{eq. G action on hom}
(g\circ f)(x)=g\cdot f(g^{-1}\cdot x)\text{ for all }g\in G,x\in B,f\in\operatorname{Hom}_H(B,C).$$
Let $P\rightarrow B$ be a projective resolution over $\mathbb{Z}G$. Then formula with $P$ in place of $B$ induces a $G/H$-action on $\operatorname{Ext}^{\ast}_H(B,C)$. In case $B=\mathbb{Z}$, gives rise to an action of $G/H$ on $\operatorname{H}^{\ast}(H;C)$.
Spectral sequences
------------------
Our main result involves spectral sequences. The reader is referred to [@rotman2008introduction; @weibel1995introduction] for an exposition of this algebraic object. In this section, we only clarify notations and terminologies.
Throughout this paper, all spectral sequences are first quadrant spectral sequences of cohomological type. We denote such a spectral sequence as $E=(E_r,d_r)_{r\geqslant a}$, where $E_r$ is the $E_r$-page of $E$ and $d_r$ is the differential on $E_r$. The notation $E^{p,q}_a\Rightarrow H^{p+q}$ indicates that $E$ converges to a graded abelian group $H=\bigoplus_{\ell\geqslant 0}H^{\ell}$.
A morphism between spectral sequences $E_1=(E_{1,r},d_{1,r})_{r\geqslant a},E_2=(E_{2,r},d_{2,r})_{r\geqslant a}$ is denoted as $$\phi=(\phi_r)_{r\geqslant b}:E_1\rightarrow E_2,$$ where $\phi_r:E_{1,r}\rightarrow E_{2,r}$ is the restriction of $\phi$ and it is understood that $b\geqslant a$ and $\phi_r$ is only defined for $r\geqslant b$. For $p,q\in\mathbb{Z}$, we denote by $$\phi^{p,q}_r:E^{p,q}_{1,r}\rightarrow E^{p,q}_{2,r},~~~~\phi^{p,\ast}_r:E^{p,\ast}_{1,r}\rightarrow E^{p,\ast}_{2,r},~~~~\phi^{\ast,q}_r:E^{\ast,q}_{1,r}\rightarrow E^{\ast,q}_{2,r}$$ the maps induced by $\phi_r$.
Also recall the notion of a Lyndon-Hochschild-Serre spectral sequence.
Let $G$ be a group, let $K$ be a normal subgroup of $G$, and let $A$ be a $\mathbb{Z}G$-module. The *Lyndon-Hochschild-Serre (LHS) spectral sequence* associated with the data $(G,K,A)$ is a spectral sequence $E=(E_r,d_r)_{r\geqslant 1}$ such that $$E^{p,q}_2=\operatorname{H}^p(G/K;\operatorname{H}^q(K;A))\Rightarrow \operatorname{H}^{p+q}(G;A).$$
Cohen-Lyndon triples {#sec.ntcgtr}
--------------------
Suppose that $G$ is a group with a family of subgroups $\{H_{\lambda}\}_{\lambda\in\Lambda}$ and for every $\lambda\in\Lambda$, $N_{\lambda}$ is a normal subgroup of $H_{\lambda}$. For future reference, we specify the following notations.
\[definition of natural maps\] Let $$\mathcal{N}=\bigcup_{\lambda\in\Lambda} N_{\lambda},~~~~{\overline{G}}=G/{\ll\mathcal{N}\rr},~~~~{\overline{H}_{\lambda}}=H_{\lambda}/N_{\lambda}.$$ If ${A}$ is a $\mathbb{Z}{\overline{G}}$-module, then there is a natural action of $G$ on $A$ which factors through ${\overline{G}}$. In particular, the action of $N_{\lambda}$ on $A$ is trivial (i.e., fixes every element of $A$). We will compute the cohomology groups $\operatorname{H}^{\ast}(G;A),\operatorname{H}^{\ast}(H_{\lambda};A),$ etc using this natural action.
For $\lambda\in\Lambda$, let $$r_{H_{\lambda}}:\operatorname{H}^{\ast}(G;A)\rightarrow \operatorname{H}^{\ast}(H_{\lambda};A),~~~~r_{N_{\lambda}}:\operatorname{H}^{\ast}({\ll\mathcal{N}\rr};A)\rightarrow \operatorname{H}^{\ast}(N_{\lambda};A)$$ be the restriction maps induced by the inclusions $H_{\lambda}\leqslant G$ and $N_{\lambda}\leqslant{\ll\mathcal{N}\rr}$, respectively. Let $$r_G:\operatorname{H}^{\ast}(G;A)\rightarrow {\prod_{\lambda\in\Lambda}}\operatorname{H}^{\ast}(H_{\lambda};A)$$ be the map induced by the maps $r_{H_{\lambda}}$. For $\lambda\in\Lambda$, the map $r_{N_{\lambda}}$ and the natural homomorphism ${\overline{H}_{\lambda}}\rightarrow{\overline{G}}$ induce a cohomology map $$\psi_{\lambda}:\operatorname{H}^{\ast}({\overline{G}};\operatorname{H}^{\ast}({\ll\mathcal{N}\rr};A))\rightarrow \operatorname{H}^{\ast}({\overline{H}_{\lambda}};\operatorname{H}^{\ast}(N_{\lambda};A)),$$ where the ${\overline{G}}$ (resp. ${\overline{H}_{\lambda}}$) cohomology is computed using the ${\overline{G}}$-action (resp. ${\overline{H}_{\lambda}}$-action) on $\operatorname{H}^{\ast}({\ll\mathcal{N}\rr};A)$ (resp. $\operatorname{H}^{\ast}(N_{\lambda};A)$) defined at the beginning of Section \[sec. preliminaries\].
Let $$\psi:\operatorname{H}^{\ast}({\overline{G}};\operatorname{H}^{\ast}({\ll\mathcal{N}\rr};A))\rightarrow {\prod_{\lambda\in\Lambda}}\operatorname{H}^{\ast}({\overline{H}_{\lambda}};\operatorname{H}^{\ast}(N_{\lambda};A))$$ be the map induced by the maps $\psi_{\lambda}$.
\[def. cohen-lyndon triple\] We call $(G,\{H_{\lambda}\}_{\lambda\in\Lambda},\{N_{\lambda}\}_{\lambda\in\Lambda})$ a *Cohen-Lyndon triple* if there exists a left transversal $T_{\lambda}\in LT(H_{\lambda}{\ll\mathcal{N}\rr},G)$ for every $\lambda\in\Lambda$ such that $${\ll\mathcal{N}\rr}=\prod^{\ast}_{\lambda\in\Lambda,t\in T_{\lambda}} tN_{\lambda}t^{-1}.$$
\[natural map embedding\] If $(G,\{H_{\lambda}\}_{\lambda\in\Lambda},\{N_{\lambda}\}_{\lambda\in\Lambda})$ is a Cohen-Lyndon triple and $N_{\lambda}\neq\{1\}$ for every $\lambda\in\Lambda$, then for all $\lambda\in\Lambda$, we have $H_{\lambda}\cap{\ll\mathcal{N}\rr}=N_{\lambda}$ [@sun2018cohomologyi Lemma 6.4] and thus the natural homomorphism ${\overline{H}_{\lambda}}\rightarrow {\overline{G}}$ is injective, identifying ${\overline{H}_{\lambda}}$ with a subgroup of ${\overline{G}}$.
Sufficiently deep Dehn fillings of weakly hyperbolically embedded subgroups are studied in [@sun2018cohomologyi], which proves the following.
\[thm. wh implies cl\] If the family $\{H_{\lambda}\}_{\lambda\in\Lambda}$ weakly hyperbolically embeds into $(G,X)$ for some $X\subset G$. Then for all sufficiently deep normal subgroups $N_{\lambda}\lhd H_{\lambda}, (G,\{H_{\lambda}\}_{\lambda\in\Lambda},\{N_{\lambda}\}_{\lambda\in\Lambda})$ is a Cohen-Lyndon triple.
A corollary to Shapiro’s lemma
------------------------------
As indicated in Section \[sec. main results\], we will prove Theorem \[thm. simple spectral sequence\] in the more general setting of a family of subgroups. It is therefore necessary to slightly generalize Shapiro’s lemma to a family of subgroups. However, we would rather call the following generalization a corollary as it actually follows directly from Shapiro’s lemma.
Let $G$ be a group, let $\{H_{\lambda}\}_{\lambda\in\Lambda}$ be a family of subgroups of $G$, and let $A_{\lambda}$ be a $\mathbb{Z}H_{\lambda}$-module for $\lambda\in\Lambda$. There is a map $$\pi_{\lambda}:\operatorname{CoInd}^G_{H_{\lambda}}A_{\lambda}\rightarrow A_{\lambda},~\pi_{\lambda}(f)=f(1).$$ The well-known Shapiro’s lemma asserts that for every $\lambda\in\Lambda$, the inclusion $H_{\lambda}\leqslant G$ and the map $\pi_{\lambda}$ induce an isomorphism $$\phi_{\lambda}:\operatorname{H}^{\ast}(G;\operatorname{CoInd}^G_{H_{\lambda}}A_{\lambda})\rightarrow \operatorname{H}^{\ast}(H_{\lambda};A_{\lambda}).$$ Let $$\phi: {\prod_{\lambda\in\Lambda}}\operatorname{H}^{\ast}(G; \operatorname{CoInd}^G_{H_{\lambda}}A_{\lambda})\rightarrow {\prod_{\lambda\in\Lambda}}\operatorname{H}^{\ast}(H_{\lambda};A_{\lambda})$$ be the map on products induced by the maps $\phi_{\lambda}$. Then $\phi$ is an isomorphism. As ${\prod_{\lambda\in\Lambda}}\operatorname{H}^{\ast}(G; \operatorname{CoInd}^G_{H_{\lambda}}A_{\lambda})$ is naturally isomorphic to $\operatorname{H}^{\ast}(G; {\prod_{\lambda\in\Lambda}}\operatorname{CoInd}^G_{H_{\lambda}}A_{\lambda})$, we can view $\phi$ as a map from $\operatorname{H}^{\ast}(G; {\prod_{\lambda\in\Lambda}}\operatorname{CoInd}^G_{H_{\lambda}}A_{\lambda})$ to ${\prod_{\lambda\in\Lambda}}\operatorname{H}^{\ast}(H_{\lambda};A_{\lambda})$ and we have the following.
\[lem. shapiro\] Let $G$ be a group, let $\{H_{\lambda}\}_{\lambda\in\Lambda}$ be a family of subgroups of $G$, and let $A_{\lambda}$ be a $\mathbb{Z}H_{\lambda}$-module for $\lambda\in\Lambda$. Then the map $\phi$ defined above is an isomorphism.
Proof of the main result {#sec. proof}
========================
In the sequel, we will study cohomology of Cohen-Lyndon triples. For simplicity, we will be frequently using notations defined in Notation \[definition of natural maps\]. Our goal is the following more general and precise version of Theorem \[thm. simple spectral sequence\].
\[main\] Let $(G,\{H_{\lambda}\}_{\lambda\in\Lambda},\{N_{\lambda}\}_{\lambda\in\Lambda})$ be a Cohen-Lyndon triple, let ${A}$ be a $\mathbb{Z}{\overline{G}}$-module, let $E_G=(E_{G,r},d_{G,r})_{r\geqslant 1}$ (resp. $E_{H_{\lambda}}=(E_{H_{\lambda},r},d_{H_{\lambda},r})_{r\geqslant 1}$) be the LHS spectral sequence associated with the data $(G,{\ll\mathcal{N}\rr},{A})$ (resp. $(H_{\lambda},N_{\lambda},{A})$), and let $E_{\mathcal{H}}=(E_{\mathcal{H},r},d_{\mathcal{H},r})_{r\geqslant 1}$ be the product of the spectral sequences $E_{H_{\lambda}}$.
Then there exists a morphism of spectral sequences ${\phi}=({\phi}_r)_{r\geqslant 2}:E_G\rightarrow E_{\mathcal{H}}$ which satisfies the following.
1. The maps ${\phi}$ and $r_G$ are compatible.
2. The map $\phi_2$ can be identified with the map $\psi$. Moreover, $\phi_2$ induces an isomorphism $\phi^{\ast,q}_2:E^{\ast,q}_{G,2}\rightarrow E^{\ast,q}_{\mathcal{H},2}$ for every $q\neq 0$.
Combining Theorems \[thm. wh implies cl\] and \[main\], we obtain:
Let $G$ be a group with a family of subgroups $\{H_{\lambda}\}_{\lambda\in\Lambda}$ weakly hyperbolically embedded into $(G,X)$ for some $X\subset G$. Then for sufficiently deep normal subgroups $N_{\lambda}\lhd H_{\lambda}$ and every $\mathbb{Z}{\overline{G}}$-module $A$, there exists a morphism of spectral sequences ${\phi}:E_G\rightarrow E_{\mathcal{H}}$ which satisfies (a) and (b) of Theorem \[main\].
Assuming Theorem \[main\], we prove Theorem \[thm. simple spectral sequence\].
Let $E_G$ (resp. $E_H$) be the LHS spectral sequence associated with the data $(G,{\langle\hspace{-.7mm}\langle }N \rr, A)$ (resp. $(H,N,A)$). Then $$E^{p,q}_{G,2}= \operatorname{H}^p({\overline{G}};\operatorname{H}^q({\langle\hspace{-.7mm}\langle }N \rr; A))\Rightarrow \operatorname{H}^{p+q}(G;{A}).$$
By Theorem \[main\], there exists a morphism ${\phi}:E_G\rightarrow E_H$ such that ${\phi}^{\ast,q}_2:E^{\ast,q}_{G,2}\rightarrow E^{\ast,q}_{H,2}$ is an isomorphism for all $q\in\mathbb{Z}\setminus\{0\}$. Replace $E^{\ast,q}_{G,2}$ with $E^{\ast,q}_{H,2}=\operatorname{H}^{\ast}({\overline{H}};\operatorname{H}^q(N;A))$ for all $q\in\mathbb{Z}\setminus\{0\}$. For $q=0$, equation tells us that we can replace $E^{\ast,0}_{G,2}$ with $\operatorname{H}^{\ast}({\overline{G}};A)$. After these replacements, we obtain the spectral sequence .
\[rem. description of the differentials\] We can describe the differentials of as follows. Let $d_r$ (resp. $d_{H,r}$) be the differential on the $E_r$-page of the spectral sequence (resp. the LHS spectral sequence $E_H$). Then $d_r$ is induced by $d_{H,r}$. More precisely, as results from replacing certain terms of $E_G$, we think of ${\phi}:E_G\rightarrow E_H$ as a morphism from the spectral sequence to $E_H$ and we have a commutative diagram $$\begin{tikzcd}[row sep=huge, column sep = huge]
E_r \arrow[r,"{\phi}"] \arrow[d,"d_r"] & E_{H,r} \arrow[d,"d_{H,r}"] \\
E_r \arrow[r,"{\phi}"] & E_{H,r}
\end{tikzcd}$$
The key to the proof of Theorem \[main\] is the following generalization of Proposition \[prop. simple coinduce\].
\[coind\] Suppose that $(G,\{H_{\lambda}\}_{\lambda\in\Lambda},\{N_{\lambda}\}_{\lambda\in\Lambda})$ is a Cohen-Lyndon triple and $N_{\lambda}\neq \{1\}$ for all $\lambda\in\Lambda$. As Remark \[natural map embedding\], think of the groups ${\overline{H}_{\lambda}}$ as subgroups of ${\overline{G}}$. Then for every $\mathbb{Z}{\overline{G}}$-module $A$, there exists a $\mathbb{Z}{\overline{G}}$-module homomorphism (where the ${\overline{G}}$-actions are defined at the beginning of Section \[sec. preliminaries\]) $$\eta: \operatorname{H}^{\ast}({\ll\mathcal{N}\rr};{A})\longrightarrow {\prod_{\lambda\in\Lambda}}{\operatorname{CoInd}^{\overline{G}}_{\overline{H}_{\lambda}}}\operatorname{H}^{\ast}(N_{\lambda};{A})$$ satisfying the following.
1. For all $q\geqslant 1$, $\eta$ maps $\operatorname{H}^q({\ll\mathcal{N}\rr};{A})$ isomorphically onto ${\prod_{\lambda\in\Lambda}}{\operatorname{CoInd}^{\overline{G}}_{\overline{H}_{\lambda}}}\operatorname{H}^q(N_{\lambda};{A})$.
2. Let $$p_{\mu}:{\prod_{\lambda\in\Lambda}}{\operatorname{CoInd}^{\overline{G}}_{\overline{H}_{\lambda}}}\operatorname{H}^{\ast}(N_{\lambda};{A})\rightarrow \operatorname{CoInd}^{{\overline{G}}}_{{\overline{H}}_{\mu}} \operatorname{H}^{\ast}(N_{\mu};{A})$$ be the coordinate projection, and define a map $$\pi_{\mu}:\operatorname{CoInd}^{{\overline{G}}}_{{\overline{H}}_{\mu}} \operatorname{H}^{\ast}(N_{\mu};{A})\rightarrow \operatorname{H}^{\ast}(N_{\mu};{A})\text{ by }\pi_{\mu}(f)=f(1).$$ Then $r_{N_{\mu}}=\pi_{\mu}\circ p_{\mu}\circ \eta$.
The proof of Proposition \[coind\] is the content of Section \[sec. proof of prop coind\]. For the moment, we assume Proposition \[coind\] and prove Theorem \[main\]. We first show the following corollary to Proposition \[coind\]. Recall that Notation \[definition of natural maps\] defines a map $\psi$.
\[simplify\] Let $(G,\{H_{\lambda}\}_{\lambda\in\Lambda},\{N_{\lambda}\}_{\lambda\in\Lambda})$ be a Cohen-Lyndon triple. Then for all $q\in\mathbb{Z}\setminus\{0\}$ and every $\mathbb{Z}{\overline{G}}$-module $A$, $\psi$ maps $\operatorname{H}^{\ast}({\overline{G}};\operatorname{H}^q({\ll\mathcal{N}\rr};A))$ isomorphically onto ${\prod_{\lambda\in\Lambda}}\operatorname{H}^{\ast}({\overline{H}_{\lambda}};\operatorname{H}^q(N_{\lambda};A))$.
We first deal with the special case where $N_{\lambda}\neq\{1\}$ for all $\lambda\in\Lambda$. In this case, Corollary \[simplify\] follows from Proposition \[coind\] and Lemma \[lem. shapiro\].
Now we reduce the general case to the special case. Let $$\Lambda'=\{\lambda\in\Lambda\mid N_{\lambda}\neq\{1\}\}.$$ Note that $(G,\{H_{\lambda}\}_{\lambda\in\Lambda'},\{N_{\lambda}\}_{\lambda\in\Lambda'})$ is again a Cohen-Lyndon tirple and, by the previous proved special case, Corollary \[simplify\] holds for $(G,\{H_{\lambda}\}_{\lambda\in\Lambda'},\{N_{\lambda}\}_{\lambda\in\Lambda'})$, which implies that Corollary \[simplify\] also holds for $(G,\{H_{\lambda}\}_{\lambda\in\Lambda},\{N_{\lambda}\}_{\lambda\in\Lambda})$.
\[lem. inclusion induces morphism of sp\] Suppose that $G$ is a group, $H$ is a subgroup of $G$, $K$ is a normal subgroup of $G$, $N=K\cap H$, and $A$ is a $\mathbb{Z}[G/K]$-module. Let $E_G=(E_{G,r},d_{G,r})_{r\geqslant 1}$ (resp. $E_H=(E_{H,r},d_{H,r})_{r\geqslant 1}$) be the LHS spectral sequence associated with the data $(G,K,A)$ (resp. $(H,N,A)$). Then there exists a morphism of spectral sequences $$\phi=(\phi_r)_{r\geqslant 1}: E_G\rightarrow E_H$$ which satsifies the following.
1. $\phi$ is compatible with $r:\operatorname{H}^{\ast}(G;A)\rightarrow \operatorname{H}^{\ast}(H;A)$, the restriction map induced by the inclusion $H\leqslant G$.
2. Let $$\psi:E_{G,2}=\operatorname{H}^{\ast}(G/K;\operatorname{H}^{\ast}(K;A))\rightarrow E_{H,2}=\operatorname{H}^{\ast}(H/N;\operatorname{H}^{\ast}(N;A))$$ be the cohomology map induced by the inclusion $N\leqslant K$ and the natural injection $H/N\hookrightarrow G/K$. Then $\psi$ can be identified with $\phi_2$.
The lemma is well-known and its proof follows from standard diagram tracing, so we only sketch the proof. The LHS spectral sequence $E_G$ is constructed as follows (for details, see [@rotman2008introduction Theorem 11.38]). Start with an injective resolution $A\rightarrow I$ of $A$ over the ring $\mathbb{Z}G$. Applying $\operatorname{Hom}_K$ to this resolution produces a cochain complex $C_G$. And then one takes an injective Cartan-Eilenberg resolution $C_G\rightarrow J_G$ over the ring $\mathbb{Z}[G/K]$. By applying $\operatorname{Hom}_{G/K}$ to $J_G$, one gets a double complex $D_G$. $E_G$ arises from the row filtration of $D_G$. The column filtration of $D_G$ also induces a spectral sequence $F_G=(F_{G,r},\delta_{G,r})_{r\geqslant 1}$ such that $F_{G,2}$ can be identified with $\operatorname{H}^{\ast}(G;A)$, telling us where $E_G$ converges to.
The construction for $E_H$ is similar. We need to start with an injective resolution of $A$ over $\mathbb{Z}H$, but as $H\leqslant G$, every injective $\mathbb{Z}G$-module is automatically an injective $\mathbb{Z}H$-module and thus we can view $A\rightarrow I$ as an injective resolution over $\mathbb{Z}H$. Applying $\operatorname{Hom}_N$ to this resolution produces a cochain complex $C_H$. And then there is an injective Cartan-Eilenberg resolution $C_H\rightarrow J_H$ over $\mathbb{Z}[H/N]$. As $H/N$ naturally embeds into $G/K$, every injective $\mathbb{Z}[G/K]$-module is automatically an injective $\mathbb{Z}[H/N]$-module and thus $C_G\rightarrow J_G$ can be regarded as a Cartan-Eilenberg resolution over $\mathbb{Z}[H/N]$. The obvious inclusion $C_G\hookrightarrow C_H$ then induces a map $J_G\rightarrow J_H$ [@weibel1995introduction Exercise 5.7.2], which further induces a map $\phi_D:D_G\rightarrow D_H$, where $D_H$ is the double complex resulting from applying $\operatorname{Hom}_{H/N}$ to $J_H$. Once again, $E_H$ arises from the row filtration of $D_H$ and we let $F_H=(F_{H,r},\delta_{H,r})_{r\geqslant 1}$ be the spectral sequence arising from the column filtration of $D_H$.
$\phi_D$ clearly respects the column and row filtrations of $D_G$ and $D_H$ and thus induces morphisms of spectral sequences $$\phi=(\phi_r)_{r\geqslant 1}:E_G\rightarrow E_H,~~~~\phi_F=(\phi_{F,r})_{r\geqslant 1}:F_G\rightarrow F_H.$$ The reason why $E_{G,2}$ (resp. $E_{H,2},F_{G,2},F_{H,2}$) can be identified with $\operatorname{H}^{\ast}(G/K;\operatorname{H}^{\ast}(K;A))$ (resp. $\operatorname{H}^{\ast}(H/N;\operatorname{H}^{\ast}(N;A)),\operatorname{H}^{\ast}(G;A),\operatorname{H}^{\ast}(H;A)$) is that $E_{G,1}$ (resp. $E_{H,1},F_{G,1},F_{H,1}$) is a cochain complex $\operatorname{Hom}_{G/K}(\mathbb{Z},I_K)$ (resp. $\operatorname{Hom}_{H/N}(\mathbb{Z},I_N),\operatorname{Hom}_G(\mathbb{Z},I),\operatorname{Hom}_H(\mathbb{Z},I)$). Here, $I_K$ (resp. $I_N$) is an injective resolution of $\operatorname{H}^{\ast}(K;A)$ (resp. $\operatorname{H}^{\ast}(N;A)$) over $\mathbb{Z}[G/K]$ (resp. $\mathbb{Z}[H/N]$). Moreover, $\phi_1$ (resp. $\phi_{F,1}$) is a map from $\operatorname{Hom}_{G/K}(\mathbb{Z},I_K)$ (resp. $\operatorname{Hom}_G(\mathbb{Z},I)$) to $\operatorname{Hom}_{H/N}(\mathbb{Z},I_N)$ (resp. $\operatorname{Hom}_H(\mathbb{Z},I)$), which can be seen, via a diagram tracing, to induce the cohomology map $\psi$ (resp. $r$), from which statement (b) (resp. (a)) follows.
Under the assumptions of Theorem \[main\], let us first construct, for every $\lambda\in\Lambda$, a morphism ${\phi}_{\lambda}=({\phi}_{\lambda,r})_{r\geqslant 2}:E_G\rightarrow E_{H_{\lambda}}$ of spectral sequences.
Let $$\Lambda'=\{\lambda\in\Lambda\mid N_{\lambda}\neq \{1\}\}.$$ Note that $(G,\{H_{\lambda}\}_{\lambda\in\Lambda'},\{N_{\lambda}\}_{\lambda\in\Lambda'})$ is a Cohen-Lyndon triple. By Remark \[natural map embedding\], for $\lambda\in\Lambda'$, we have $H_{\lambda}\cap{\ll\mathcal{N}\rr}=N_{\lambda}$. Apply Lemma \[lem. inclusion induces morphism of sp\] and let $${\phi}_{\lambda}:E_G\rightarrow E_{H_{\lambda}}$$ be the morphism given by that lemma.
Let $\lambda\in\Lambda\setminus\Lambda'$. Then Lemma \[lem. inclusion induces morphism of sp\] is not applicable, but we can construct a morphism by hand. For $r\geqslant 2$, as $H^0(\{1\};A)$ is naturally isomorphic to $A$, we have $$\begin{aligned}
E^{p,q}_{H_{\lambda},r}=
\begin{cases}
\operatorname{H}^p(H_{\lambda};A)&,\text{ if }q=0\\
0&,\text{ if }q\neq 0.
\end{cases}\end{aligned}$$
We define a family of maps ${\phi}^{\ast,q}_{\lambda,r}:E^{\ast,q}_{G,r}\rightarrow E^{\ast,q}_{H_{\lambda},r},q\in\mathbb{Z},$ by the following.
1. ${\phi}^{\ast,q}_{\lambda,r}:E^{\ast,q}_{G,r}\rightarrow E^{\ast,q}_{H_{\lambda}}$ is the zero map for all $q\in\mathbb{Z}\setminus\{0\}$.
2. For $p\in\mathbb{Z}$, let $R>r$ be sufficiently large such that $E^{p,0}_{G,R}$ naturally embeds into $\operatorname{H}^p(G;{A})$ (such an $R$ exists as $E^{k,\ell}_{G,2}\Rightarrow \operatorname{H}^{k+\ell}(G;{A})$). By the definition of spectral sequences, there is a quotient map $E^{p,0}_{G,r}\rightarrow E^{p,0}_{G,R}$. Let ${\phi}^{p,0}_{\lambda,r}$ be the composition $$E^{p,0}_{G,r}\rightarrow E^{p,0}_{G,R}\rightarrow \operatorname{H}^p(G;{A})\xrightarrow{r_{H_{\lambda}}} \operatorname{H}^p(H_{\lambda};{A})= E^{p,0}_{H_{\lambda},r}$$ (the definition of ${\phi}^{p,0}_{\lambda,r}$ does not depend on the choice of $R$).
The maps ${\phi}^{\ast,q}_{\lambda,r},q\in\mathbb{Z},r\geqslant 2$, form a morphism ${\phi}_{\lambda}:E_G\rightarrow E_{H_{\lambda}}$ between spectral sequences.
**Claim.** *For every $\lambda\in\Lambda$,*
1. ${\phi}_{\lambda}$ is compatible with the map $r_{H_{\lambda}}$;
2. $\phi_{\lambda,2}$ can be identified with the map $\psi_{\lambda}$.
If $\lambda\in\Lambda'$, then statements (i) and (ii) follow from Lemma \[lem. inclusion induces morphism of sp\]. If $\lambda\in\Lambda\setminus\Lambda'$, then (i) and (ii) follow directly from the construction of ${\phi}_{\lambda}$.
It follows from the claim that ${\phi}$ is compatible with $r_G$ and $\phi_2$ can be identified with $\psi$. For all $q\in\mathbb{Z}\setminus\{0\}$, Corollary \[simplify\] implies that $\psi$ maps $\operatorname{H}^{\ast}({\overline{G}};\operatorname{H}^q({\ll\mathcal{N}\rr};A))$ isomorphically onto ${\prod_{\lambda\in\Lambda}}\operatorname{H}^{\ast}({\overline{H}_{\lambda}};\operatorname{H}^q(N_{\lambda};A))$ and thus $\phi^{\ast,q}_2$ is an isomorphism for all $q\in\mathbb{Z}\setminus\{0\}$.
To finish this paper, it remains to prove Proposition \[coind\].
Ext-groups of coset rings {#sec. ext of coset rings}
-------------------------
Our approach to Proposition \[coind\] goes as follows. Under the assumptions of Proposition \[coind\], we show that for every $q\geqslant 1$, there is a sequence of $\mathbb{Z}{\overline{G}}$-module isomorphisms $$\label{eq. sequence of isomorphisms}
\operatorname{H}^q({\ll\mathcal{N}\rr};A) \cong {\prod_{\lambda\in\Lambda}}\operatorname{Ext}^q_{{\ll\mathcal{N}\rr}}(\mathbb{Z}[G/H_{\lambda}],A) \cong {\prod_{\lambda\in\Lambda}}{\operatorname{CoInd}^{\overline{G}}_{\overline{H}_{\lambda}}}\operatorname{H}^q(N_{\lambda};A)$$ whose composition is the map $\eta$ claimed by Proposition \[coind\].
Let us start with the second isomorphism of . It suffices to prove $$\label{eq. component-wise isomorphism of ext and coind}
\operatorname{Ext}^{\ast}_{{\ll\mathcal{N}\rr}}(\mathbb{Z}[G/H_{\lambda}],A) \cong {\operatorname{CoInd}^{\overline{G}}_{\overline{H}_{\lambda}}}\operatorname{H}^{\ast}(N_{\lambda};A)$$ as $\mathbb{Z}{\overline{G}}$-modules.
Fix $\lambda$. Let $A\rightarrow I$ be an injective resolution over $\mathbb{Z}{\ll\mathcal{N}\rr}$. Then $\operatorname{Ext}^{\ast}_{{\ll\mathcal{N}\rr}}(\mathbb{Z}[G/H_{\lambda}],A)$ is the cohomology group of the cochain complex $$\label{eq. cochain for coset}
\operatorname{Hom}_{{\ll\mathcal{N}\rr}}(\mathbb{Z}[G/H_{\lambda}],I).$$
As $N_{\lambda}\leqslant {\ll\mathcal{N}\rr}$, every injective $\mathbb{Z}{\ll\mathcal{N}\rr}$-module is automatically an injective $\mathbb{Z}N_{\lambda}$-module and thus $A\rightarrow I$ is also an injective resolution over $\mathbb{Z}N_{\lambda}$. It is easy to see that ${\operatorname{CoInd}^{\overline{G}}_{\overline{H}_{\lambda}}}\operatorname{H}^{\ast}(N_{\lambda};A)$ is naturally isomorphic to the cohomology group of the cochain complex $$\label{eq. cochain for coind}
{\operatorname{CoInd}^{\overline{G}}_{\overline{H}_{\lambda}}}\operatorname{Hom}_{N_{\lambda}}(\mathbb{Z},I),$$ whose differential is induced by the differential of $I$. Indeed, given a function $f\in{\operatorname{CoInd}^{\overline{G}}_{\overline{H}_{\lambda}}}\operatorname{H}^{\ast}(N_{\lambda};A)$, lift $f$ to a function $\widetilde{f}\in{\operatorname{CoInd}^{\overline{G}}_{\overline{H}_{\lambda}}}\operatorname{Hom}_{N_{\lambda}}(\mathbb{Z},I)$. It is easy to check that $\widetilde{f}$ represents an element $[\widetilde{f}]$ of the cohomology group of . The correspondence $f\leftrightarrow [\widetilde{f}]$ then provides the desired isomorphism.
Now we construct a chain map from to . Suppose that a function $f\in\operatorname{Hom}_{{\ll\mathcal{N}\rr}}(\mathbb{Z}[G/H_{\lambda}],I)$ is given. Define a function $\overline{f}\in{\operatorname{CoInd}^{\overline{G}}_{\overline{H}_{\lambda}}}\operatorname{Hom}_{N_{\lambda}}(\mathbb{Z},I)$ by the following rule. Let $T_{\lambda}$ be a left transversal in $LT(H_{\lambda}{\ll\mathcal{N}\rr},G)$. Denote the image of elements $t\in T_{\lambda}$ under the quotient map $G\rightarrow {\overline{G}}$ by $\overline{t}$ and denote the image of $\overline{t}^{-1}$ under $\overline{f}$ by $\overline{f}_{\overline{t}^{-1}}$. Then $\overline{f}_{\overline{t}^{-1}}$ is a function from $\mathbb{Z}$ to $I$, and we demand that $\overline{f}_{\overline{t}^{-1}}(1)=t^{-1}f(tH_{\lambda})$. This uniquely determines the function $\overline{f}$.
Let $$\phi_{\lambda}:\operatorname{Hom}_{{\ll\mathcal{N}\rr}}(\mathbb{Z}[G/H_{\lambda}],I)\rightarrow {\operatorname{CoInd}^{\overline{G}}_{\overline{H}_{\lambda}}}\operatorname{Hom}_{N_{\lambda}}(\mathbb{Z},I)$$ be the function sending every $f\in\operatorname{Hom}_{{\ll\mathcal{N}\rr}}(\mathbb{Z}[G/H_{\lambda}],I)$ to the corresponding $\overline{f}$. Direct computation shows that $\phi_{\lambda}$ is a chain map and an isomorphism of abelian groups.
Recall that there are ${\overline{G}}$-actions on and defined at the beginning of Section \[sec. preliminaries\]. Direct computation also shows that $\phi_{\lambda}$ is ${\overline{G}}$-equivariant. It follows that $\phi_{\lambda}$ is also a $\mathbb{Z}{\overline{G}}$-module isomorphism and thus is a chain isomorphism. We conclude with:
\[lem. component-wise isomorphism of ext and coind\] The map $\phi_{\lambda}$ defined above induces a $\mathbb{Z}{\overline{G}}$-module isomorphism .
Proof of Proposition \[coind\] {#sec. proof of prop coind}
------------------------------
Consider the ring ${\bigoplus_{\lambda\in\Lambda}}\mathbb{Z}[G/H_{\lambda}]$. There is an augmentation homomorphism $$\epsilon:{\bigoplus_{\lambda\in\Lambda}}\mathbb{Z}[G/H_{\lambda}]\rightarrow\mathbb{Z},~\epsilon(gH_{\lambda})=1\text{ for all }g\in G\text{ and }\lambda\in\Lambda.$$
$\epsilon$ induces a $\mathbb{Z}{\overline{G}}$-module homomorphism of Ext-groups $$\epsilon^{\ast}:\operatorname{H}^{\ast}({\ll\mathcal{N}\rr};A)\rightarrow \operatorname{Ext}^{\ast}_{{\ll\mathcal{N}\rr}}\left({\bigoplus_{\lambda\in\Lambda}}\mathbb{Z}[G/H_{\lambda}],A\right).$$
As $\operatorname{Ext}^{\ast}_{{\ll\mathcal{N}\rr}}\left({\bigoplus_{\lambda\in\Lambda}}\mathbb{Z}[G/H_{\lambda}],A\right)$ is naturally isomorphic to ${\prod_{\lambda\in\Lambda}}\operatorname{Ext}^{\ast}_{{\ll\mathcal{N}\rr}}\left(\mathbb{Z}[G/H_{\lambda}],A\right)$, $\epsilon^{\ast}$ can be regarded as a map from $\operatorname{Hom}^{\ast}({\ll\mathcal{N}\rr};A)$ to ${\prod_{\lambda\in\Lambda}}\operatorname{Ext}^{\ast}_{{\ll\mathcal{N}\rr}}\left(\mathbb{Z}[G/H_{\lambda}],A\right)$.
For every $\lambda\in\Lambda$, the map $\phi_{\lambda}$ constructed in Section \[sec. ext of coset rings\] induces an isomorphism (still denoted by) $$\phi_{\lambda}:\operatorname{Ext}^{\ast}_{{\ll\mathcal{N}\rr}}\left(\mathbb{Z}[G/H_{\lambda}],A\right)\rightarrow{\operatorname{CoInd}^{\overline{G}}_{\overline{H}_{\lambda}}}\operatorname{H}^{\ast}(N_{\lambda};A).$$ Let $$\phi:{\prod_{\lambda\in\Lambda}}\operatorname{Ext}^{\ast}_{{\ll\mathcal{N}\rr}}\left(\mathbb{Z}[G/H_{\lambda}],A\right)\rightarrow{\prod_{\lambda\in\Lambda}}{\operatorname{CoInd}^{\overline{G}}_{\overline{H}_{\lambda}}}\operatorname{H}^{\ast}(N_{\lambda};A)$$ be the map on products induced by the maps $\phi_{\lambda}$.
Let $\eta=\phi\circ\epsilon^{\ast}$. Then diagram tracing shows $r_{N_{\mu}}=\pi_{\mu}\circ p_{\mu}\circ \eta$ for all $\mu\in\Lambda$. To finish the proof of Proposition \[coind\], it suffices to show
\[lem. normal closure to ext of coset rings\] $\phi$ maps $\operatorname{H}^q({\ll\mathcal{N}\rr};A)$ isomorphically onto $\operatorname{Ext}^q_{{\ll\mathcal{N}\rr}}\left({\bigoplus_{\lambda\in\Lambda}}\mathbb{Z}[G/H_{\lambda}],A\right)$ for all $q\geqslant 1$.
We approach Lemma \[lem. normal closure to ext of coset rings\] via specific free resolutions. We construct a free resolution $P\rightarrow{\bigoplus_{\lambda\in\Lambda}}\mathbb{Z}[G/H_{\lambda}]$ over $\mathbb{Z}{\ll\mathcal{N}\rr}$ and show that $P$ is almost a free resolution of $\mathbb{Z}$ over $\mathbb{Z}{\ll\mathcal{N}\rr}$. Formally, we construct a free resolution $R\rightarrow\mathbb{Z}$ over $\mathbb{Z}{\ll\mathcal{N}\rr}$ and prove that $P$ and $R$ coincide at dimension $1$ and beyond, which will imply Lemma \[lem. normal closure to ext of coset rings\].
As $(G,\{H_{\lambda}\}_{\lambda\in\Lambda},\{N_{\lambda}\}_{\lambda\in\Lambda})$ is a Cohen-Lyndon triple, there exist transversals $T_{\lambda}\in LT(H_{\lambda}{\ll\mathcal{N}\rr},G)$ such that $$\label{eq. factorization of normal closure}
{\ll\mathcal{N}\rr}=\prod^{\ast}_{\lambda\in\Lambda,t\in T_{\lambda}}tN_{\lambda}t^{-1}.$$
The resolution $P\rightarrow {\bigoplus_{\lambda\in\Lambda}}\mathbb{Z}[G/H_{\lambda}]$ results from picking one free resolution $P_{\lambda}\rightarrow\mathbb{Z}[G/H_{\lambda}]$ for each $\lambda\in\Lambda$ and then taking direct sum of these resolutions. Formally, let $F\rightarrow\mathbb{Z}$ be the bar resolution of $G$. For $\lambda\in\Lambda$, consider the abelian group tensor product $F\otimes\mathbb{Z}[G/H_{\lambda}]$. The actions of $G$ on $F$ and $F\otimes\mathbb{Z}[G/H_{\lambda}]$ are diagonal: $$g\cdot (g_0,...,g_n)=(gg_0,...,gg_n),~~~~g\cdot \left((g_0,...,g_n)\otimes g_{n+1}H_{\lambda}\right)=(gg_0,...,gg_n)\otimes gg_{n+1}H_{\lambda}.$$
Let $P_{\lambda}$ be the $\mathbb{Z}{\ll\mathcal{N}\rr}$-submodule of $F\otimes\mathbb{Z}[G/H_{\lambda}]$ generated by elements of the form $p\otimes tH_{\lambda}$, where $t$ ranges over elements of $T_{\lambda}$ and $p$ ranges over all tuples of elements of the set $tN_{\lambda}$.
As a $\mathbb{Z}G$-module, $F$ is freely generated by the set $$\{(1,g_1,...,g_q)\mid q\geqslant 0, g_1,...,g_q\in G\}.$$ Therefore, $F\otimes \mathbb{Z}[G/H_{\lambda}]$, as a $\mathbb{Z}G$-module, is freely generated by the set $$\{(1,g_1,...,g_q)\otimes gH_{\lambda}\mid q\geqslant 0, g_1,...,g_q\in G, gH_{\lambda} \text{ ranges over left cosets of }H_{\lambda}\}.$$ Thus,
\[lem. free basis of Q lambda\] $P_{\lambda}$, as a $\mathbb{Z}{\ll\mathcal{N}\rr}$-module, is freely generated by the set $$\{(t,tn_1,...,tn_q)\otimes tH_{\lambda} \mid q\geqslant 0, n_1,...,n_q\in N_{\lambda}, t\in T_{\lambda}\}.$$
The boundary operator of $P\otimes\mathbb{Z}[G/H_{\lambda}]$ restricts to a boundary operator on $P_{\lambda}$, which turns $P_{\lambda}$ into a resolution. Moreover, the augmentation map of $F\otimes \mathbb{Z}[G/H_{\lambda}]$ maps $P_{\lambda}$ onto $\mathbb{Z}[{\ll\mathcal{N}\rr}T_{\lambda}H_{\lambda}/H_{\lambda}]$. As $T_{\lambda}\in LT(H_{\lambda}{\ll\mathcal{N}\rr},G)$, we have ${\ll\mathcal{N}\rr}T_{\lambda}H_{\lambda}=G$. It follows that $P_{\lambda}\rightarrow\mathbb{Z}[G/H_{\lambda}]$ is a free resolution over $\mathbb{Z}{\ll\mathcal{N}\rr}$. Let $$P={\bigoplus_{\lambda\in\Lambda}}P_{\lambda}.$$ Then $P\rightarrow{\bigoplus_{\lambda\in\Lambda}}\mathbb{Z}[G/H_{\lambda}]$ is a free resolution over $\mathbb{Z}{\ll\mathcal{N}\rr}$.
The resolution $R\rightarrow\mathbb{Z}$ results from modifying $P\rightarrow{\bigoplus_{\lambda\in\Lambda}}\mathbb{Z}[G/H_{\lambda}]$ at dimension $0$. Let $R=\bigoplus_{q\geqslant 0}R_q$ be the graded abelian group such that
1. $R_0=\mathbb{Z}{\ll\mathcal{N}\rr}$;
2. $R_q=P_q$ for all $q\geqslant 1$, where $P_q$ is the component of $P$ at dimension $q$.
For $q\geqslant 1$, we demand that the boundary map $R_{q+1}\rightarrow R_q$ is the same as the boundary map $P_{q+1}\rightarrow P_q$. To define the boundary map $\partial^R_1: R_1\rightarrow R_0$, it suffices to specify the values of $\partial^R_1$ on the set $$\{(t,tn)\otimes tH_{\lambda} \mid \lambda\in\Lambda, n\in N_{\lambda}, t\in T_{\lambda}\}$$ by Lemma \[lem. free basis of Q lambda\], and we require that $$\partial^R_1\left((t,tn)\otimes tH_{\lambda}\right)=1-tnt^{-1}\text{ for all }\lambda\in\Lambda, n\in N_{\lambda}, t\in T_{\lambda}.$$ Let $\epsilon^{{\ll\mathcal{N}\rr}}:R_0=\mathbb{Z}{\ll\mathcal{N}\rr}\rightarrow\mathbb{Z}$ be the augmentation map of $\mathbb{Z}{\ll\mathcal{N}\rr}$. Our goal is the following.
\[lem. key\] $R\xrightarrow{\epsilon^{{\ll\mathcal{N}\rr}}}\mathbb{Z}$ is a free resolution over $\mathbb{Z}{\ll\mathcal{N}\rr}$. Moreover, the augmentation map $\epsilon:{\bigoplus_{\lambda\in\Lambda}}\mathbb{Z}[G/H_{\lambda}]\rightarrow \mathbb{Z}$ induces a chain map
$$\label{eq. chain map epsilon}
\begin{tikzcd}
\cdot\cdot\cdot\arrow[r] & P_2 \arrow[r] \arrow[d,"\epsilon_2"] & P_1 \arrow[r] \arrow[d,"\epsilon_1"] & P_0 \arrow[r] \arrow[d,"\epsilon_0"] & {\bigoplus_{\lambda\in\Lambda}}\mathbb{Z}[G/H_{\lambda}] \arrow[r] \arrow[d,"\epsilon"] & 0\\
\cdot\cdot\cdot\arrow[r] & R_2 \arrow[r] & R_1 \arrow[r] & R_0 \arrow[r] & \mathbb{Z} \arrow[r] & 0
\end{tikzcd}$$
such that for all $q\geqslant1$, $\epsilon_q$ is the identity map.
Lemma \[lem. key\] clearly implies that $\epsilon^{\ast}$ maps $\operatorname{H}^q({\ll\mathcal{N}\rr};A)$ isomorphically onto $\operatorname{Ext}^q_{{\ll\mathcal{N}\rr}}\left({\bigoplus_{\lambda\in\Lambda}}\mathbb{Z}[G/H_{\lambda}],A\right)$ for all $q\geqslant 1$ and thus finishes the proof of Proposition \[coind\].
To prove Lemma \[lem. key\], for $q\geqslant 1$, let $\epsilon_q$ be the identity map. To define $\epsilon_0$, it suffices to specify the values of $\epsilon_0$ on the set $$\{t\otimes tH_{\lambda}\mid \lambda\in\Lambda, t\in T_{\lambda}\}$$ by Lemma \[lem. free basis of Q lambda\], and we demand that $$\epsilon_0\left(t\otimes tH_{\lambda}\right)=1 \text{ for all }\lambda\in\Lambda \text{ and } t\in T_{\lambda}.$$ Clearly, $(\epsilon_q)_{q\geqslant 0}$ is a chain map from $P$ to $R$.
Note that the $\mathbb{Z}{\ll\mathcal{N}\rr}$-module $\operatorname{im}(\partial^R_1)$ is generated by the set $$\{1-tnt^{-1}\mid \lambda\in\Lambda,n\in N_{\lambda},t\in T_{\lambda}\}$$ and hence is the augmentation ideal of $R_0=\mathbb{Z}{\ll\mathcal{N}\rr}$. To finish the prove of Lemma \[lem. key\], it suffices to show the following.
\[kernel\] $\ker(\partial^R_1)=\ker(\partial^P_1)$, where $\partial^P_1$ is the boundary map from $P_1$ to $P_0$.
Lemma \[kernel\] implies that the second row of is exact and thus $R\rightarrow \mathbb{Z}$ is a free resolution.
As $\partial^R_1=\epsilon_0\circ\partial^P_1$, we have $\ker(\partial^P_1)\subset\ker(\partial^R_1)$. In order to prove the converse containment, we factorize elements of ${\ll\mathcal{N}\rr}$. By equation , every element $m\neq 1$ of ${\ll\mathcal{N}\rr}$ can be uniquely factorized as $$\label{eq factorization}
m=\prod_{i=1}^k t_i n_i t^{-1}_i,$$ where
1. $t_i\in T_{\lambda_i},n_i\in N_{\lambda_i}\setminus\{1\}$, and $\lambda_1,...,\lambda_k\in\Lambda$;
2. for $i=1,...,k-1$, either $\lambda_i\neq\lambda_{i+1}$ or $t_i\neq t_{i+1}$.
Equation is called the *factorization* of $m$. *The number of factors of* $m$, denoted as $\omega(m)$, is the number $k$ in . The *factorization* of $1\in{\ll\mathcal{N}\rr}$ is just $1=1$ and we let $\omega(1)=0$.
For every $\lambda\in\Lambda$ and every $t\in T_{\lambda}$, let $X_{\lambda,t}$ be the subset of ${\ll\mathcal{N}\rr}$ consisting of elements whose factorizations do not end with a factor from $tN_{\lambda}t^{-1}$.
We construct a basis for the abelian group $P_1$. For $\lambda\in\Lambda$, let ${E}_{\lambda}$ be a set of pairs of elements of $N_{\lambda}$ such that
1. for every $n\in N_{\lambda}$, the pair $(n,n)$ belongs to ${E}_{\lambda}$;
2. if $n_1,n_2$ are distinct elements of $N_{\lambda}$, then ${E}_{\lambda}$ contains exactly one of $(n_1,n_2)$ and $(n_2,n_1)$.
Let $$S=\{(xtn_1,xtn_2)\otimes xtH_{\lambda}\mid \lambda\in\Lambda,t\in T _{\lambda},x\in X_{\lambda,t},(n_1,n_2)\in {E}_{\lambda}\}\subset R_1.$$ Then $S$ is a basis for the abelian group $P_1=R_1$.
For every $$s=(xtn_1,xtn_2)\otimes xtH_{\lambda}\in S,$$ let $$\Omega(s)=\max\{\omega(xtn_1t^{-1}),\omega(xtn_2t^{-1})\}.$$
Every $r\in R_1$ can be uniquely written in the form $$r=\sum_{s\in S} C_{r,s}s$$ where $C_{r,s}\in\mathbb{Z}$. The above sum makes sense as there are only finitely many non-zero terms.
We call the number $C_{r,s}$ in the above equation the *coefficient of* $r$ with respect to $s$. Let $\operatorname{rank}:R_1\rightarrow \mathbb{N}$ be the function summing the absoute values of the coefficients: $$\operatorname{rank}(r)=\sum_{s\in S}|C_{r,s}|.$$
Suppose $r\in \ker(\partial^R_1)$. We prove $r\in \ker(\partial^P_1)$ by an induction on $\operatorname{rank}(r)$. The base case $\operatorname{rank}(r)=0$ implies $r=0$ and thus $r\in \ker(\partial^P_1)$. So let us suppose that $\operatorname{rank}(r)>0$ and that, for all $r'\in \ker(\partial^R_1)$ with $\operatorname{rank}(r')<\operatorname{rank}(r)$, we have $r'\in \ker(\partial^P_1)$.
Let $$s_0=(x_0t_0n_1,x_0t_0n_2)\otimes x_0t_0H_{\lambda_0}\in S$$ such that $C_{r,s_0}\neq 0$ and
1. if $s\in S$ satisfying $C_{r,s}\neq 0$, then $\Omega(s_0)\geqslant \Omega(s)$.
If $n_1=n_2$, consider the element $r'\in R_1$ such that $C_{r',s}=C_{r,s}$ for $s\in S\setminus\{s_0\}$ and $C_{r',s_0}=0$. Direct computation shows $$\operatorname{rank}(r')<\operatorname{rank}(r),~~~~\partial^P_1(r-r')=\partial^R_1(r-r')=0.$$ Thus, $\partial^R_1(r')=0$ and the induction hypothesis implies $\partial^P_1(r')=0$. It follows that $$\partial^P_1(r)=\partial^P_1(r-r')+\partial^P_1(r')=0.$$ Therefore, $r\in \ker(\partial^P_1)$.
Thus, without loss of generality, let us assume $n_1\neq n_2$. It follows that at least one of $n_1$ and $n_2$ is not the identity of $G$. Without loss of generality, we may further assume $n_1\neq 1$ (the case $n_2\neq 1$ is similar), in which case $$\Omega(s_0)=\omega(x_0t_0 n_1t^{-1}_0).$$
Let us also assume $C_{r,s}>0$ (otherwise, consider $-r$). Note that $$\label{boundary}
\partial^R_1(r)=\sum_{s\in S}C_{r,s}\partial^R_1(s).$$
On the right-hand side of , $$C_{r,s_0}\partial^R_1(s_0)=C_{r,s_0}(x_0t_0n_2t^{-1}_0-x_0t_0n_1t^{-1}_0).$$ Thus, $s_0$ contributes a negative number of $x_0t_0n_1t^{-1}_0$ to $\partial^R_1(r)$. As $\partial^R_1(r)=0$, there exists some $s_1\in S$ which contributes a positive number of $x_0t_0n_1t^{-1}_0$ to $\partial^R_1(r)$. In other words, one of the following cases happens
1. $s_1=(x_1t_1n_3,x_1t_1n_4)\otimes x_1t_1H_{\lambda_1}$ with $C_{r,s_1}<0$, $n_3\neq n_4$, and $x_1t_1 n_3t^{-1}_1=x_0t_0n_1t^{-1}_0$.
2. $s_1=(x_1t_1n_3,x_1t_1n_4)\otimes x_1t_1H_{\lambda_1}$ with $C_{r,s_1}>0$, $n_3\neq n_4$, and $x_1t_1n_4t^{-1}_1=x_0t_0n_1t^{-1}_0$.
Let us suppose that Case (1) happens (Case (2) can be treated in the same manner). Note that $n_3\neq 1$ in this case. Indeed, if $n_3=1$, then $n_4\neq 1$ since $n_4\neq n_3$. It follows that $$\begin{aligned}
\Omega(s_1)&>\omega(x_1t_1n_3t^{-1}_1)&&\text{as }n_3=1,n_4\neq 1,x_1\in X_{\lambda_1,t_1}&\\
=& \omega(x_0t_0n_1t^{-1}_0)&&\text{as }x_1t_1 n_3t^{-1}_1=x_0t_0n_1t^{-1}_0&\\
=& \Omega(s_0),\end{aligned}$$ which contradicts condition (max $\Omega$). Thus, $n_3\neq 1$, which, together with the assumption $x_1\in X_{\lambda_1,t_1}$, implies that the factorization of $x_1t_1 n_3t^{-1}_1$ ends with $t_1 n_3t^{-1}_1$.
As $x_0\in X_{\lambda_0,t_0}$ and $n_1\neq 1$, the factorization of $x_0t_0n_1t^{-1}_0$ ends with $t_0n_1t^{-1}_0$. Since $x_1t_1 n_3t^{-1}_1=x_0t_0n_1t^{-1}_0$, we have $$t_0n_1t^{-1}_0=t_1 n_3 t^{-1}_1\in t_0N_{\lambda_0}t^{-1}_0\cap t_1N_{\lambda_1}t^{-1}_1.$$ As $n_1\neq 1$, we have $$\label{non-trivial intersection}
t_0N_{\lambda_0}t^{-1}_0\cap t_1N_{\lambda_1}t^{-1}_1\neq\{1\}.$$
Equations and imply $\lambda_1=\lambda_0,t_1=t_0$, which, together with $x_1t_1 n_3t^{-1}_1=x_0t_0n_1t^{-1}_0$, imply $n_1=n_0,x_1=x_0$ and thus $$s_1=(x_0t_0n_1,x_0t_0n_4)\otimes x_0t_0H_{\lambda_0}.$$
Exactly one of $(n_2,n_4)$ and $(n_4,n_2)$ is in ${E}_{\lambda_0}$. Without loss of generality, we assume that $(n_2,n_4)\in {E}_{\lambda_0}$ (the other case is similar). Let $$s_2=(x_0t_0n_2,x_0t_0n_4)\otimes x_0t_0H_{\lambda_0},$$ and let $r'\in R_1$ such that $C_{r',s}=C_{r,s}$ for $s\in S\setminus\{s_0,s_1,s_2\}$ and $$C_{r',s_0}=C_{r,s_0}-1,~~~~C_{r',s_1}=C_{r,s_1}+1,~~~~C_{r',s_2}=C_{r,s_2}-1.$$ As $C_{r,s_0}>0, C_{r,s_1}<0$, and $\partial^R_1(r)=0$, direct computation shows $$\operatorname{rank}(r')<\operatorname{rank}(r),~~~~\partial^P_1(r-r')=\partial^R_1(r-r')=0.$$ Thus, $\partial^R_1(r')=0$ and the induction hypothesis implies $\partial^P_1(r')=0$. It follows that $$\partial^P_1(r)=\partial^P_1(r-r')+\partial^P_1(r')=0,$$ that is, $r\in \ker(\partial^P_1)$.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We consider a Dirac one-electron atom placed in a weak, static, uniform magnetic field. We show that, to the first order in the strength $B$ of the perturbing field, the only electric multipole moment induced by the field in the ground state of the atom is the quadrupole one. Using the Sturmian expansion of the generalized Dirac–Coulomb Green function \[R. Szmytkowski, J. Phys. B 30 (1997) 825; erratum 30 (1997) 2747\], we derive a closed-form expression for the induced electric quadrupole moment. The result contains the generalized hypergeometric function $_{3}F_{2}$ of the unit argument. Earlier calculations by other authors, based on the non-relativistic model of the atom, predicted in the low-field region the quadratic dependence of the induced electric quadrupole moment on $B$. 3ex **Key words:** electric quadrupole moment; magnetic field; Zeeman effect; Dirac one-electron atom; Dirac–Coulomb Green function; Sturmian functions 1ex **PACS:** 31.15.ap, 32.10.Dk, 32.60.+i'
author:
- |
Rados[ł]{}aw Szmytkowski[^1] and Patrycja Stefa[ń]{}ska\
Atomic Physics Division, Department of Atomic Physics and Luminescence,\
Faculty of Applied Physics and Mathematics, Gda[ń]{}sk University of Technology,\
Narutowicza 11/12, 80–233 Gda[ń]{}sk, Poland
title: |
Magnetic field-induced electric quadrupole moment\
in the ground state of the relativistic hydrogen-like atom:\
Application of the Sturmian expansion of the generalized Dirac–Coulomb Green function
---
Introduction {#I}
============
In the mid 1950’s, Coulson and Stephen [@Coul56] pointed out that the uniform magnetic field should induce the electric quadrupole moment (EQM) in the hydrogen-like atom. They used the perturbation theory and found that in the case of the non-relativistic atom, for field strengths $B$ corresponding to the complete Paschen–Back effect, the leading term in the expansion of the induced EQM in powers of the field strength is *quadratic* in $B$. For the atomic ground state, their result, after being translated into the form conforming both to the definition of the EQM (cf. Sec.\[III\] below) and to the notation used in the present work, is $$\mathcal{Q}_{20}\simeq\mathcal{Q}_{20}^{(2)}
=-\frac{5}{16}\frac{\alpha^{4}ea_{0}^{2}}{Z^{6}}
\frac{B^{2}}{B_{0}^{2}},
\label{1.1}$$ where $Ze$ is the nuclear charge, $a_{0}$ is the Bohr radius, $\alpha$ is the Sommerfeld fine-structure constant, whereas $$B_{0}=\frac{\mu_{0}}{4\pi}\frac{\mu_{\mathrm{B}}}{a_{0}^{3}}
=\frac{\alpha^{2}\hbar}{2ea_{0}^{2}}\simeq6.26\,\mathrm{T}
\label{1.2}$$ is the atomic unit of the magnetic induction ($\mu_{0}$ is the vacuum permeability and $\mu_{\mathrm{B}}$ is the Bohr magneton). Later, Turbiner [@Turb87 Sec. 5] arrived at the same expression for $\mathcal{Q}_{20}^{(2)}$ for the Schr[ö]{}dinger one-electron atom with $Z=1$ and found also explicitly the next non-vanishing term (being $\mathcal{Q}_{20}^{(4)}\sim B^{4}$) in the expansion of $\mathcal{Q}_{20}$ in powers of $B$. Moreover, using the variational technique, he determined the function $\mathcal{Q}_{20}(B)$ for magnetic fields ranging from vanishing to superstrong. A decade ago, Potekhin and Turbiner [@Pote01] calculated $\mathcal{Q}_{20}(B)$, over a still broader range of $B$, by two different methods, one being the variational approach with a more sophisticated trial function than the one used in Ref. [@Turb87], the other being based on the expansion of the perturbed electronic wave function in the Landau orbitals. In the low-field limit, the results of Ref.[@Pote01] reproduced the quadratic dependence of the induced EQM on $B$ predicted in Refs. [@Coul56; @Turb87].
In all the aforementioned works, the atom has been described non-relativistically. In the present paper, we show that if the relativity is taken into account and the atomic model adopted is the one based on the Dirac equation for the electron, it appears that the weak, static, uniform magnetic field induces in the ground state of the atom the EQM which, to the lowest order, is *linear* in the perturbing field strength $B$; no other electric multipole moments are induced in the system to the first order in $B$. Exploiting the Sturmian expansion of the generalized (or reduced) Dirac–Coulomb Green function, found by one of us in Ref.[@Szmy97] and subsequently successfully used in analytical calculations of various electromagnetic properties of the Dirac one-electron atom [@Szmy02; @Szmy04; @Miel06; @Szmy11; @Stef11], in Sec.\[IV\] we derive a closed-form expression for the induced EQM, in terms of the generalized hypergeometric function $_{3}F_{2}$ with the unit argument.
Preliminaries {#II}
=============
It has been already stated in the introduction that the system we shall be concerned with in the present work is the Dirac one-electron atom. Its nucleus will be assumed to be point-like, infinitely heavy, spinless, and to carry the electric charge $Ze$. In the absence of external perturbations, the atomic ground-state energy level $$E^{(0)}=mc^{2}\gamma_{1},
\label{2.1}$$ where $$\gamma_{\kappa}=\sqrt{\kappa^{2}-(\alpha Z)^{2}},
\label{2.2}$$ is two-fold degenerate, the two associated Hamiltonian eigenfunctions, orthonormal in the sense of $$\int_{\mathbb{R}^{3}}\mathrm{d}^{3}\boldsymbol{r}\:
\Psi_{\mu}^{(0)\dag}(\boldsymbol{r})\Psi_{\mu'}^{(0)}(\boldsymbol{r})
=\delta_{\mu\mu'},
\label{2.3}$$ being $$\Psi_{\mu}^{(0)}(\boldsymbol{r})
=\frac{1}{r}
\left(
\begin{array}{c}
P^{(0)}(r)\Omega_{-1\mu}(\boldsymbol{n}_{r}) \\
\mathrm{i}Q^{(0)}(r)\Omega_{1\mu}(\boldsymbol{n}_{r}) \\
\end{array}
\right)
\qquad ({\textstyle\mu=\pm\frac{1}{2}}).
\label{2.4}$$ The radial functions appearing in Eq. (\[2.4\]), normalized according to $$\int_{0}^{\infty}\mathrm{d}r
\left\{[P^{(0)}(r)]^{2}+[Q^{(0)}(r)]^{2}\right\}=1,
\label{2.5}$$ are explicitly given by
$$P^{(0)}(r)=-\sqrt{\frac{Z}{a_{0}}\frac{1+\gamma_{1}}
{\Gamma(2\gamma_{1}+1)}}
\left(\frac{2Zr}{a_{0}}\right)^{\gamma_{1}}\mathrm{e}^{-Zr/a_{0}},
\label{2.6a}$$
$$Q^{(0)}(r)=\sqrt{\frac{Z}{a_{0}}\frac{1-\gamma_{1}}
{\Gamma(2\gamma_{1}+1)}}
\left(\frac{2Zr}{a_{0}}\right)^{\gamma_{1}}\mathrm{e}^{-Zr/a_{0}},
\label{2.6b}$$
\[2.6\]
whereas $\Omega_{\kappa\mu}(\boldsymbol{n}_{r})$, with $\boldsymbol{n}_{r}=\boldsymbol{r}/r$, are the orthonormal spherical spinors defined as in Ref. [@Szmy07].
In the presence of a weak, static, uniform magnetic field $\boldsymbol{B}=B\boldsymbol{n}_{z}$, the level $E^{(0)}$ splits into two. Their energies, to the first order in $\boldsymbol{B}$, are given by $$E_{\mu}\simeq E^{(0)}+E_{\mu}^{(1)}
\qquad ({\textstyle\mu=\pm\frac{1}{2}}),
\label{2.7}$$ with $$E_{\mu}^{(1)}={\mathop{\textrm{sgn}}}(\mu)\frac{2\gamma_{1}+1}{3}\mu_{\mathrm{B}}B.
\label{2.8}$$ The corresponding wave functions, to the same approximation order, are $$\Psi_{\mu}(\boldsymbol{r})\simeq\Psi_{\mu}^{(0)}(\boldsymbol{r})
+\Psi_{\mu}^{(1)}(\boldsymbol{r})
\qquad ({\textstyle\mu=\pm\frac{1}{2}}).
\label{2.9}$$ Here, the zeroth-order component $\Psi_{\mu}^{(0)}(\boldsymbol{r})$ is given by Eq. (\[2.4\]) (the space quantization axis is now chosen to be directed along $\boldsymbol{B}$). The correction $\Psi_{\mu}^{(1)}(\boldsymbol{r})$ solves the inhomogeneous differential equation $$\left[-\mathrm{i}c\hbar\boldsymbol{\alpha}\cdot\boldsymbol{\nabla}
+\beta mc^{2}-\frac{Ze^{2}}{(4\pi\epsilon_{0})r}-E^{(0)}\right]
\Psi_{\mu}^{(1)}(\boldsymbol{r})
=-\left[\frac{1}{2}ec\boldsymbol{\alpha}
\cdot(\boldsymbol{B}\times\boldsymbol{r})-E_{\mu}^{(1)}\right]
\Psi_{\mu}^{(0)}(\boldsymbol{r})
\label{2.10}$$ ($\boldsymbol{\alpha}$ and $\beta$ are the standard Dirac matrices), subject to the usual regularity conditions and the orthogonality constraint $$\int_{\mathbb{R}^{3}}\mathrm{d}^{3}\boldsymbol{r}\:
\Psi_{\mu}^{(0)\dag}(\boldsymbol{r})
\Psi_{\mu'}^{(1)}(\boldsymbol{r})=0
\qquad ({\textstyle\mu,\mu'=\pm\frac{1}{2}}).
\label{2.11}$$ The integral representation of $\Psi_{\mu}^{(1)}(\boldsymbol{r})$ is $$\Psi_{\mu}^{(1)}(\boldsymbol{r})=-\frac{1}{2}ec\boldsymbol{B}
\cdot\int_{\mathbb{R}^{3}}\mathrm{d}^{3}\boldsymbol{r}'\:
\bar{G}\mbox{}^{(0)}(\boldsymbol{r},\boldsymbol{r}')
(\boldsymbol{r}'\times\boldsymbol{\alpha})
\Psi_{\mu}^{(0)}(\boldsymbol{r}'),
\label{2.12}$$ where $\bar{G}\mbox{}^{(0)}(\boldsymbol{r},\boldsymbol{r}')$ is the generalized Dirac–Coulomb Green function associated with the ground-state energy level (\[2.1\]).
Analysis of electric multipole moments of the atom in the magnetic field {#III}
========================================================================
After these preparatory steps, we set the problem: which electric multipole moments, apart from the monopole one, characterize the electronic cloud of the atom in the perturbed state $\Psi_{\mu}(\boldsymbol{r})$? If $\rho_{\mu}(\boldsymbol{r})$ is the electronic charge density for that state, the spherical components of the $L$-th order electric multipole moment tensor are defined as $$\mathcal{Q}_{LM\mu}=\sqrt{\frac{4\pi}{2L+1}}
\int_{\mathbb{R}^{3}}\mathrm{d}^{3}\boldsymbol{r}\:
r^{L}Y_{LM}^{*}(\boldsymbol{n}_{r})\rho_{\mu}(\boldsymbol{r}),
\label{3.1}$$ where the asterisk denotes the complex conjugation and $Y_{LM}(\boldsymbol{n}_{r})$ is the normalized spherical harmonic defined according to the Condon–Shortley phase convention [@Vars75 chapter 5]. For the atom being in the state $\Psi_{\mu}(\boldsymbol{r})$, the density $\rho_{\mu}(\boldsymbol{r})$ is given by $$\rho_{\mu}(\boldsymbol{r})=\frac{-e\Psi_{\mu}^{\dag}(\boldsymbol{r})
\Psi_{\mu}(\boldsymbol{r})}{\int_{\mathbb{R}^{3}}
\mathrm{d}^{3}\boldsymbol{r}'\:\Psi_{\mu}^{\dag}(\boldsymbol{r}')
\Psi_{\mu}(\boldsymbol{r}')}.
\label{3.2}$$ Using Eqs. (\[2.9\]) and (\[2.11\]), to the first order in the perturbing field, one has $$\rho_{\mu}(\boldsymbol{r})\simeq\rho_{\mu}^{(0)}(\boldsymbol{r})
+\rho_{\mu}^{(1)}(\boldsymbol{r}),
\label{3.3}$$ with $$\rho_{\mu}^{(0)}(\boldsymbol{r})
=-e\Psi_{\mu}^{(0)\dag}(\boldsymbol{r})
\Psi_{\mu}^{(0)}(\boldsymbol{r})
\label{3.4}$$ and $$\rho_{\mu}^{(1)}(\boldsymbol{r})
=-e\left[\Psi_{\mu}^{(1)\dag}(\boldsymbol{r})
\Psi_{\mu}^{(0)}(\boldsymbol{r})+\Psi_{\mu}^{(0)\dag}(\boldsymbol{r})
\Psi_{\mu}^{(1)}(\boldsymbol{r})\right].
\label{3.5}$$ Accordingly, it follows that $$\mathcal{Q}_{LM\mu}\simeq\mathcal{Q}_{LM\mu}^{(0)}
+\mathcal{Q}_{LM\mu}^{(1)},
\label{3.6}$$ where $$\mathcal{Q}_{LM\mu}^{(0)}=-e\sqrt{\frac{4\pi}{2L+1}}
\int_{\mathbb{R}^{3}}\mathrm{d}^{3}\boldsymbol{r}\:
\Psi_{\mu}^{(0)\dag}(\boldsymbol{r})
r^{L}Y_{LM}^{*}(\boldsymbol{n}_{r})\Psi_{\mu}^{(0)}(\boldsymbol{r})
\label{3.7}$$ is the multipole moment for the unperturbed atom, whereas $$\mathcal{Q}_{LM\mu}^{(1)}=\widetilde{\mathcal{Q}}_{LM\mu}^{(1)}
+(-)^{M}\widetilde{\mathcal{Q}}_{L,-M\mu}^{(1)*},
\label{3.8}$$ with $$\widetilde{\mathcal{Q}}_{LM\mu}^{(1)}=-e\sqrt{\frac{4\pi}{2L+1}}
\int_{\mathbb{R}^{3}}\mathrm{d}^{3}\boldsymbol{r}\:
\Psi_{\mu}^{(0)\dag}(\boldsymbol{r})
r^{L}Y_{LM}^{*}(\boldsymbol{n}_{r})\Psi_{\mu}^{(1)}(\boldsymbol{r}),
\label{3.9}$$ is the first-order correction induced by the perturbing magnetic field. To arrive at Eq. (\[3.8\]), we have exploited the well-known identity $$Y_{LM}(\boldsymbol{n}_{r})=(-)^{M}Y_{L,-M}^{*}(\boldsymbol{n}_{r}).
\label{3.10}$$ If the representation (\[2.12\]) of $\Psi_{\mu}^{(1)}(\boldsymbol{r})$ is plugged into Eq. (\[3.9\]), this gives $\widetilde{\mathcal{Q}}_{LM\mu}^{(1)}$ in the form of the double integral over $\mathbb{R}^{3}$: $$\widetilde{\mathcal{Q}}_{LM\mu}^{(1)}
=\frac{1}{2}\sqrt{\frac{4\pi}{2L+1}}\,
e^{2}cB\int_{\mathbb{R}^{3}}\mathrm{d}^{3}\boldsymbol{r}
\int_{\mathbb{R}^{3}}\mathrm{d}^{3}\boldsymbol{r}'\:
\Psi_{\mu}^{(0)\dag}(\boldsymbol{r})
r^{L}Y_{LM}^{*}(\boldsymbol{n}_{r})
\bar{G}\mbox{}^{(0)}(\boldsymbol{r},\boldsymbol{r}')
\boldsymbol{n}_{z}\cdot(\boldsymbol{r}'\times\boldsymbol{\alpha})
\Psi_{\mu}^{(0)}(\boldsymbol{r}').
\label{3.11}$$ Using Eq. (\[2.4\]) and the explicit representations of the spherical spinors $\Omega_{\mp1\mu}(\boldsymbol{n}_{r})$, it is easy to show that the electronic charge density in the unperturbed atom, $\rho_{\mu}^{(0)}(\boldsymbol{r})$, is spherically symmetric and the integral in Eq. (\[3.7\]) differs from zero only if $L=0$ and $M=0$. Hence, for the ground state of the unperturbed atom it holds that $$\mathcal{Q}_{LM\mu}^{(0)}=-e\delta_{L0}\delta_{M0},
\label{3.12}$$ i.e., all permanent electric multipole moments of the electronic cloud other than the monopole one do vanish.
Next, we proceed to the analysis of the first-order induced multipole moments $\mathcal{Q}_{LM\mu}^{(1)}$. To this end, we have to invoke the multipole representation of the generalized Green function, which is $$\begin{aligned}
\bar{G}\mbox{}^{(0)}(\boldsymbol{r},\boldsymbol{r}')
&=& \frac{4\pi\epsilon_{0}}{e^{2}}
\sum_{\substack{\kappa=-\infty \\ (\kappa\neq0)}}^{\infty}
\sum_{m=-|\kappa|+1/2}^{|\kappa|-1/2}\frac{1}{rr'}
\nonumber \\
&& \hspace*{-5em}
\times\left(
\begin{array}{cc}
\bar{g}\mbox{}^{(0)}_{\kappa,(++)}(r,r')
\Omega_{\kappa m}(\boldsymbol{n}_{r})
\Omega_{\kappa m}^{\dag}(\boldsymbol{n}_{r}^{\prime}) &
-\mathrm{i}\bar{g}\mbox{}^{(0)}_{\kappa,(+-)}(r,r')
\Omega_{\kappa m}(\boldsymbol{n}_{r})
\Omega_{-\kappa m}^{\dag}(\boldsymbol{n}_{r}^{\prime}) \\
\mathrm{i}\bar{g}\mbox{}^{(0)}_{\kappa,(-+)}(r,r')
\Omega_{-\kappa m}(\boldsymbol{n}_{r})
\Omega_{\kappa m}^{\dag}(\boldsymbol{n}_{r}^{\prime}) &
\bar{g}\mbox{}^{(0)}_{\kappa,(--)}(r,r')
\Omega_{-\kappa m}(\boldsymbol{n}_{r})
\Omega_{-\kappa m}^{\dag}(\boldsymbol{n}_{r}^{\prime})
\end{array}
\right).
\nonumber \\
&&
\label{3.13}\end{aligned}$$ If Eqs. (\[2.4\]) and (\[3.13\]) are inserted into Eq.(\[3.11\]), the angular integration in the resulting formula may be conveniently carried out with the aid of the identity [@Szmy07 Eq.(3.1.6)] $$\begin{aligned}
\boldsymbol{n}_{z}
\cdot(\boldsymbol{n}_{r}\times\boldsymbol{\sigma})\,
\Omega_{\kappa\mu}(\boldsymbol{n}_{r})
&=& \mathrm{i}\frac{4\mu\kappa}{4\kappa^{2}-1}
\Omega_{-\kappa\mu}(\boldsymbol{n}_{r})
+\mathrm{i}\frac{\sqrt{(\kappa+\frac{1}{2})^{2}-\mu^{2}}}{|2\kappa+1|}
\Omega_{\kappa+1,\mu}(\boldsymbol{n}_{r})
\nonumber \\
&& -\,\mathrm{i}
\frac{\sqrt{(\kappa-\frac{1}{2})^{2}-\mu^{2}}}{|2\kappa-1|}
\Omega_{\kappa-1,\mu}(\boldsymbol{n}_{r})
\label{3.14}\end{aligned}$$ and the relation $$\begin{aligned}
&& \hspace*{-3em}
\sqrt{\frac{4\pi}{2L+1}}
\oint_{4\pi}\mathrm{d}^{2}\boldsymbol{n}_{r}\:
\Omega_{\kappa\mu}^{\dag}(\boldsymbol{n}_{r})
Y_{LM}^{*}(\boldsymbol{n}_{r})
\Omega_{\kappa'\mu'}(\boldsymbol{n}_{r})
\nonumber \\
&& =\,(-)^{\mu'+1/2}2\sqrt{|\kappa\kappa'|}
\left(
\begin{array}{ccc}
|\kappa|-\frac{1}{2} & L & |\kappa'|-\frac{1}{2} \\
-\frac{1}{2} & 0 & \frac{1}{2}
\end{array}
\right)
\left(
\begin{array}{ccc}
|\kappa|-\frac{1}{2} & L & |\kappa'|-\frac{1}{2} \\
\mu & M & -\mu'
\end{array}
\right)
\Pi(l,L,l'),
\nonumber \\
&&
\label{3.15}\end{aligned}$$ with $$\Pi(l,L,l')
=\left\{
\begin{array}{ll}
1 & \textrm{for $l+L+l'$ even} \\
0 & \textrm{for $l+L+l'$ odd}.
\end{array}
\right.
\label{3.16}$$ In Eq. (\[3.15\]), [${\textstyle
\left(\begin{array}{ccc}j_{a} & j_{b} & j_{c} \\
m_{a} & m_{b} & m_{c}\end{array}\right)}$]{} denotes the Wigner’s 3$j$ coefficient, whereas $$l=\left|\kappa+\frac{1}{2}\right|-\frac{1}{2},
\label{3.17}$$ and similarly for $l'$. Exploiting the selection rules embodied in Eq. (\[3.16\]) and properties of the 3$j$ coefficients, one deduces that the only case when $\widetilde{\mathcal{Q}}_{LM\mu}^{(1)}$ does not vanish is the one with $L=2$ and $M=0$. Since $\widetilde{\mathcal{Q}}_{20\mu}^{(1)}$ is real, from Eq. (\[3.8\]) one has $$\mathcal{Q}_{LM\mu}^{(1)}
=\mathcal{Q}_{20\mu}^{(1)}\delta_{L2}\delta_{M0},
\label{3.18}$$ with $\mathcal{Q}_{20\mu}^{(1)}$ being given in the form of the following double radial integral: $$\mathcal{Q}_{20\mu}^{(1)}={\mathop{\textrm{sgn}}}(\mu)\frac{2}{15}(4\pi\epsilon_{0})cB
\int_{0}^{\infty}\mathrm{d}r\int_{0}^{\infty}\mathrm{d}r'\:
\left(
\begin{array}{cc}
P^{(0)}(r) & Q^{(0)}(r)
\end{array}
\right)
r^{2}\bar{\mathsf{G}}\mbox{}^{(0)}_{2}(r,r')r'
\left(
\begin{array}{c}
Q^{(0)}(r') \\
P^{(0)}(r')
\end{array}
\right),
\label{3.19}$$ where $$\bar{\mathsf{G}}\mbox{}^{(0)}_{\kappa}(r,r')
=\left(
\begin{array}{cc}
\bar{g}\mbox{}^{(0)}_{\kappa,(++)}(r,r') &
\bar{g}\mbox{}^{(0)}_{\kappa,(+-)}(r,r') \\
\bar{g}\mbox{}^{(0)}_{\kappa,(-+)}(r,r') &
\bar{g}\mbox{}^{(0)}_{\kappa,(--)}(r,r')
\end{array}
\right)
\label{3.20}$$ is the radial generalized Green function.
Evaluation of the induced electric quadrupole moment {#IV}
====================================================
In the preceding section, we have proved that, to the first order in $B$, the only electric multipole moment induced in the electronic cloud of the atom by the weak, static, uniform magnetic field is the quadrupole ($L=2$) one, and that its only non-vanishing spherical component is the one with $M=0$. Now, the Cartesian components of the traceless tensor of the induced quadrupole moment, $$\mathcal{Q}_{ij\mu}^{(1)}
=\int_{\mathbb{R}^{3}}\mathrm{d}^{3}\boldsymbol{r}\:
\frac{1}{2}\left(3r_{i}r_{j}-r^{2}\delta_{ij}\right)
\rho_{\mu}^{(1)}(\boldsymbol{r})
\qquad (i,j\in\{x,y,z\}),
\label{4.1}$$ are related to its spherical components through [$$\begin{aligned}
\left(
\begin{array}{ccc}
\mathcal{Q}_{xx\mu}^{(1)} & \mathcal{Q}_{xy\mu}^{(1)} &
\mathcal{Q}_{xz\mu}^{(1)} \\
\mathcal{Q}_{yx\mu}^{(1)} & \mathcal{Q}_{yy\mu}^{(1)} &
\mathcal{Q}_{yz\mu}^{(1)} \\
\mathcal{Q}_{zx\mu}^{(1)} & \mathcal{Q}_{zy\mu}^{(1)} &
\mathcal{Q}_{zz\mu}^{(1)}
\end{array}
\right)
\nonumber \\
&& \hspace*{-12em}
=\left(
\begin{array}{ccc}
-\frac{1}{2}\mathcal{Q}_{20\mu}^{(1)}+\sqrt{\frac{3}{8}}\,
[\mathcal{Q}_{22\mu}^{(1)}+\mathcal{Q}_{2-2\mu}^{(1)}] &
\mathrm{i}\sqrt{\frac{3}{8}}\,
[\mathcal{Q}_{22\mu}^{(1)}-\mathcal{Q}_{2-2\mu}^{(1)}] &
-\sqrt{\frac{3}{8}}\,
[\mathcal{Q}_{21\mu}^{(1)}-\mathcal{Q}_{2-1\mu}^{(1)}] \\
\mathrm{i}\sqrt{\frac{3}{8}}\,
[\mathcal{Q}_{22\mu}^{(1)}-\mathcal{Q}_{2-2\mu}^{(1)}] &
-\frac{1}{2}\mathcal{Q}_{20\mu}^{(1)}-\sqrt{\frac{3}{8}}\,
[\mathcal{Q}_{22\mu}^{(1)}+\mathcal{Q}_{2-2\mu}^{(1)}] &
-\mathrm{i}\sqrt{\frac{3}{8}}\,
[\mathcal{Q}_{21\mu}^{(1)}+\mathcal{Q}_{2-1\mu}^{(1)}] \\
-\sqrt{\frac{3}{8}}\,
[\mathcal{Q}_{21\mu}^{(1)}-\mathcal{Q}_{2-1\mu}^{(1)}] &
-\mathrm{i}\sqrt{\frac{3}{8}}\,
[\mathcal{Q}_{21\mu}^{(1)}+\mathcal{Q}_{2-1\mu}^{(1)}] &
\mathcal{Q}_{20\mu}^{(1)}
\end{array}
\right).
\nonumber \\
&&
\label{4.2}\end{aligned}$$ ]{} Hence, it follows that in the Cartesian basis all the off-diagonal elements of the quadrupole moment induced in the system under study do vanish: $$\mathcal{Q}_{ij\mu}^{(1)}=0
\qquad (\textrm{$i\neq j$; $i,j\in\{x,y,z\}$}),
\label{4.3}$$ while the diagonal elements are given by $$\mathcal{Q}_{xx\mu}^{(1)}=\mathcal{Q}_{yy\mu}^{(1)}
=-\frac{1}{2}\mathcal{Q}_{zz\mu}^{(1)}
=-\frac{1}{2}\mathcal{Q}_{20\mu}^{(1)}.
\label{4.4}$$ Such a structure of the Cartesian representation of the quadrupole moment tensor is characteristic for systems possessing the rotational symmetry around the $z$ axis (recall that in our case the $z$ axis is directed along the magnetic field).
It remains to evaluate the double integral in Eq. (\[3.19\]). For this purpose, we shall make use of the Sturmian expansion of the pertinent generalized radial Dirac–Coulomb Green function for the atomic ground state, which is [@Szmy97] $$\bar{\mathsf{G}}\mbox{}^{(0)}_{\kappa}(r,r')
=\sum_{n_{r}=-\infty}^{\infty}
\frac{1}{\mu_{n_{r}\kappa}^{(0)}-1}
\left(
\begin{array}{c}
S_{n_{r}\kappa}^{(0)}(r) \\
T_{n_{r}\kappa}^{(0)}(r)
\end{array}
\right)
\left(
\begin{array}{cc}
\mu_{n_{r}\kappa}^{(0)}S_{n_{r}\kappa}^{(0)}(r')
& T_{n_{r}\kappa}^{(0)}(r')
\end{array}
\right)
\qquad (\kappa\neq-1).
\label{4.5}$$ Here
$$\begin{aligned}
S_{n_{r}\kappa}^{(0)}(r)
&=& \sqrt{\frac{(1+\gamma_{1})(|n_{r}|+2\gamma_{\kappa})|n_{r}|!}
{2ZN_{n_{r}\kappa}(N_{n_{r}\kappa}-\kappa)
\Gamma(|n_{r}|+2\gamma_{\kappa})}}
\nonumber \\
&& \times\left(\frac{2Zr}{a_{0}}\right)^{\gamma_{\kappa}}
\textrm{e}^{-Zr/a_{0}}
\left[L_{|n_{r}|-1}^{(2\gamma_{\kappa})}\left(\frac{2Zr}{a_{0}}\right)
+\frac{\kappa-N_{n_{r}\kappa}}{|n_{r}|+2\gamma_{\kappa}}
L_{|n_{r}|}^{(2\gamma_{\kappa})}\left(\frac{2Zr}{a_{0}}\right)\right]
\label{4.6a}\end{aligned}$$
and $$\begin{aligned}
T_{n_{r}\kappa}^{(0)}(r)
&=& \sqrt{\frac{(1-\gamma_{1})(|n_{r}|+2\gamma_{\kappa})|n_{r}|!}
{2ZN_{n_{r}\kappa}(N_{n_{r}\kappa}-\kappa)
\Gamma(|n_{r}|+2\gamma_{\kappa})}}
\nonumber \\
&& \times\left(\frac{2Zr}{a_{0}}\right)^{\gamma_{\kappa}}
\textrm{e}^{-Zr/a_{0}}
\left[L_{|n_{r}|-1}^{(2\gamma_{\kappa})}
\left(\frac{2Zr}{a_{0}}\right)
-\frac{\kappa-N_{n_{r}\kappa}}{|n_{r}|+2\gamma_{\kappa}}
L_{|n_{r}|}^{(2\gamma_{\kappa})}\left(\frac{2Zr}{a_{0}}\right)\right]
\label{4.6b}\end{aligned}$$ \[4.6\]
are the radial Dirac–Coulomb Sturmian functions associated with the hydrogenic ground-state energy level, whereas $$\mu_{n_{r}\kappa}^{(0)}
=\frac{|n_{r}|+\gamma_{\kappa}+N_{n_{r}\kappa}}{\gamma_{1}+1},
\label{4.7}$$ with $$N_{n_{r}\kappa}
=\pm\sqrt{(|n_{r}|+\gamma_{\kappa})^{2}+(\alpha Z)^{2}}
=\pm\sqrt{|n_{r}|^{2}+2|n_{r}|\gamma_{\kappa}+\kappa^{2}}
\label{4.8}$$ being the ‘apparent principal quantum number’ (notice that it may assume positive as well as negative values!); the following sign convention applies to the definition (\[4.8\]): the plus sign should be chosen for $n_{r}>0$ and the minus one for $n_{r}<0$; for $n_{r}=0$ one chooses the plus sign if $\kappa<0$ and the minus sign if $\kappa>0$. Insertion of the separable expansion (\[4.5\]) into the right-hand side of the formula in Eq. (\[3.19\]) leads to the following expression for $\mathcal{Q}_{20\mu}^{(1)}$: $$\begin{aligned}
\mathcal{Q}_{20\mu}^{(1)}
&=& {\mathop{\textrm{sgn}}}(\mu)\frac{2}{15}(4\pi\epsilon_{0})cB
\sum_{n_{r}=-\infty}^{\infty}\frac{1}{\mu_{n_{r}2}^{(0)}-1}
\int_{0}^{\infty}\mathrm{d}r\:r^{2}
\left[P^{(0)}(r)S_{n_{r}2}^{(0)}(r)
+Q^{(0)}(r)T_{n_{r}2}^{(0)}(r)\right]
\nonumber \\
&& \times\int_{0}^{\infty}\mathrm{d}r'\:r'
\left[\mu_{n_{r}2}^{(0)}Q^{(0)}(r')S_{n_{r}2}^{(0)}(r')
+P^{(0)}(r')T_{n_{r}2}^{(0)}(r')\right].
\label{4.9}\end{aligned}$$ The radial integrals in Eq. (\[4.9\]) may be taken after one makes use of Eq. (\[4.7\]) and of the explicit representations of the radial functions $P^{(0)}(r)$, $Q^{(0)}(r)$ and the radial Sturmians $S_{n_{r}2}^{(0)}(r)$, $T_{n_{r}2}^{(0)}(r)$, given by Eqs.(\[2.6\]) and (\[4.6\]), respectively. Exploiting the known integral formula [@Grad07 Eq. (7.414.11)] $$\int_{0}^{\infty}\mathrm{d}x\:
x^{\gamma}\mathrm{e}^{-x}L_{n}^{(\alpha)}(x)
=\frac{\Gamma(\gamma+1)\Gamma(n+\alpha-\gamma)}
{n!\Gamma(\alpha-\gamma)}
\qquad [{\mathop{\textrm{Re}}}(\gamma)>-1]
\label{4.10}$$ and the trivial but extremely useful identity $$\gamma_{2}^{2}=\gamma_{1}^{2}+3,
\label{4.11}$$ one arrives at $$\begin{aligned}
\mathcal{Q}_{20\mu}^{(1)}
&=& -\,{\mathop{\textrm{sgn}}}(\mu)\frac{\alpha^{2}ea_{0}^{2}}{Z^{4}}
\frac{B}{B_{0}}\frac{\Gamma^{2}(\gamma_{1}+\gamma_{2}+3)}
{480(4\gamma_{1}+1)\Gamma(2\gamma_{1}+1)
\Gamma^{2}(\gamma_{2}-\gamma_{1}-2)}
\nonumber \\
&& \times\sum_{n_{r}=-\infty}^{\infty}
\frac{\Gamma(|n_{r}|+\gamma_{2}-\gamma_{1}-3)
\Gamma(|n_{r}|+\gamma_{2}-\gamma_{1}-2)}
{|n_{r}|!\Gamma(|n_{r}|+2\gamma_{2}+1)}
\frac{N_{n_{r}2}-2}{N_{n_{r}2}}
\nonumber \\
&& \times
\frac{(|n_{r}|+\gamma_{2}-3\gamma_{1}-3-\gamma_{1}N_{n_{r}2})
(3|n_{r}|+\gamma_{1}+3\gamma_{2}+1+3N_{n_{r}2})}
{|n_{r}|+\gamma_{2}-\gamma_{1}-1+N_{n_{r}2}}.
\label{4.12}\end{aligned}$$ The above result may be simplified considerably if in the series $\sum_{n_{r}=-\infty}^{\infty}(\cdots)$ one collects together terms with the same absolute value of the summation index $n_{r}$ (the Sturmian radial quantum number). Proceeding in that way, after much labor, using Eq. (\[4.8\]) and again the identity (\[4.11\]), one finds that $$\begin{aligned}
\mathcal{Q}_{20\mu}^{(1)}
&=& {\mathop{\textrm{sgn}}}(\mu)\frac{\alpha^{2}ea_{0}^{2}}{Z^{4}}
\frac{B}{B_{0}}\frac{\Gamma^{2}(\gamma_{1}+\gamma_{2}+3)}
{240(4\gamma_{1}+1)\Gamma(2\gamma_{1}+1)
\Gamma^{2}(\gamma_{2}-\gamma_{1}-2)}
\nonumber \\
&& \times\sum_{n_{r}=0}^{\infty}
\frac{\Gamma(n_{r}+\gamma_{2}-\gamma_{1}-3)
\Gamma(n_{r}+\gamma_{2}-\gamma_{1}-2)}
{n_{r}!(n_{r}+\gamma_{2}-\gamma_{1})\Gamma(n_{r}+2\gamma_{2}+1)}
\nonumber \\
&& \times[(2\gamma_{1}-1)(n_{r}+\gamma_{2}-\gamma_{1}-3)
(n_{r}+\gamma_{2}-\gamma_{1})
+2(\gamma_{1}+1)(2\gamma_{1}+1)
(n_{r}+\gamma_{2}-\gamma_{1})
\nonumber \\
&& \qquad -\,6(\gamma_{1}+1)].
\label{4.13}\end{aligned}$$ It is possible to achieve a further simplification. To this end, we express the right-hand side of Eq. (\[4.13\]) in terms of the hypergeometric functions $_{2}F_{1}$ and $_{3}F_{2}$ of the unit argument. Since it holds that $$\sum_{n=0}^{\infty}\frac{\Gamma(n+a_{1})\Gamma(n+a_{2})}
{n!\Gamma(n+b)}
=\frac{\Gamma(a_{1})\Gamma(a_{2})}{\Gamma(b)}\,
{}_{2}F_{1}
\left(
\begin{array}{c}
a_{1},a_{2} \\
b
\end{array}
;1
\right)
\label{4.14}$$ and $$\sum_{n=0}^{\infty}
\frac{\Gamma(n+a_{1})\Gamma(n+a_{2})}{n!(n+a_{3})\Gamma(n+b)}
=\frac{\Gamma(a_{1})\Gamma(a_{2})}{a_{3}\Gamma(b)}\,
{}_{3}F_{2}
\left(
\begin{array}{c}
a_{1},a_{2},a_{3} \\
a_{3}+1,b
\end{array}
;1
\right),
\label{4.15}$$ Eq. (\[4.13\]) becomes $$\begin{aligned}
\mathcal{Q}_{20\mu}^{(1)}
&=& -\,{\mathop{\textrm{sgn}}}(\mu)\frac{\alpha^{2}ea_{0}^{2}}{Z^{4}}
\frac{B}{B_{0}}\frac{\Gamma(\gamma_{1}+\gamma_{2}+3)
\Gamma(\gamma_{1}+\gamma_{2}+4)}
{1440(\gamma_{1}+1)(4\gamma_{1}+1)\Gamma(2\gamma_{1}+1)
\Gamma(2\gamma_{2}+1)}
\nonumber \\
&& \times\left[
(2\gamma_{1}-1)(\gamma_{2}-\gamma_{1}-3)\,
{}_{2}F_{1}\left(
\begin{array}{c}
\gamma_{2}-\gamma_{1}-2,\gamma_{2}-\gamma_{1}-2 \\
2\gamma_{2}+1
\end{array}
;1
\right)
\right.
\nonumber \\
&& \qquad +\,2(\gamma_{1}+1)(2\gamma_{1}+1)\,
{}_{2}F_{1}\left(
\begin{array}{c}
\gamma_{2}-\gamma_{1}-3,\gamma_{2}-\gamma_{1}-2 \\
2\gamma_{2}+1
\end{array}
;1
\right)
\nonumber \\
&& \qquad -\left.
\frac{6(\gamma_{1}+1)}{\gamma_{2}-\gamma_{1}}\,
{}_{3}F_{2}\left(
\begin{array}{c}
\gamma_{2}-\gamma_{1}-3,\gamma_{2}-\gamma_{1}-2,
\gamma_{2}-\gamma_{1} \\
\gamma_{2}-\gamma_{1}+1,2\gamma_{2}+1
\end{array}
;1
\right)
\right].
\label{4.16}\end{aligned}$$ The two $_{2}F_{1}$ functions may be then eliminated with the aid of the Gauss’ identity [@Grad07 Eq. (9.122.1)] $${}_{2}F_{1}
\left(
\begin{array}{c}
a_{1},a_{2} \\
b
\end{array}
;1
\right)
=\frac{\Gamma(b)\Gamma(b-a_{1}-a_{2})}
{\Gamma(b-a_{1})\Gamma(b-a_{2})}
\qquad [{\mathop{\textrm{Re}}}(b-a_{1}-a_{2})>0].
\label{4.17}$$ After some simple algebra, we obtain $$\begin{aligned}
\mathcal{Q}_{20\mu}^{(1)}
&=& {\mathop{\textrm{sgn}}}(\mu)\frac{\alpha^{2}ea_{0}^{2}}{Z^{4}}
\frac{B}{B_{0}}\frac{\Gamma(2\gamma_{1}+5)}
{720(4\gamma_{1}+1)\Gamma(2\gamma_{1}+1)}
\bigg[-2(2\gamma_{1}^{2}+3\gamma_{1}+4)
\nonumber \\
&& +\,\frac{(\gamma_{1}+\gamma_{2})\Gamma(\gamma_{1}+\gamma_{2}+3)
\Gamma(\gamma_{1}+\gamma_{2}+4)}
{\Gamma(2\gamma_{1}+5)\Gamma(2\gamma_{2}+1)}\,
{}_{3}F_{2}
\left(
\begin{array}{c}
\gamma_{2}-\gamma_{1}-3,\gamma_{2}-\gamma_{1}-2,
\gamma_{2}-\gamma_{1} \\
\gamma_{2}-\gamma_{1}+1,2\gamma_{2}+1
\end{array}
;1
\right)
\bigg].
\nonumber \\
&&
\label{4.18}\end{aligned}$$ Various other equivalent representations of $\mathcal{Q}_{20\mu}^{(1)}$ may be derived from Eq. (\[4.18\]) with the help of recurrence relations obeyed by the $_{3}F_{2}(1)$ function. For instance, if one uses repeatedly the relation $$\begin{aligned}
&& {}_{3}F_{2}
\left(
\begin{array}{c}
a_{1},a_{2},a_{3} \\
a_{3}+1,b
\end{array}
;1
\right)
=-\frac{a_{3}}{a_{1}-a_{3}}
\frac{\Gamma(b)\Gamma(b-a_{1}-a_{2})}
{\Gamma(b-a_{1})\Gamma(b-a_{2})}
+\frac{a_{1}}{a_{1}-a_{3}}\,
{}_{3}F_{2}
\left(
\begin{array}{c}
a_{1}+1,a_{2},a_{3} \\
a_{3}+1,b
\end{array}
;1
\right)
\nonumber \\
&& \hspace*{20em}
[{\mathop{\textrm{Re}}}(b-a_{1}-a_{2})>0],
\label{4.19}\end{aligned}$$ and its analogue with $a_{1}$ and $a_{2}$ interchanged, Eq.(\[4.18\]) is casted into the slightly more compact expression $$\begin{aligned}
\mathcal{Q}_{20\mu}^{(1)}
&=& {\mathop{\textrm{sgn}}}(\mu)\frac{\alpha^{2}ea_{0}^{2}}{Z^{4}}
\frac{B}{B_{0}}\frac{\Gamma(2\gamma_{1}+5)}{1440\,\Gamma(2\gamma_{1})}
\left[-1+\frac{(\gamma_{1}+1)(\gamma_{1}+\gamma_{2})
\Gamma(\gamma_{1}+\gamma_{2}+2)\Gamma(\gamma_{1}+\gamma_{2}+3)}
{\gamma_{1}\Gamma(2\gamma_{1}+5)\Gamma(2\gamma_{2}+1)}\right.
\nonumber \\
&& \qquad \times\left.
{}_{3}F_{2}
\left(
\begin{array}{c}
\gamma_{2}-\gamma_{1}-2,\gamma_{2}-\gamma_{1}-1,
\gamma_{2}-\gamma_{1} \\
\gamma_{2}-\gamma_{1}+1,2\gamma_{2}+1
\end{array}
;1
\right)
\right].
\label{4.20}\end{aligned}$$ Before concluding, it seems worthwhile to investigate the approximation to Eq. (\[4.20\]) for . Then one has $$\gamma_{\kappa}\simeq|\kappa|-\frac{(\alpha Z)^{2}}{2|\kappa|}
\qquad (\kappa\in\mathbb{Z}\setminus\{0\}),
\label{4.21}$$ and consequently $$\begin{aligned}
\Gamma(\gamma_{\kappa}+\gamma_{\kappa'}+k)
&\simeq& (|\kappa|+|\kappa'|+k-1)!
\left[1-\frac{1}{2}\left(\frac{1}{|\kappa|}+\frac{1}{|\kappa'|}\right)
\psi\left(|\kappa|+|\kappa'|+k\right)(\alpha Z)^{2}\right]
\nonumber \\
&& \hspace*{15em}
(\textrm{$\kappa,\kappa'\in\mathbb{Z}\setminus\{0\}$,
$k\in\mathbb{N}$}),
\label{4.22}\end{aligned}$$ where $$\psi(\zeta)=\frac{1}{\Gamma(\zeta)}
\frac{\mathrm{d}\Gamma(\zeta)}{\mathrm{d}\zeta}
\label{4.23}$$ is the digamma function. Using Eq. (\[4.22\]) and the recurrence relation $$\psi(\zeta+1)=\psi(\zeta)+\frac{1}{\zeta}
\label{4.24}$$ yields $$\frac{(\gamma_{1}+1)(\gamma_{1}+\gamma_{2})
\Gamma(\gamma_{1}+\gamma_{2}+2)\Gamma(\gamma_{1}+\gamma_{2}+3)}
{\gamma_{1}\Gamma(2\gamma_{1}+5)\Gamma(2\gamma_{2}+1)}
\simeq1+\frac{13}{60}(\alpha Z)^{2}.
\label{4.25}$$ Since at the same time, to the second order in $\alpha Z$, it holds that $${}_{3}F_{2}
\left(
\begin{array}{c}
\gamma_{2}-\gamma_{1}-2,\gamma_{2}-\gamma_{1}-1,
\gamma_{2}-\gamma_{1} \\
\gamma_{2}-\gamma_{1}+1,2\gamma_{2}+1
\end{array}
;1
\right)
\simeq1-\frac{(\alpha Z)^{2}}{40},
\label{4.26}$$ the sought approximation to $\mathcal{Q}_{20\mu}^{(1)}$ is $$\mathcal{Q}_{20\mu}^{(1)}\simeq{\mathop{\textrm{sgn}}}(\mu)
\frac{23}{240}\frac{\alpha^{4}ea_{0}^{2}}{Z^{2}}\frac{B}{B_{0}}
\qquad (\alpha Z\ll1).
\label{4.27}$$ It is seen from Eqs. (\[4.27\]) and (\[1.1\]) that for the hydrogen atom ($Z=1)$ and for the perturbing magnetic field comparable with the intra-atomic magnetic field, i.e., for $B\simeq
B_{0}$, the first-order quadrupole moment $\mathcal{Q}_{20\mu}^{(1)}$ predicted by the relativistic formalism is of the same order of magnitude as the second-order moment $\mathcal{Q}_{20}^{(2)}$ obtained from the non-relativistic theory.
Conclusions {#V}
===========
Earlier calculations of the magnetic field-induced electric quadrupole moment in the ground state of the hydrogen-like atom, based on the non-relativistic atomic model, predicted the quadratic dependence of that moment on the field strength in the low-field regime. In the present paper, we have shown that if the relativity is taken into account and considerations are based on the Dirac rather than the Schr[ö]{}dinger or the Pauli equation for the electron, the leading term in the expansion of the induced electric quadrupole moment in powers of the field strength appears to be linear, not quadratic. The calculations of the actual value of that moment we have carried out in Sec. \[IV\] provide a still another example of the usefulness of the Sturmian expansion of the generalized Dirac–Coulomb Green function [@Szmy97] for analytical determination of electromagnetic properties of the relativistic one-electron atom.
[99]{} C. A. Coulson, M. J. Stephen, Proc. Phys. Soc. A 69 (1956) 777 A. V. Turbiner, Yad. Fiz. 46 (1987) 204 A. Y. Potekhin, A. V. Turbiner, Phys. Rev. A 63 (2001) 065402 R. Szmytkowski, J. Phys. B 30 (1997) 825 \[erratum J. Phys. B 30 (1997) 2747; addendum arXiv:physics/9902050\] R. Szmytkowski, J. Phys. B 35 (2002) 1379 R. Szmytkowski, K. Mielewczyk, J. Phys. B 37 (2004) 3961 K. Mielewczyk, R. Szmytkowski, Phys. Rev. A 73 (2006) 022511 \[erratum Phys. Rev. A 73 (2006) 039908\] R. Szmytkowski, P. Stefa[ń]{}ska, arXiv:1102.1811 P. Stefa[ń]{}ska, R. Szmytkowski, Int. J. Quantum Chem. doi:10.1002/qua.23108 \[preprint arXiv:1102.3853\] R. Szmytkowski, J. Math. Chem. 42 (2007) 397 \[preprint arXiv:1011.3433\] D. A. Varshalovich, A. N. Moskalev, V. K. Khersonskii, Quantum Theory of Angular Momentum, Nauka, Leningrad, 1975 (in Russian) W. Magnus, F. Oberhettinger, R. P. Soni, Formulas and Theorems for the Special Functions of Mathematical Physics, 3rd ed., Springer, Berlin, 1966 I. S. Gradshteyn, I. M. Ryzhik, Table of Integrals, Series, and Products, 7th ed., Elsevier, Amsterdam, 2007
[^1]: Corresponding author. Email: [email protected]
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We describe the equivariant Chow ring of the wonderful compactification $X$ of a symmetric space of minimal rank, via restriction to the associated toric variety $Y$. Also, we show that the restrictions to $Y$ of the tangent bundle $T_X$ and its logarithmic analogue $S_X$ decompose into a direct sum of line bundles. This yields closed formulae for the equivariant Chern classes of $T_X$ and $S_X$, and, in turn, for the Chern classes of reductive groups considered by Kiritchenko.'
address:
- |
Michel Brion\
Université de Grenoble–I\
Institut Fourier, BP 74\
38402 Saint-Martin d’Hères Cedex\
France
- |
Roy Joshua\
Department of Mathematics\
Ohio State University\
231 W 18th Avenue\
Columbus, OH 43210\
USA
author:
- 'M. Brion and R. Joshua'
title: Equivariant Chow ring and Chern classes of wonderful symmetric varieties of minimal rank
---
[^1]
Introduction {#sec:introduction}
============
The purpose of this article is to describe the equivariant intersection ring and equivariant Chern classes of a class of almost homogeneous varieties, namely, wonderful symmetric varieties of minimal rank; these include the wonderful compactifications of semi-simple groups of adjoint type.
The main motivation comes from questions of enumerative geometry on a spherical homogeneous space $G/K$. As shown by De Concini and Procesi, these questions find their proper setting in the ring of conditions $C^*(G/K)$, isomorphic to the direct limit of cohomology rings of $G$-equivariant compactifications $X$ of $G/K$ (see [@DP83; @DP85]). Recently, the Euler characteristic of any complete intersection of hyper-surfaces in $G/K$ has been expressed by Kiritchenko (see [@Ki06]), in terms of the Chern classes of the logarithmic tangent bundle $S_X$ of any “regular” compactification $X$. As shown in [@Ki06], these Chern classes are independent of the choice of $X$, and hence yield elements of $C^*(G/K)$; moreover, their determination may be reduced to the case where $X$ is a “wonderful variety”.
In fact, it is more convenient to work with the rational equivariant cohomology ring $H^*_G(X)$, from which the ordinary rational cohomology ring $H^*(X)$ is obtained by killing the action of generators of the polynomial ring $H^*(BG)$; the Chern classes of $S_X$ have natural representatives in $H^*_G(X)$, the equivariant Chern classes. When $X$ is a complete symmetric variety, the ring $H^*_G(X)$ admits algebraic descriptions by work of Bifet, De Concini, Littelman, and Procesi (see [@BDP90; @LP90]).
Here we consider the case where $X$ is the wonderful compactification of a symmetric space $G/K$ of minimal rank, that is, $G$ is semi-simple of adjoint type and ${\operatorname{rk}}(G/K) = {\operatorname{rk}}(G) - {\operatorname{rk}}(K)$; the main examples are the groups $G = (G \times G)/ {\operatorname{diag}}(G)$ and the spaces ${\operatorname{PSL}}(2n)/{\operatorname{PSp}}(2n)$. Moreover, we follow a purely algebraic approach: we work over an arbitrary algebraically closed field, and replace the equivariant cohomology ring with the equivariant intersection ring $A^*_G(X)$ of [@EG98] (for wonderful varieties over the complex numbers, both rings are isomorphic over the rationals).
We show in Theorem \[thm:eq\] that the pull-back map $$r : A^*_G(X) \to A^*_T(Y)^{W_K}$$ is an isomorphism over the rationals. Here $T \subset G$ denotes a maximal torus containing a maximal torus $T_K \subset K$ with Weyl group $W_K$, and $Y$ denotes the closure in $X$ of $T/T_K \subset G/K$, so that $Y$ is the toric variety associated with the Weyl chambers of the restricted root system of $G/K$.
We also determine the images under $r$ of the equivariant Chern classes of the tangent bundle $T_X$ and its logarithmic analogue $S_X$. For this, we show in Theorem \[thm:split\] that the normal bundle $N_{Y/X}$ decomposes (as a $T$-linearized bundle) into a direct sum of line bundles indexed by certain roots of $K$; moreover, any such line bundle is the pull-back of $O_{{{\mathbb P}}^1}(1)$ under a certain $T$-equivariant morphism $Y \to {\mathbb P}^1$. By Proposition \[prop:prod\], the product of these morphisms yields a closed immersion of the toric variety $Y$ into a product of projective lines, indexed by the restricted roots.
In the case of regular compactifications of reductive groups, Theorem \[thm:eq\] is due to Littelmann and Procesi for equivariant cohomology rings (see [@LP90]); it has been adapted to equivariant Chow ring in [@Br98]. Here, as in the latter paper, we rely on a precise version of the localization theorem in equivariant intersection theory inspired, in turn, by a similar result in equivariant cohomology, see [@GKM99]. The main ingredient is the finiteness of $T$-stable points and curves in $X$; this also plays an essential role in Tchoudjem’s description of cohomology groups of line bundles on wonderful varieties of minimal rank, see [@Tc05].
For wonderful group compactifications, a more precise, “additive” description of the equivariant cohomology ring is due to Strickland, see [@St06]; an analogous description of the equivariant Grothendieck group has been obtained by Uma in [@Um05]. Both results may be generalized to our setting of minimal rank. However, determining generators and relations for the equivariant cohomology or Grothendieck ring is still an open question; see [@Br04; @Um05] for some steps in this direction.
Our determination of the equivariant Chern classes seems to be new, already in the group case; it yields a closed formula for the image under $r$ of the equivariant Todd class of $X$, analogous to the well-known formula expressing the Todd class of a toric variety in terms of boundary divisors. The toric variety $Y$ associated to Weyl chambers is considered in [@Pr90; @DL94], where its cohomology is described as a graded representation of the Weyl group; its realization as a general orbit closure in a product of projective lines seems to have been unnoticed.
Our results extend readily to all regular compactifications of symmetric spaces of minimal rank. Specifically, the description of the equivariant Chow ring holds unchanged, with a similar proof, and the determination of equivariant Chern classes follows from the wonderful case by the results of [@Ki06 Sec. 5]. Another direct generalization concerns the spherical (not necessarily symmetric) varieties of minimal rank considered in [@Tc05]. Indeed, the structure of such varieties may be reduced to the symmetric case, as shown by Ressayre in [@Re04].
This article is organized as follows. Section 1 gathers preliminary notions and results on symmetric spaces, their wonderful compactifications, and the associated toric varieties. In particular, for a symmetric space $G/K$ of minimal rank, we study the relations between the root systems and Weyl groups of $G$, $K$, and $G/K$; these are our main combinatorial tools. In Section 2, we first describe the $T$-stable points and curves in a wonderful symmetric variety $X$ of minimal rank; then we obtain our main structure result for $A^*_G(X)$, and some useful complements as well. Section 3 contains the decompositions of $N_{Y/X}$ and of the restrictions $T_X\vert_Y$, $S_X \vert_Y$, together with their applications to equivariant Chern and Todd classes.
Throughout this article, we consider algebraic varieties over an algebraically closed field $k$ of characteristic $\neq 2$; by a point of such a variety, we mean a closed point. As general references, we use [@Ha77] for algebraic geometry, and [@Sp98] for algebraic groups.
Preliminaries
=============
The toric variety associated with Weyl chambers {#subsec:tv}
-----------------------------------------------
Let $\Phi$ be a root system in a real vector space $V$ (we follow the conventions of [@Bo81] for root systems; in particular, $\Phi$ is finite but not necessarily reduced). Let $W$ be the Weyl group, $Q$ the root lattice in $V$, and $Q^{\vee}$ the dual lattice (the co-weight lattice) in the dual vector space $V^*$. The Weyl chambers form a subdivision of $V^*$ into rational polyhedral convex cones; let $\Sigma$ be the fan of $V^*$ consisting of all Weyl chambers and their faces. The pair $(Q^{\vee},\Sigma)$ corresponds to a toric variety $$Y = Y(\Phi)$$ equipped with an action of $W$ via its action on $Q^{\vee}$ which permutes the Weyl chambers. The group $W$ acts compatibly on the associated torus $$T := {\operatorname{Hom}}(Q,{\mathbb G}_m) = Q^{\vee} \otimes_{{\mathbb Z}} {\mathbb G}_m.$$ Thus, $Y$ is equipped with an action of the semi-direct product $T \, W$. Note that the character group ${\mathcal X}(T)$ is identified with $Q$; in particular, we may regard each root $\alpha$ as a homomorphism $$\alpha: T \to {\mathbb G}_m.$$
The choice of a basis of $\Phi$, $$\Delta = \{\alpha_1,\ldots,\alpha_r\},$$ defines a positive Weyl chamber, the dual cone to $\Delta$. Let $Y_0 \subset Y$ be the corresponding $T$-stable open affine subset. Then $Y_0$ is isomorphic to the affine space ${\mathbb A}^r$ on which $T$ acts linearly with weights $-\alpha_1,\ldots,-\alpha_r$. Moreover, the translates $w \cdot Y_0$, where $w \in W$, form an open covering of $Y$.
In particular, the variety $Y$ is nonsingular. Also, $Y$ is projective, as $\Sigma$ is the normal fan to the convex polytope with vertices $w \cdot v$ ($w \in W$), where $v$ is any prescribed regular element of $V$. The following result yields an explicit projective embedding of $Y$:
\[prop:prod\] [(i)]{} For any $\alpha \in \Phi$, the morphism $\alpha : T \to {\mathbb G}_m$ extends to a morphism $$f_{\alpha} : Y \to {\mathbb P}^1.$$ Moreover, $f_{\alpha}$ and $f_{-\alpha}$ differ by the inverse map ${\mathbb P}^1 \to {\mathbb P}^1$, $z \mapsto z^{-1}$.
[(ii)]{} The product morphism $$f := \prod_{\alpha \in \Phi} f_{\alpha} : Y \to
\prod_{\alpha \in \Phi} {\mathbb P}^1$$ is a closed immersion. It is equivariant under $T \, W$, where $T$ acts on the right-hand side via its action on each factor ${\mathbb P}^1_{\alpha}$ through the character $\alpha$, and $W$ acts via its natural action on the set $\Phi$ of indices.
[(iii)]{} Conversely, the $T$-orbit closure of any point of $\prod_{\alpha \in \Phi} ({\mathbb P}^1 \setminus \{0,\infty \})$ is isomorphic to $Y$.
[(iv)]{} Any non-constant morphism $F : Y \to C$, where $C$ is an irreducible curve, factors through $f_\alpha : Y \to {\mathbb P}^1$ where $\alpha$ is an indivisible root, unique up to sign. Moreover, $$\label{eqn:fxp}
(f_\alpha)_* {\mathcal O}_Y = {\mathcal O}_{{\mathbb P}^1}.$$
\(i) Since $\alpha$ has a constant sign on each Weyl chamber, it defines a morphism of fans from $\Sigma$ to the fan of ${\mathbb P}^1$, consisting of two opposite half-lines and the origin. This implies our statement.
\(ii) The equivariance property of $f$ is readily verified. Moreover, the product map $$\prod_{i=1}^r f_{\alpha_i} : Y \to ({\mathbb P}^1)^r$$ restricts to an isomorphism $Y_0 \to ({\mathbb P}^1 \setminus \{\infty\})^r$, since each $f_{\alpha_i}$ restricts to the $i$-th coordinate function on $Y_0 \cong {\mathbb A}^r$. Since $Y = W \cdot Y_0$, it follows that $f$ is a closed immersion.
\(iii) follows from (ii) by using the action of tuples $(t_\alpha)_{\alpha \in \Phi}$ of non-zero scalars, via component-wise multiplication.
\(iv) Taking the Stein factorization, we may assume that $F_* {\mathcal O}_Y = {\mathcal O}_C$. Then $C$ is normal, and hence nonsingular. Moreover, the action of $T$ on $Y$ descends to a unique action on $C$ such that $F$ is equivariant (indeed, $F$ equals the canonical morphism $Y \to {\operatorname{Proj}}R(Y,F^* {\mathcal L})$, where ${\mathcal L}$ is any ample invertible sheaf on $C$, and $R(Y, F^*{\mathcal L})$ denotes the section ring $\bigoplus_n \Gamma(Y,F^*{\mathcal L}^n)$. Furthermore, $F^*{\mathcal L}$ admits a $T$-linearization, and hence $T$ acts on $R(Y,{\mathcal L})$). It follows that $C \cong {\mathbb P}^1$ where $T$ acts through a character $\chi$, uniquely defined up to sign. Thus, $F$ induces a morphism from the fan of $Y$ to the fan of ${\mathbb P}^1$; this morphism is given by the linear map $\chi : V^* \to {\mathbb R}$. In other words, $\chi$ has a constant sign on each Weyl chamber. Thus, $\chi$ is an integral multiple of an indivisible root $\alpha$, uniquely defined up to sign. Since $F$ has connected fibers, then $\chi = \pm \alpha$.
Conversely, if $\alpha$ is an indivisible root, then the fibers of the morphism $\alpha : T \to {\mathbb G}_m$ are irreducible. This implies (\[eqn:fxp\]).
Next, for later use, we determine the divisor of each $f_\alpha$ regarded as a rational function on $Y$. Since $f_\alpha$ is a $T$-eigenvector, its divisor is a linear combination of the $T$-stable prime divisors $Y_1,\ldots,Y_m$ of the toric variety $Y$, also called its *boundary divisors*. Recall that $Y_1,\ldots,Y_m$ correspond bijectively to the rays of the Weyl chambers, i.e., to the $W$-translates of the fundamental co-weights $\omega_1^{\vee},\ldots,\omega_r^{\vee}$ (which form the dual basis of the basis of simple roots). The isotropy group of each $\omega_i^{\vee}$ in $W$ is the maximal parabolic subgroup $W_i$ generated by the reflections associated with the simple roots $\alpha_j$, $j \neq i$. Thus, the orbit $W \omega_i^{\vee} \cong W/W_i$ is in bijection with the subset $$W^i := \{ w\in W ~\vert~ w\alpha_j \in \Phi^+ \text{ for all }j \neq i \}$$ of minimal representatives for the coset space $W/W_i$. So the boundary divisors are indexed by the set $$E(\Phi) := \{(i,w) ~\vert~ 1 \leq i \leq r, ~ w \in W^i\}
\cong \{ w\omega_i^{\vee} ~\vert~ 1 \leq i \leq r, ~ w \in W\};$$ we will denote these divisors by $Y_{i,w}$. Furthermore, we have $$\label{eqn:div}
{\operatorname{div}}(f_\alpha) = \sum_{(i,w) \in E(\Phi)}
\langle \alpha, w \omega_i^{\vee} \rangle \;Y_{i,w}$$ by Proposition \[prop:prod\] and the classical formula for the divisor of a character in a toric variety (see e.g. [@Od88 Prop. 2.1]). Also, note that $\langle \alpha, w \omega_i^{\vee} \rangle$ is the $i$-th coordinate of $w^{-1}\alpha$ in the basis of simple roots.
Symmetric spaces {#subsec:ss}
----------------
Let $G$ be a connected reductive algebraic group, and $$\theta: G \to G$$ an involutive automorphism. Denote by $$K = G^{\theta} \subset G$$ the subgroup of fixed points; then the homogeneous space $G/K$ is a *symmetric space*.
We now collect some results on the structure of symmetric spaces, referring to [@Ri82; @Sp85] for details and proofs. The identity component $K^0$ is reductive, and non-trivial unless $G$ is a $\theta$-*split torus*, i.e., a torus where $\theta$ acts via the inverse map $g \mapsto g^{-1}$.
A parabolic subgroup $P \subseteq G$ is said to be $\theta$-*split* if the parabolic subgroup $\theta(P)$ is opposite to $P$. The minimal $\theta$-split parabolic subgroups are all conjugate by elements of $K^0$; we choose such a subgroup $P$ and put $$L := P \cap \theta(P),$$ a $\theta$-stable Levi subgroup of $P$. The intersection $L \cap K$ contains the derived subgroup $[L,L]$; thus, every maximal torus of $L$ is $\theta$-stable. We choose such a torus $T$, so that $$\label{eqn:dec}
T = T^{\theta} T^{-\theta} \quad \text{and} \quad
T^{\theta} \cap T^{-\theta} \text{ is finite.}$$ Moreover, the identity component $$A := T^{-\theta,0}$$ is a maximal $\theta$-split subtorus of $G$. All such subtori are conjugate in $K^0$; their common dimension is the *rank* of the symmetric space $G/K$, denoted by ${\operatorname{rk}}(G/K)$. Moreover, $$C_G(A) = L = (L \cap K) A$$ (where $C_G(A)$ denotes the centralizer of $A$ in $G$), and $(L \cap K) \cap A = A \cap K$ consists of all elements of order $2$ of $A$.
The product $P K^0 \subseteq G$ is open, and equals $PK$; thus, $PK/K$ is an open subset of $G/K$, isomorphic to $P/P\cap K = P/L \cap K$. Let $P_u$ be the unipotent radical of $P$, so that $P = P_u L$. Then the map $$\label{eqn:id}
\iota : P_u \times A/A \cap K \to PK/K, \quad (g,x) \mapsto g \cdot x$$ is an isomorphism.
The character group ${\mathcal X}(A/A \cap K)$ may be identified with the subgroup $2 {\mathcal X}(A)\subset {\mathcal X}(A)$. On the other hand, $A/A \cap K \cong T/T \cap K$ and hence ${\mathcal X}(A/A \cap K)$ may be identified with the subgroup of ${\mathcal X}(T)$ consisting of those characters that vanish on $T \cap K$, i.e., $$\label{eqn:char}
{\mathcal X}(A/A \cap K) = \{ \chi - \theta(\chi) ~\vert~ \chi \in {\mathcal X}(T)\}.$$ Here $\theta$ acts on ${\mathcal X}(T)$ via its action on $T$.
Denote by $$\Phi_G \subset {\mathcal X}(T)$$ the root system of $(G,T)$, with Weyl group $$W_G = N_G(T)/T.$$ Choose a basis $\Delta_G$ consisting of roots of $P$. Let $\Phi^+_G \subset \Phi_G$ be the corresponding subset of positive roots and let $\Delta_L \subset \Delta_G$ be the subset of simple roots of $L$. The natural action of the involution $\theta$ on $\Phi$ fixes point-wise the subroot system $\Phi_L$. Moreover, $\theta$ exchanges the subsets $\Phi^+_G \setminus \Phi^+_L$ and $\Phi^-_G \setminus \Phi^-_L$ (the sets of roots of $P_u$ and of $\theta(P_u) = \theta(P)_u$).
Also, denote by $$p : {\mathcal X}(T) \to {\mathcal X}(A)$$ the restriction map from the character group of $T$ to that of $A$. Then $p(\Phi_G) \setminus \{0\}$ is a (possibly non-reduced) root system called the *restricted root system*, that we denote by $\Phi_{G/K}$. Moreover, $$\Delta_{G/K} := p(\Delta_G \setminus \Delta_L)$$ is a basis of $\Phi_{G/K}$. The corresponding Weyl group is $$\label{eqn:wg}
W_{G/K} = N_G(A)/C_G(A) \cong N_{K^0}(A)/C_{K^0}(A).$$ Also, $W_{G/K} \cong N_W(A)/C_W(A)$, and $N_W(A) = W_G^{\theta}$ whereas $C_W(A) = W_L$. This yields an exact sequence $$\label{eqn:ext}
1 \to W_L \to W_G^{\theta} \to W_{G/K} \to 1.$$
The wonderful compactification of an adjoint symmetric space {#subsec:wc}
------------------------------------------------------------
We keep the notation and assumptions of Subsec. \[subsec:ss\] and we assume, in addition, that $G$ is semi-simple and adjoint; equivalently, $\Delta_G$ is a basis of ${\mathcal X}(T)$. Then the symmetric space $G/K$ is said to be *adjoint* as well.
By [@DP83; @DS99], $G/K$ admits a canonical compactification: the *wonderful compactification* $X$, which satisfies the following properties.
\(i) $X$ is a nonsingular projective variety.
\(ii) $G$ acts on $X$ with an open orbit isomorphic to $G/K$.
\(iii) The complement of the open orbit is the union of $r = {\operatorname{rk}}(G/K)$ nonsingular prime divisors $X_1, \ldots, X_r$ with normal crossings.
\(iv) The $G$-orbit closures in $X$ are exactly the partial intersections $$X_I := \bigcap_{i \in I} X_i$$ where $I$ runs over the subsets of $\{1,\ldots,r\}$.
\(v) The unique closed orbit, $X_1 \cap \cdots \cap X_r$, is isomorphic to $G/P \cong G/\theta(P)$.
We say that $X$ is a *wonderful symmetric variety* with *boundary divisors* $X_1,\ldots,X_r$. By (iii) and (iv), each orbit closure $X_I$ is nonsingular.
Let $Y$ be the closure in $X$ of the subset $$A/A \cap K \cong AK/K = LK/K \subseteq G/K.$$ Then $Y$ is stable under the action of the subgroup $L N_K(A) \subseteq G$. Since $L \cap N_K(A) = L \cap K = C_K(A)$, and $N_K(A)/C_K(A) \cong W_{G/K}$ by (\[eqn:wg\]), we obtain an exact sequence $$1 \to L \to L N_K(A) \to W_{G/K} \to 1.$$ Moreover, since $Y$ is fixed point-wise by $L \cap K$, the action of $L N_K(A)$ factors through an action of the semi-direct product $$(L/L \cap K) \, W_{G/K} \cong (A/A \cap K) \, W_{G/K}.$$
The adjointness of $G$ and (\[eqn:char\]) imply that ${\mathcal X}(A/A \cap K)$ is the restricted root lattice, with basis $$\Delta_{G/K} = \{ \alpha - \theta(\alpha) ~\vert~
\alpha \in \Delta_G \setminus \Delta_L \}.$$ Moreover, $Y$ is the toric variety associated with the Weyl chambers of the restricted root system $\Phi_{G/K}$ as in Subsec. \[subsec:tv\]. This defines the open affine toric subvariety $Y_0 \subset Y$ associated with the positive Weyl chamber dual to $\Delta_{G/K}$. Note that $$\label{eqn:ywy}
Y = W_{G/K} \cdot Y_0.$$
Also, recall the local structure of the wonderful symmetric variety $X$: the subset $$X_0 := P \cdot Y_0 = P_u \cdot Y_0$$ is open in $X$, and the map $$\label{eqn:ls}
\iota: P_u \times Y_0 \to X_0, \quad (g,x) \mapsto g \cdot x$$ is a $P$-equivariant isomorphism. Moreover, any $G$-orbit in $X$ meets $X_0$ along a unique orbit of $P$, and meets transversally $Y_0$ along a unique orbit of $A/A\cap K$.
It follows that the $G$-orbit structure of $X$ is determined by that of the associated toric variety $Y$: any $G$-orbit in $X$ meets transversally $Y$ along a disjoint union of orbit closures of $A/A \cap K$, permuted transitively by $W_{G/K}$. As another consequence, $X_0 \cap G/K = PK/K$ and $Y_0 \cap G/K = AK/K$, so that $\iota$ restricts to the isomorphism (\[eqn:id\]).
Finally, the closed $G$-orbit $X_1 \cap \cdots \cap X_r$ meets $Y_0$ transversally at a unique point $z$, the $T$-fixed point in $Y_0$. The isotropy group $G_z$ equals $\theta(P)$, and the normal space to $G \cdot z$ at $z$ is identified with the tangent space to $Y$ at that point. Hence the weights of $T$ in the tangent space to $X$ at $z$ are the positive roots $\alpha \in \Phi^+_G \setminus \Phi^+_L$ (the contribution of the tangent space to $G \cdot z$), and the simple restricted roots $\gamma = \alpha - \theta(\alpha)$, where $\alpha \in \Delta_G \setminus \Delta_L$ (the contribution of the tangent space to $Y$).
Symmetric spaces of minimal rank {#subsec:ssm}
--------------------------------
We return to the setting of Subsect. \[subsec:ss\]. In particular, we consider a connected reductive group $G$ equipped with an involutive automorphism $\theta$, and the fixed point subgroup $K = G^{\theta}$.
Let $T$ be any $\theta$-stable maximal torus of $G$. Then (\[eqn:dec\]) implies that $${\operatorname{rk}}(G) \geq {\operatorname{rk}}(K) + {\operatorname{rk}}(G/K)$$ with equality if and only if the identity component $T^{\theta,0}$ is a maximal torus of $K^0$, and $T^{-\theta,0}$ is a maximal $\theta$-split subtorus. We then say that the symmetric space $G/K$ is *of minimal rank*; equivalently, all $\theta$-stable maximal tori of $G$ are conjugate in $K^0$.
We refer to [@Br04 Subsec. 3.2] for the proof of the following auxiliary result, where we put $$T_K := T^{\theta,0} = (T \cap K)^0.$$
\[lem:ro\] [(i)]{} The roots of $(K^0, T_K)$ are exactly the restrictions to $T_K$ of the roots of $(G,T)$.
[(ii)]{} The Weyl group of $(K^0,T_K)$ may be identified with $W^{\theta}_G$.
In particular, $C_G(T_K) = T$ by (i) (this may also be seen directly). We put $$W_K := W^{\theta}_G \quad \text{and} \quad
N_K := N_{K^0}(T) = N_{K^0}(T_K).$$ By (ii), this yields an exact sequence $$\label{eqn:ex}
1 \to T_K \to N_K \to W_K \to 1.$$ Moreover, by (\[eqn:ext\]), $W_K$ fits into an exact sequence $$\label{eqn:exa}
1 \to W_L \to W_K \to W_{G/K} \to 1.$$ The group $W_K$ acts on $\Phi_G$ and stabilizes the subset $\Phi_G^{\theta} = \Phi_L$; the restriction map $$q : {\mathcal X}(T) \to {\mathcal X}(T_K)$$ is $W_K$-equivariant and $\theta$-invariant. Denoting by $\Phi_K$ the root system of $(K^0,T_K)$, Lemma \[lem:ro\] (i) yields the equality $$\Phi_K = q(\Phi_G).$$
We now obtain two additional auxiliary results:
\[lem:fi\] Let $\beta \in \Phi_K$. Then one of the following cases occurs:
[(a)]{} $q^{-1}(\beta)$ consists of a unique root, $\alpha \in \Phi_L$.
[(b)]{} $q^{-1}(\beta)$ consists of two strongly orthogonal roots $\alpha$, $\theta(\alpha)$, where $\alpha \in \Phi^+_G \setminus \Phi^+_L$ and $\theta(\alpha) \in \Phi^-_G \setminus \Phi^-_L$.
In particular, $q$ induces a bijection $q^{-1}q(\Phi_L) = \Phi_L \cong q(\Phi_L)$. Moreover, $\alpha$ and $\theta(\alpha)$ are strongly orthogonal for any $\alpha \in \Phi_G \setminus \Phi_L$; then $s_\alpha s_{\theta(\alpha)} \in W_G^{\theta} = W_K$ is a representative of the reflection of $W_{G/K}$ associated with the restricted root $\alpha - \theta(\alpha)$.
Note that any $\alpha \in q^{-1}(\beta)$ is a root of $(C_G(S),T)$, where $S \subseteq T_K$ denotes the identity component of the kernel of $\beta$, and $C_G(S)$ stands for the centralizer of $S$ in $G$. Also, $C_G(S)$ is a connected reductive $\theta$-stable subgroup of $G$, and the symmetric space $C_G(S)/C_K(S)$ is of minimal rank. Moreover, $\theta$ yields an involution of the quotient group $C_G(S)/S$, and the corresponding symmetric space is still of minimal rank. So we may reduce to the case where $S$ is trivial, i.e., $K$ has rank $1$. Since $\beta$ is a root of $K^0$, it follows that $K^0 \cong {\operatorname{SL}}(2)$ or ${\operatorname{PSL}}(2)$. Together with the minimal rank assumption, it follows that one of the following cases occurs, up to an isogeny of $G$:
\(a) $K^0 = G = {\operatorname{PSL}}(2)$; then $\theta$ is trivial.
\(b) $K^0 = {\operatorname{PSL}}(2)$ and $G = {\operatorname{PSL}}(2) \times {\operatorname{PSL}}(2)$; then $\theta$ exchanges both factors.
This implies our assertions.
We will identify $q(\Phi_L)$ with $\Phi_L$ in view of Lemma \[lem:fi\].
\[lem:res\] [(i)]{} $q$ induces bijections $$\Phi_G^+ \setminus \Phi_L^+ \cong \Phi_K \setminus \Phi_L
\cong \Phi_G^- \setminus \Phi_L^- .$$
[(ii)]{} $\Phi_K \setminus \Phi_L$ is a root system, stable under the action of $W_K$ on $\Phi_K$.
[(iii)]{} The restricted root system $\Phi_{G/K}$ is reduced.
\(i) follows readily from Lemma \[lem:fi\].
\(ii) Since $\Phi_L$ is stable under $W_K$, then so is $\Phi_K \setminus \Phi_L$. In particular, the latter is stable under any reflection $s_\beta$, where $\beta \in \Phi_K \setminus \Phi_L$. It follows that $\Phi_K \setminus \Phi_L$ is a root system.
\(iii) We have to check the non-existence of roots $\alpha_1, \alpha_2 \in \Phi_G \setminus \Phi_L$ such that $\alpha_2 - \theta(\alpha_2) = 2(\alpha_1 - \theta(\alpha_1))$. Considering the identity component of the intersection of kernels of $\alpha_1$, $\theta(\alpha_1)$ and $\alpha_2$ and arguing as in the proof of Lemma \[lem:fi\], we may assume that ${\operatorname{rk}}(G) \leq 3$. Clearly, we may further assume that $G$ is semi-simple and adjoint. Then the pair $(G,K)$ is either $({\operatorname{PSL}}(2) \times {\operatorname{PSL}}(2), {\operatorname{PSL}}(2))$ or $({\operatorname{PSL}}(4),{\operatorname{PSp}}(4))$, and the assertion follows by inspection.
Also, note that the pull-back under $q$ of any system of positive roots of $\Phi_K$ is a $\theta$-stable system of positive roots of $\Phi_G$. Let $\Sigma_G$ be the corresponding basis of $\Phi_G$; then $q(\Sigma_G)$ is a basis of $\Phi_K$. If $G$ is semi-simple, then the involution $\theta$ is uniquely determined by the its restriction to $\Sigma_G$, and the latter is an involution of the Dynkin diagram of $G$. It follows that the symmetric spaces of minimal rank under an adjoint semi-simple group are exactly the products of symmetric spaces that occur in the following list.
1\) Let $G = {\mathbf G}\times {\mathbf G}$, where ${\mathbf G}$ is an adjoint semi-simple group, and let $\theta$ be the involution of $G$ such that $\theta(x,y) = (y,x)$; then $K = {\operatorname{diag}}({\mathbf G})$ and ${\operatorname{rk}}(G/K) = {\operatorname{rk}}({\mathbf G})$.
The maximal $\theta$-stable subtori $T \subset G$ are exactly the products ${\mathbf T}\times {\mathbf T}$, where ${\mathbf T}$ is a maximal torus of ${\mathbf G}$; then $T_K = {\operatorname{diag}}({\mathbf T})$, and $A = \{(x,x^{-1}) ~\vert~ x \in {\mathbf T}\}$. Thus, $L = T$, and $W_K = W_{G/K} = {\operatorname{diag}}(W_{{\mathbf G}}) \subset W_{{\mathbf G}} \times W_{{\mathbf G}} = W_G$.
Moreover, $\Phi_G$ is the disjoint union of two copies of $\Phi_{{\mathbf G}}$, each of them being mapped isomorphically to $\Phi_K = \Phi_{G/K}$ by $q$.
2\) Consider the group $G = {\operatorname{PSL}}(2n)$ and the involution $\theta$ associated with the symmetry of the Dynkin diagram; then $K = {\operatorname{PSp}}(2n)$ and ${\operatorname{rk}}(G/K) = n-1$.
The Levi subgroup $L$ is the image in $G = {\operatorname{PGL}}(2n)$ of the product ${\operatorname{GL}}(2) \times \cdots \times {\operatorname{GL}}(2)$ ($n$ copies). The Weyl group $W_G$ is the symmetric group $S_{2n}$, and $W_K$ is the subgroup preserving the partition of the set $\{1, 2, \ldots,2n \}$ into the $n$ subsets $\{1,2 \}$, $\{3,4 \}$, $\ldots$, $\{ 2n-1,2n \}$. Thus, $W_K$ is the semi-direct product of $S_2 \times \cdots \times S_2$ ($n$ copies) with $S_n$. Moreover, $W_L = S_2 \times \cdots \times S_2$, so that $W_{G/K} = S_n$.
For the root systems, we have $\Phi_G = A_{2n-1}$, $\Phi_K = C_n$, $\Phi_L = A_1 \times \cdots \times A_1$ (the subset of long roots of $\Phi_K$), $\Phi_K \setminus \Phi_L = D_n$ (the short roots), and $\Phi_{G/K} = A_{n-1}$.
3\) Consider the group $G = {\operatorname{PSO}}(2n)$ and the involution $\theta$ associated with the symmetry of the Dynkin diagram (of type $D_n$); then $K = {\operatorname{PSO}}(2n-1)$ and ${\operatorname{rk}}(G/K) = 1$.
The Levi subgroup $L$ is the image in $G$ of ${\operatorname{SO}}(2) \times {\operatorname{SO}}(2n-2)$. The Weyl group $W_G$ is the semi-direct product of $\{\pm 1 \}^{n-1}$ with $S_n$, and $W_K$ is the semi-direct product of $\{\pm 1 \}^{n-1}$ with $S_{n-1}$. Moreover, $W_L$ is the semi-direct product of $\{\pm 1 \}^{n-2}$ with $S_{n-1}$, so that $W_{G/K} = \{ \pm 1 \}$.
We have $\Phi_G = D_n$, $\Phi_K = B_{n-1}$, $\Phi_L = D_{n-1}$ (the subset of long roots of $\Phi_K$), $\Phi_K \setminus \Phi_L = A_1 \times \cdots \times A_1$ ($n-1$ copies), and $\Phi_{G/K} = A_1$.
4\) Let $G$ be an adjoint simple group of type $E_6$, and $\theta$ the involution associated with the symmetry of the Dynkin diagram; then $K$ is a simple group of type $F_4$, and ${\operatorname{rk}}(G/K) = 2$.
The Levi subgroup $L$ has type $D_4$, and $W_{G/K} = S_3$. We have $\Phi_G = E_6$, $\Phi_K = F_4$, $\Phi_L = D_4$ (the subset of long roots of $\Phi_K$), $\Phi_K \setminus \Phi_L = D_4$, and $\Phi_{G/K} = A_2$.
Equivariant Chow ring {#sec:gws}
=====================
Wonderful symmetric varieties of minimal rank {#subsec:fpc}
---------------------------------------------
From now on, we consider an adjoint semi-simple group $G$ equipped with an involutive automorphism $\theta$ such that the corresponding symmetric space $G/K$ is of minimal rank. Then the group $K$ is connected, semi-simple and adjoint; see [@Br04 Lem. 5].
We choose a $\theta$-stable maximal torus $T \subseteq G$, so that $A := T^{-\theta,0}$ is a maximal $\theta$-split subtorus. Also, we put $T_K := T^{\theta}$; this group is connected by [@Br04 Lem. 5] again. Thus, $T_K$ is a maximal torus of $K$. In agreement with the notation of Subsec. \[subsec:ssm\], we denote by $N_K$ the normalizer of $T_K$ in $K$, and by $W_K$ the Weyl group of $(K,T_K)$.
As in Subsec. \[subsec:wc\], we denote by $X$ the wonderful compactification of $G/K$, also called a *wonderful symmetric variety of minimal rank*. The associated toric variety $Y$ is the closure in $X$ of $T/T_K \cong A/A \cap K$. Recall that $Y$ is stable under the subgroup $L N_K \subseteq G$, and fixed point-wise by $L \cap K$. Thus, $L N_K$ acts on $Y$ via its quotient group $$L N_K / (L \cap K) \cong (T/T_K) \, W_{G/K} \cong T N_K/T_K.$$ We will mostly consider $Y$ as a $TN_K$-variety.
By [@Tc05 Sec. 10], $X$ contains only finitely many $T$-stable curves. We now obtain a precise description of all these curves, and of those that lie in $Y$. This may be deduced from the results of \[loc. cit.\], which hold in the more general setting of wonderful varieties of minimal rank, but we prefer to provide direct, somewhat simpler arguments.
\[lem:fix\] [(i)]{} The $T$-fixed points in $X$ (resp. $Y$) are exactly the points $w \cdot z$, where $w \in W$ (resp. $W_K$), and $z$ denotes the unique $T$-fixed point of $Y_0$. These fixed points are parametrized by $W_G/W_L$ (resp. $W_K/W_L \cong W_{G/K}$).
[(ii)]{} For any $\alpha \in \Phi^+_G \setminus \Phi^+_L$, there exists a unique irreducible $T$-stable curve $C_{z,\alpha}$ which contains $z$ and on which $T$ acts through its character $\alpha$. The $T$-fixed points in $C_{z,\alpha}$ are exactly $z$ and $s_\alpha \cdot z$.
[(iii)]{} For any $\gamma = \alpha - \theta(\alpha) \in \Delta_{G/K}$, there exists a unique irreducible $T$-stable curve $C_{z,\gamma}$ which contains $z$ and on which $T$ acts through its character $\gamma$. The $T$-fixed points in $C_{z,\gamma}$ are exactly $z$ and $s_\alpha s_{\theta(\alpha)} \cdot z$.
[(iv)]{} The irreducible $T$-stable curves in $X$ are the $W_G$-translates of the curves $C_{z,\alpha}$ and $C_{z,\gamma}$. They are all isomorphic to ${\mathbb P}^1$.
[(v)]{} The irreducible $T$-stable curves in $Y$ are the $W_{G/K}$-translates of the curves $C_{z,\gamma}$.
The assertions on the $T$-fixed points in $X$ are proved in [@Br04 Lem. 6]. And since $Y$ is the toric variety associated with the Weyl chambers of $\Phi_{G/K}$, the group $W_{G/K}$ acts simply transitively on its $T$-fixed points. This proves (i).
Let $C \subset X$ be an irreducible $T$-stable curve. Replacing $C$ with a $W_G$-translate, we may assume that it contains $z$. Then $C \cap X_0$ is an irreducible $T$-stable curve in $X_0$, an affine space where $T$ acts linearly with weights the positive roots $\alpha \in \Phi^+_G \setminus \Phi^+_L$, and the simple restricted roots $\gamma = \alpha - \theta(\alpha)$, $\alpha \in \Delta_G \setminus \Delta_L$. Since all these weights have multiplicity $1$, it follows that $C \cap X_0$ is a coordinate line in $X_0$. Thus, $C$ is isomorphic to ${\mathbb P}^1$ where $T$ acts through $\alpha$ or $\gamma$. In the former case, $C$ is contained in the closed $G$-orbit $G \cdot z$; it follows that its other $T$-fixed point is $s_\alpha \cdot z$. In the latter case, $C$ is contained in $Y$, and hence its other $T$-fixed point corresponds to a simple reflection in $W_{G/K}$. By considering the weight of the $T$-action on $C$, this simple reflection must be the image in $W_{G/K}$ of $s_{\alpha} s_{\theta(\alpha)} \in W_K$. This implies the remaining assertions (ii)-(v).
Structure of the equivariant Chow ring {#subsec:eq}
--------------------------------------
We will obtain a description of the $G$-equivariant Chow ring of $X$ with rational coefficients. For this, we briefly recall some properties of equivariant intersection theory, referring to [@Br97; @EG98] for details.
To any nonsingular variety $Z$ carrying an action of a linear algebraic group $H$, one associates the equivariant Chow ring $A^*_H(Z)$. This is a positively graded ring with degree-$0$ part ${\mathbb Z}$, and degree-$1$ part the equivariant Picard group ${\operatorname{Pic}}_H(Z)$ consisting of isomorphism classes of $H$-linearized invertible sheaves on $Z$.
Every closed $H$-stable subvariety $Y \subseteq Z$ of codimension $n$ yields an equivariant class $$[Y]_H \in A^n_H(Z).$$ The class $[Z]_H$ is the unit element of $A^*_H(Z)$.
Any equivariant morphism $f: Z \to Z'$, where $Z'$ is a nonsingular $H$-variety, yields a pull-back homomorphism $$f^* : A^*_H(Z') \to A^*_H(Z).$$ In particular, $A^*_H(Z)$ is an algebra over $A^*_H(pt)$, where $pt$ denotes ${\operatorname{Spec}}k$.
The equivariant Chow ring of $Z$ is related to the ordinary Chow ring $A^*(Z)$ via a homomorphism of graded rings $$\varphi_H: A^*_H(Z) \to A^*(Z)$$ which restricts trivially to the ideal of $A^*_H(Z)$ generated by $A^+_H(pt)$ (the positive part of $A^*_H(pt)$). If $H$ is connected, then $\varphi_H$ induces an isomorphism over the rationals: $$A^*_H(Z)_{{\mathbb Q}}/ A^+_H(pt) A^*_H(Z)_{{\mathbb Q}} \cong A^*(Z)_{{\mathbb Q}}.$$
More generally, there is a natural homomorphism of graded rings $$\varphi_H^{H'}: A^*_H(Z) \to A^*_{H'}(Z)$$ for any closed subgroup $H' \subset H$. If $H' = H^0$, the neutral component of $H$, then the group of components $H/H^0$ acts on the graded ring $A^*_{H^0}(Z)$, and the image of $\varphi_H^{H'}$ is contained in the invariant subring $A^*_{H^0}(Z)^{H/H^0}$. Moreover, $\varphi_H^{H^0}$ induces an isomorphism of rational equivariant Chow rings $$A^*_H(Z)_{{\mathbb Q}} \cong A^*_{H^0}(Z)^{H/H^0}_{{\mathbb Q}}.$$
If $H$ is a connected reductive group, and $T \subseteq H$ is a maximal torus with normalizer $N$ and associated Weyl group $W$, then the composite of the canonical maps $$A^*_H(Z) \to A^*_N(Z) \to A^*_T(Z)^W$$ is an isomorphism over the rationals. In particular, we obtain an isomorphism $$A^*_H(pt)_{{\mathbb Q}} \cong A^*_T(pt)_{{\mathbb Q}}^W.$$ Furthermore, $A^*_T(pt)$ is canonically isomorphic to the symmetric algebra (over the integers) of the character group ${\mathcal X}(T)$. This algebra will be denoted by $S_T$, or just by $S$ if this yields no confusion.
Returning to the $G$-variety $X$, we may now state our structure result:
\[thm:eq\] The map $$\label{eqn:iso}
r: A^*_G(X) \to A^*_T(X)^{W_G} \to A^*_T(X)^{W_K} \to
A^*_T(Y)^{W_K}$$ obtained by composing the canonical maps, is an isomorphism over the rationals.
We adapt the arguments of [@Br98 Sec. 3.1] regarding regular compactifications of reductive groups; our starting point is the precise version of the localization theorem obtained in [@Br97 Sec. 3.4]. Together with Lemma \[lem:fix\], it implies that the $T$-equivariant Chow ring $A^*_T(X)$ may be identified as an $S$-algebra to the space of tuples $(f_{w \cdot z})_{w \in W_G/W_L}$ of elements of $S$ such that $$f_{v\cdot z} \equiv f_{w \cdot z} \quad ({\operatorname{mod}}\chi)$$ whenever the $T$-fixed points $v \cdot z$ and $w \cdot z$ are joined by an irreducible $T$-stable curve where $T$ acts through its character $\chi$. This identification is obtained by restricting to the fixed points. The ring structure on the above space of tuples is given by pointwise addition and multiplication; moreover, $S$ is identified with the subring of constant tuples $(f)$.
It follows that $A^*_G(X)_{{\mathbb Q}} \cong A^*_T(X)_{{\mathbb Q}}^{W_G}$ may be identified, via restriction to $z$, with the subring of $S^{W_L}_{{\mathbb Q}}$ consisting of those $f$ such that $$\label{eqn:cong}
v^{-1} \cdot f \equiv w^{-1} \cdot f \quad ({\operatorname{mod}}\chi)$$ for all $v$, $w$ and $\chi$ as above. By Lemma \[lem:fix\] again, it suffices to check the congruences (\[eqn:cong\]) when $v = 1$. Then either we are in case (ii) of that lemma, and $w = s_\alpha$, or we are in case (iii) and $w = s_{\alpha} s_{\theta(\alpha)}$. In the former case, (\[eqn:cong\]) is equivalent to the congruence $$f \equiv s_{\alpha} \cdot f \quad ({\operatorname{mod}}\alpha),$$ which holds for any $f \in S_{{\mathbb Q}}$. In the latter case, we obtain $$\label{eqn:scong}
f \equiv s_{\alpha} s_{\theta(\alpha)} \cdot f \quad
({\operatorname{mod}}\alpha - \theta(\alpha)).$$ Thus, $A^*_G(X)_{{\mathbb Q}}$ is identified with the subring of $S^{W_L}_{{\mathbb Q}}$ defined by the congruences (\[eqn:scong\]) for all $\alpha \in \Delta_G \setminus \Delta_L$.
On the other hand, we may apply the same localization theorem to the $T$-variety $Y$. Taking invariants of $W_K$ and using the exact sequence (\[eqn:exa\]), we see that $A^*_T(Y)^{W_K}_{{\mathbb Q}}$ may be identified with the same subring of $S_{{\mathbb Q}}$, by restricting to the same point $z$. This implies our statement.
Further developments {#subsec:fd}
--------------------
We will obtain a more precise description of the ring $A^*_T(Y)^{W_K}_{{\mathbb Q}}$ that occurs in Theorem \[thm:eq\]. This will not be used in the sequel of this article, but has its own interest.
\[prop:dec\] [(i)]{} We have compatible isomorphisms of graded rings $$A^*_T(Y)^{W_K}_{{\mathbb Q}} \cong (S_{T_K}^{W_L} \otimes
A^*_{T/T_K}(Y))^{W_{G/K}}_{{\mathbb Q}}$$ and $$A^*_T(pt)^{W_K}_{{\mathbb Q}} \cong S^{W_K}_{T,{\mathbb Q}}
\cong (S_{T_K}^{W_L} \otimes S_{T/T_K})^{W_{G/K}}_{{\mathbb Q}}.$$
[(ii)]{} The image in $A^*_G(X)_{{\mathbb Q}} \cong A^*_T(Y)^{W_K}_{{\mathbb Q}}$ of the subring $$A^*_K(pt)_{{\mathbb Q}} \cong S_{T_K,{\mathbb Q}}^{W_K} \cong
(S_{T_K}^{W_L} \otimes {\mathbb Q})^{W_{G/K}}
\subseteq A^*_T(pt)^{W_K}_{{\mathbb Q}}$$ is mapped isomorphically to $A^*_G(G/K)_{{\mathbb Q}} \cong A^*_K(pt)_{{\mathbb Q}}$ under the pull-back from $X$ to the open orbit $G/K$.
[(iii)]{} We have isomorphisms $$\label{eqn:is}
{\operatorname{Pic}}(X)_{{\mathbb Q}} \cong {\operatorname{Pic}}_G(X)_{{\mathbb Q}} \cong
{\operatorname{Pic}}_T(Y)^{W_K}_{{\mathbb Q}} \cong {\operatorname{Pic}}_{T/T_K}(Y)^{W_K}_{{\mathbb Q}}$$ that identify the class $[X_i]$ of any boundary divisor with $$\label{eqn:sum}
[X_i \cap Y]_{T/T_K} = \sum_{w \in W_{G/K}^i} [Y_{i,w}]_{T/T_K}$$ where $Y_{i,w}$ denote the boundary divisors of $Y$, indexed as in Subsec. \[subsec:tv\].
\(i) Lemma \[lem:rel\] below yields a $W_K$-equivariant isomorphism of graded $S_T$-algebras $$A^*_T(Y) \cong S_T \otimes_{S_{T/T_K}} A^*_{T/T_K}(Y),$$ where $W_K$ acts on $S_T$ via its action on $T$, and on $A^*_{T/T_K}(Y)$ via its compatible actions on $T/T_K$ and $Y$. Moreover, ${\mathcal X}(T)_{{\mathbb Q}} \cong {\mathcal X}(T_K)_{{\mathbb Q}} \oplus {\mathcal X}(T/T_K)_{{\mathbb Q}}$ as $W_K$-modules, so that $$S_{T,{\mathbb Q}} \cong S_{T_K,{\mathbb Q}} \otimes S_{T/T_K, {\mathbb Q}}$$ as graded $W_K$-algebras. It follows that $$A^*_T(Y)_{{\mathbb Q}} \cong S_{T_K,{\mathbb Q}} \otimes A^*_{T/T_K}(Y)_{{\mathbb Q}}$$ as graded $S_{T,{\mathbb Q}}$-$W_K$-algebras. Taking $W_K$-invariants and observing that the action of $W_L \subseteq W_K$ on the right-hand side fixes pointwise $A^*_{T/T_K}(Y)_{{\mathbb Q}}$, we obtain the desired isomorphisms in view of the exact sequence (\[eqn:exa\]).
\(ii) Since $(G/K) \cap Y = T/T_K$, we obtain a commutative square $$\CD
A^*_G(X) @>{r}>> A^*_T(Y)^{W_K} \\
@VVV @V{t}VV \\
A^*_G(G/K) @>{s}>> A^*_T(T/T_K)^{W_K} \\
\endCD$$ where the vertical arrows are pull-backs, and $s$ is defined analogously to $r$. Moreover, $A^*_G(G/K) \cong A^*_K(pt)$, $A^*_T(T/T_K) \cong A^*_{T_K}(pt)$, and this identifies $s_{{\mathbb Q}}$ with the isomorphism $A^*_K(pt)_{{\mathbb Q}} \to S_{T_K,{\mathbb Q}}^{W_K}$. Likewise, $t_{{\mathbb Q}}$ is identified with the map $$(S_{T_K}^{W_L} \otimes A^*_{T/T_K}(Y))^{W_{G/K}}_{{\mathbb Q}}
\to
(S_{T_K}^{W_L} \otimes {\mathbb Q})^{W_{G/K}}$$ induced by the natural map $ A^*_{T/T_K}(Y)_{{\mathbb Q}} \to {\mathbb Q}$. This implies our assertion.
\(iii) Let ${\mathcal L}$ be an invertible sheaf on $X$, then some positive tensor power ${\mathcal L}^n$ admits a $G$-linearization, and such a linearization is unique since $G$ is semi-simple. This implies the first isomorphism of (\[eqn:is\]). The second isomorphism is a consequence of Theorem \[thm:eq\]. To show the third isomorphism, recall that ${\operatorname{Pic}}^T(Y) \cong A^1_T(Y)$, so that $${\operatorname{Pic}}_T(Y)^{W_K}_{{\mathbb Q}} \cong
({\mathcal X}(T_K)^{W_L} \oplus {\operatorname{Pic}}_{T/T_K}(Y))^{W_{G/K}}_{{\mathbb Q}}
\cong
{\mathcal X}(T_K)^{W_K}_{{\mathbb Q}} \oplus {\operatorname{Pic}}_{T/T_K}(Y)^{W_{G/K}}_{{\mathbb Q}}$$ by (i); moreover, ${\mathcal X}(T_K)^{W_K}_{{\mathbb Q}} =0$ since the group $K$ is semi-simple. Finally, (\[eqn:sum\]) follows from the decomposition of $X_i \cap Y$ into irreducible components $Y_{i,w}$, each of them having intersection multiplicity one.
\[lem:rel\] Let $Z$ be a nonsingular variety carrying an action of a torus $T$ and let $T' \subset T$ be a closed subgroup acting trivially on $Z$. Then there is a natural isomorphism of graded $S_T$-algebras $$A^*_T(Z) \cong S_T \otimes_{S_{T/T'}} A^*_{T/T'}(Z),$$ where $S_{T/T'}$ is identified with a subring of $S_T$ via the inclusion of ${\mathcal X}(T/T')$ into ${\mathcal X}(T)$.
In particular, if $T'$ is finite then there is a natural isomorphism of graded algebras over $S_{T,{\mathbb Q}} \cong S_{T/T',{\mathbb Q}}$: $$A^*_T(Z)_{{\mathbb Q}} \cong A^*_{T/T'}(Z)_{{\mathbb Q}}.$$
We begin by constructing a morphism of graded $S_{T/T'}$-algebras $$f: A^*_{T/T'}(Z) \to A^*_T(Z)$$ such that $f([Y]_{T/T'}) = [Y]_T$ for any $T$-stable subvariety $Y \subseteq Z$.
For this, we work in a fixed degree $n$ and consider a pair $(V,U)$, where $V$ is a finite-dimensional $T$-module, $U \subset V$ is a $T$-stable open subset such that the quotient $U \to U/T$ is a principal $T$-bundle, and the codimension of $V \setminus U$ is sufficiently large; see [@EG98] for details. Then we can form the mixed quotient $$Z \times^T U := (Z \times U)/T,$$ and we obtain $$\label{eqn:isom}
A^n_T(Z) = A^n(Z \times^T U) \cong A^n(Z \times^{T/T'} U/T')
\cong A^n_{T/T'}(Z \times U/T'),$$ where the latter isomorphism follows from the freeness of the diagonal $T/T'$-action on $Z \times U/T'$. Composing (\[eqn:isom\]) with the pull-back under the projection $Z \times U/T' \to Z$ yields a morphism $$f_n: A^n_{T/T'}(Z) \to A^n_T(Z).$$ One may check as in [@EG98] that $f_n$ is independent of the choice of $(V,U)$, and hence yields the desired morphism $f$.
Using the description of the $S_{T/T'}$-module $A^*_{T/T'}(Z)$ and of the $S$-module $A^*_T(Z)$ in terms of invariant cycles (see [@Br97 Thm. 2.1]), we obtain an isomorphism of graded $S_T$-modules $${\operatorname{id}}\otimes f : S_T \otimes_{S_{T/T'}} A^*_{T/T'}(Z) \to A^*_T(Z).$$
Next, we show that $A^*_G(X)_{{\mathbb Q}}$ is a free module over a big polynomial subring:
\[prop:hsp\] Let $R$ denote the ${\mathbb Q}$-subalgebra of $A^*_G(X)_{{\mathbb Q}}$ generated by the image of $A^*_K(pt)$ (defined in Proposition \[prop:dec\](ii)) and by the equivariant classes $[X_1]_G,\ldots,[X_r]_G$ of the boundary divisors. Then $R$ is a graded polynomial ring, and the $R$-module $A^*_G(X)_{{\mathbb Q}}$ is free of rank $\vert W_{G/K} \vert$.
By [@Br97 6.7 Corollary] and Lemma \[lem:fix\], $A^*_G(X)_{{\mathbb Q}}$ is a free module over $A^*_G(pt)_{{\mathbb Q}}$ of rank being the index $\vert W_G : W_L \vert$. As a consequence, the ring $A^*_G(X)_{{\mathbb Q}}$ is Cohen–Macaulay of dimension ${\operatorname{rk}}(G)$.
On the other hand, the $R$-module $A^*_G(X)_{{\mathbb Q}}$ is finite by [@BP02 Lemma 6] (the latter result is proved there in the setting of equivariant cohomology, but the arguments may be readily adapted to equivariant intersection theory).
Since $R$ is a quotient of a polynomial ring in ${\operatorname{rk}}(K) + r = {\operatorname{rk}}(G)$ variables, it follows that $R$ equals this polynomial ring. Moreover, $A^*_G(X)_{{\mathbb Q}}$ is a free $R$-module, since it is Cohen–Macaulay. This proves all assertions except that on the rank of the $R$-module $A^*_G(X)_{{\mathbb Q}}$, which may be checked by adapting the Poincaré series arguments of [@BP02].
Equivariant Chern classes {#sec:ecc}
=========================
The normal bundle of the associated toric variety {#subsec:nb}
-------------------------------------------------
We maintain the notation and assumptions of Subsec. \[subsec:fpc\]. In particular, $X$ denotes a wonderful symmetric variety of minimal rank, with associated toric variety $Y$.
Let ${\mathcal N}_{Y/X}$ denote the normal sheaf to $Y$ in $X$. This is a $LN_K$-linearized locally free sheaf on $Y$, which fits into an exact sequence of such sheaves $$\label{eqn:tn}
0 \to {\mathcal T}_Y \to {\mathcal T}_X\vert_Y \to {\mathcal N}_{Y/X} \to 0.$$ Here ${\mathcal T}_X$ denotes the tangent sheaf to $X$ (this is a $G$-linearized locally free sheaf on $X$) and ${\mathcal T}_X\vert_Y$ denotes its pull-back to $Y$.
The action of $G$ on $X$ yields a morphism of $G$-linearized sheaves $$\label{eqn:op}
op_X: {\mathcal O}_X \otimes {\mathfrak g}\to {\mathcal T}_X,$$ where ${\mathfrak g}$ denotes the Lie algebra of $G$. In turn, this yields a morphism of $T$-linearized sheaves ${\mathcal O}_Y \otimes {\mathfrak g}\to {\mathcal N}_{Y/X}$ which factors through another such morphism $$\label{eqn:gg}
\varphi : {\mathcal O}_Y \otimes {\mathfrak g}/{\mathfrak l}\to {\mathcal N}_{Y/X}$$ (where ${\mathfrak l}$ denotes the Lie algebra of $L$), since $Y$ is stable under $L$. Also, note the isomorphism of $T$-modules $${\mathfrak g}/{\mathfrak l}\cong \bigoplus_{\alpha \in \Phi_G \setminus \Phi_L} {\mathfrak g}_{\alpha}.$$ We may now formulate a splitting theorem for ${\mathcal N}_{Y/X}$:
\[thm:split\] [(i)]{} We have a decomposition of $T$-linearized sheaves $$\label{eqn:split}
{\mathcal N}_{Y/X} = \bigoplus_{\beta \in \Phi_K \setminus \Phi_L} {\mathcal L}_{\beta}$$ where each ${\mathcal L}_{\beta}$ is an invertible sheaf on which $T_K$ acts via its character $\beta$. The action of $N_K$ on ${\mathcal N}_{Y/X}$ permutes the ${\mathcal L}_{\beta}$’s according to the action of $W_K$ on $\Phi_K \setminus \Phi_L$.
[(ii)]{} The map (\[eqn:gg\]) restricts to surjective maps $$\varphi_{\beta}: {\mathcal O}_Y \otimes
({\mathfrak g}_{\alpha} \oplus {\mathfrak g}_{\theta(\alpha)}) \to {\mathcal L}_{\beta} \quad
(\beta =q(\alpha), ~\alpha \in \Phi_G^+ \setminus \Phi^+_L)$$ which induce isomorphisms of $T$-modules $$\Gamma(Y, {\mathcal L}_{\beta}) \cong {\mathfrak g}_{\alpha} \oplus {\mathfrak g}_{\theta(\alpha)}.$$ In particular, $\varphi$ is surjective, and each invertible sheaf ${\mathcal L}_{\beta}$ is generated by its global sections. Moreover, the corresponding morphism $$F_\beta: Y \to {\mathbb P}({\mathfrak g}_{\alpha} \oplus {\mathfrak g}_{\theta(\alpha)})^*
\cong {\mathbb P}^1$$ equals the morphism $f_{\alpha - \theta(\alpha)}$ (defined in Prop. \[prop:prod\]).
\(i) Since $Y$ is connected and fixed pointwise by $T_K$, each fiber ${\mathcal N}_{Y/X}(y)$, $y \in Y$, is a $T_K$-module, independent of the point $y$. Considering the base point of $G/K$ and denoting by ${\mathfrak t}$ (resp. ${\mathfrak k}$) the Lie algebra of $T$ (resp. $K$), we obtain an isomorphism of $T_K$-modules $${\mathcal N}_{Y/X}(y) \cong {\mathfrak g}/({\mathfrak t}+ {\mathfrak k}) \cong
\bigoplus_{\alpha \in \Phi^+_G \setminus \Phi_L^+} {\mathfrak g}_{\alpha}$$ which yields the decomposition (\[eqn:split\]) of the normal sheaf regarded as a $T_K$-linearized sheaf. Since the summands ${\mathcal L}_{\beta}$ are exactly the $T_K$-eigenspaces, they are stable under $T$, and permuted by $N_K$ according to their weights.
\(ii) Consider the restriction $$\varphi_0 : {\mathcal O}_{Y_0} \otimes {\mathfrak g}\to {\mathcal N}_{Y_0/X_0}.$$ By the isomorphism (\[eqn:ls\]), the composite map $$\label{eqn:lt}
{\mathcal O}_{Y_0} \otimes {\mathfrak p}_u \to {\mathcal O}_{Y_0} \otimes {\mathfrak g}\to {\mathcal N}_{Y_0/X_0}$$ is an isomorphism. Thus, $\varphi_0$ is surjective; by $N_K$-equivariance, it follows that $\varphi$ is surjective as well. Considering $T_K$-eigenspaces, this implies in turn the surjectivity of each $\varphi_{\beta}$.
Thus, the $T$-linearized sheaf ${\mathcal L}_{\beta}$ is generated by a $2$-dimensional $T$-module of global sections with weights $\alpha$ and $\theta(\alpha)$. This yields a $T$-equivariant morphism $F_{\beta}: Y \to {\mathbb P}^1$. Its restriction to the open orbit $T/T_K$ is equivariant of weight $\alpha - \theta(\alpha)$ for the action of $T$ by left multiplication; thus, we may identify this restriction with the character $\alpha - \theta(\alpha)$. Now Proposition \[prop:prod\] implies that $F_\beta = f_{\alpha - \theta(\alpha)}$. By (\[eqn:fxp\]) and the projection formula, it follows that the map $${\mathfrak g}_{\alpha} \oplus {\mathfrak g}_{\theta(\alpha)} =
\Gamma({\mathbb P}^1,{\mathcal O}_{{\mathbb P}^1}(1)) \to
\Gamma(Y,F_{\beta}^*{\mathcal O}_{{\mathbb P}^1}(1)) =
\Gamma(Y,{\mathcal L}_{\beta})$$ is an isomorphism.
The (logarithmic) tangent bundle {#subsec:tb}
--------------------------------
Recall that ${\mathcal T}_X$ denotes the tangent sheaf of $X$, consisting of all $k$-derivations of ${\mathcal O}_X$. Let ${\mathcal S}_X \subseteq {\mathcal T}_X$ be the subsheaf consisting of derivations preserving the ideal sheaf of the boundary $\partial X$. Since $\partial X$ is a divisor with normal crossings, the sheaf ${\mathcal S}_X$ is locally free; it is called the *logarithmic tangent sheaf* of the pair $(X,\partial X)$, also denoted by ${\mathcal T}_X (- \log \partial X)$.
Since $G$ acts on $X$ and preserves $X$, the map $op_X$ of (\[eqn:op\]) factors through a map $$\label{eqn:opr}
op_{X, \partial X} : {\mathcal O}_X \otimes {\mathfrak g}\to {\mathcal S}_X.$$ In fact, $op_{X,\partial X}$ is surjective; this follows, e.g., from the local structure of $X$, see [@BB96 Prop. 2.3.1] for details. In other words, ${\mathcal S}_X$ is the subsheaf of ${\mathcal T}_X$ generated by the derivations arising from the $G$-action.
Clearly, the sheaf ${\mathcal T}_X$ is $G$-linearized compatibly with the subsheaf ${\mathcal S}_X$. Moreover, the natural maps ${\mathcal T}_X \to {\mathcal N}_{X_i/X} \cong {\mathcal O}_X(X_i)\vert_{X_i}$ (where $X_1,\ldots,X_r$ denote the boundary divisors) fit into an exact sequence of $G$-linearized sheaves $$\label{eqn:st}
0 \to {\mathcal S}_X \to {\mathcal T}_X \to
\bigoplus_{i=1}^r {\mathcal N}_{X_i/X} \to 0,$$ see e.g. [@BB96 Prop. 2.3.2].
The pull-backs of $T_X$ and ${\mathcal S}_X$ to $Y$ are described by the following:
\[prop:ex\] [(i)]{} The exact sequence of $T N_K$-linearized sheaves $$0 \to {\mathcal T}_Y \to {\mathcal T}_X\vert_Y \to {\mathcal N}_{Y/X} \to 0$$ admits a unique splitting.
[(ii)]{} We also have a uniquely split exact sequence of $T N_K$-linearized sheaves $$\label{eqn:sn}
0 \to {\mathcal S}_Y \to {\mathcal S}_X \vert_Y \to
{\mathcal N}_{Y/X} \to 0,$$ where ${\mathcal S}_Y$ denotes the logarithmic tangent sheaf of the pair $(Y,\partial Y)$. Moreover, the $T N_K$-linearized sheaf ${\mathcal S}_Y$ is isomorphic to ${\mathcal O}_Y \otimes {\mathfrak a}$, where $T N_K$ acts on ${\mathfrak a}$ (the Lie algebra of $A$) via the natural action of its quotient $W_{G/K}$.
\(i) is checked by considering the $T_K$-eigenspaces as in the proof of Theorem \[thm:split\]. Specifically, the $T_K$-fixed part of ${\mathcal T}_X\vert_Y$ is exactly ${\mathcal T}_Y$, while the sum of all the $T_K$-eigenspaces with non-zero weights is mapped isomorphically to ${\mathcal N}_{Y/X}$.
\(ii) First, note that the natural map $$\label{eqn:ltb}
{\mathcal O}_Y \otimes {\mathfrak a}\to {\mathcal S}_Y$$ is an isomorphism, since $Y$ is a nonsingular toric variety under the torus $A/A \cap K$; see e.g. [@Od88 Prop. 3.1].
Next, consider the map ${\mathcal T}_X\vert_Y \to {\mathcal N}_{Y/X}$ and its restriction $$\pi : {\mathcal S}_X\vert_Y \to {\mathcal N}_{Y/X}.$$ Clearly, the kernel of $\pi$ contains the image of the natural map $$i: {\mathcal O}_Y \otimes {\mathfrak a}\to {\mathcal S}_X\vert_Y.$$ We claim that the resulting complex of $T N_K$-linearized sheaves $$\label{eqn:cplx}
{\mathcal O}_Y\otimes {\mathfrak a}\to {\mathcal S}_X\vert_Y \to {\mathcal N}_{Y/X}$$ is exact. By equivariance, it suffices to check this on $Y_0$. Then the local structure (\[eqn:ls\]) yields an exact sequence of $P$-linearized sheaves $$0 \to {\mathcal O}_{X_0} \otimes {\mathfrak p}_u \to
{\mathcal S}_X\vert_{X_0}
\to {\mathcal O}_{X_0}\otimes {\mathfrak a}\to 0,$$ see [@BB96 Prop. 2.3.1]. This yields, in turn, an isomorphism $${\mathcal O}_{Y_0}\otimes ({\mathfrak p}_u \oplus {\mathfrak a}) \cong
{\mathcal S}_X\vert_{Y_0}$$ which implies our claim by using the isomorphisms (\[eqn:lt\]) and (\[eqn:ltb\]).
In turn, this implies the exact sequence (\[eqn:sn\]); its splitting is shown by arguing as in (i).
\[cor:split\] We have isomorphisms of $T N_K$-linearized sheaves $${\mathcal T}_X\vert_Y \cong {\mathcal T}_Y \oplus
\bigoplus_{\beta \in \Phi_K \setminus \Phi_L} {\mathcal L}_{\beta},
\quad
{\mathcal S}_X\vert_Y \cong ({\mathcal O}_Y\otimes {\mathfrak a}) \oplus
\bigoplus_{\beta \in \Phi_K \setminus \Phi_L} {\mathcal L}_{\beta},$$ and an exact sequence of $T N_K$-linearized sheaves $$0 \to {\mathcal O}_Y \otimes {\mathfrak a}\to {\mathcal T}_Y \to
\bigoplus_{j=1}^m {\mathcal O}_Y(Y_j)\vert_{Y_j} \to 0,$$ where $Y_1, \ldots, Y_m$ denote the boundary divisors of the toric variety $Y$.
Equivariant Chern polynomials {#subsec:ec}
-----------------------------
By [@EG98], any $G$-linearized locally free sheaf ${\mathcal E}$ on $X$ yields *equivariant Chern classes* $$c_i^G({\mathcal E}) \in A^i_G(X) \quad (i= 0, 1,\ldots, {\operatorname{rk}}({\mathcal E}))$$ which we may encode by the *equivariant Chern polynomial* $$c_t^G({\mathcal E}) := \sum_{i=0}^{{\operatorname{rk}}({\mathcal E})} c_i^G({\mathcal E}) \; t^i.$$ The map $$r : A^*_G(X) \to A^*_T(Y)^{W_K}$$ of Theorem \[thm:eq\] sends $c_t^G({\mathcal E})$ to $c_t^T({\mathcal E}\vert_Y)$, by functoriality of Chern classes.
Together with the decompositions of the restrictions ${\mathcal T}_X\vert_Y$ and ${\mathcal S}_X\vert_Y$ (Corollary \[cor:split\]), this yields product formulae for the equivariant Chern polynomials of the $G$-linearized sheaves ${\mathcal T}_X$ and ${\mathcal S}_X$:
\[ecp\] With the above notation, we have equalities in $A^*_T(Y)$: $$r(c_t^G({\mathcal S}_X)) = \prod_{\beta \in \Phi_K \setminus \Phi_L}
(1 + t \; c_1^T({\mathcal L}_{\beta})),$$ $$r(c_t^G({\mathcal T}_X)) = \prod_{j=1}^m (1 + t \; [Y_j]_T) \times
\prod_{\beta \in \Phi_K \setminus \Phi_L} (1 + t \; c_1^T({\mathcal L}_{\beta})),$$ where $c_1^T({\mathcal L}_{\beta})\in {\operatorname{Pic}}_T(Y)$ denotes the equivariant Chern class of the $T$-linearized invertible sheaf ${\mathcal L}_{\beta}$, and $[Y_j]_T \in {\operatorname{Pic}}_T(Y)$ denotes the equivariant class of the boundary divisor $Y_j$.
(Note that the above products are all $W_K$-invariant, but their linear factors are not.)
Likewise, we may express the image under $r$ of the *equivariant Todd classes* ${\operatorname{td}}^G({\mathcal T}_X)$ and ${\operatorname{td}}^G({\mathcal S}_X)$: $$r({\operatorname{td}}^G({\mathcal S}_X)) = \prod_{\beta \in \Phi_K \setminus \Phi_L}
\frac{c_1^T({\mathcal L}_{\beta})}{1 - \exp(-c_1^T({\mathcal L}_{\beta}))},$$ $$r({\operatorname{td}}^G({\mathcal T}_X)) =
\prod_{j=1}^m
\frac{[Y_j]_T}{1 - \exp(-[Y_j]_T)} \times
\prod_{\beta \in \Phi_K \setminus \Phi_L}
\frac{c_1^T({\mathcal L}_{\beta})}{1 - \exp(-c_1^T({\mathcal L}_{\beta}))}.$$
Finally, we determine the equivariant Chern classes $c_1^T({\mathcal L}_{\beta}) \in {\operatorname{Pic}}^T(Y)$ in terms of the boundary divisors of $Y$, indexed as in Subsec. \[subsec:tv\]. Here $\beta \in \Phi_K \setminus \Phi_L$ and hence $\beta = q(\alpha)$ for a unique $\alpha \in \Phi^+_G \setminus \Phi^+_L$.
\[prop:rel\] With the preceding notation, we have $$\label{eqn:big}
c_1^T({\mathcal L}_{\beta}) = \alpha +
\sum_{i,w} \langle \alpha - \theta(\alpha), w \omega_i^{\vee} \rangle
\; [Y_{i,w}]_T,$$ where $\omega_i^{\vee}$ denote the fundamental co-weights of the restricted root system $\Phi_{G/K}$, and the sum runs over those pairs $(i,w) \in E(\Phi_{G/K})$ such that $w^{-1}(\alpha - \theta(\alpha)) \in \Phi^+_{G/K}$.
Recall from Theorem \[thm:split\] that the $T$-module of global sections of ${\mathcal L}_\beta$ is isomorphic to ${\mathfrak g}_{\alpha} \oplus {\mathfrak g}_{\theta(\alpha)}$. Let $s$ be the section of ${\mathcal L}_\beta$ associated with a generator of the line ${\mathfrak g}_\alpha$. Then $c_1^T({\mathcal L}_\beta) = \alpha + {\operatorname{div}}_T(s)$, since $s$ is a $T$-eigenvector of weight $\alpha$. Moreover, ${\operatorname{div}}_T(s)$ is the divisor of zeroes of the character $\alpha - \theta(\alpha)$. Together with (\[eqn:div\]), this implies the equation (\[eqn:big\]).
[100]{}
E. Bifet, C. De Concini and C. Procesi, *Cohomology of regular embeddings*, Adv. Math. **82** (1990), 1–34.
F. Bien and M. Brion, *Automorphisms and deformations of regular varieties*, Compositio Math. **104** (1996), 1–26.
N. Bourbaki, *Groupes et algèbres de Lie. Chapitres 4, 5 et 6*, Masson, Paris, 1981.
M. Brion, *Equivariant Chow groups for torus actions*, Transformation Groups **2** (1997), 225–267.
M. Brion, *The behaviour at infinity of the Bruhat decomposition*, Comment. Math. Helv. **73** (1998), 137–174.
M. Brion and E. Peyre, *The virtual Poincaré polynomials of homogeneous spaces*, Compositio Math. **134** (2002), 319–335.
M. Brion, *Construction of equivariant vector bundles*, preprint, arXiv: math.AG/0410039; to appear in the Proceedings of the International Conference on Algebraic Groups (Mumbai, 2004).
C. De Concini and C. Procesi, *Complete symmetric varieties I*, pp. 1–44 in: Lecture Note in Math. **996**, Springer–Verlag, New York, 1983.
C. De Concini and C. Procesi, *Complete symmetric varieties II. Intersection theory*, pp. 481–513 in: Advanced Studies in Pure Math. **6**, North Holland, Amsterdam, 1985.
I. Dolgachev and V. Lunts, *A character formula for the representation of a Weyl group in the cohomology of the associated toric variety*, J. Algebra [**168**]{} (1994), 741–772.
C. De Concini and T. A. Springer, *Compactification of symmetric varieties*, Transformation Groups **4** (1999), 273–300.
D. Edidin and W. Graham, *Equivariant intersection theory*, Invent. math. [**131**]{} (1998), 595–644.
M. Goresky, R. Kottwitz and R. MacPherson, *Equivariant cohomology, Koszul duality, and the localization theorem*, Invent. Math. **131** (1999), 25–83.
R. Hartshorne, *Algebraic Geometry*, Grad. Text Math. **52**, Springer–Verlag, New York, 1977.
V. Kiritchenko, *Chern classes of reductive groups and an adjunction formula*, Ann. Inst. Fourier (Grenoble) **56** (2006), 1225–1256.
P. Littelmann and C. Procesi, *Equivariant cohomology of wonderful compactifications* in: Operator Algebras, Unitary Representations, Enveloping Algebras, and Invariant Theory, Progress in Mathematics **92**, Birkhäuser, 1990.
T. Oda, *Convex Bodies and Algebraic Geometry*, Ergebnisse der Math. **15**, Springer-Verlag, 1988.
C. Procesi, *The toric variety associated to Weyl chambers*, Mots, 153–161, Lang. Raison. Calc., Hermès, Paris, 1990.
N. Ressayre, *Spherical homogeneous spaces of minimal rank*, preprint (2004), available at www.math.uni-montp2.fr/\~ressayre/spherangmin.pdf
R. W. Richardson, *Orbits, invariants and representations associated to involutions of reductive groups*, Invent. math. [**66**]{} (1982), 287–312.
T. A. Springer, *Some results on algebraic groups with involutions*, 525–543, Adv. Stud. Pure Math. **6**, North Holland, Amsterdam, 1985.
T. A. Springer, *Linear Algebraic Groups*, second edition, Progress in Mathematics **9**, Birkhäuser, 1998.
E. Strickland, *Equivariant cohomology of the wonderful group compactification*, J. Algebra **306** (2006), 610–621.
A. Tchoudjem, *Cohomologie des fibrés en droites sur les variétés magnifiques de rang minimal*, preprint, arXiv: math.AG/0507581.
V. Uma, *Equivariant $K$-theory of compactifications of algebraic groups*, preprint, arXiv: math.AG/0512187; to appear in Transformation Groups.
[^1]: The second author thanks the IHES, the MPI and the NSA for support
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Using numerical arguments we find that for $N$ = 306 a tetrahedral configuration ($T_h$) and for $N=542$ a dihedral configuration ($D_5$) are likely the global energy minimum for Thomson’s problem of minimizing the energy of $N$ unit charges on the surface of a unit conducting sphere. These would be the largest $N$ by far, outside of the icosadeltahedral series, for which a global minimum for Thomson’s problem is known. We also note that the current theoretical understanding of Thomson’s problem does not rule out a symmetric configuration as the global minima for $N=306$ and 542. We explicitly find that analogues of the tetrahedral and dihedral configurations for $N$ larger than 306 and 542, respectively, are not global minima, thus helping to confirm the theory of Dodgson and Moore (Phys. Rev. B 55, 3816 (1997)) that as $N$ grows dislocation defects can lower the lattice strain of symmetric configurations and concomitantly the energy. As well, making explicit previous work by ourselves and others, for $N<1000$ we give a full accounting of icosadeltahedral configuration which are not global minima and those which appear to be, and discuss how this listing and our results for the tetahedral and dihedral configurations may be used to refine theoretical understanding of Thomson’s problem.'
author:
- Eric Lewin Altschuler
- 'Antonio Pérez–Garrido'
title: 'New Global Minima for Thomson’s Problem of Charges on a Sphere '
---
What configuration of $N$ unit charges on the (surface) of a unit conducting sphere minimizes the Coulombic energy $\sum_{i\ne j}1/r_{ij}$ [@1]? Beyond physics this question has utility in assembly of biological [@2] and chemical structures [@3; @4], to benchmark optimization algorithms, and, as well, in mathematics Smale [@5] has noted this question to be a [*Hilbert problem*]{} for the Twenty-First Century. For $2\le
N\le 100$, the original question asked by J.J. Thomson a century ago, and a bit beyond, there is agreement of all numerical [@6; @7; @8; @9; @10] and theoretical [@11] methods suggesting that the global minimum configuration has been found. However, for larger $N$ owing to an exponential growth in good local minima [@7], finding global minima, general principles or insights for minimization, or even methods or cases to test hypotheses has proven extremely difficult. For $N = 10(h^2 + k^2 + hk) + 2$ highly symmetric icosadeltahedral configurations can be constructed. While it was initially thought that such configurations might be global minima [@12], for large $N$ adding defects to the icosadeltahedral lattice lowers the energy [@13; @14; @15; @Toomre; @Bowick]. Here we note a tetrahedral configuration for $N = 306$ and a dihedral configuration for $N=542$ which based on numerical arguments appear to be global minima, the largest such $N$ by far, aside from the icosadeltahedral series, for which a global energy minimum configuration is known. Study of tetrahedral analogues larger than 306 and dihedral analogues larger than 542 help confirm the theory [@16] on defects lowering lattice strain and energy. As well, we note that the fact that lattice configurations fail to be global minima for $N > 800$, may help explain why Mo$_{154}$ anions self-assemble into a spherical superstructure with a non-lattice number of subunits [@4].
For $N = 78$, as originally suggested by Edmunson [@11], the presumed global minimum configuration has tetrahedral ($T_h$) symmetry (Figure 1$a$) [@7; @8; @9; @10; @11]. We had previously suggested that an analogue of this configuration with 306 charges (Figure 1$b$, see below for method of construction of this analogue) also might be a global minimum and done limited numerical testing of this idea [@12]. Now, here we have extensively checked numerically on over one thousand runs, and have found no configuration of lower energy. (Numerically to look for non-lattice configurations with energies lower than the lattice configuration we used random initial configurations followed by a local steepest descent method. For $N = 306$ we could not find a configuration in one thousand runs with an energy lower than the lattice energy, though for many other $N$ considered in this paper including $N =$ 1218 and 4866, and smaller $N$ with icosadeltahedral lattices (see Tables I and II, and Figure 2) configurations with energies lower than the lattice energy could be found in only fifteen runs.) But one thousand runs is orders of magnitude less than the estimated [@7] over 1.5 million local minima for $N = 306$. As well, without an analytic proof we could not be certain the tetrahedral configuration in Figure 1$b$ is the one of minimum energy for $N = 306$. Current theoretical understanding of Thomson’s problem [@13; @Toomre; @Bowick; @16; @17], to be discussed below, does not exclude the possibility of a symmetric configuration as the global energy minimum for $N=306$. Of course, our configuration stands open to challenge.
Using a genetic algorithm Morris, Deaven and Ho [@9] confirmed previously found [@10] presumed global minima for $N\le 112$ and gave their likely global minima for $N\le 200$. Fitting the energy of these minima for $N\le 200$ to the function: $E=N^2/2\left( 1+aN^{-1/2}+bN^{-3/2} \right)$, (see [@9] and refs. therein for an explanation of why this function was used) for $N > $100 they found particularly deep minima with respect to this function for icosadeltahedral configurations for $N =$ 122, 132 192 and for dihedral $D_5$ configurations for $N =$ 137, 182 and 187 and $D_2$ configuration for N = 146. By our numerical testing as well these dihedral configurations appear to be global minima, though we have no analytic proof and numerically we are orders of magnitude short \[7\] to even sample a majority of local minima. We next looked at the higher split analogues of these dihedral configurations with $4N- 6$ charges. (See below for the method of construction of such analogues.) The next larger dihedral analogues for $N =$ 146, 182 and 187–$N =$ 578, 722, and 742 respectively–are found not to be global minima after only a few runs (E.L.A. and A.P.G. data not shown). However, the next larger dihedral ($D_5$) analogue of $N =$ 137–$N =$ 542–after over one thousand runs appears possibly to be a global minima. As for the tetrahedral configuration for $N =$ 306, we have no analytic proof of this proposition and given the huge number of local minima for an $N$ this large [@7] our numerical runs only begin to address the question. Conversely, we easily found that the next analogue of $N =$ 137, $N =$ 2162, is explicitly not a global energy minimum. See Figure 3 for dihedral configurations for $N =$ 137, 542 and 2162.
Euler’s theorem asserts that when a (convex) polyhedron is constructed by joining points on a sphere the number of faces ($F$) plus the number of vertices ($V$) is equal to the number of edges $(E) + 2; F + V = E + 2$. For $N> 12$ this has the result that in addition to the sixfold coordinated points of a planar two-dimensional lattice (hexamers), there must be at least twelve points of fivefold coordination (pentamers). The tetrahedral configurations for $N = 78$ and 306, dihedral configurations for $N=137$ and 542, and icosadeltahedral configurations for $N = 10(h^2 + k^2 + hk) + 2$ have exactly twelve pentamers and the rest hexamers (see, e.g., Figure 1$a$, $b$, $c$, $e$, Figure 2$a$, $b$ and Figure 3$a$, $b$, $c$). Larger analogues ([*split*]{} configurations) of the tetrahedral configuration for $N = 78$ (for $N = $306, 1218, 4886 and for the dihedral configuration for $N =$ 137 ($N =$ 542, 2162) (Figure 1$b$, $c$, $e$ and Figure 3$b$, $c$) are made as follows: in addition to the $N$ charges preexisting place one charge at the center of each of the $3N-6$ edges. (If all the charges were sixfold coordinated hexamers there would be $3N$ edges, six must be subtracted from this to take into account the twelve fivefold coordinated pentamers.) Then relax to the final position by a local gradient method. The resulting configuration has $N + (3N-6) = 4N-6 $ charges. (See Figure 1$b, c, e$). Some icosadeltahedral configurations can be made ([*split*]{}) as analogues of smaller ones (see e.g. $N = 1242$ in Figure 2$b$, a larger analogue of $N = 312$ in Figure 2$a$). A method for making icosadeltahedral configurations [*de novo*]{} has been discussed previously [@12].
Conversely, to the cases of $N = 78$ and 306, and $N=$ 137 and 542 for the larger analogues we have studied we find that adding dislocation defects to the lattice produces a configuration with lower energy (Fig. 1$c-f$ and Fig. 3$c,d$ ). Similarly, for $N > 792$ icosadeltahedral configurations with dislocation defects, additional fivefold coordinated points and then necessary also sevenfold coordinated points (septamers), have lower energy than symmetric lattice configurations, while for smaller $N$ the symmetric lattice configurations appear to be global minima (see Tables I and II, see Figure 2 for an example), though further numerical testing may show that some such configurations are not global minima.
These numerical results on the $N$ at which tetrahedral, dihedral and icosadeltahedral configurations fail to be global energy minima are in remarkable concordance with a theory given by Dodgson and Moore [@13] originally for icosadeltahedral configurations: Using continuum elasticity theory [@16] they studied the energy cost of a pair of pentamers, compared with a pure hexagonal lattice, and suggested that dislocation defects–extra fivefold coordinated points, with (necessarily) paired sevenfold coordinated points–would lower lattice strain and energy for $N$ in the $\approx$ 500–1000 range. Similar reasoning should apply to the tetrahedral configurations in the $N = 78$ series and dihedral configurations in the $N=137$ series. Our results given here are strong confirmation of Dodgson and Moore’s theory [@13].
The fact that apparently for $N > 792$ all symmetric tetrahedral, dihedral or icosadeltahedral lattices are not global minima, along with the exponential growth in good local minima may help explain why the number of Mo$_{154}$ anions which self-assemble into a spherical superstructure [@4] is a non–lattice number 1165, rather than, for example, 1172 the closest icosadeltahedral lattice or the tetrahedral lattice at 1218, while for small $N$ self-assembly often produces a symmetric lattice configuration [@2; @3].
For icosadeltahedral configurations for $N\le $ 792 whether or not a lattice configuration is a potential global minimum depends not only on the magnitude of $N$, but also apparently on the details of the lattice itself. All lattice $N$ are listed in Tables I and II. As can be seen for $N$ = 42, 92, 162, 252, 362, 432, 492, 572, 642, 732 the icosahedral lattice configuration is manifestly not the global minimum, while for the other $N$, the lattice configuration appears to be so.
In an icosadeltahedral lattice $N$ = 10($h^2 + k^2 + hk$) + 2, to go from the center of one pentamer to the center of an adjacent pentamer one moves $h$ steps along one basis vector fo the lattice, and then $k$ steps in the other. We noted previously [@12] (also discussed in ref. [@17]) that the energy in an icosadeltahedral lattice configuration with a large ratio of $h$ to $k$ ($h\ge k$) may be increased due to the vertices of the pentamers being closely aliged (or perfectly aligned in a lattice with $k = 0$). It has previously been noted that as $N$ grows the icosadeltahedral lattice configuration may not be the global minimum [@13; @14; @15; @Toomre; @Bowick], though we have not seen any explicit published accounting of the $N$ for which the lattice fails to be a global minimum. This is given in Tables I and II (in addition to the rule that for $N > 792$ the lattice is not the global minimum). A clear pattern emerges: for an icosadeltahedral configuration with $k$ = 0, besides the extremely exceptional case of $N$ =12, the icosadeltahedral lattice configuration is not the global minimum. For the three small $N=$ 42, 92 and 162 the apparent global minimum configuration has exactly twelve pentamers, and no dislocation defects, but arranged in an non-icosadeltahedral configuration, likely lowering the energy cost of having the vertices of pentamers aligned. For $N$ larger than 162 in the $k$ = 0 series, the apparent global minimum incorporates dislocation defects. For $N = 432$ and higher for lattices with $k = 1$, the lattice is also not the global minimum in accordance with the notion stated above that the energy of the lattice is increased by relative alignment of the vertices of the pentamers. The data in Tables I and II may be useful in refining theoretical predictions for global enery minima: The current theory [@16; @13] correctly predicts that for $N>1000$ icosadeltahedral and tetrahedral configurations will not be global minima, but does not yet account for various cases for $N<1000$. As well, for $k=0$ other theoretical work [@Bowick; @17] predicts dislocations lowering the energy for $N>300$, but the first instance of this is found for $N=252$.
For tetrahedral or dihedral lattices we have not yet been able to find an obvious rule or principle to predict for which $N$ the lattice configuration is a global minimum. Indeed, while all groups using a variety of different methods find the $T_h$ configuration a global minimum for $N = 78$ [@10], and we have found similarly for $N = 306$, there are a number of $N < 100$ for which $T_h$ configurations are not the global minima [@10; @11], and also we have found that the next analogue of the global minimum for $N = 100$ which has $T$ (though not $T_h$) symmetry, $N = 394$, is not a global minimum (E.L.A. and A.P.G., data not shown). And we see no obvious difference between the $N = 78$ lattice and the others to explain why not only for $N = 78$, but the next higher analogue the lattice appears to be the global minimum. Similarly, we have appreciated no obvious reason why the next $D_2$ analogue of 146–578–or the next $D_5$ analogues of 182 and 187–722 and 742–are not global minima, but the next $D_5$ analogue of 137–542–appears to also be a global minimum. The theory of Dodgson and Moore [@13] does not predict [*a priori*]{} that defects would lower a lattice energy for an $N$ of 306 and 542, as it does for the next analogues 1218 and 2162. Perhaps, whatever still unknown reasons that explains the good minimum for the $T_h$ lattice for $N = 78$ and $D_5$ lattice for $N=137$ also permit the $N = 306$ and 542 analogues to be global minima.
We thank the anonymous reviewers for comments extremely helpful in revising the manuscript and also inspirational leading to finding a new possible global minimum. A.P.G. would like to aknowledge financial support from Spanish MCyT under grant No. MAT2003–04887.
[99]{} J.J. Thomson, Philosophical Magazine 7, 237-265 (1904). D. Caspar, and A. Klug, Cold Spring Harb. Symp. Quant. Biol. 27, 1 (1962). H.W. Kroto, J.R. Heath, S.C. O’Brien, R.F. Curl, and R.E. Smalley, Nature 318, 162 (1985). T. Liu, E. Diemann, H. Li, A.W. Dress, and A. Muller, Nature 426, 59 (2003). S. Smale, Math. Intelligencer 20 (2), 7 (1998). E.L. Altschuler, T.J. Williams, E.R. Ratner, F. Dowla, and F. Wooten, Phys. Rev. Let. 72, 2671 (1994). T. Erber, and G.M. Hockney, G. M. Phys. Rev. Let. 74, 1482 (1995). A. Pérez–Garrido, M. Ortuño, E. Cuevas and J. Ruiz, J. Phys. A 29, 1973 (1996). J.R. Morris, D.M. Deaven and K.M. Ho, Phys. Rev. B 53, R1740 (1996). T. Erber T. and G.M. Hockney, Adv Chem Phys. 98, 495 (1997). J.R. Edmundson, Acta Cryst. A49, 648 (1993). E.L. Altschuler, T.J. Williams, E.R. Ratner, R. Tipton, R. Stong, F. Dowla and F. Wooten, Phys. Rev. Lett. 78, 2681 (1997). M.J.W. Dodgson and M.A. Moore, Rev. B 55, 3816 (1997). A. Pérez–Garrido, M.J.W. Dodgson, M.A. Moore, M. Ortuño and A. Díaz–Sánchez, Phys. Rev. Lett. 79, 1417 (1997). A. Pérez–Garrido, M.J.W. Dodgson and M.A. Moore, Phys. Rev. B 56, 3640 (1997). A. Toomre, unpublished results, discussed in \[17,19\] M. J. Bowick, D. R. Nelson and A. Travesset, Phys. Rev. B 62, 8738 (2000) M.J.W. Dodgson, J. Phys. A 29, 2499 (1996). M. Bowick, A. Cacciuto, D.R. Nelson, and A. Travesset, Phys. Rev. Lett. 89, 185502 (2002).
=.5
=.5
=.5
Charges Energy
--------- ------------------------ ------------------
12 h= 1 k= 0 49.1652530580000
32 h= 1 k= 1 412.261274651000
42 h= 2 k= 0 732.256241038000
\*Non–icosadeltahedral 732.078107551000
72 h= 2 k= 1 2255.00119099000
92 h= 3 k= 0 3745.61873908500
\*Non–icosadeltahedral 3745.29163624500
122 h= 2 k= 2 6698.37449926100
132 h= 3 k= 1 7875.04534281600
162 h= 4 k= 0 11984.5514338730
\*Non–icosadeltahedral 11984.0503358310
192 h= 3 k= 2 16963.3383864710
212 h= 4 k= 1 20768.0530859690
252 h= 5 k= 0 29544.2821928610
\* w/defects 29543.7859569610
272 h= 3 k= 3 34515.1932926880
282 h= 4 k= 2 37147.2944184740
312 h= 5 k= 1 45629.3627238190
362 h= 6 k= 0 61720.0233978130
\* w/defects 61719.3090545160
372 h= 4 k= 3 65230.0271225660
392 h= 5 k= 2 72546.2583708950
: Energy of icosadeltahedral configurations with $N<400$. An ‘\*’ indicates a configuration of lower energy, though not necessarily the global minimum. For each $N$, here and Table II, we tried fifteen runs–random initial configurations followed by a local gradient descent–to find a configuration with an energy lower than that of the icosadeltahedral lattice. For $N = 42$, 92 and 162 the best known configuration has exactly twelve pentamers (and thus no dislocation defects), but does not have icosadeltahedral symmetry. We note that the “split” larger analogue of the global minimum configuration for $N = 42$ is not the global minimum configuration for $N = 162$, and neither are the split larger analogues of the $N = 92$ and 162 global minima (data not shown).
-1cm
------ -------------- ------------------
432 h= 6 k= 1 88354.2293807250
\* w/defects 88354.1906652260
482 h= 4 k= 4 110318.139920155
492 h= 7 k= 0 115006.982258289
\*h= 5 k= 3 115005.255889700
522 h= 6 k= 2 129655.833007858
572 h= 7 k= 1 156037.879346228
\* w/defects 156037.316647696
612 h= 5 k= 4 178910.494981768
642 h= 8 k= 0 197100.363816212
\* w/defects 197098.637958037
672 h= 7 k= 2 216171.432658341
732 h= 8 k= 1 256975.527362500
\* w/defects 256974.262894426
752 h= 5 k= 5 271362.588212841
762 h= 6 k= 4 278704.548700071
792 h= 7 k= 3 301321.818305597
812 h= 9 k= 0 316895.372099956
\* w/defects 316892.668538128
842 h= 8 k= 2 340988.383415978
\* w/defects 340987.675098937
912 h= 9 k= 1 400662.383224662
h= 6 k= 5 400660.132041002
\* w/defects 400659.747279004
932 h= 7 k= 4 418596.898209635
\* w/defects 418595.636527970
972 h= 8 k= 3 455654.618623736
\* w/defects 455653.441695822
1082 h= 6 k= 6 565703.908873765
\* w/defects 565703.766602964
------ -------------- ------------------
: Energy of icosadeltahedral configurations with $N>400$. An ‘\*’ indicates a configuration of lower energy, though not necessarily the global minimum. Thus, for $N < 792$ we also cannot be certain that the symmetric icosadeltahedral configurations are the global minima.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We present a cooling method that should be generally applicable to atoms with narrow optical transitions. This technique uses velocity-selective pulses to drive atoms towards a zero-velocity dark state and then quenches the excited state to increase the cooling rate. We demonstrate this technique of quenched narrow-line cooling by reducing the 1-D temperature of a sample of neutral $^{40}$Ca atoms. We velocity select and cool with the $^1$S$_0$(4s$^2$)$\rightarrow$$^3$P$_1$(4s4p) 657 nm intercombination line and quench with the $^3$P$_1$(4s4p)$\rightarrow$$^1$S$_0$(4s5s) intercombination line at 553 nm, which increases the cooling rate eight-fold. Limited only by available quenching laser power, we have transferred 18 % of the atoms from our initial 2 mK velocity distribution and achieved temperatures as low as 4 $\mu$K, corresponding to a v$_{rms}$ of 2.8 cm/s or 2 recoils at 657 nm. This cooling technique, which is closely related to Raman cooling, can be extended to three dimensions.'
address: |
Time and Frequency Division, National Institute of Standards and Technology,\
325 Broadway, Boulder, Colorado 80305
author:
- 'E. A. Curtis$^*$, C. W. Oates, and L. Hollberg'
title: 'Quenched Narrow-Line Laser Cooling of $^{40}$Ca to Near the Photon Recoil Limit '
---
Following the great success of sub-Doppler laser cooling techniques with alkali atoms, several groups are now striving to achieve similar cooling results for the alkaline-earth atoms, with applications ranging from Bose-Einstein condensation to atomic interferometry and optical frequency standards. Unfortunately, the lack of magnetic and hyperfine sub-structure in the ground states of the predominant alkaline-earth isotopes precludes the usual sub-Doppler mechanisms such as polarization-gradient cooling [@Dalibard; @Chu], velocity-selective coherent population trapping [@Aspect], or Raman cooling [@Kasevich]. Furthermore, the broad cooling transitions for the alkaline-earth atoms have large Doppler limits, and typically yield magneto-optic trap (MOT) temperatures of more than 1 mK [@Shimizu; @Mg; @PTBMOT]. These complications have necessitated the development of alternative cooling strategies to reach the microkelvin temperatures “routinely" achieved with several alkali species. This paper presents a new approach, termed “quenched narrow-line laser cooling" (QNLC), which uses a narrow optical transition to provide increased velocity selectivity for laser cooling, and quenching of the upper state of this transition to enhance the cooling efficiency. To demonstrate the potential of this technique, we have used repeated velocity-selective stepwise excitation with two intercombination lines of neutral $^{40}$Ca to transfer as much as 18 % of the atoms in an initial velocity distribution with a 2 mK temperature into a narrow peak. The lowest 1-D temperature observed was 4 $\mu$K with a corresponding transfer efficiency of 7 %. Moreover, these results can be significantly improved simply by increasing the quenching power, and can be readily extended to two or three dimensions. These greatly reduced temperatures should significantly reduce the uncertainty in Ca-based optical frequency standards, for which a recent measurement of the absolute frequency was principally limited by residual atomic velocity [@Hg-Ca].
Our approach to second-stage cooling takes advantage of the two-electron structure of these atoms, which gives rise to narrow intercombination lines. In principle, these transitions can serve as excellent frequency or velocity (via the first-order Doppler shift) discriminators, which are ideal for sub-recoil laser cooling. The spectacular results achieved with the intercombination line of $^{88}$Sr, for which temperatures of 700 nK have been attained, confirm the potential of these transitions [@Katori; @Kurt]. Unfortunately, for several species, such as Ca and Mg, the intercombination line is too weak to effectively cool a distribution with an initial temperature of several millikelvins. Nonetheless, Binnewies *et al.* [@Maxwell] were able to use the velocity selectivity of the Ca intercombination line at 657 nm to demonstrate a new cooling mechanism, Maxwell-demon cooling. They were able to achieve a net transfer of 5 % of the initial distribution into a peak with a 1-D temperature of $<$ 10 $\mu$K. However, this technique does not seem to lend itself to 3-D cooling, which is critical for many applications.
A more versatile approach is to increase the cooling efficiency by quenching the upper level of the narrow transition through excitation to another state that decays more quickly to the ground state. In Fig. 1 we show how we quench the metastable $^3$P$_1$ state in Ca (lifetime = 400 $\mu$s) through excitation to the $^1$S$_0$ (4s5s) level (lifetime = 30 ns) via the intercombination line at 553 nm. The enhancement of narrow-line cooling rates through quenching was first demonstrated in the context of sideband cooling of trapped ions [@IonCool1; @IonCool2]. When applied to neutral atoms, quenching can increase the cooling efficiency not only by speeding up the rate at which one repeats cooling cycles, but also by giving the atoms additional momentum kicks towards zero velocity. These advantages were used to increase the force on a metastable He atomic beam by excitation of a two-photon transition with simultaneous light fields [@Wilbert].
To take advantage of the velocity selectivity of the narrow transition, however, it is essential to excite the atoms in a stepwise manner with time-separated pulses. One can then implement the velocity-selective excitation strategies that have been used with great success in Raman cooling. In the first demonstration of Raman cooling, Kasevich and Chu used a Raman transition between hyperfine ground states to cool Na atoms to a 1-D temperature of 100 nK and a 3-D temperature of 4.3 $\mu$K [@Kasevich; @Chu3D]. Later Reichel *et al.* used square Raman-cooling pulses to achieve a 1-D temperature of 3 nK for Cs [@French3D].
For quenching times short compared to excitation times used for the narrow transition, we propose the following QNLC procedure and illustrate it with the Ca system. We start by pumping a velocity slice towards zero velocity using light whose frequency is tuned slightly red of the $^1$S$_0$$\rightarrow$$^3$P$_1$ intercombination line at 657 nm (see Fig. 1). Due to the narrow linewidth of this transition (400 Hz), the spectra of our red pulses are Fourier-transform limited, so we can easily control the width and shape of the velocity group excited. We then quench the $^3$P$_1$ population with a 553 nm light pulse (copropagating with the red pump beam), which moves the population to the $^1$S$_0$(4s5s) state and gives the velocity-selected atoms a second momentum kick towards zero velocity. From this state the atoms quickly return to the ground state via two cascaded decays (see Fig. 1) with their associated (randomly directed) recoils. Next we use a similar sequence of red and green pulses from the opposite direction to pump a second velocity slice, symmetrically located on the opposite side of the distribution, towards zero. A single cooling cycle can thus reduce atomic velocities on both sides of the distribution by two-photon momenta (analogous to the case of Raman cooling). We repeat this cooling cycle many times, driving the atoms towards zero velocity, which coincides with a zero in our sinc$^2$ Fourier-transform-limited excitation function. In this way, atoms that accumulate around zero velocity have only a small probability of being pumped away, so in principle, extremely low temperatures can be achieved. As in the case of Raman cooling, the cooling limit is set by the cooling time and the width of the small transition-probability region around zero velocity, but not by the random recoils. Narrower sinc$^2$ functions (i.e. longer red pulses in the time domain) yield colder temperatures, but they address a smaller range of velocities, so there is a compromise between temperature and number, unless cooling with a sequence of pulse widths is used [@Kasevich; @Chu3D; @French3D].
In our experimental demonstration of this technique, inadequate light power at 553 nm (limited by what could be transferred through a 180 m fiber) led to quenching times that were an order of magnitude longer than the velocity-selective cooling pulses. Under these conditions it was more efficient to use a slightly modified version of the procedure just described, in which we use a standing-wave pulse of 553 nm light to quench the $^3$P$_1$ state after each pair of counter-propagating red pulse excitations. The drawback to this approach is the randomness introduced in the direction of the momentum kick from the quenching pulse. This loss, however, is more than offset in our case by the increased intensity of a standing (rather than travelling) wave that reduces the quenching time and the number of quenching pulses (by two) used per cooling cycle.
Our experimental realization of this cooling method (see Fig. 2) begins with an $\sim$ 7 ms loading cycle, during which we load 10$^7$ atoms from a Ca beam into a MOT (see Ref. [@NISTCaMOT] for details of our diode-laser based MOT) using the 423 nm $^1$S$_0$$\leftrightarrow$$^1$P$_1$ cooling transition (see Fig. 1). After this first cooling stage the temperature of the atomic sample is about 2 mK (corresponding to a v$_{rms}$ of 0.6 m/s), slightly above the 0.8 mK Doppler limit. To avoid large light shifts of the ground state that would compromise the velocity selectivity of the red cooling transition, we turn off the blue trapping light for the duration of the second-stage cooling. We then commence QNLC with a pair of counter-propagating red pulses (2.5 $\mu$s square-shaped $\pi$-pulses), separated by 2 $\mu$s in time. This 657 nm light is spatially filtered with an optical fiber and collimated to a diameter of 4 mm (8 mW in each beam). In order to excite the m=0$\rightarrow$m=0 transition, the light is linearly polarized parallel to the dominant B-field direction (due to a trap imbalance, the atoms rest at a 6 G point in our magnetic field gradient, which remains on during the entire measurement). We tune the laser to a frequency $\sim$ 350 kHz (corresponding to a velocity of 23 cm/s) red of resonance so that the first zero of the sinc$^2$ frequency spectrum is within 5 kHz of resonance for atoms at rest. The frequency of the red light is stabilized to an environmentally isolated, high-finesse Fabry-Perot cavity resonance, whose drift is cancelled to less than 0.3 kHz/minute (typical data averaging times were about 1 minute). After the first pair of counter-propagating red pulses, atoms that were transferred to the excited state are pumped to the $^1$S$_0$(4s5s) m=0 state by a 553 nm pulse (duration of $\sim$ 50 $\mu$s) with an efficiency of $\sim$ 50 %. As a result of the small transition rate (we estimate $\Gamma$($^1$S$_0$(4s5s)$\rightarrow$$^3$P$_1$(4s4p)) $\sim$ 2440(600) s$^{-1}$), we could only attain a 1/e quenching time of 55 $\mu$s, with 17 mW of green light collimated to a standing wave diameter of 3 mm [@PTBQuench]. The frequency-stabilized green light is generated by a dye laser in another laboratory and sent to us via optical fiber. This light intersects the atoms at an angle of $\sim$ 8$^o$ relative to the red cooling beams. From the $^1$S$_0$(4s5s) state the atoms decay rapidly ($<$ 35 ns) to the ground state by way of the $^1$P$_1$ state. Then follows a second pair of counter-propagating red pulses, but this time in reverse order (see Fig. 2) to make the overall cooling process more symmetric. Due to the lobes of the sinc$^2$ frequency spectrum there is some probability that the second pulse of a pulse pair can transfer atoms excited by the first red pulse back to the ground state, making the cooling due to the first pulse slightly less efficient. After the second pair of red pulses, a second green quenching pulse follows to complete the cooling cycle. We then repeat the cooling cycle (consisting of two pairs of counter-propagating pulses, each followed by a standing-wave quenching pulse) 8 to 20 times. At the end of this sequence we clean out the $^3$P$_1$ state with an extra green pulse lasting 100 to 150 $\mu$s that leaves less than 3 % of the population in the excited state. We then measure the velocity distribution.
Our velocity probe consists of a single 657 nm pulse (10 or 20 $\mu$s in duration, depending on our desired spectral resolution) that excites a narrow velocity slice of atoms to the excited state. It excites the m=0$\rightarrow$m=0 transition, and is collinear with the red cooling beams. While continuously cycling the complete measurement sequence, we slowly sweep the frequency of the velocity probe over the atomic velocity range to generate our distributions. We typically average 60 sweeps of 0.5 s duration to generate a data set. In order to achieve a good signal-to-noise ratio for these velocity distribution measurements, we use two 423 nm probe pulses in a normalized shelving detection scheme that we developed for our optical clock (see Fig. 2) [@NISTCaMOT].
In Figure 3 we show the effect of 15 cooling cycles taken with a red pulse length of 2.5 $\mu$s and a green pulse length of 50 $\mu$s. We see that the QNLC process transfers most of the atoms in the range of ±30 cm/s into the narrow peak at zero velocity, consistent with the expected momentum transfer from our cooling pulses. A Gaussian fit to the central peak yielded a v$_{rms}$ of 6 cm/s, or about 4 recoils at 657 nm, corresponding to a 1-D temperature of 17 $\mu$K. The transfer efficiency can be estimated by comparing the area of the peak to the area of the initial distribution - in this case we find 30 % transfer. However, since approximately half of the atoms have escaped the interaction region since we first turned off the trapping beams, we have a net efficiency closer to 15 % (we have seen net efficiencies as high as 18 %). Fast ballistic expansion due to our warm initial temperature causes the atoms to move into weaker parts of the laser beams during the cooling and probe periods. Because of laser power constraints, we are unable to increase the beam sizes, thus the number of cooling cycles we can use is severely limited.
Indeed, when more cooling cycles are used, we see increased transfer efficiency and narrower distributions (as was demonstrated in Raman cooling experiments [@Kasevich; @Chu3D; @French3D]). However the longer cooling time required means more atoms are lost transverse to the cooling direction. We experimented with using more cooling cycles with less efficient quenching and found that the central peak was fairly insensitive over the range from 18 cycles with 30 $\mu$s quenching time to 8 cycles with 70 $\mu$s quenching time (keeping the total cooling time approximately fixed at 1.5 ms).
The 14 cm/s width of the central peak in Fig. 3 is a result of the convolution of the velocity distribution and the velocity probe. Data taken at higher probe resolution indicate that v$_{rms}$ for these conditions is less than 4 cm/s (7.5 $\mu$K). To strive for colder temperatures, we increased the red pulse length to 5 $\mu$s to give a sharper velocity discriminator. Figure 4 shows a velocity distribution resulting from 10 cooling cycles using these longer cooling pulses. To keep the 657 nm pulse area constant for these measurements, we chose to double the size of the red beams rather than reduce their power in an effort to increase their pumping efficiency. We also adjusted the detuning to $\sim$ -173 kHz (corresponding to 11.4 cm/s) to place the zero of the excitation function at zero velocity. Indeed, we see narrower distributions under these conditions (we estimate a deconvolved value of v$_{rms}$ $<$ 2.8 cm/s, corresponding to temperature of 3.7 $\mu$K) and the expected smaller fraction of atoms in the peak (7 % net efficiency) due to the reduced range of velocities covered. We also see quite clearly several other nodes of the sinc$^2$ function that serve as additional dark velocities where the atoms can accumulate. We fully expect that at higher resolutions we can achieve sub-recoil temperatures, as was seen in the Raman cooling experiments. Monte Carlo simulations of the cooling process (including the recoils at 553 nm, 1.03 $\mu$m, and 423 nm) support this expectation.
From these results, a strategy that can cool a large fraction of the initial distribution to very cold temperatures becomes evident, namely to use a series of pulses at different resolutions and detunings [@Kasevich; @Chu3D; @French3D]. To be able to exploit such a strategy, however, we need to accelerate the cooling process. With simple scaling of the 553 nm power, we calculate that a fifteen-fold increase (dye lasers regularly supply 20x the power we currently send to our atoms) would enable us to cool the atoms nine times faster. Results from a Monte Carlo simulation assuming a 1/e pumping time of 4 $\mu$s and using multiple red cooling pulses ranging from 2.5 to 5 $\mu$s in duration indicate that we can expect to transfer 50 % of the atoms into a peak with sub-recoil temperature.
Moreover, one can envision a feasible 3-D strategy in which one cools alternately in three dimensions [@Chu3D], with each reduction in temperature effectively increasing the available cooling time. Due to the multi-recoils involved in this experiment, it would be hard to achieve 3-D sub-recoil temperatures due to recoil heating in transverse dimensions. However, with appropriate shaping of our red pulses [@Chu3D; @French3D] a limit of several recoils seems quite feasible, especially in light of the near-recoil temperatures achieved with 3-D Raman cooling [@Chu3D]. Furthermore, as one nears the recoil limit and presumably has increased the cooling time available, one could remove the quenching and use the decay of the $^3$P$_1$ state itself to reduce the recoil effect. Regardless, 3-D temperatures of a few microkelvins rather than millikelvins would greatly benefit the Ca optical frequency standard, atomic interferometry, and other applications.
In conclusion, we have proposed a second-stage cooling scheme based on quenched narrow-line laser cooling. We have used this technique to reduce the temperature of a $^{40}$Ca atomic cloud in 1-D by a factor of 500. With ample quenching power, near sub-recoil temperatures in 3-D should be possible. Moreover, this technique should be applicable to other atoms (such as Mg) with narrow optical transitions and available quenching transitions.
We thank J. Bergquist and U. Tanaka, who generously supplied the 553 nm light used in this work.
\*Also at Phys. Dept., University of Colorado at Boulder, Boulder, CO 80309.
[99]{}
J. Dalibard and C. Cohen-Tannoudji, J. Opt. Soc. Am. B. [**6**]{}, 2023 (1989).
P.J. Ungar [*et al.*]{}, J. Opt. Soc. Am. B. [**6**]{}, 2058 (1989).
A. Aspect [*et al.*]{}, Phys. Rev. Lett. [**61**]{}, 826 (1988).
M. Kasevich and S. Chu, Phys. Rev. Lett. [**69**]{}, 1741 (1992).
T. Kurosu and F. Shimizu, Japanese J. of Appl. Phys. [**29**]{}, L2127 (1990).
K. Sengstock [*et al.*]{}, Appl. Phys. B [**59**]{}, 99 (1994).
Th. Kisters [*et al.*]{}, Appl. Phys. B [**59**]{}, 89 (1994).
Th. Udem [*et al.*]{}, accepted for publication in Phys. Rev. Lett.
H. Katori [*et al.*]{}, Phys. Rev. Lett. [**82**]{}, 1116 (1999).
K. R. Vogel [*et al.*]{}, IEEE Trans. Instr. and Meas. [**48**]{}, 618 (1999).
T. Binnewies [*et al.*]{}, Phys. Rev. A [**62**]{} , 1601 (2000).
F. Diedrich [*et al.*]{}, Phys. Rev. Lett. [**62**]{}, 403 (1989).
Ch. Roos [*et al.*]{}, Phys. Rev. Lett. [**83**]{}, 4713 (1999).
W. Rooijakkers [*et al.*]{}, Phys. Rev. Lett. [**74**]{}, 3348 (1995).
N. Davidson [*et al.*]{}, Phys. Rev. Lett. [**72**]{}, 3158 (1994).
J. Reichel [*et al.*]{}, Phys. Rev. Lett. [**75**]{}, 4575 (1995).
C.W. Oates [*et al.*]{}, Eur. J. of Phys. D [**7**]{}, 449 (1999).
The Ca group at PTB has also demonstrated quenching of this level through the $^3$P$_1$(4s4p)$\rightarrow$$^1$D$_2$(4s4d) transition at 453 nm and achieved a time constant of 70 $\mu$s. F. Riehle, Proc. SPIE Vol. 4269, “Laser Frequency Stabilization, Standards, Measurement, and Applications," San Jose, CA, Jan. 25-26, 2001, eds. John L. Hall and Jun Ye, p. 22.
![Relevant energy level diagram for QNLC in $^{40}$Ca. The 553 nm quenching beam transfers population from the long-lived (400 $\mu$s) $^3$P$_1$(4s4p) state to the ground state via the short-lived (30 ns) $^1$S$_0$(4s5s) state. []{data-label="Fig1"}](Fig1.eps){width="8cm"}
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We prove that to find optimal positional strategies for stochastic mean payoff games when the value of every state of the game is known, in general, is as hard as solving such games tout court. This answers a question posed by Daniel Andersson and Peter Bro Miltersen.'
address: |
Laboratoire d’Informatique de l’École Polytechnique (<span style="font-variant:small-caps;">lix</span>)\
Bâtiment Alan Turing\
1 rue Honoré d’Estienne d’Orves, Campus de l’École Polytechnique, 91120 Palaiseau, France.
author:
- Marcello Mamino
title: Strategy Recovery for Stochastic Mean Payoff Games
---
=1
In this note, we consider perfect information $0$-sum stochastic games, which, for short, we will just call [*stochastic games*]{}. For us, a stochastic game is a finite directed graph whose vertices we call [*states*]{} and whose edges we call [*transitions*]{}, multiple edges and loops are allowed but no state can be a sink. To each state $s$ is associated an [*owner*]{} $o(s)$ which is one of the two players and . Each transition $s\trans Ap t$ has an [*action*]{} $A$ and a [*probability*]{} $p\in\Q\cap\intervalcc01$, with the condition that, for each state $s$, the probabilities of the transitions exiting $s$ associated to the same action must sum to $1$. We say that the action $A$ is [*available*]{} at state $s$ if one of the transitions exiting $s$ is associated to $A$. Furthermore to each action $A$ is associated a [*reward*]{} $r(A)\in\Q$.
A play of a stochastic game $G$ begins in some state $s_0$ and produces an unending sequence of states $\{s_i\}_{i\in\N}$ and actions $\{A_i\}_{i\in\N}$. At move $i$, the owner of the current state $s_i$ chooses an action $A_i$ among those available at $s_i$, then one of the transitions exiting $s_i$ with action $A_i$ is selected at random according to their respective probabilities, and the next state $s_{i+1}$ is the destination of the chosen transition. A play can be evaluated according to the $\beta$-discounted payoff criterion $$\v_\beta(A_0,A_1\dotsc) = (1-\beta) \sum_{i=0}^{\infty} r(A_i)\beta^i$$ for $\beta\in\intervalco01$. Or it can be evaluated according to the mean payoff criterion $$\v_1(A_0,A_1\dotsc) = \liminf_{n\to\infty} \frac{1}{n+1} \sum_{i=0}^{n} r(A_i)$$ The goal of is to maximize the evaluation, that of is to minimize it. It is known that for both criteria there are optimal strategies which are [*positional*]{} [@Gil57; @LiLi69], namely such that the action chosen at $s_i$ depends only on the state $s_i$ – an not, for instance, on the preceding states in the play, on $i$, or on a random choice. Given two positional strategies $\sigma$ and $\tau$ for and respectively, and given $\beta\in\intervalcc01$, we denote $\v_\beta(G,s_0,\sigma,\tau)$ the expected value of $\v_\beta$ on all plays generated by $\sigma$ and $\tau$ starting from $s_0$. We write $\v_\beta(G,s_0)$ for $\v_\beta(G,s_0,\sigma,\tau)$ with $\sigma$ and $\tau$ optimal. For basic information on stochastic games one may refer to the book [@Filar-Vrieze].
Given a stochastic game with probabilities and rewards encoded in binary, and a value of $\beta$ also encoded in binary, it makes sense to study the computational complexity of the task of solving the game. Strategically solving a game, as defined in [@AnMil09], means to find a pair of optimal strategies. Quantitatively solving $G$ means to find $\v_\beta(G,s)$ for all states $s$. In general, the second task is easier than the first. The [*strategy recovery problem*]{} is, given the quantitative solution of a game, to produce a strategic solution. It has been observed in [@AnMil09] that this task can be performed trivially in linear time for discounted payoff games, and also, but not trivially, for terminal payoff and simple stochastic games, hence it was asked whether the same could be done for stochastic mean payoff games (this is, indeed, the only missing element to complete Andersson and Miltersen’s picture). Our aim is to prove that the strategy recovery problem for stochastic mean payoff games is as hard as it possibly can.
\[th-main\] The strategy recovery problem for stochastic mean payoff games is equivalent, modulo polynomial time Turing reductions, to the task of strategically solving mean payoff games.
We will combine the reduction from stochastic mean payoff to discounted payoff games proven in [@AnMil09] with a new reduction from discounted to mean payoff games of a special form that we call $\beta$-recurrent. Then we will show that $\beta$-recurrent mean payoff games can be turned into [*strategically equivalent*]{} mean payoff games having the additional property that all states have value $0$. For this latter class of games, the strategy recovery problem is obviously equivalent to solving the games strategically.
Let $G$ be a stochastic game and $s_0$ one of the states of $G$. We define the [*$\beta$-recurrent game associated to $G$ and $s_0$*]{}, denoted $G_{\beta,s_0}$. The game $G_{\beta,s_0}$ has the same state-space as $G$. Each transition $a\trans Ap b$ in $G$ is replaced by two new transitions in $G_{\beta,s_0}$, namely $a\trans A{\beta p} b$ and $a\trans A {(1-\beta)p} s_0$. The first of these new transitions will be called [*of the first kind*]{}, the second [*of the second kind*]{}. We say that a game is [*$\beta$-recurrent*]{} if it results from the construction just defined, for some $G$.
Notice that our $\beta$-recurrent games are ergodic in the sense of [@BEGM10]. The complexity of ergodic games has been settled in a recent work [@ChaIJ14] (see the full version [@ChaIJ14-full]), however we need for our reduction the extra properties of $\beta$-recurrent games. Interestingly, the definition of ergodic in [@ChaIJ14] is more restrictive than that in [@BEGM10], and, in particular, in this stronger sense, a $\beta$-recurrent game may not be ergodic, nor an ergodic game needs to be $\beta$-recurrent.
\[th-lemma1\] The task of quantitatively solving stochastic discounted payoff games is polynomial time Turing reducible to quantitatively solving $\beta$-recurrent stochastic mean payoff games.
Consider a stochastic game $G$ and discount factor $\beta$. Let $s_0$ denote a state of $G$. We will show that $$\v_\beta(G,s_0) = \v_1(G_{\beta,s_0},s_0)$$ Intuitively, an infinite play of $G_{\beta,s_0}$ can be seen as a sequence of finite sub-plays, each of which lasts until a transition of the second kind is taken and the game is reset to the initial state $s_0$. Each sub-play lasts at least one move, but a second move is played only with probability $\beta$, a third one with probability $\beta^2$, and so on, thus imitating the discounted payoff situation.
In order to prove the proposition, it suffices to show that, for any pair of positional strategies $\sigma$ and $\tau$ for and respectively, one has $$\tag{$\star$}
\v_1(G_{\beta,s_0},s_0,\sigma,\tau) = \v_\beta(G,s_0,\sigma,\tau)$$ In fact, it follows from this equation that $\sigma$ and $\tau$ are a pair of optimal positional strategies for $G_{\beta,s_0}$ if and only if they are a pair of optimal positional strategies for $G$ with starting position $s_0$.
It remains to prove equation $(\star)$. For each state $s$ of $G$, call $A_{\sigma,\tau}(s)$ the action chosen by either $\sigma$ or $\tau$ (according to the owner of $s$) at the state $s$. The $\beta$-discounted values of the states of $G$ are determined by the condition $$\v_\beta(G,s,\sigma,\tau) = (1-\beta) r( A_{\sigma,\tau}(s) ) + \sum_{t\in G}
\beta p_{\sigma,\tau}(s\to t) \v_\beta(G,t,\sigma,\tau)$$ where $p_{\sigma,\tau}(v\to w)$ denotes the probability that, from state $s$, a transition to state $t$ is chosen when playing strategy $\sigma$ against $\tau$. If we call $s_0\dotsc s_n$ the states of $G$ and $\bar v_\beta =
(\v_\beta(G,s_i,\sigma,\tau))_{i=1\dotsc n}$ the value vector of $G$, then the condition above can be rewritten in the form $$\bar v_\beta = (1-\beta) \bar r + \beta P \bar v_\beta$$ where $\bar r$ is the vector of the rewards $\bar r_i = r(
A_{\sigma,\tau}(s_i))$, and $P$ denotes the matrix of the transition probabilities $P_{i,j} = p_{\sigma,\tau}(s_i \to s_j)$. Hence $$\bar v_\beta = (1-\beta) (I - \beta P)^{-1} \bar r$$ where $I$ denotes the $n\times n$ identity matrix.
Now we turn our attention to the mean payoff of the pair of strategies $\sigma$ and $\tau$ in $G_{\beta,s_0}$. We can compute $\v_1(G_{\beta,s_0},s_0,\sigma,\tau)$ averaging the rewards over the stable distribution of the Markov chain induced by these strategies on the states of $G$. This stable distribution $\mu$ must be unique, because, by virtue of $G_{\beta,s_0}$ being $\beta$-recurrent, the Markov chain is connected. Moreover $\mu$ is determined by the condition $$\mu(s) = (1-\beta)\delta_{s_0}(s) + \sum_{t\in G} \beta p_{\sigma,\tau}(t,s)\mu(t)$$ where $\delta_{s_0}(s)$ is $1$ if $s=s_0$ and $0$ otherwise. Rewriting as above, we get $$\bar \mu = (1-\beta)e_0 + \beta P^T \bar\mu$$ where $e_0$ is the first element of the canonical basis and $\bar \mu_i =
\mu(s_i)$. Hence $$\bar \mu = (1-\beta) (I - \beta P^T)^{-1} e_0$$ Now, computing the average $$\begin{aligned}
\v_1(G_{\beta,s_0},s_0,\sigma,\tau) &= \sum_{s\in G} \mu(s) r(
A_{\sigma,\tau}(s)) \\
&= {\bar \mu}^T \bar r \\
&= e_0^T (1-\beta) (I - \beta P)^{-1} \bar r \\
&= e_0^T \bar v_\beta \\
&= \v_\beta(G,s_0,\sigma,\tau) \qedhere\end{aligned}$$
\[th-lemma2\] The task of strategically solving $\beta$-recurrent stochastic mean payoff games is polynomial time many-one reducible to the strategy recovery problem for stochastic mean payoff games.
Let $G_{\beta,s_0}$ be a $\beta$-recurrent stochastic game. As we noticed, all the states of $G_{\beta,s_0}$ have the same value. Nevertheless, we have no obvious way to determine this value in order to complete the reduction. Instead, we choose to construct a new mean payoff game $G'$ in such a way that all the states of $G'$ get mean payoff value equal to $0$, and nonetheless a pair of optimal strategies for $G_{\beta,s_0}$ can be recovered from a pair of optimal strategies for $G'$. This is clearly sufficient to establish the lemma.
The game $G'$ is constructed as two chained copies $G^1$ and $G^2$ of $G_{\beta,s_0}$, redirecting all the transitions of the second kind in each instance – that go to the state corresponding to $s_0$ in that instance – to the $s_0$-state in the other. The states of $G^1$ have the same owner as in $G_{\beta,s_0}$, and the transitions originating in $G^1$ are associated to the same actions with the same rewards as in $G_{\beta,s_0}$. In $G_2$, however, the owners are switched and the signs of the rewards exchanged (formally we replace each action $A$ with a new one $A'$ having $r(A')=-r(A)$). If both players play optimally, we may expect each to win in $G_1$ precisely as much as he loses in $G_2$, hence, arguably the value of $G'$ should be $0$. On the other hand, in order to play optimally in $G'$, one should play optimally in both the components, so we should be able to extract optimal positional strategies for $G_{\beta,s_0}$ from optimal positional strategies for $G'$ by mere restriction to the component $G^1$. We will now proceed to prove our statement.
Let us denote by $s^1$ and $s^2$ respectively the states of $G^1$ and $G^2$ corresponding to a given state $s$ of $G_{\beta,s_0}$. First observe that a play of $G'$, almost surely, will eventually reach state $s_0^1$, from this follows that all the states of $G'$ must have the same value ($G'$ is ergodic). A positional strategy $\sigma$ for in $G'$ can be seen as a pair of positional strategies $(\sigma^1,\sigma^2)$ where $\sigma^1$ is the strategy for in $G_{\beta,s_0}$ that we get restricting $\sigma$ to $G^1$, and $\sigma^2$ is the strategy for in $G_{\beta,s_0}$ that we get from the restriction of $\sigma$ to $G^2$ (remember that in $G^2$ the players are switched). Similarly a strategy $\tau$ for in $G'$ can be seen as a pair of strategies $(\tau^1,\tau^2)$ in $G_{\beta,s_0}$, the first one for and the second for . We will prove that for any $\sigma$ and $\tau$ $$\tag{$\star\star$}
\v_1(G',\cdot,\sigma,\tau) =
\frac{1}{2} \v_1(G,\cdot,\sigma^1,\tau^1)
-
\frac{1}{2} \v_1(G,\cdot,\tau^2,\sigma^2)$$ From this equation, it follows at once that $\sigma$ is an optimal strategy for $G'$ if and only if $(\sigma^1,\sigma^2)$ is a pair of optimal strategies for $G_{\beta,s_0}$, and, in particular, the value of $G'$ is $0$.
We turn now to the proof of equation ($\star\star$). Consider the unique stable distribution $\mu$ of the Markov process induced by $\sigma$ and $\tau$. Observe that, independently from $\sigma$ and $\tau$, at any given state, our Markov chain has probability $\beta$ of transitioning to a state belonging to the same component, and probability $1-\beta$ of switching component. It follows that the sequence of the components must obey the law of a two-state Markov chain with transition matrix $$(
\begin{matrix}
\beta & 1-\beta \\
1-\beta & \beta
\end{matrix}
)$$ Hence $\mu(G^1)=\mu(G^2)=1/2$. It suffices to prove that the probability distributions $\mu^1$ and $\mu^2$ defined on the states of $G_{\beta,s_0}$ by $\mu^1(s) = 2 \mu(s^1)$ and $\mu^2 = 2 \mu(s^2)$ are the stable distributions induced on $G_{\beta,s_0}$ by the pairs of strategies $(\sigma^1,\tau^1)$ and $(\tau^2,\sigma^2)$ respectively.
By symmetry, we can concentrate on $\mu^1$. Let $p_{\sigma,\tau}(t,s)$ denote the probability of the transition $t\to s$ in the Markov process induced by the strategies $\sigma$ and $\tau$. Since all states of $G^1$ except $s_0^1$ are only reachable from within $G^1$ itself, the consistency equation for $\mu$ being a stable distribution on $G'$ $$\mu(s) = \sum_{t\in G'} p_{\sigma,\tau}(t,s)\mu(t)$$ implies the same condition for $\mu^1$ at all states except $s_0$. At $s_0$ one concludes by direct computation observing that the component of the sum on the right hand side due to transitions of the second kind must be $$(1-\beta)\mu(G^2) = \frac{1-\beta}{2} = (1-\beta)\mu(G^1) \qedhere$$
By [@AnMil09 Theorem 1], solving stochastic mean payoff games strategically is reducible to solving stochastic discounted payoff games quantitatively, which reduces, by Lemma \[th-lemma1\], to solving $\beta$-recurrent stochastic mean payoff games quantitatively. In turn, solving such $\beta$-recurrent games quantitatively is reducible to solving the same strategically, just because they are, in particular, stochastic mean payoff games. By Lemma \[th-lemma2\], this final task is reducible to the strategy recovery problem for stochastic mean payoff games.
Finally, we would like to remark that our construction relies on the interpretation of [*strategic solution*]{} as requiring optimal [*positional*]{} strategies. Were a more general class of strategies available, then the problem of finding an optimal one would become easier. In particular, the games produced by Lemma \[th-lemma2\] happen to be symmetric under switching the players and the signs of the rewards. Under this circumstance, it would not be surprising if one could play optimally by some form of strategy stealing technique.
Acknowledgements {#acknowledgements .unnumbered}
================
We would like to express gratitude to Manuel Bodirsky and Eleonora Bardelli for interesting discussions.
[BEGM10]{}
Daniel Andersson and Peter Bro Miltersen. The complexity of solving stochastic games on graphs. In [*Algorithms and computation*]{}, volume 5878 of [*Lecture Notes in Comput. Sci.*]{}, pages 112–121. Springer, Berlin, 2009.
Endre Boros, Khaled Elbassioni, Vladimir Gurvich, and Kazuhisa Makino. A pumping algorithm for ergodic stochastic mean payoff games with perfect information. In [*Integer programming and combinatorial optimization*]{}, volume 6080 of [*Lecture Notes in Comput. Sci.*]{}, pages 341–354. Springer, Berlin, 2010.
Krishnendu Chatterjee and Rasmus Ibsen-Jensen. The complexity of ergodic mean-payoff games. In Javier Esparza, Pierre Fraigniaud, Thore Husfeldt, and Elias Koutsoupias, editors, [*Automata, Languages, and Programming*]{}, volume 8573 of [*Lecture Notes in Computer Science*]{}, pages 122–133. Springer Berlin Heidelberg, 2014.
Krishnendu Chatterjee and Rasmus Ibsen-Jensen. The complexity of ergodic mean-payoff games. , abs/1404.5734, 2014.
Jerzy Filar and Koos Vrieze. . Springer-Verlag, New York, 1997.
Dean Gillette. Stochastic games with zero stop probabilities. In [*Contributions to the theory of games, vol. 3*]{}, Annals of Mathematics Studies, no. 39, pages 179–187. Princeton University Press, Princeton, N. J., 1957.
Thomas M. Liggett and Steven A. Lippman. Stochastic games with perfect information and time average payoff. , 11:604–607, 1969.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: '[**Abstract**]{} In the paper we consider the problem of valuation of American options written on dividend-paying assets whose price dynamics follow the classical multidimensional Black and Scholes model. We provide a general early exercise premium representation formula for options with payoff functions which are convex or satisfy mild regularity assumptions. Examples include index options, spread options, call on max options, put on min options, multiply strike options and power-product options. In the proof of the formula we exploit close connections between the optimal stopping problems associated with valuation of American options, obstacle problems and reflected backward stochastic differential equations.'
author:
- |
Tomasz Klimsiak and Andrzej Rozkosz\
[Faculty of Mathematics and Computer Science, Nicolaus Copernicus University]{},\
[Chopina 12/18, 87-100 Toruń, Poland]{}
title: The early exercise premium representation for American options on multiply assets
---
[**Keywords**]{} American option, multiply assets, early exercise premium, backward stochastic differential equation, optimal stopping, obstacle problem.\
[**Mathematics Subject Classification (2000)**]{} 91B28, 60H10, 65M06.
Introduction
============
In the paper we study American options written on dividend-paying assets. We assume that the underlying assets dynamics follow the classical multidimensional Black and Scholes model. It is now well known that the arbitrage-free value of American options can be expressed in terms of the optimal stopping problem (Bensoussan [@Be], Karatzas [@K]; see also [@KS] for nice exposition and additional references), in terms of variational inequalities (Jaillet, Lamberton and Lapeyre [@JLL]) and in terms of solutions of reflected BSDEs (El Karoui and Quenez [@EQ]). Although these approaches provide complete characterization of the option value (see Section \[sec2\] for a short review), the paper by Broadie and Detemple [@BD2] shows that it is of interest to provide alternative representation, which expresses the value of an American option as the value of the corresponding European option plus the gain from early exercise. The main reason is that the representation, called the early exercise premium formula, gives useful information on the determinants of the option value. The formula was proved first by Kim [@Ki] in the case of standard American put option on a single asset. Another important contributions in the case of single asset include Broadie and Detemple [@BD1], El Karoui and Karatzas [@EK] and Jacka [@J] (see also [@D; @KS] and the references therein). The case of options on multiply dividend-paying assets is more difficult and has received rather little attention in the literature. In the important paper [@BD2] and next in Detemple, Feng and Tian [@DFT] (see also [@D]) the early exercise premium formula was established for concrete classes of options on multiply assets. Note that in the last paper call on min option, i.e. option with nonconvex payoff function is investigated. A subclass of call on min options consisting of capped options is studied in [@BD1; @BD2; @BD3] (see also [@D]).
In the present paper we provide a unified way of treating a wide variety of seemingly disparate examples. It allows us to prove a general exercise premium formula for options with convex payoff functions satisfying the polynomial growth condition or payoff function satisfying quite general condition considered in Laurence and Salsa [@LS]. Verifying the last condition requires knowledge of the payoff function and the structure of the exercise set. Therefore it is a complicated task in general. Fortunately, in most interesting cases one can easily check convexity of the payoff function or check some simpler condition implying the general condition from [@LS]. The class of options covered by our formula includes index options, spread options, call on max options, put on min options, multiply strike options, power-product options and others.
In the proof of the exercise premium formula we rely on some results on reflected BSDEs and their links with optimal stopping problems (see [@EQ]) and with parabolic variational inequalities established in Bally, Caballero, Fernandez and El Karoui [@BCFE]. We also use classical results on regularity of the solution of the Cauchy problem for parabolic operator with constant coefficients, and in case of convex payoffs, some fine properties of convex functions. Perhaps it is worth mentioning that we do not use any regularity results on the free boundary problem for an American option. The basic idea of the proof comes from our earlier paper [@KR:BPAN] devoted to standard American call and put options on single asset.
Preliminaries {#sec2}
=============
We will assume that under the risk-neutral probability measure the underlying assets prices $X^{s,x,1},\dots,X^{s,x,n}$ evolve on the time interval $[s,T]$ according to stochastic differential equation of the form $$\label{eq1.1}
X^{s,x,i}_{t}=x_i+\int_{s}^{t}(r-d_{i})X^{s,x,i}_{\theta}\,d\theta
+\sum_{j=1}^n\int_{s}^{t}\sigma_{ij}X^{s,x,i}_{\theta}\,dW^j_{\theta},
\quad t\in[s,T].$$ Here $W=(W^1,\dots,W^n)$ is a standard $n$-dimensional Wiener process, $r\ge0$ is the rate of interest, $d_i\ge0$ is the dividend rate of the asset $i$ and $\sigma=\{\sigma_{ij}\}$ is the $n$-dimensional volatility matrix. We assume that $a=\sigma\cdot\sigma^*$ is positive definite. Since the distributions of the processes $X^{s,x,i}$ depend $\sigma$ only through $a$, we may and will assume that $\sigma$ is a symmetric square root of $a$. As for the payoff function $\psi$ we will assume that it satisfies the assumptions:
1. $\psi$ is a nonnegative continuous function on ${{\mathbb R}}^n$ with polynomial growth,
2. For every $t\in (0,T)$, $\psi$ is a smooth function on $\{\psi=u\}\cap\bar Q_{t}$, i.e. there exists an open set $U\subset {{\mathbb R}}^n$ such that $\{u=\psi\}\cap\bar Q_{t}\subset
[0,t]\times U$ and $\psi$ is smooth on $U$ (Here $Q_t\equiv[0,t)\times{{\mathbb R}}^n$, $\bar Q_t\equiv[0,t]\times{{\mathbb R}}^n$ and $u$ is the value of an option with payoff $\psi$; see (\[eq1.2\]) and (\[eq2.4\]) below)
or
1. $\psi$ is a nonnegative convex function on ${{\mathbb R}}^n$ with polynomial growth.
Note that convex functions are locally Lipschitz, so assumption (A3) implies (A1). Assumption (A2) is considered in [@LS]. It is satisfied for instance if
1. The region where $\psi$ is strictly positive is the union of several connected components in which $\psi$ is smooth.
Following [@LS] let us also note that unlike (A2$'$) or (A3), condition (A2) cannot be verified by appealing to the structure of the payoff alone. Verifying (A2) requires additional knowledge of the structure of the exercise set $\{u=\psi\}$.
Let $\Omega=C([0,T];{{\mathbb R}}^n)$ and let $X$ be the canonical process on $\Omega$. For $(s,x)\in Q_{T}$ let $P_{s,x}$ denote the law of the process $X^{s,x}=(X^{s,x,1},\dots,X^{s,x,n})$ defined by (\[eq1.1\]) and let $\{{{\mathcal{F}}}^s_t\}$ denote the completion of $\sigma(X_{\theta};\theta\in[s,t])$ with respect to the family $\{P_{s,\mu};\mu$ a finite measure on ${{\mathcal{B}}}({{\mathbb R}}^n)\}$, where $P_{s,\mu}(\cdot)=\int_{{{\mathbb R}}^n}P_{s,x}(\cdot)\,\mu(dx)$. Then for each $s\in[0,T)$, ${{\mathbb X}}=(\Omega,({{\mathcal{F}}}^s_t)_{t\in[s,T]},X,P_{s,x})$ is a Markov process on $[0,T]$.
Let $I=\{0,1\}^{n}$. For $\iota=(i_1,\dots,i_n)\in I$ we set $D_{\iota}=\{x\in{{\mathbb R}}^n; (-1)^{i_{k}}x_{k}>0,k=1,\dots,n\}$, $P=\bigcup_{\iota\in I} D_{\iota}$, $P_{T}=[0,T)\times P$. By Itô’s formula, $$\label{eq2.02} X^{s,x,i}_t=x^i\exp\big((r-d_i-a_{ii})(t-s)
+\sum^n_{j=1}\sigma_{ij}(W^j_t-W^j_s)\big),\quad t\in[s,T].$$ Therefore if $s\in[0,T)$ and $x\in D_{\iota}$ for some $\iota\in
I$ then $P_{s,x}(X_{t}\in D_{\iota},\, t\ge s)=1$. From this and the fact that $a$ is positive definite it follows that if $x\in
P_T$ then $\det\sigma(X_t)>0$, $P_{s,x}$-a.s. for every $t\ge s$, where $\sigma(x)=\{\sigma_{ij}x_i\}_{i,j=1,\dots, n}$. Moreover, $[s,T]\ni t\mapsto\sigma^{-1}(X_t)$ is a continuous process. Therefore, if $x\in P_T$ then by Lévy’s theorem the process $B_{s,\cdot}$ defined as $B_{s,t}=\int^t_s\sigma^{-1}(X_{\theta})\,dM_{\theta}$, where $M^i_t=X^i_t-X^i_0-\int^t_0(r-d_i)X^i_{\theta}\,d\theta$, $t\in
[s,T]$, is under $P_{s,x}$ a standard $n$-dimensional $\{{{\mathcal{F}}}^s_t\}$-Wiener process on $[s,T]$ and $$\label{eq2.05}
X^i_t-x^i=\int_{s}^{t}(r-d_{i})X^{i}_{\theta}\,d\theta
+\sum_{j=1}^n\int_{s}^{t}\sigma_{ij}X^{i}_{\theta}\,dB^j_{s,\theta},
\quad t\in[s,T],\quad P_{s,x}\mbox{-a.s.},$$ i.e. $$\label{eq2.04} X^{i}_t=x^i\exp\big((r-d_i-a_{ii})(t-s)
+\sum^n_{j=1}\sigma_{ij}B_{s,t}\big),\quad t\in[s,T],\quad
P_{s,x}\mbox{-a.s.}$$ The above forms of the assets price dynamics will be more convenient for us than (\[eq1.1\]) or (\[eq2.02\]). Note that from the definition of the process $B_{s,\cdot}$ and (\[eq2.04\]) it follows that $\sigma(X_{\theta};\theta\in[s,t])=\sigma(B_{s,\theta};\theta\in[s,t])$ for $s\in[0,T)$, so for every $s\in[0,T)$ the filtration $\{{{\mathcal{F}}}^s_t\}$ is the completion of the Brownian filtration.
In Bensoussan [@Be] and Karatzas [@K] (see also Section 2.5 in [@KS]) it is shown that under (A1) the arbitrage-free value $V$ of an American option with payoff function $\psi$ and expiration time $T$ is given by the solution of the stopping problem $$\label{eq1.2} V(s,x)=\sup_{\tau\in\mathcal{T}_{s}}
E_{s,x}\big(e^{-r(\tau-s)}\psi(X_{\tau})\big),$$ where the supremum is taken over the set $\mathcal{T}_{s}$ of all $\{{{\mathcal{F}}}^s_t\}$-stopping times $\tau$ with values in $[s,T]$.
From the results proved in [@EKPPQ] it follows that under (A1) for every $(s,x)$ there exists a unique solution $(Y^{s,x},Z^{s,x},K^{s,x})$, on the space $(\Omega,{{\mathcal{F}}}^s_T,P_{s,x})$, to the reflected BSDE with terminal condition $\psi(X_T)$, coefficient $f:{{\mathbb R}}\rightarrow{{\mathbb R}}$ defined as $f(y)=-ry$, $y\in{{\mathbb R}}$, and barrier $\psi(X)$ (RBSDE${}_{s,x}(\psi,-r y,\psi)$ for short). This means that the processes $Y^{s,x},Z^{s,x},K^{s,x}$ are $\{{{\mathcal{F}}}^s_t\}$-progressively measurable, satisfy some integrability conditions and $P_{s,x}$-a.s., $$\label{eq2.1} \left\{
\begin{array}{l}
Y^{s,x}_t=\psi(X_T)-\int^T_tr Y^{s,x}_{\theta}\,d\theta
+K^{s,x}_T-K^{s,x}_t
-\int^T_tZ^{s,x}_{\theta}\,dB_{s,\theta},\quad
t\in[s,T],\medskip\\
Y^{s,x}_t\ge \psi(X_t),\quad t\in[s,T],\medskip \\
K^{s,x}\mbox{ is increasing, continuous, }K^{s,x}_s=0,\,\,
\int^T_s(Y^{s,x}_t-\psi(X_t))\,dK^{s,x}_t=0.
\end{array}
\right.$$ In [@EKPPQ] it is also proved that for every $(s,x)\in Q_T$, $$\label{eq2.2} Y^{s,x}_t=u(t,X_t),\quad t\in[s,T],\quad
P_{s,x}\mbox{-a.s.},$$ where $u$ is a viscosity solution to the obstacle problem $$\label{eq2.3} \left\{
\begin{array}{ll}
\min(u(s,x)-\psi(x),-u_s-L_{BS}u(s,x)+r u(s,x))=0, &
(s,x)\in Q_T,\medskip\\
u(T,x)=\psi(x), & x\in{{\mathbb R}}^n
\end{array}
\right.$$ with $$L_{BS}u=\sum^n_{i=1}(r-d_{i})x_{i}u_{x_{i}} +\frac12\sum^n_{i,j=1}
a_{ij}x_ix_ju_{x_{i}x_{j}}.$$ From [@EKPPQ; @EQ] we know that $V$ defined by (\[eq1.2\]) is equal to $Y^{s,x}_s$. Hence $$\label{eq2.4}
V(s,x)=Y^{s,x}_s=u(s,x),\quad(s,x)\in[0,T]\times{{\mathbb R}}^n.$$ In the next section we analyze $V$ via (\[eq2.4\]) but as a matter of fact instead of viscosity solutions of (\[eq2.3\]) we consider variational solutions which provide more information on the value function $V$.
Obstacle problem for the Black and Scholes equation {#sec3}
===================================================
Assume that $\psi:{{\mathbb R}}^n\rightarrow{{\mathbb R}}_+$ is continuous and satisfies the polynomial growth condition. Let $L^2_{\varrho}=L^2({{\mathbb R}}^n;\varrho^2\,dx)$, $H^{1}_{\varrho}=\{u\in
L^2_{\varrho}:\sum^n_{j=1} \sigma_{ij}x_iu_{x_j}\in
L^{2}({{\mathbb R}}^n;\varrho^2\,dx),\,i=1,\dots,n\}$ and $W_{\varrho}=\{u\in L^2(0,T;H^1_{\varrho}):u_t\in
L^2(0,T;H^{-1}_{\varrho})\}$, where $u_t,u_{x_i}$ denote the partial derivatives in the distribution sense, $\varrho(x)=(1+|x|^2)^{-\gamma}$ and $\gamma>0$ is chosen so that $\int_{{{\mathbb R}}^n}\varrho^2(x)\,dx<\infty$ and $\int_{{{\mathbb R}}^n}\psi^{2}(x)\varrho^2(x)\,dx<\infty$. Following [@BCFE; @KR:BPAN] we adopt the following definition.
\(a) A pair $(u,\mu)$ consisting of $u\in W_{\varrho}\cap C(\bar
Q_T)$ and a Radon measure $\mu$ on $Q_T$ is a variational solution to (\[eq2.3\]) if $$u(T,\cdot)=\psi,\quad u\ge
\psi,\quad \int_{Q_T}(u-\psi)\varrho^2\,d\mu=0$$ and the equation $$u_t+L_{BS}u=ru-\mu$$ is satisfied in the strong sense, i.e. for every $\eta\in
C_{0}^{\infty}(Q_{T})$, $$\langle u_t,\eta\rangle_{\varrho,T}+ \langle
L_{BS}u,\eta\rangle_{\varrho,T} =r\langle
u,\eta\rangle_{2,\varrho,T}-\int_{Q_{T}}\eta\varrho^{2}\,d\mu,$$ where $$\langle L_{BS}u,\eta\rangle_{\varrho,T}
=\sum^n_{i=1}\langle(r-d_i)x_iu_{x_i},\eta\rangle_{2,\varrho,T}
-\frac12\sum^n_{i,j=1}a_{ij}\langle u_{x_i},
(x_ix_j\eta\varrho^2)_{x_j}\rangle_{2,T}.$$ Here $\langle\cdot,\cdot\rangle_{\varrho,T}$ stands for the duality pairing between $L^2(0,T;H^1_{\varrho})$ and $L^2(0,T;H^{-1}_{\varrho})$, $\langle\cdot,\cdot\rangle_{2,\varrho,T}$ is the usual scalar product in $L^2(0,T;L^2_{\varrho})$ and $\langle\cdot,\cdot\rangle_{2,T}
=\langle\cdot,\cdot\rangle_{2,\varrho,T}$ with $\varrho\equiv1$.\
(b) If $\mu$ in the above definition admits a density (with respect to the Lebesgue measure) of the form $\Phi_u(t,x)=\Phi(t,x,u(t,x))$ for some measurable $\Phi:\bar
Q_T\times{{\mathbb R}}\rightarrow{{\mathbb R}}_+$, then we say that $u$ is a variational solution to the semilinear problem $$\label{eq3.10} u_t+L_{BS}u=ru-\Phi_u,\quad u(T,\cdot)=\psi,\quad
u\ge\psi.$$
In our main theorems below we show that if $\psi$ satisfies (A1) and (A2) or (A3) then the measure $\mu$ is absolutely continuous with respect to the Lebesgue measure and its density has the form $\Phi_u(t,x)=\mathbf{1}_{\{u(t,x)=\psi(x)\}}\Psi^{-}(x)$, where $\Psi^{-}=\max\{-\Psi,0\}$ and $\Psi$ is determined by $\psi$ and the parameters $r,d,a$. In the next section we compute $\Psi$ for some concrete options.
Payoffs satisfying (A1), (A2)
-----------------------------
One can check that if $u$ is a solution to (\[eq3.10\]) then $v$ defined as $$v(t,x)=u(T-t,(-1)^{i_1}e^{x_1},\dots,(-1)^{i_n}e^{x_n})\equiv
u(T-t,e^x)$$ for $t\in[0,T]$, $x=(x_1,\dots,x_n)\in D_{\iota}$, $\iota\in I$ ($D_{\iota}$ is defined in Section \[sec2\]) is a variational solution of the Cauchy problem $$v_{t}-Lv=-r v+\bar\Phi,\quad v\ge\bar\psi,\quad
v(0,\cdot)=\bar\psi,$$ where $$Lv=\sum^n_{i=1}(r-d_{i}-\frac12 \sigma_{ii}^{2})v_{x_{i}}+
\frac12\sum^n_{i,j=1}a_{ij}v_{x_{i}x_{j}}$$ and $\bar\Phi(t,x)=\Phi_u(T-t,e^x)$, $\bar\psi(t,x)=\psi(T-t,e^x)$. Furthermore, a simple calculation shows that if $\eta$ is a smooth function on ${{\mathbb R}}^n$ with compact support and $U\subset{{\mathbb R}}^n$ is a bounded open set such that supp$[\eta]\subset U$ then $\tilde{v}=v\eta$ is a solution of the Cauchy-Dirichlet problem $$\tilde{v}_{t}-\tilde{L}\tilde{v}=-r\tilde{v}+f,\quad \quad
\tilde{v}(0,\cdot)=\tilde{\psi},\quad \tilde
v_{|[0,T]\times\partial U}=0,$$ where $\tilde{\psi}=\bar\psi\eta$, $\tilde{L}$ is some uniformly elliptic operator with smooth coefficients not depending on $t$ and $f\in L^{2}(0,T;L^2(U))$. By classical regularity results (see, e.g., Theorem 5 in §7.1 in [@E]), $\tilde v\in
L^2(0,T;H^2(U))\cap L^{\infty}(0,T;H^1_0(U))$ and $\tilde v_t\in
L^2(0,T;L^2(U))$. From this and the construction of $\tilde v$ we infer that the regularity properties of $\tilde v$ are retained by $u$. It follows in particular that $$\label{eq3.11} u_t+L_{BS}u=ru-\Phi_u\quad\mbox{a.e. on }P_T.$$
\[tw2.1\] Assume .
1. $u$ defined by is a variational solution of the semilinear Cauchy problem $$\label{eq2.9} u_{t}+L_{BS}u=r u-\Phi^{-}_u,\quad u(T,\cdot)=\psi$$ with $$\Phi_u(t,x)=\mathbf{1}_{\{u(t,x)=\psi(x)\}}\Psi(x),\quad (t,x)\in
Q_T,$$ where for $x\in{{\mathbb R}}^n$ such that $(t,x)\in\{u=\psi\}$, $$\Psi(x)=-r\psi(x)+L_{BS}\psi(x).$$
2. Set $\sigma(x)=\{\sigma_{ij}x_i\}_{i,j=1,\dots,n}$ and $$\label{eq3.1}
K_{s,t}=\int_{s}^{t}\Phi^{-}_u(\theta,X_{\theta})\,d\theta,\quad
t\in [s,T].$$ Then for every $(s,x)\in P_{T}$ the triple $(u(\cdot,X),\sigma(X)u_x(\cdot,X),K_{s,\cdot})$ is a unique solution of RBSDE$_{s,x}(\psi,-r y,\psi)$.
Fix $(s,x)\in P_{T}$. Let $(Y^{s,x},Z^{s,x},K^{s,x})$ be a solution of RBSDE$_{s,x}(\psi,-r y,\psi)$ and let $u$ be a viscosity solution of (\[eq2.3\]). For $t_0\in(s,T)$ let $U\subset{{\mathbb R}}^n$ be an open set of assumption (A2). Then there exists $\eta\in C^{\infty}({{\mathbb R}}^n)$ such that $\eta\ge0$, $\eta\equiv1$ on $\{u=\psi\}\cap Q_{t_0}$ and $\eta\equiv0$ on $U^{c}$ (we make the convention that $\eta(t,x)=\eta(x)$). Of course $(Y^{s,x},Z^{s,x}, K^{s,x})$ is a solution of RBSDE$_{s,x}(Y^{s,x}_{t_{0}},-r y,\psi)$ on $[s,t_{0}]$. It is also a solution of RBSDE$_{s,x}(Y^{s,x}_{t_{0}},-ry,
\tilde{\psi})$ on $[s,t_{0}]$ with $\tilde{\psi}(x)=\eta(x)\psi(x)$, because $\tilde\psi\le\psi$ and by (\[eq2.1\]) and (\[eq2.2\]), $$\begin{aligned}
\int^{t_0}_s(Y^{s,x}_t-\tilde\psi(X_t))\,dK^{s,x}_t
&=\int^{t_0}_s(u(t,X_t)-\tilde\psi(X_t))
\mathbf{1}_{\{u(t,X_t)=\psi(X_t)\}}\,dK^{s,x}_t\\
&=\int^{t_0}_s(u(t,X_t)-\psi(X_t))\,dK^{s,x}_t=0.\end{aligned}$$ Since $\tilde{\psi}$ is smooth, applying Itô’s formula yields $$\tilde{\psi}(X_{t}) =\tilde\psi(X_s)+
\sum^n_{i=1}\int_{s}^{t}\tilde{\psi}_{x_i}(X_{\theta})\,dX^i_{\theta}
+\frac12\sum^n_{i,j=1}\int_{s}^{t}a_{ij}X^i_{\theta}X^j_{\theta}
\tilde{\psi}_{x_{i}x_{j}}(X_{\theta})\,d\theta.$$ From the above, (\[eq2.2\]) and [@EKPPQ Remark 4.3] it follows that there exists a predictable process $\alpha^{s,x}$ such that $0\le \alpha^{s,x}\le 1$ and $$\begin{aligned}
dK^{s,x}_{t}=\alpha^{s,x}_{t}\mathbf{1}_{\{u={\psi}\}}(X_{t})
\Big(-r\tilde{\psi}(X_t)
&+\sum^n_{i=1}(r-d_{i})X^{i}_{t}\tilde{\psi}_{x_{i}}(X_{t})\\
&\qquad+\frac12\sum^n_{i,j=1} a_{ij}
X^i_tX^j_t\tilde{\psi}_{x_{i}x_{j}}(X_{t})\Big)^{-}\,dt\end{aligned}$$ on $[s,t_{0}]$. Thus $$\label{eq3.20}
dK^{s,x}_{t}=\alpha^{s,x}_{t}\mathbf{1}_{\{u(t,X_t)=\psi(X_t)\}}
\Psi^{-}(X_t)\,dt$$ on $[s,t_{0}]$ for every $t_0\in[s,T)$. Consequently, the above equation is satisfied on $[s,T]$. Since the coefficients of the stochastic differential equation (\[eq2.05\]) satisfy the assumptions of the “equivalence of norm" result proved in [@BM] (see [@BM Proposition 5.1]), it follows from [@BCFE Theorem 3] that there exists a function $\alpha$ on $Q_{T}$ such that $0\le\alpha\le 1$ a.e. and for a.e. $(s,x)\in
Q_{T}$, $$\label{eq3.2} \alpha^{s,x}_{t}=\alpha(t,X_{t}),\,\quad dt\otimes
P_{s,x}\mbox{-a.s.}$$ Moreover, $u\in C(\bar Q_T)$ by [@EKPPQ Lemma 8.4] and from [@BCFE Theorem 3] it follows that $u\in W_{\varrho}$ and $u$ is a variational solution of the Cauchy problem $$u_{t}+L_{BS}u=r u-\alpha\mathbf{1}_{\{u=\psi\}}\Psi^{-},\quad
u(T,\cdot)=\psi.$$ By the above and (\[eq3.11\]), $$u_{t}+L_{BS}u=r u -\alpha\mathbf{1}_{\{u=\psi\}}\Psi^{-}\quad
\mbox{a.e. on } Q_T,$$ so by Lemma A.4 in Chapter II in [@KiS], $$\psi_{t}+L_{BS}\psi=r\psi-\alpha\Psi^{-}\quad \mbox{a.e. on }
\{u=\psi\}.$$ On the other hand, by the definition of $\Psi$, $$\psi_{t}+L_{BS}\psi=L_{BS}\psi=r\psi+\Psi\quad \mbox{on }
\{u=\psi\}.$$ Thus $\Psi=-\alpha\Psi^{-}$ a.e. on $\{u=\psi\}$, which implies that $\alpha\Psi=\Psi$ a.e. on $\{u=\psi\}$, and hence that $$\label{eq2.10}
\mathbf{1}_{\{u=\psi\}}\alpha\Psi^{-}
=\mathbf{1}_{\{u=\psi\}}\Psi^{-}\quad\mbox{a.e.}$$ Accordingly (\[eq2.9\]) is satisfied. From (\[eq2.02\]) it is clear that if $s\in[0,T)$ and $x\in D_{\iota}$ for some $\iota\in
I$ then $P_{s,x}(X_{t}\in D_{\iota},\, t\ge s)=1$ and for every $t\in(s,T]$ the random variable $X_t$ has strictly positive density on $D_{\iota}$ under $P_{s,x}$. From this and (\[eq2.10\]) it follows that $$\label{eq3.3}
\mathbf{1}_{\{u=\psi\}}(t,X_{t})\alpha(t,X_{t})\Psi^{-}(X_t)
=\mathbf{1}_{\{u=\psi\}}(t,X_{t})\Psi^{-}(X_t),\quad dt\otimes
P_{s,x}\mbox{-a.s.}$$ for every $(s,x)\in P_{T}$. In [@Kl:PA1] it is proved that the function $\mathbf{1}_{\{u=\psi\}}\alpha$ is a weak limit in $L^{2}(Q_{T})$ of some sequence $\{\alpha_{n}\}$ of nonnegative functions bounded by $1$ and such that $\alpha_{n}(t,X_{t})\rightarrow \alpha^{s,x}_{t}$ weakly in $L^{2}([0,T]\times\Omega;dt\otimes P_{s,x})$ for every $(s,x)\in
Q_{T}$. Therefore using once again the fact that for every $(s,x)\in P_{T}$ the process $X$ has a strictly positive transition density under $P_{s,x}$ we conclude that (\[eq3.2\]) holds for every $(s,x)\in P_{T}$, which when combined with (\[eq3.3\]) implies (\[eq3.1\]). What is left is to prove that for every $(s,x)\in P_{T}$, $$\label{eq3.4} Z^{s,x}_{t}=\sigma(X_t)u_x (t,X_{t}),\quad dt\otimes
P_{s,x}\mbox{-a.s.}$$ From the results proved in [@EKPPQ Section 6] it follows that for every $(s,x)\in Q_{T}$, $$\label{eq3.5} E_{s,x}\sup_{s\le t\le T}
|Y^{s,x,n}_t-Y^{s,x}_{t}|^{2}
+E_{s,x}\int_{s}^{T}|Z^{s,x,n}_{t}-Z^{s,x}_{t}|^{2}\,dt\rightarrow0,$$ where $(Y^{s,x,n},Z^{s,x,n})$ is a solution of the BSDE $$\begin{aligned}
\label{eq2.18} Y^{s,x,n}_{t}&=\psi(X_{T})-\int_{t}^{T}r
Y^{s,x,n}_{\theta}\,d\theta \nonumber\\
&\quad+\int_{t}^{T}n(Y^{s,x,n}_{\theta}-\psi(X_{\theta}))^{-}\,d\theta
-\int_{t}^{T}Z^{s,x,n}_{\theta}\,dB_{s,\theta}.\end{aligned}$$ It is known (see [@Pardoux]) that $$\label{eq3.8} Y^{s,x,n}_{t}=u_{n}(t,X_{t}),\quad t\in [s,T],\quad
P_{s,x}\mbox{-a.s.},$$ where $u_{n}$ is a viscosity solution of the Cauchy problem $$(u_n)_t+L_{BS}u_n=-r u_n+n(u_n-\psi)^{-},\quad u_n(T,\cdot)=\psi.$$ We know that $P_{s,x}(X_{t}\in D_{\iota},\, t\ge s)=1$ if $x\in
D_{\iota}$. Moreover, by classical regularity results (see, e.g., [@F Theorem 1.5.9] and Remark preceding Theorem \[tw2.1\]), $u_{n}\in C^{1,2}(P_{T})$. Therefore applying Itô’s formula shows that (\[eq2.18\]) holds true with $Z^{s,x,n}_{\theta}$ replaced by $\sigma(X_{\theta})(u_n)_x(\theta,X_{\theta})$. Since (\[eq2.18\]) has a unique solution (see [@EKPPQ Corollary 3.7]), it follows that $$\label{eq3.29} Z^{s,x,n}_{t}=\sigma(X_t)(u_{n})_x(t,X_{t}), \quad
dt\otimes P_{s,x}\mbox{-a.s.}$$ for every $(s,x)\in P_{T}$. By (\[eq3.5\]) and (\[eq3.8\]), $u_n\rightarrow u$ pointwise in $Q_T$. Moreover, from (\[eq3.8\]), (\[eq3.29\]) and standard estimates for solutions of BSDEs (see, e.g., [@EKPPQ Section 6]) it follows that there is $C>0$ such that for any $(s,x)\in P_T$, $$\label{eq3.30} E_{s,x}\sup_{s\le t\le T}
|u_n(t,X_t)|^2+E_{s,x}\int^T_s|\sigma(X_t)(u_n)_x(t,X_t)|^2\,dt
\le CE_{s,x}|\psi(X_T)|^2,$$ while from (\[eq3.5\]), (\[eq3.29\]) it follows that $$\label{eq3.31} E_{s,x}\sup_{s\le t\le T}
\int^T_s|\sigma(X_t)((u_n)_x-(u_m)_x)(t,X_t)|^2\,dt \rightarrow0$$ as $n,m\rightarrow\infty$. From (\[eq3.30\]) one can deduce that $u_n\in L^2(0,T;H_{\varrho})$ and then, by using (\[eq3.31\]), that $u_n\rightarrow u$ in $L^2(0,T;H_{\varrho})$ (see the arguments following (2.12) in the proof of [@KR:BPAN Theorem 2.3]). From the last convergence and (\[eq3.5\]), (\[eq3.29\]) it may be concluded that $$E_{s,x}\int^T_s|\sigma(X_t)(u_n)_x(t,X_t)-Z^{s,x}_t|^2\,dt=0$$ for $(s,x)\in P_T$, which implies (\[eq3.4\]).
Convex payoffs
--------------
Assume that $\psi:{{\mathbb R}}^n\rightarrow{{\mathbb R}}$ is convex. Let $m$ denote the Lebesgue measure on ${{\mathbb R}}^n$, $\nabla_i\psi$ denote the usual partial derivative with respect to $x_i$, $i=1\dots,n$, and let $E$ be set of all $x\in{{\mathbb R}}^n$ for which the gradient $$\nabla\psi(x)=(\nabla_1\psi(x),\dots,\nabla_n\psi(x))$$ exists. Since $\psi$ is locally Lipschitz function, $m(E^c)=0$ and $\nabla\psi=(\psi_{x_1},\dots,\psi_{x_n})$ a.e. (recall that $\psi_{x_i}$ stands for the partial derivative in the distribution sense). Moreover, for a.e. $x\in E$ there exists an $n$-dimensional symmetric matrix $\{H(x)=\{H_{ij}(x)\}$ such that $$\label{eq3.26} \lim_{E\ni y\rightarrow x}
\frac{\nabla\psi(y)-\nabla\psi(x) -H(x)(y-x)}{|y-x|}=0,$$ i.e. $H_{ij}(x)$ are defined as limits through the set where $\nabla_i\psi$ exists (see, e.g., [@AA Section 7.9]). By Alexandrov’s theorem (see, e.g., [@AA Theorem 7.10]), if $x\in E$ is a point where (\[eq3.26\]) holds then $\psi$ has second order differential at $x$ and $H(x)$ is the hessian matrix of $\psi$ at $x$, i.e. $H(x)=\{\nabla^2_{ij}\psi(x)\}$.
The second order derivative of $\psi$ in the distribution sense $D^2\psi=\{\psi_{x_ix_j}\}$ is a matrix of real-valued Radon measures $\{\mu_{ij}\}$ on ${{\mathbb R}}^n$ such that $\mu_{ij}=\mu_{ji}$ and for each Borel set $B$, $\{\mu_{ij}(B)\}$ is a nonnegative definite matrix (see, e.g., [@EG Section 6.3]). Let $\mu_{ij}=\mu^a_{ij}+\mu^s_{ij}$ be the Lebesgue decomposition of $\mu_{ij}$ into the absolutely continuous and singular parts with respect to $m$. By Theorem 1 in Section 6.4 in [@EG], $$\label{eq3.17} \mu^a_{ij}(dx)=\nabla^2_{ij}\psi(x)\,dx.$$
For $R>0$ set $D_R=P\cap\{x\in{{\mathbb R}}^n:|x|<R\}$ and $\tau_R=\inf\{t\ge s:X_t\notin D_R\}$. Let $\tilde L_{BS}$ denote the operator formally adjoint to $L_{BS}$. By [@Pi Theorem 4.2.5] for a sufficiently large $\alpha>0$ there exist the Green’s functions $G^{\alpha}_R$, $\tilde G^{\alpha}_R$ for $\alpha-L_{BS}$ and $\alpha-\tilde L_{BS}$ on $D_R$. Let $A$ be a continuous additive functional of ${{\mathbb X}}$ and let $\nu$ denote the Revuz measure of $A$ (see, e.g., [@Re]). By the theorem proved in Section V.5 of [@Re], for every nonnegative $f\in
C_0({{\mathbb R}}^d)$, $$E_{s,x}\int^{\tau_R}_se^{-\alpha t}f(X_t)\,dA^{\nu}_t
=\int_{{{\mathbb R}}^n}G^{\alpha}_R(x,y)f(y)\,\nu(dy).$$ Since $G^{\alpha}_R(x,y)=\tilde G^{\alpha}_R(y,x)$ by [@Pi Corollary 4.2.6], it follows that $$\label{eq3.14} E_{s,g\cdot m}\int^{\tau_R}_se^{-\alpha t}
f(X_t)\,dA^{\nu}_t=\int_{{{\mathbb R}}^n}\tilde
G^{\alpha}_Rg(y)f(y)\,\nu(dy)$$ for any nonnegative $g\in C_0(D_R)$, where $E_{s,g\cdot m}$ denotes the expectation with respect to the measure $P_{s,g\cdot
m}(\cdot)=\int P_{s,x}(\cdot)g(x)\,dx$ and $$\tilde G^{\alpha}_Rg(y)=\int_{G_R}\tilde
G^{\alpha}_R(y,x)g(x)\,dx.$$ Note that if $g$ is not identically equal to zero then $\tilde
G^{\alpha}_Rg$ is strictly positive (see [@Pi Theorem 4.2.5]).
Set $${{\mathcal{L}}}_{BS}=\sum^n_{i=1}(r-d_{i})x_{i}\nabla_i +\frac12\sum^n_{i,j=1}
a_{ij}x_ix_j\nabla^2_{ij}\,.$$
\[tw3.2\] Assume . Then assertions of Theorem \[tw2.1\] hold true with $L_{BS}$ replaced by ${{\mathcal{L}}}_{BS}$.
We use the notation of Theorem \[tw2.1\]. Fix $s\in[0,T)$. Since $\psi$ is a continuous convex function, from Itô’s formula proved in [@Bo] it follows that there exists a continuous increasing process $A$ such that for $x\in{{\mathbb R}}^n$, $$\label{eq3.9} \psi(X_{t})=\psi(X_{s})+A_{t}
+\int_{s}^{t}\nabla\psi (X_{\theta})\, dX_{\theta},\quad t\in
[s,T],\quad P_{s,x}\mbox{-a.s.}$$ From (\[eq3.9\]) it follows that $A$ is a positive continuous additive functional (PCAF for short) of ${{\mathbb X}}$. Let $\nu$ denote the Revuz measure of $A$. We are going to show that $\mathbf{1}_P\cdot\nu=\mathbf{1}_P\cdot\mu $ where $\mu$ is the measure on ${{\mathbb R}}^n$ defined as $$\mu(dx)=\sum^n_{i,j=1}a_{ij}x_ix_j\,\mu_{ij}(dx).$$ To this end, let us set $$\mu^{\varepsilon}_{ij}=\frac{\partial^2\psi}{\partial x_i\partial
x_j}*\rho_{\varepsilon}\,,\quad \mu^{\varepsilon}(dx)
=\sum^n_{i,j=1}a_{ij}x_ix_j\,\mu^{\varepsilon}_{ij}(dx),$$ where $\{\rho_{\varepsilon}\}_{\varepsilon>0}$ is some family of mollifiers. Fix a nonnegative $g\in C_0(D_R)$ such that $g(x)>0$ for some $x\in D_R$ and denote by $A^{\varepsilon}$ the PCAF of ${{\mathbb X}}$ in Revuz correspondence with $\mu_{\varepsilon}$. Then for a sufficiently large $\alpha>0$, $$\label{eq3.15} E_{s,g\cdot m}\int^{\tau_R}_se^{-\alpha t}
f(X_t)\,dA^{\varepsilon}_t =\int_{{{\mathbb R}}^n}\tilde
G^{\alpha}_Rg(y)f(y)\,\mu^{\varepsilon}(dy)$$ for all nonnegative $f\in C_0({{\mathbb R}}^d)$. By [@CP Theorem 2], $E_{s,x}\sup_{t\ge s}
|A^{\varepsilon}_{t\wedge\tau_R}-A_{t\wedge\tau_R}|\rightarrow0$ as $\varepsilon\downarrow0$ for every $x\in{{\mathbb R}}^d$. Hence $\int^{\tau_R}_se^{-\alpha t}f(X_t)\,dA^{\varepsilon}_t
\rightarrow \int^{\tau_R}_se^{-\alpha t}f(X_t)\,dA_t$ weakly under $P_{s,x}$ for $x\in{{\mathbb R}}^d$. Since $$A^{\varepsilon}_{t\wedge\tau_R}=\psi_{\varepsilon}(X_{t\wedge\tau_R})
-\psi_{\varepsilon}(X_s)-\int^{t\wedge\tau_R}_s
\nabla\psi_{\varepsilon}(X_{\theta})\,dX_{\theta}$$ and $\sup_{\varepsilon>0}\sup_{|x|\le R}|
\nabla\psi_{\varepsilon}(x)|\le C(R)<\infty$ by Lemma in [@CP], it follows that for every compact subset $K\subset{{\mathbb R}}^n$, $\sup_{x\in K}
\sup_{\varepsilon>0}E_{s,x}|A^{\varepsilon}_{t\wedge\tau_R}|^2<\infty$. Therefore $$\label{eq3.16} E_{s,g\cdot m}\int^{\tau_R}_se^{-\alpha t}
f(X_t)\,dA^{\varepsilon}_t\rightarrow E_{s,g\cdot m}
\int^{\tau_R}_se^{-\alpha t}f(X_t)\,dA_t$$ as $\varepsilon\downarrow0$. On the other hand, since $\mu^{\varepsilon}_{ij}\rightarrow\mu_{ij}$ weakly${}^{*}$ for $i,j=1,\dots,n$ and, by [@Pi Theorem 4.2.5], $f\tilde
G^{\alpha}_Rg\in C_0(G_R)$, we have $$\sum^n_{i,j=1}\int_{{{\mathbb R}}^n} \tilde
G^{\alpha}_Rg(y)f(y)a_{ij}y_iy_j\, \mu^{\varepsilon}_{ij}(dy)
\rightarrow \sum^n_{i,j=1}\int_{{{\mathbb R}}^n} \tilde
G^{\alpha}_Rg(y)f(y)\, a_{ij}y_iy_j\mu_{ij}(dy).$$ Combining this with (\[eq3.14\]), (\[eq3.15\]), (\[eq3.16\]) we see that for every $f\in C_0({{\mathbb R}}^n)$, $$\int_{{{\mathbb R}}^n}\tilde G^{\alpha}_Rg(y)f(y)\mu(dy)=\int_{{{\mathbb R}}^n}\tilde
G^{\alpha}_Rg(y)f(y)\,\nu(dy).$$ Since $\tilde G^{\alpha}_Rg$ is strictly positive on $D_R$, we conclude from the above that $\mu=\nu$ on $D_R$ for each $R>0$. Consequently, $\mu=\nu$ on $P$. For $x\in P$, $P_{s,x}(X_t\in{{\mathbb R}}^n\setminus P)=0$ for $t\ge s$. Hence $$\label{eq3.18}
A^{\nu}_t=\int^t_s\mathbf{1}_P(X_s)\,dA^{\nu}_s=A^{\mathbf{1}_P\cdot\nu}_t
=A^{\mathbf{1}_P\cdot\mu}_t,\quad t\ge s,\quad P_{s,x}\mbox{-a.s.}$$ for $x\in P$. Let $\mu^a$ denote the absolutely continuous part in the Lebesgue decomposition of $\mathbf{1}_P\cdot\mu$. By (\[eq3.17\]), $\mu^a(dx)=\sum^n_{i,j=1}\mathbf{1}_P(x)a_{ij}x_ix_j\nabla^2_{ij}\psi(x)\,dx$. Hence $$\label{eq3.19}
A^{\mu^a}_t=\sum^n_{i,j=1}\int^t_sa_{ij}X^i_{\theta}X^j_{\theta}
\nabla^2_{ij}\psi(X_{\theta})\,d\theta, \quad t\ge s, \quad
P_{s,x}\mbox{-a.s.}$$ for $x\in P$. From (\[eq3.9\]), (\[eq3.18\]), (\[eq3.19\]) and [@EKPPQ Remark 4.3] it follows that $$\begin{aligned}
dK^{s,x}_{t}=\alpha^{s,x}_{t}\mathbf{1}_{\{u=\psi\}}
(X_{t})\Big(-r\psi(X_{t})
&+\sum_{i=1}^{n}(r-d_{i})X^{i}_{t}\nabla_i\psi(X_{t})\\
&\qquad+\frac12\sum^n_{i,j=1} a_{ij}
X^i_tX^j_t\nabla^2_{ij}\psi(X_{t})\Big)^{-}\,dt.\end{aligned}$$ Let $u$ be a viscosity solution of (\[eq2.3\]). From the above and the results proved in [@BCFE] (see the reasoning following (\[eq3.20\])) we conclude that $u\in W_{\varrho}\cap C(\bar
Q_{T})$ and there is a function $\alpha$ on $Q_T$ such that $0\le\alpha\le 1$ a.e., (\[eq3.2\]) is satisfied and $u$ is a variational solution of the Cauchy problem $$\label{eq3.24} u_{t}+L_{BS}u=r u
-\alpha\mathbf{1}_{\{u=\psi\}}\Psi^{-},\quad u(T,\cdot)=\psi$$ with $$\label{eq3.22} \Psi=-r\psi+{{\mathcal{L}}}_{BS}\psi\quad \mbox{on }
\{u=\psi\}.$$ By Remark preceding Theorem \[tw2.1\], $u(t,\cdot)\in
H^2_{loc}({{\mathbb R}}^n)$. Therefore by Remark (ii) following Theorem 4 in Section 6.1 in [@EG] the distributional derivatives $u_{x_i}$, $u_{x_ix_j}$ are a.e. equal to the approximate derivatives $\nabla^{ap}_iu$, $(\nabla^{ap})^2_{ij}u$. Let ${{\mathcal{L}}}^{ap}_{BS}$ denote the operator defined as ${{\mathcal{L}}}_{BS}$ but with $\nabla_i$, $\nabla_{ij}$ replaced by $\nabla^{ap}_i$, $(\nabla^{ap})^2_{ij}$. Then $u$ is a variational solution of (\[eq3.24\]) with $L_{BS}$ replaced by ${{\mathcal{L}}}^{ap}_{BS}$ and (\[eq3.11\]) holds with $L_{BS}$ replaced by ${{\mathcal{L}}}^{ap}_{BS}$. Hence $$u_{t}+{{\mathcal{L}}}^{ap}_{BS}u=ru
-\alpha\mathbf{1}_{\{u=\psi\}}\Psi^{-}\quad\mbox{ a.e. on }Q_T.$$ On the other hand, since $\psi$ is convex, $\psi\in
BV_{loc}({{\mathbb R}}^n)$ as a locally Lipschitz continuous function and, by Theorem 3 in Section 6.3 in [@EG], $\psi_{x_i}\in
BV_{loc}({{\mathbb R}}^n)$, $i=1,\dots,n$. Therefore $\psi$ is twice approximately differentiable a.e. by Theorem 4 in Section 6.1 in [@EG]. It follows now from Theorem 3 in Section 6.1 in [@EG] that ${{\mathcal{L}}}^{ap}u={{\mathcal{L}}}^{ap}\psi$ a.e. on $\{u=\psi\}$. Consequently, $$\label{eq3.23}
{{\mathcal{L}}}^{ap}_{BS}\psi=r\psi-\alpha\Psi^{-}\quad\mbox{a.e. on }
\{u=\psi\}.$$ Moreover, since $\psi$ is convex, ${{\mathcal{L}}}_{BS}\psi={{\mathcal{L}}}^{ap}_{BS}\psi$ a.e. on ${{\mathbb R}}^n$ by Remark (i) following Theorem 4 in Section 6.1 in [@EG]. Therefore combining (\[eq3.22\]) with (\[eq3.23\]) we see that $\Psi=-\alpha\Psi^{-}$ a.e. on $\{u=\psi\}$ from which as in the proof of Theorem \[tw2.1\] we get (\[eq3.3\]). To complete the proof it suffices now to repeat step by step the arguments following (\[eq3.3\]) in the proof of Theorem \[tw2.1\].
The early exercise premium representation
=========================================
Let $\xi$ denote the payoff process for an American option with payoff function $\psi$, i.e. $$\xi_t=e^{-r(t-s)}\psi(X_t),\quad t\in[s,T],$$ and let $\eta$ denote the Snell envelope for $\xi$, i.e. the smallest supermartingale which dominates $\xi$. It is known (see, e.g., Section 2.5 in [@KS]) that $$\eta_t=e^{-r(t-s)}V(t,X_t),\quad t\in[s,T].$$ Assume (A1), (A2) or (A3). Applying Itô’s formula and using Theorem \[tw2.1\] or \[tw3.2\] we get $$\begin{aligned}
\eta_t=e^{-r(t-s)}Y^{s,x}_t&=e^{-r(T-s)}\psi(X_T) +\int^T_t
e^{-r(\theta-s)}\Phi^{-}(X_{\theta},Y^{s,x}_{\theta})\,d\theta\\
&\quad-\int^T_te^{-r(\theta-s)}Z^{s,x}_{\theta}\,dW_{\theta},\quad
t\in[s,T],\quad P_{s,x}\mbox{-a.s.},\end{aligned}$$ which leads to the following corollary.
For every $(s,x)\in Q_T$ the Snell envelope admits the representation $$\label{eq4.1} \eta_t=E_{s,x}\Big(e^{-r(T-s)}\psi(X_T)
+\int^T_te^{-r(\theta-s)}
\Phi^{-}(X_{\theta},Y^{s,x}_{\theta})\,d\theta\,|{{\mathcal{F}}}_t\Big),\quad
t\in[s,T].$$
Taking $t=s$ in (\[eq4.1\]) and using (\[eq2.2\]) we get the early exercise premium representation for the value function.
For every $(s,x)\in Q_T$ the value function $V$ admits the representation $$V(s,x)=V^E(s,x)+E_{s,x}\int^T_s e^{-r(t-s)}
{\mathbf{1}}_{\{V(t,X_t)=\psi(X_t)\}}\Psi^{-}(X_t)\,dt,$$ where $$V^E(s,x)=E_{s,x}\big(e^{-r(T-s)}\psi(X_T)\big)$$ is the value of the European option with payoff function $\psi$ and expiration time $T$.
In closing this section we show by examples that for many options $\Psi^{-}$ can be explicitly computed. Using results of §4 and §5 in [@Ro] one can check that the payoff functions $\psi$ in examples 1–4 below satisfy (A3). It is also easy to see that the payoff function $\psi$ in example 5 satisfies (A2$'$). Note that the payoff function in example 1 also satisfies (A2$'$) and, by [@BD2; @LS], the payoff functions in examples 2–4 satisfy (A2). We would like to stress that the last assertion is by no means evident. On the other hand, the convexity of $\psi$ in examples 2–4 is readily checked.
In all the examples we have computed the corresponding functions $\Psi^{-}$ on the region $\{u=\psi\}$. When computing $\Psi$ we keep in mind that $\{u=\psi\}\subset[0,T]\times\{\psi>0\}$.
1. $$\psi(x)=\big(\sum_{i=1}^{n} w_{i}x_i-K\big)^{+},\quad
\Psi^-(x)=\big(\sum_{i=1}^{n} w_{i}d_{i}x_i-r K\big)^{+}\quad
\mbox{(call)}$$ $$\psi(x)=\big(K-\sum_{i=1}^{n} w_{i}x_i\big)^{+},\quad
\Psi^-(x)=\big(r K-\sum_{i=1}^{n} w_{i}d_{i}x_i\big)^{+}\quad
\mbox{(put)}$$ (Here $w_i\in{{\mathbb R}}$ for $i=1,\dots,n$).
2. $$\psi(x)=(\max\{x_1,\dots,x_n\}-K)^{+} \quad \mbox{(call on max)}$$ $$\Psi^{-}(x)=\big(\sum_{i=1}^{n} d_{i}\mathbf{1}_{B_{i}}(x)x_i-r
K\big)^{+},$$ where $B_{i}=\{x\in{{\mathbb R}}^n; x_{i}>x_{j},\, j\neq i\}$.
3. $$\psi(x)=(K-\min\{x_1,\dots,x_n\})^{+}\quad \mbox{(put on min)}$$ $$\Psi^{-}(x)=\big(r K
-\sum_{i=1}^{n}d_{i}\mathbf{1}_{C_{i}}(x)x_i)^{+},$$ where $C_{i}=\{x\in{{\mathbb R}}^n; x_{i}<x_{j},\, j\neq i\}$.
4. $$\psi(x)=(\max\{x_1-K_{1},\dots, x_n-K_{n}\})^{+},$$ $$\Psi^{-}(x)=\big(\sum_{i=1}^{n} \mathbf{1}_{B_{i}}(x-K)(d_{i}x_i-r
K_{i})\big)^{+},$$ where $K=(K_{1},\dots,K_{n})$.
5. $$\psi(x)=(|x_1\cdot\ldots\cdot x_n|^{\gamma}-K)^+\quad \mbox{for
some }\gamma>0.$$ If $x\in D_{\iota}$ with $\iota=(i_1,\dots,i_n)\in\{0,1\}^n$ then $$\Psi^{-}(x)=\big((r-\gamma\sum_{i=1}^n(r-d_i-a_{ii})
-\gamma^2\sum^n_{i,j=1}a_{ij})f(x)-r K\big)^{+},$$ where $f(x)=((-1)^{|\iota|}\,x_1\cdot\ldots\cdot x_n)^{\gamma}$ and $|\iota|=i_1+\ldots+ i_n$.
[25]{}
Alberti, G., Ambrosio, L.: A geometrical approach to monotone functions in ${{\mathbb R}}^n$. Math. Z. [**230**]{}, 259–316 (1999)
Bally, V., Caballero, M.E., Fernandez, B., El Karoui, N.: Reflected BSDE’s, PDE’s and Variational Inequalities. Report No. 4455, pp. 1–32, INRIA, Le Chesnay, France (2002)
Bally, V., Matoussi, A.: Weak solutions for SPDEs and backward doubly stochastic differential equations. J. Theoret. Probab. [**14**]{}, 125–164 (2001)
Bensoussan, A.: On the theory of option pricing. Acta Appl. Math. [**2**]{}, 139–158 (1984)
Bouleau, N.: Formules de changement de variables. Ann. Inst. H. Poincaré Probab. Statist. [**20**]{} 133–145 (1984)
Broadie, M., Detemple, J.: American capped call options on dividend-paying assets. Rev. Financial Stud. [**8**]{}, 161–191 (1995)
Broadie, M., Detemple, J.: The valuation of American options on multiply assets. Math. Finance [**17**]{}, 241–286 (1997)
M. Broadie, J. Detemple, American options on dividencd-paying assets. In: G. Chichilnisky (ed.) Topology and Markets, pp. 69–97, Fields Inst. Commun. [**22**]{}, Amer. Math. Soc., Providence, RI (1999)
Carlen, E., Protter, P.: On semimartingale decompositions of convex functions of semimartingales. Illinois J. Math. [**36**]{}, 420–427 (1992)
Detemple, J.: American-style derivatives. Valuation and computation. Chapman & Hall/CRC, Boca Raton, FL (2006)
Detemple, J., Feng, S., Tian, W.: The valuation of American call options on the minimum of two dividend-paying assets. Ann. Appl. Probab. [**13**]{}, 953–983 (2003)
El Karoui, N., Kapoudjian, C., Pardoux, E., Peng, S., Quenez, M.C.: Reflected solutions of backward SDEs, and related obstacle problems for PDE’s. Ann. Probab. [**25**]{}, 702–737 (1997)
El Karoui, N., Karatzas, I.: The optimal stopping problem for a general American put-option. In: M. Davis et al. (eds): Mathematical Finance. IMA Vol. [**65**]{}, pp. 77–88, Springer, New York (1995)
El Karoui, N., Quenez, M.C.: Non-linear pricing theory and backward stochastic differential equations. Lecture Notes in Math. [**1656**]{}, 191–246 (1997)
Evans L.C.: Partial Differential Equations. American Mathematical Society, Providence, RI (1998)
Evans L.C., Gariepy R.F.: Measure Theorey and Fine Properties of Functions. CRC Press, Boca Raton, Fl (1992)
Friedman, A.: Partial differential equations of parabolic type. Prentice-Hall, Inc., Englewood Cliffs, N.J. (1964)
Jacka, S.: Optimal stopping and the American put. Math. Finance [**1**]{}, 1–14 (1991)
Jaillet, P., Lamberton, D., Lapeyre, B.: Variational Inequalities and the Pricing of American Options. Acta. Appl. Math. [**21**]{}, 263–289 (1990)
Karatzas, I.: On the pricing of American options. Appl. Math. Optim. [**17**]{}, 37–60 (1988)
Karatzas, I., Shreve, S.E.: Methods of Mathematical Finance. Springer, New York (1998)
Kim, I.J.: The analytic valuation of American options. Rev. Financial Studies [**3**]{}, 547–572 (1990)
Kinderlehrer, D., Stampacchia, G.: An introduction to variational inequalities and their applications. Academic Press, New York-London (1980)
Klimsiak, T.: Strong solutions of semilinear parabolic equations with measure data and generalized backward stochastic differential equations. Potential Anal. [**36**]{}, 373–404 (2012)
Klimsiak, T., Rozkosz, A.: On backward stochastic differential equations approach to valuation of American options. Bull. Pol. Acad. Sci. Math. [**59**]{}, 275–288 (2011)
Laurence, P., Salsa, S.: Regularity of the free boundary of an American option on several assets. Comm. Pure Appl. Math. [**62**]{}, 969–994 (2009)
Pardoux, É.: Backward Stochastic Differential Equations and Viscosity Solutions of Systems of Semilinear Parabolic and Elliptic PDEs of Second Order. In: L. Decreusefond at al. (eds.): Stochastic Analysis and Related Topics VI (The Geilo Workshop, 1996), pp. 79–129, Birkhauser, Boston (1998)
Pinsky, R.G.: Positive harminic functions and diffusion. Cambridge University Press, Cambridge (1995)
Revuz, D.: Measures associées aux fonctionelles additives de Markov I, Trans. Amer. Math. Soc. [**148**]{}, 501–531 (1970)
Rockafellar, R.T.: Convex anlysis. Princeton University Press, Princeton, N.J. (1970)
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'For a congruence subgroup $\Gamma$, we define the notion of $\Gamma$-equivalence on binary quadratic forms which is the same as proper equivalence if $\Gamma = \mathrm{SL}_2(\mathbb Z)$. We develop a theory on $\Gamma$-equivalence such as the finiteness of $\Gamma$-reduced forms, the isomorphism between $\Gamma_0(N)$-form class group and the ideal class group, $N$-representation of integers, and $N$-genus of binary quadratic forms. As an application, we deal with representations of integers by binary quadratic forms under certain congruence condition on variables.'
address: 'Department of Mathematics, Dongguk University-Seoul, 30 Pildong-ro 1-gil, Jung-gu, Seoul, 04620, Republic of Korea'
author:
- Bumkyu Cho
title: 'On the $\Gamma$-equivalence of binary quadratic forms'
---
[^1]
Introduction
============
Let $Q(x, y)$, $Q'(x, y)$ be two binary quadratic forms with coefficients in $\mathbb Z$ and let $\Gamma$ be a congruence subgroup of $\mathrm{SL}_2(\mathbb Z)$. If $Q'(x, y) = Q(ax + by, cx + dy)$ for some $(\begin{smallmatrix} a & b \\ c & d \end{smallmatrix}) \in \Gamma$, then we say that $Q'(x, y)$ is $\Gamma$-equivalent to $Q(x, y)$.
The $\mathrm{SL}_2(\mathbb Z)$-equivalence is clearly the same as the usual proper equivalence. The purpose of this article is to develop a theory of $\Gamma$-equivalence just as the theory of proper equivalence. To do this, we mainly follow the course given in Cox’s book [@Cox].
The beginning of this work is based on the following observation. Suppose that $Q(x, y)$ and $Q'(x, y)$ are $\Gamma$-equivalent.
i\) If $\Gamma = \Gamma_0(N)$, then $$\{ Q(x, y) \in \mathbb Z \, | \, x, y \in \mathbb Z, \, y \equiv 0 \, (\bmod \, N) \} \, = \, \{ Q'(x, y) \in \mathbb Z \, | \, x, y \in \mathbb Z, \, y \equiv 0 \, (\bmod \, N) \}.$$
ii\) If $\Gamma = \Gamma_1(N)$ and $r \in \mathbb Z$, then $$\{ Q(x, y) \in \mathbb Z \, | \, x \equiv r \, (\bmod \, N), \, y \equiv 0 \, (\bmod \, N) \} \, = \, \{ Q'(x, y) \in \mathbb Z \, | \, x \equiv r \, (\bmod \, N), \, y \equiv 0 \, (\bmod \, N) \}.$$
iii\) If $\Gamma = \Gamma(N)$ and $r, s \in \mathbb Z$, then $$\{ Q(x, y) \in \mathbb Z \, | \, x \equiv r \, (\bmod \, N), \, y \equiv s \, (\bmod \, N) \} \, = \, \{ Q'(x, y) \in \mathbb Z \, | \, x \equiv r \, (\bmod \, N), \, y \equiv s \, (\bmod \, N) \}.$$
In Section 3, we deal with representations of integers by binary quadratic forms under the congruence condition given in Case i) above.
The $\Gamma$-equivalence of binary quadratic forms
==================================================
We denote by $\mathrm{QF}$ the set of all primitive positive-definite binary quadratic forms, and denote by $\mathrm{QT}$ the set of all points $\tau \in \mathbb H$ such that $[\mathbb Q(\tau) : \mathbb Q] = 2$. Then a congruence subgroup $\Gamma$ of $SL_2(\mathbb Z)$ acts on both $\mathrm{QF}$ and $\mathrm{QT}$ in a natural way as $$\begin{aligned}
Q \cdot \gamma & = & (x \ y) \tensor*[^t]\gamma{} A_Q \gamma \left(\begin{smallmatrix} x \\ y \end{smallmatrix}\right), \\
\gamma(\tau) & = & \frac{a\tau + b}{c \tau + d},\end{aligned}$$ where $Q \in \mathrm{QF}$, $\gamma = \left(\begin{smallmatrix} a & b \\ c & d \end{smallmatrix}\right) \in \Gamma$, $\tau \in \mathrm{QT}$, and $A_Q$ is a symmetric matrix associated with $Q$. By definition, a form $Q'$ is $\Gamma$-equivalent to a form $Q$ if and only if $Q' = Q\cdot \gamma$ for some $\gamma \in \Gamma$.
Given a form $Q \in \mathrm{QF}$, let $\tau_Q$ denote the unique solution of $Q(x, 1) = 0$ in $\mathbb H$. Then it is not difficult to see that the map defined by $Q \mapsto \tau_Q$ is a one-to-one correspondence between $\mathrm{QF}$ and $\mathrm{QT}$, and moreover $$\tau_{Q \cdot \gamma} \, = \, \gamma^{-1}(\tau_Q).$$
Let $\mathfrak F_\Gamma$ be a fundamental region for a congruence subgroup $\Gamma$.
\(1) A form $Q \in \mathrm{QF}$ is said to be $\Gamma$-reduced if $\tau_Q \in \mathfrak F_\Gamma$. The set of all $\Gamma$-reduced forms in $\mathrm{QF}$ is denoted by $\Gamma\text{-}\mathrm{RF}$.
\(2) Given a negative integer $D \equiv 0, 1 \pmod 4$, we denote by $\mathrm{QF}(D)$ the set of all forms in $\mathrm{QF}$ of discriminant $D$, and denote by $\Gamma\text{-}\mathrm{RF}(D)$ the set of all $\Gamma$-reduced forms in $\mathrm{QF}(D)$.
It is worthwhile to remark that the definition of $\Gamma$-reduced forms depends on the choice of the fundamental region $\mathfrak F_\Gamma$.
As is well known, a fundamental region for $\mathrm{SL}_2(\mathbb Z)$ is $$\begin{aligned}
\mathfrak F_{\mathrm{SL}_2(\mathbb Z)} & = & \big\{ \tau \in \mathbb H \, \big| \, |\mathrm{Re}(\tau)| \leq 1/2, \ |\tau| \geq 1, \\
& & \qquad \quad \ \ \big( |\mathrm{Re}(\tau)| = 1/2 \mbox{ or } |\tau| = 1 \big) \Longrightarrow \mathrm{Re}(\tau) \leq 0 \big\}.\end{aligned}$$ With respect to this choice of $\mathfrak F_{\mathrm{SL}_2(\mathbb Z)}$, a form $ax^2 + bxy + cy^2 \in \mathrm{QF}$ is $\mathrm{SL}_2(\mathbb Z)$-reduced if and only if it satisfies
\(1) $|b| \leq a \leq c$
\(2) $(|b| = a \mbox{ or } a = c) \Longrightarrow b \geq 0$.\
Notice that it coincides with the usual definition of reduced forms for the proper equivalence. We comment that this relation between reduced forms and fundamental domains is not new to mathematicians (see [@Coh Section 5.3.1] and [@Cox Exercise 11.4]).
For $p = 2, 3$ it is not difficult to see that $$\begin{aligned}
\mathfrak F_{\Gamma_0(p)} & = & \big\{ \tau \in \mathbb H \, \big| \, |\mathrm{Re}(\tau)| \leq 1/2, \ |\tau - 1/p| \geq 1/p, \ |\tau + 1/p| \geq 1/p, \\
& & \qquad \quad \ \ \big( |\mathrm{Re}(\tau) = 1/2| \mbox{ or } |\tau - 1/p| = 1/p \mbox{ or } |\tau + 1/p| = 1/p \big) \Longrightarrow \mathrm{Re}(\tau) < 0 \big\}.\end{aligned}$$ Therefore $ax^2 + bxy + cy^2$ is defined to be $\Gamma_0(p)$-reduced if
\(1) $|b| \leq a$, $|b| \leq pc$
\(2) $(|b| = a \mbox{ or } |b| = pc) \Longrightarrow b > 0$.\
Under this defintion, we easily compute that $$\begin{aligned}
\Gamma_0(2)\text{-}\mathrm{RF}(-3) & = & \{ x^2 + xy + y^2 \}, \\
\Gamma_0(2)\text{-}\mathrm{RF}(-4) & = & \{ x^2 + y^2, \ 2x^2 + 2xy + y^2 \}, \\
\Gamma_0(2)\text{-}\mathrm{RF}(-7) & = & \{ x^2 + xy + 2y^2, \ 2x^2 \pm xy + y^2 \}, \\
\Gamma_0(2)\text{-}\mathrm{RF}(-8) & = & \{ x^2 + 2y^2, \ 2x^2 + y^2, \ 3x^2 + 2xy + y^2 \},
$$ and so on.
The reader will see soon that the coefficient of $x^2$ must be relatively prime to $N$ in order for $\Gamma_0(N)\text{-}\mathrm{RF}(D)$ to be an abelian group under the Dirichlet composition.
Let $p \geq 5$ be a prime number. Then $ax^2 + bxy + cy^2 \in \mathrm{QF}$ is defined to be $\Gamma_0(p)$-reduced if it satisfies the following properties:
1. $|b| \leq a$
2. $|b| \leq \frac{p^2c + (k^2 - 1)a}{p \cdot |k|}$ for all $k \in S_p$
3. $|b| = a \ \Longrightarrow \ b = a$
4. $b \neq -pc$
5. $b = - \frac{p^2c + (k^2 - 1)a}{pk}$ for some $k \in E_p^{(2)}$ $\Longrightarrow$ $b \geq -\frac{2ka}{p}$
6. $b = - \frac{p^2c + (k^2 - 1)a}{pk}$ for some $k \in S_p - (\{ \pm 1 \} \cup E_p^{(2)})$ $\Longrightarrow$ $b \geq -\frac{(2k_{(2)} + 1)a}{p}$
7. $b^2 - 4ac = - \frac{3a^2}{p^2}$ $\Longrightarrow$ $b \neq \frac{(1 - 2k)a}{p}$ for all $k \in S_p - (\{ 1 \} \cup E_p^{(3)})$ with $k \neq k_{(3)}$,
where the definitions of $S_p$, $E_p^{(2)}$, $k_{(2)}$, $E_p^{(3)}$, and $k_{(3)}$ are given in Section 4. The definition of $\Gamma_0(p)$-reduction above corresponds to the fundamental region for $\Gamma_0(p)$ described explicitly in Section 4.
From this definition, we can directly deduce that there are at most finitely many $\Gamma_0(p)$-reduced forms of a given discriminant $D$. In fact, it will turn out that the conditions (1) and (2) are good enough to imply the finiteness of $\Gamma_0(p)\text{-}\mathrm{RF}(D)$. Note that $$\begin{aligned}
& & |b| \leq \frac{p^2c + (k^2 - 1)a}{p \cdot |k|} \ \mbox{ for all } k \in S_p \\
& \Longrightarrow & \big|\tau_Q - k/p\big| \geq 1/p \ \mbox{ for all } k \in S_p, \mbox{ where } \tau_Q = (-b + \sqrt{D})/2a \\
& \Longrightarrow & \mathrm{Im}(\tau_Q) \geq \sqrt{3}/2 \mbox{ or } \big(|\mathrm{Re}(\tau)| \leq 1/2p, \ |\tau_Q - 1/p| \geq 1/p, \ |\tau_Q + 1/p| \geq 1/p \big) \\
& \Longrightarrow & a \leq \sqrt{-D/3} \mbox{ or } \big(|b| \leq a/p, \ |b| \leq pc \big).\end{aligned}$$
Fix a negative discriminant $D$.
$|b| \leq a$, $a \leq \sqrt{-D/3}$
Clearly, there are at most finitely many triplets $(a, b, c)$ of integers satisfying these conditions.
$|b| \leq a/p$, $|b| \leq pc$
Since $b^2 = |b| \cdot |b| \leq \frac{a}{p} \cdot pc = ac$, the finiteness follows immediately from the fact that $$\begin{aligned}
3b^2 \leq 4ac - b^2 = -D & \Longrightarrow & |b| \leq \sqrt{-D/3}, \\
4ac = b^2 - D \leq -4D/3 & \Longrightarrow & ac \leq -D/3.\end{aligned}$$
We now prove the finiteness of $\Gamma$-reduced forms of a given discriminant for all $\Gamma$.
Let $\Gamma\text{-}\mathrm{RF}(D)$ be the set of all $\Gamma$-reduced forms of discriminant $D$ with respect to $\mathfrak F_\Gamma$. Then we have the following.
\(1) Any two distinct $\Gamma$-reduced forms in $\Gamma\text{-}\mathrm{RF}(D)$ are not $\Gamma$-equivalent.
\(2) Every form $Q \in \mathrm{QF}(D)$ is $\Gamma$-equivalent to a unique $\Gamma$-reduced form in $\Gamma\text{-}\mathrm{RF}(D)$.
\(3) The cardinality of $\Gamma\text{-}\mathrm{RF}(D)$ is independent of the choice of $\mathfrak F_\Gamma$.
\(4) The cardinality $|\Gamma\text{-}\mathrm{RF}(D)|$ is finite.
\(1) Suppose that $Q, Q' \in \Gamma\text{-}\mathrm{RF}(D)$ are $\Gamma$-equivalent to each other. Since $Q' = Q \cdot \gamma$ for some $\gamma \in \Gamma$, we have $\tau_{Q'} = \gamma^{-1}(\tau_Q)$. This says that $\tau_Q, \tau_{Q'} \in \mathfrak F_\Gamma$ are in the same $\Gamma$-orbit, from which we deduce that $\tau_Q = \tau_{Q'}$. Appealing to the one-to-one correspondence between $\mathrm{QF}$ and $\mathrm{QT}$, it follows that $Q =Q'$.
\(2) Given a form $Q \in \mathrm{QF}(D)$, there exists $\gamma \in \Gamma$ such that $\tau_{Q \cdot \gamma} = \gamma^{-1}(\tau_Q)$ is contained in $\mathfrak F_\Gamma$. By definition, $Q$ is $\Gamma$-equivalent to $Q \cdot \gamma \in \Gamma\text{-}\mathrm{RF}(D)$. The uniqueness is an immediate consequence of (1).
\(3) Let $\Gamma\text{-}\mathrm{RF}(D)$ and $\Gamma\text{-}\mathrm{RF}(D)'$ be the sets of $\Gamma$-reduced forms in $\mathrm{QF}(D)$ with respect to $\mathfrak F_\Gamma$ and $\mathfrak F_\Gamma'$, respectively. According to (2) there is a function $f:\Gamma\text{-}\mathrm{RF}(D) \rightarrow \Gamma\text{-}\mathrm{RF}(D)'$ by mapping $Q$ to a unique $Q'$ in $\Gamma\text{-}\mathrm{RF}(D)'$ which is $\Gamma$-equivalent to $Q$. Note that $f$ is injective by (1). For if $f(Q_1) = f(Q_2)$, then $Q_1, Q_2 \in \Gamma\text{-}\mathrm{RF}(D)$ are $\Gamma$-equivalent, so we have $Q_1 = Q_2$. Now the symmetry immediately implies our desired assertion.
\(4) Let $\gamma_1, \ldots, \gamma_n$ be all the distinct left coset representatives of $\Gamma$ in $\mathrm{SL}_2(\mathbb Z)$, where $n = [\mathrm{SL}_2(\mathbb Z) : \Gamma]$. Then we can choose $\mathfrak F_\Gamma$ so that $\mathfrak F_\Gamma$ is contained in $$\bigcup_{i=1}^n \gamma_i^{-1}(\mathfrak F_{\mathrm{SL}_2(\mathbb Z)}).$$ Under this choice of $\mathfrak F_\Gamma$ we have $$\begin{aligned}
\Gamma\text{-}\mathrm{RF}(D) & = & \{ Q \in \mathrm{QF}(D) \, | \, \tau_Q \in \mathfrak F_\Gamma \} \\
& \subset & \{ Q \in \mathrm{QF}(D) \, | \, \gamma_i(\tau_Q) \in \mathfrak F_{\mathrm{SL}_2(\mathbb Z)} \mbox{ for some } i = 1, \ldots, n \} \\
& = & \{ Q \in \mathrm{QF}(D) \, | \, Q \cdot \gamma_i^{-1} \mbox{ is $\mathrm{SL}_2(\mathbb Z)$-reduced for some } i = 1, \ldots, n \} \\
& = & \bigcup_{i=1}^n \{ Q' \cdot \gamma_i \in \mathrm{QF}(D) \, | \, Q' \in \mathrm{SL}_2(\mathbb Z)\text{-}\mathrm{RF}(D) \}.\end{aligned}$$ Because $|\mathrm{SL}_2(\mathbb Z)\text{-}\mathrm{RF}(D)|$ is finite, we can derive that $|\Gamma\text{-}\mathrm{RF}(D)|$ is also finite.
For a negative discriminant $D$ and a positive integer $N$ we define $$\mathcal C(D, \Gamma_0(N)) \ = \ \{ ax^2 + bxy + cy^2 \in \mathrm{QF}(D) \, | \, \gcd(a, N) = 1 \} / \sim$$ where $\sim$ denotes the $\Gamma_0(N)$-equivalence. We simply denote $\mathcal C(D, \Gamma_0(1))$ by $\mathcal C(D)$ in accordance with standard notation.
Given an imaginary quadratic field $K$, let $d_K$ denote the discriminant of $K$, and $\mathcal O_K = \mathbb Z + \mathbb Z \frac{d_K + \sqrt{d_K}}{2}$ the ring of integers in $K$, and $\mathcal O := \mathbb Z + f\mathcal O_K$ an order of conductor $f$ in $K$. Then $\mathcal O$ has discriminant $D := f^2 d_K$. Every fractional $\mathcal O$-ideal is of the form $\alpha \mathfrak a$ for some $\alpha \in K^\ast$ and some $\mathcal O$-ideal $\mathfrak a$, and vice versa. A fractional $\mathcal O$-ideal $\mathfrak a$ is said to be proper if it satisfies $$\mathcal O \ = \ \{ \alpha \in K \, | \, \alpha \mathfrak a \subset \mathfrak a \}.$$ Then it is easy to see that all principal $\mathcal O$-ideals and all $\mathcal O_K$-ideals are proper. An $\mathcal O$-ideal $\mathfrak a$ is said to be relatively prime to $N \in \mathbb N$ if $\mathfrak a + N\mathcal O = \mathcal O$. Then we see from the proof of [@Cox Lemma 7.18] that $\mathfrak a$ is relatively prime to $N$ if and only if its norm $\mathrm{N}(\mathfrak a) := |\mathcal O / \mathfrak a|$ is relatively prime to $N$, and that every $\mathcal O$-ideal relatively prime to $f$ is proper.
Let $\mathcal I(\mathcal O)$ denote the group of all proper fractional $\mathcal O$-ideals, and $\mathcal I(\mathcal O, N)$ the subgroup of $\mathcal I(\mathcal O)$ generated by all proper $\mathcal O$-ideals relatively prime to $N$. We also denote by $\mathcal P(\mathcal O, N)$ the subgroup of $\mathcal I(\mathcal O)$ generated by all principal $\mathcal O$-ideals that are relativel prime to $N$, and we denote by $\mathcal P_0(\mathcal O, N)$ the subgroup of $\mathcal I(\mathcal O)$ generated by all principal ideals $\alpha \mathcal O$ where $\alpha \in \mathcal O$, $\alpha \equiv a \ (\bmod \ N\mathcal O)$ for some $a \in \mathbb Z$ with $(a, N) = 1$. Then $\mathcal P_0(\mathcal O, N) \subset \mathcal P(\mathcal O, N) \subset \mathcal I(\mathcal O, N)$ and hence we may consider the quotient groups $$\begin{aligned}
\mathcal C(\mathcal O, N) & := & \mathcal I(\mathcal O, N)/\mathcal P(\mathcal O, N), \\
\mathcal C_0(\mathcal O, N) & := & \mathcal I(\mathcal O, N)/\mathcal P_0(\mathcal O, N).\end{aligned}$$ We employ the standard notation $\mathcal C(\mathcal O)$ and $\mathcal P(\mathcal O)$ instead of $\mathcal C(\mathcal O, 1)$ and $\mathcal P(\mathcal O, 1)$, respectively. It is trivial that $\mathcal I(\mathcal O) = \mathcal I(\mathcal O, 1)$, $\mathcal P(\mathcal O) = \mathcal P_0(\mathcal O, 1)$, and $\mathcal C(\mathcal O) = \mathcal C_0(\mathcal O, 1)$. Note also that $\mathcal I(\mathcal O_K, N)$ and $\mathcal P_0(\mathcal O_K, N)$ are equal to the standard notation $\mathcal I_K(N)$ and $\mathcal P_{K, \mathbb Z}(N)$ in Cox’s book, respectively.
\[Proposition - bijection\] If $\mathcal O$ is the order of discriminant $D$ in an imaginary quadratic field, then the map $$\Phi: \mathcal C(D, \Gamma_0(N)) \longrightarrow \mathcal C_0(\mathcal O, N)$$ taking $[Q]$ to $[\mathbb Z + \mathbb Z \tau_Q]$ is a bijection.
(well-defined) Let $Q(x, y) = ax^2 + bxy + cy^2 \in \mathrm{QF}(D)$ satisfy $(a, N) = 1$. Then $\mathcal O = \mathbb Z + \mathbb Z a \tau_Q$ and $\mathbb Za + \mathbb Z a\tau_Q$ is a proper $\mathcal O$-ideal. Because its norm $$\mathrm{N}(\mathbb Za + \mathbb Z a\tau_Q) \ = \ |\mathcal O / (\mathbb Za + \mathbb Z a \tau_Q)| = a$$ is relatively prime to $N$, we see that $\mathbb Z + \mathbb Z\tau_Q \in \mathcal I(\mathcal O, N)$.
Now assume that $Q(x, y)$ and $Q'(x, y)$ are $\Gamma_0(N)$-equivalent to each other. Then we have $\tau_{Q'} = \gamma (\tau_Q)$ for some $\gamma = ( \begin{smallmatrix} a' & b' \\ c' & d' \end{smallmatrix} ) \in \Gamma_0(N)$. Note that $$(c'\tau_Q + d')(\mathbb Z + \mathbb Z \tau_{Q'}) \ = \ \mathbb Z(c'\tau_Q + d') + \mathbb Z(a'\tau_Q + b') \ = \ \mathbb Z + \mathbb Z \tau_Q$$ and that $$(c'\tau_Q + d')\mathcal O \ = \ (ac'\tau_Q + ad')a^{-1}\mathcal O \ \in \ \mathcal P_0(\mathcal O, N)$$ because $c'a\tau_Q + ad' \in \mathcal O$, $c'a\tau_Q + ad' \equiv a \pmod{N\mathcal O}$, and $(a, N) = 1$. Thus we have $[\mathbb Z + \mathbb Z \tau_Q] = [\mathbb Z + \mathbb Z \tau_{Q'}]$ in $\mathcal C_0(\mathcal O, N)$.
(1-to-1) Assume that $[\mathbb Z + \mathbb Z \tau_Q] = [\mathbb Z + \mathbb Z \tau_{Q'}]$, i.e. that $\alpha(\mathbb Z + \mathbb Z \tau_{Q'}) = \mathbb Z + \mathbb Z \tau_Q$ for some $\alpha \mathcal O \in \mathcal P_0(\mathcal O, N)$. Then we see that $\alpha \tau_{Q'} = a \tau_Q + b$ and $\alpha = c\tau_Q + d$ for some $a, b, c, d \in \mathbb Z$, from which we obtain that $\tau_{Q'} = \gamma(\tau_Q)$ with $\gamma := (\begin{smallmatrix} a & b \\ c & d \end{smallmatrix}) \in M_2(\mathbb Z)$. Since $$\mathbb Z + \mathbb Z \tau_Q \ = \ \alpha(\mathbb Z + \mathbb Z \tau_{Q'}) \ = \ \mathbb Z(c\tau_Q + d) + \mathbb Z(a\tau_Q + b)$$ and $\tau_Q, \tau_{Q'} \in \mathbb H$, we derive that $\gamma \in SL_2(\mathbb Z)$.
Now we show that $\gamma \in \Gamma_0(N)$. By definition, we may set $\alpha = \beta \delta^{-1}$, where $\beta, \delta \in \mathcal O$ and $\beta \equiv b' \pmod {N\mathcal O}$, $\delta \equiv d' \pmod {N\mathcal O}$ and $b', d' \in \mathbb Z$, $(b'd', N) = 1$. Let $a' = Q(1, 0)$. Then $\mathcal O = \mathbb Z + \mathbb Z a'\tau_Q$. Since $a'\alpha = c a' \tau_Q + a'd \in \mathcal O$, we have $$\begin{aligned}
a'\alpha \delta & \equiv & c d' a' \tau_Q + a'dd' \pmod{N\mathcal O},\\
a'\alpha \delta & = & a'\beta \ \equiv \ a'b' \pmod{N\mathcal O},\end{aligned}$$ and hence $c \equiv 0 \pmod N$. This says that $Q(x, y)$ is $\Gamma_0(N)$-equivalent to $Q'(x, y)$.
(onto) Let $\mathfrak a \in \mathcal I(\mathcal O, N)$. Then we may write $\mathfrak a = s^{-1} \mathfrak b$, where $\mathfrak b$ is a proper $\mathcal O$-ideal relatively prime to $N$ and $s$ is an integer relatively prime to $N$. Let $t = \min \{ k \in \mathbb N \, | \, k \in \mathfrak b \}$. Since $\mathfrak b \bar{\mathfrak b} = \mathrm{N}(\mathfrak b) \mathcal O$, we have $\mathrm{N}(\mathfrak b) \in \mathfrak b$ and so $t \leq \mathrm{N}(\mathfrak b)$. We write $\mathrm{N}(\mathfrak b) = tq + r$ for some $0 \leq r < t$. Since $r = \mathrm{N}(\mathfrak b) - tq \in \mathfrak b$, we deduce that $r = 0$. Thus $t$ divides $\mathrm{N}(\mathfrak b)$ and hence $(t, N) = 1$.
Write $\mathfrak b = \mathbb Z \alpha + \mathbb Z \beta$ for some $\alpha, \beta \in \mathcal O$ and put $t = u \alpha + v \beta$ for some $u, v \in \mathbb Z$. Notice that $(u, v) = 1$ by the minimality of $t$. So there exist $w, z \in \mathbb Z$ such that $uz - vw = 1$. Since $(\begin{smallmatrix} u & v \\ w & z \end{smallmatrix}) \in \mathrm{SL}_2(\mathbb Z)$, we have $$\mathfrak b \ = \ \mathbb Z \alpha + \mathbb Z \beta \ = \ \mathbb Z(u\alpha + v\beta) + \mathbb Z(w\alpha + z\beta) \ = \ t(\mathbb Z + \mathbb Z \tau),$$ where $\tau = \frac{w\alpha + z\beta}{t}$. Replacing $\tau$ by $-\tau$ if necessary, we may assume that $\tau \in \mathbb H$. Let $ax^2 + bx + c \in \mathbb Z[x]$ be the minimal polynomial having $\tau$ as its zero, where $(a, b, c) = 1$ and $a > 0$. Clearly, $Q(x, y) := ax^2 + bxy + cy^2$ is a primitive positive-definite quadratic form. Because $\mathbb Z + \mathbb Z\tau$ is a proper fractional ideal of both $\mathcal O$ and $\mathbb Z + \mathbb Z a\tau$, we see that $\mathcal O = \mathbb Z + \mathbb Z a\tau$ and hence that $Q$ has discriminant $D$. Moreover, we deduce that $(a, N) = 1$ because $t \tau \in \mathfrak b \subset \mathcal O = \mathbb Z + \mathbb Z a\tau$ implies that $a$ divides $t$. Finally, since $st^{-1}\mathcal O \in \mathcal P_0(\mathcal O, N)$, we have $$\Phi([Q(x, y)]) \ = \ [\mathbb Z + \mathbb Z \tau] \ = \ [ st^{-1} \mathfrak a] \ = \ [\mathfrak a].$$ This completes the proof.
\[Lemma - relatively prime for quadratic forms\] For any $Q \in \mathrm{QF}(D)$ and $M, N \in \mathbb N$ with $(M, N) = 1$, there exists $Q' \in \mathrm{QF}(D)$ such that $Q'$ is $\Gamma_0(N)$-equivalent to $Q$ and $Q'(1, 0)$ is relatively prime to $M$.
For now, let us assume that for any prime divisor $p_k$ $(1 \leq k \leq s)$ of $M$ there is a pair $(a_k, c_k)$ of integers such that $\mathrm{gcd}(a_k, c_k) = 1$ and $p_k \nmid Q(a_k, c_k)$. Applying the Chinese remainder theorem, we can find $c \in \mathbb Z$ such that $c \equiv 0 \pmod N$ and $c \equiv c_k \pmod{p_k}$ for all $k = 1, \ldots, s$. Now, relabeling the indices if necessary, we may assume that $p_i \nmid a_i$ for $1 \leq i \leq r$ and that $p_j|a_j$ for $r < j \leq s$. Clearly, $p_j \nmid c_j$ for $r < j \leq s$. According to both the Chinese remainder theorem and the Dirichlet theorem on primes in arithmetic progressions, there is a prime $q$ such that $q \equiv [p_{r+1} \cdots p_s]_{p_i}^{-1} a_i \pmod {p_i}$ for all $i = 1, \ldots, r$ and $q \nmid c$.
Set $a = p_{r+1} \cdots p_s q$. Note that $\mathrm{gcd}(a, c) = 1$ because $q \nmid c$ and $\mathrm{gcd}(p_j, c) = \mathrm{gcd}(p_j, c_j) = 1$ for $r < j \leq s$. Moreover, we see that $\mathrm{gcd}(M, Q(a, c)) = 1$ because $Q(a, c) \equiv Q(a_i, c_i) \not \equiv 0 \pmod{p_i}$ for $1 \leq i \leq r$ and $Q(a, c) \equiv Q(0, c_j) \equiv Q(a_j, c_j) \not \equiv 0 \pmod {p_j}$ for $r < j \leq s$. Since $\mathrm{gcd}(a, c) = 1$, we have $\gamma := (\begin{smallmatrix} a & b \\ c & d \end{smallmatrix}) \in \Gamma_0(N)$ for some $b, d \in \mathbb Z$. Put $Q' = Q \cdot \gamma$. Then we have shown that $Q'(1, 0) = Q(a, c)$ is relatively prime to $M$.
Finally, it remains to show that for any prime $p$ there is a pair $(a_0, c_0)$ of integers such that $\mathrm{gcd}(a_0, c_0) = 1$ and $p \nmid Q(a_0, c_0)$. Since $Q$ is primitive, it is readily seen that $\mathrm{gcd}(Q(1, 0),$ $Q(1, 1),$ $Q(0, 1)) = 1$, from which it follows directly that at least one of $Q(1, 0)$, $Q(1, 1)$, and $Q(0, 1)$ is not divisible by $p$.
We now review the formula for the Dirichlet composition. Given two forms $Q(x, y) = ax^2 + bxy + cy^2$, $Q'(x, y) = a'x^2 +b'xy + c'y^2$ $\in$ $\mathrm{QF}(D)$ satisfying $\gcd(a, a', \frac{b+b'}{2}) = 1$, there exists a unique integer $B$ modulo $2aa'$ such that $$\begin{aligned}
B & \equiv & b \pmod {2a} \\
B & \equiv & b' \pmod {2a'} \\
B^2 & \equiv & D \pmod {4aa'}.\end{aligned}$$ Then the Dirichlet composition of $Q(x, y)$ and $Q'(x, y)$ is defined to be the form $$Q''(x, y) \ = \ aa' x^2 + Bxy + \frac{B^2 - D}{4aa'} y^2 \ \in \ \mathrm{QF}(D).$$
We will show that the Dirichlet composition makes $\mathcal C(D, \Gamma_0(N))$ an abelian group. Let $[Q(x, y)]$, $[Q'(x, y)]$ $\in$ $\mathcal C(D, \Gamma_0(N))$ be given as $Q(x, y) = ax^2 + bxy + cy^2$, $Q'(x, y) = a'x^2 + b'xy + c'y^2$. According to the previous lemma we may assume that $(a, a') = 1$. Then the binary operation on $\mathcal C(D, \Gamma_0(N))$ is defined by $$[Q(x, y)] \cdot [Q'(x, y)] \ = \ [Q''(x, y)],$$ where $Q''(x, y)$ is the Dirichlet composition of $Q(x, y)$ and $Q'(x, y)$.
\[Theorem - isomorphism\] The set $\mathcal C(D, \Gamma_0(N))$ is an abelian group under the operation defined as above and the map $\Phi: \mathcal \mathcal C(D, \Gamma_0(N)) \rightarrow \mathcal C_0(\mathcal O, N)$ in Proposition \[Proposition - bijection\] is an isomorphism.
We need to show that the operation induced from the Dirichlet composition is well defined. Let $Q$, $Q'$, and $Q''$ be given as above. First of all, the operation is independent of the choice of $B$. For, if $B' \equiv B \pmod {2aa'}$, then $aa' x^2 + B'xy + \frac{(B')^2 - D}{4aa'} y^2$ is $\Gamma_0(N)$-equivalent to $Q''$ because $\Gamma_0(N)$ contains $(\begin{smallmatrix} 1 & 1 \\ 0 & 1 \end{smallmatrix})$.
Now we consider the three proper $\mathcal O$-ideals $$\mathbb Za + \mathbb Z\frac{-b + \sqrt{D}}{2}, \ \mathbb Za' + \mathbb Z\frac{-b' + \sqrt{D}}{2}, \ \mathbb Zaa' + \mathbb Z\frac{-B + \sqrt{D}}{2},$$ which are associated with $Q$, $Q'$, and $Q''$, respectively. We claim that $$(\mathbb Za + \mathbb Z\frac{-b + \sqrt{D}}{2}) \cdot (\mathbb Za' + \mathbb Z\frac{-b' + \sqrt{D}}{2}) \ = \ \mathbb Zaa' + \mathbb Z\frac{-B + \sqrt{D}}{2}$$ in $\mathcal I(\mathcal O, N)$. Setting $\Delta = \frac{-B + \sqrt{D}}{2}$ and making use of $B \equiv b \pmod{2a}$ and $B \equiv b' \pmod{2a'}$, the above three ideals can be written as $\mathbb Za + \mathbb Z\Delta$, $\mathbb Za' + \mathbb Z\Delta$, and $\mathbb Zaa' + \mathbb Z\Delta$, respectively. Since $\Delta^2 \equiv -B\Delta \ (\bmod \ {aa'})$ by the congruence equation $B^2 \equiv D \ (\bmod \ {4aa'})$, we have $$\begin{aligned}
(\mathbb Za + \mathbb Z\Delta) \cdot (\mathbb Za' + \mathbb Z\Delta) & = & \mathbb Z aa' + \mathbb Z a \Delta + \mathbb Za'\Delta + \mathbb Z\Delta^2 \\
& = & \mathbb Z aa' + \Delta(\mathbb Za + \mathbb Za' + \mathbb ZB) \\
& = & \mathbb Zaa' + \mathbb Z\Delta.\end{aligned}$$ Here we have used the property that $(a, a', B) = 1$.
Let $[Q] = [Q_0]$ and $[Q'] = [Q'_0]$, where $Q_0 = a_0x^2 + b_0xy + c_0y^2$ and $Q'_0 = a'_0x^2 + b'_0xy + c'_0y^2$ with $(a_0, a'_0, \frac{b_0 + b_0'}{2}) = 1$. Let $Q''_0 = a_0a'_0x^2 + B_0xy + \frac{B_0^2 - D}{4a_0a'_0}y^2$ be the Dirichlet composition of $Q_0$ and $Q'_0$. To complete the proof of the well-definedness, it suffices to show that $\Phi([Q'']) = \Phi([Q''_0])$ as $\Phi$ is bijective. Note that $$\begin{aligned}
\Phi([Q'']) & = & [\mathbb Z + \mathbb Z \tau_{Q''}] \ = \ \lbrack \mathbb Zaa' + \mathbb Z\Delta \rbrack \ = \ \lbrack \mathbb Za + \mathbb Z\Delta \rbrack \cdot \lbrack \mathbb Za' + \mathbb Z \Delta \rbrack \\
& = & \Phi([Q]) \cdot \Phi([Q']) \ = \ \Phi([Q_0]) \cdot \Phi([Q'_0]) \\
& = & \lbrack \mathbb Z a_0 + \mathbb Z\frac{-b_0 + \sqrt{D}}{2} \rbrack \cdot \lbrack \mathbb Z a'_0 + \mathbb Z\frac{-b'_0 + \sqrt{D}}{2} \rbrack \\
& = & \lbrack \mathbb Z a_0 a'_0 + \mathbb Z \frac{-B_0 + \sqrt{D}}{2} \rbrack \\
& = & \Phi([Q''_0]).\end{aligned}$$ Here we have employed the claim above. Therefore the operation is well defined and automatically it has been proved that $\mathcal C(D, \Gamma_0(N))$ is an abelian group and that $\Phi$ is an isomorphism.
The identity element of $\mathcal C(D, \Gamma_0(N))$ is $$\begin{array}{ll}
\lbrack x^2 - \frac{D}{4} y^2 \rbrack & \mbox{if } D \equiv 0 \ (\bmod \ 4), \\
\lbrack x^2 + xy + \frac{1 - D}{4} y^2 \rbrack & \mbox{if } D \equiv 1 \ (\bmod \ 4),
\end{array}$$ and the inverse element of $[ax^2 + bxy + cy^2]$ in $\mathcal C(D, \Gamma_0(N))$ is $[ax^2 - bxy + cy^2]$.
It follows immediately from the facts that $$\begin{array}{ll}
\Phi(\lbrack x^2 - \frac{D}{4} y^2 \rbrack) \ = \ \lbrack \mathbb Z + \mathbb Z \frac{\sqrt{D}}{2} \rbrack \ = \ \lbrack \mathcal O \rbrack & \mbox{if } D \equiv 0 \ (\bmod \ 4), \\
\Phi(\lbrack x^2 + xy + \frac{1 - D}{4} y^2 \rbrack) \ = \ \lbrack \mathbb Z + \mathbb Z \frac{-1 + \sqrt{D}}{2} \rbrack \ = \ \lbrack \mathcal O \rbrack \quad & \mbox{if } D \equiv 1 \ (\bmod \ 4),
\end{array}$$ and that $$\begin{aligned}
& & \hspace{-.6cm} \Phi(\lbrack ax^2 + bxy + cy^2 \rbrack) \cdot \Phi(\lbrack ax^2 - bxy + cy^2 \rbrack) \\
& = & \lbrack \mathbb Za + \mathbb Z\frac{-b + \sqrt{D}}{2} \rbrack \cdot \lbrack \mathbb Za + \mathbb Z\frac{b + \sqrt{D}}{2} \rbrack \\
& = & \lbrack \mathbb Za^2 + \mathbb Z ab + \mathbb Z ac + \mathbb Z \frac{-ab + a\sqrt{D}}{2} \rbrack \\
& = & \lbrack a \cdot (\mathbb Z + \mathbb Z\frac{-b + \sqrt{D}}{2}) \rbrack \\
& = & [ \mathcal O ].\end{aligned}$$
The following proposition is a slight generalization of Proposition 7.20 in Cox’s book [@Cox] and can be proved by the same argument as in the proof of Cox’s. However, we present here the proof for the sake of completeness.
\[Proposition - norm-preserving isomorphism\] Let $\mathcal O$ be an order of conductor $f$ in an imaginary quadratic field $K$ and let $N$ denote a positive integer. Then the map $\mathfrak a$ $\longmapsto$ $\mathfrak a \cap \mathcal O$ induces a norm-preserving isomorphism from $\mathcal I(\mathcal O_K, Nf)$ to $\mathcal I(\mathcal O, Nf)$ and its inverse map is given by $\mathfrak b$ $\longmapsto$ $\mathfrak b \mathcal O_K$.
Let $\mathfrak a$ be an $\mathcal O_K$-ideal relatively prime to $Nf$. Then the natural map $$\mathcal O / (\mathfrak a \cap \mathcal O) \longrightarrow \mathcal O_K / \mathfrak a$$ is an injective homomorphism. Moreover, this map is also surjective since $\mathfrak a + Nf\mathcal O_K = \mathcal O_K$ and $f \mathcal O_K \subset \mathcal O$. Thus we have $\mathrm{N}(\mathfrak a) = \mathrm{N}(\mathfrak a \cap \mathcal O)$ and so $\mathfrak a \cap \mathcal O \in \mathcal I(\mathcal O, Nf)$. We further deduce that $(\mathfrak a \cap \mathcal O)\mathcal O_K = \mathfrak a$ because $(\mathfrak a \cap \mathcal O)\mathcal O_K \subset \mathfrak a$ is clear and $$\mathfrak a \ = \ \mathfrak a \mathcal O \ = \ \mathfrak a(\mathfrak a \cap \mathcal O + Nf\mathcal O) \ \subset \ (\mathfrak a \cap \mathcal O)\mathcal O_K + f \mathfrak a \ \subset \ (\mathfrak a \cap \mathcal O)\mathcal O_K.$$
Now let $\mathfrak b$ be an $\mathcal O$-ideal relatively prime to $Nf$. Then we see $$\mathfrak b \mathcal O_K + Nf \mathcal O_K \ = \ (\mathfrak b + Nf\mathcal O)\mathcal O_K \ = \ \mathcal O \mathcal O_K \ = \ \mathcal O_K,$$ whence $\mathfrak b\mathcal O_K$ is an $\mathcal O_K$-ideal relatively prime to $Nf$. We further deduce that $\mathfrak b \mathcal O_K \cap \mathcal O = \mathfrak b$ because $\mathfrak b \subset \mathfrak b \mathcal O_K \cap \mathcal O$ is clear and $$\begin{aligned}
\mathfrak b \mathcal O_K \cap \mathcal O & = & (\mathfrak b \mathcal O_K \cap \mathcal O) \mathcal O \ = \ (\mathfrak b \mathcal O_K \cap \mathcal O)(\mathfrak b + Nf\mathcal O) \\
& \subset & \mathfrak b + Nf(\mathfrak b\mathcal O_K \cap \mathcal O) \ \subset \ \mathfrak b + Nf \mathfrak b \mathcal O_K \\
& \subset & \mathfrak b + N \mathfrak b \ \subset \ \mathfrak b.\end{aligned}$$
This shows that we get a bijection between the monoids of $\mathcal O_K$-ideals and $\mathcal O$-ideals that are relatively prime to $Nf$. Since the inverse map is obviously multiplicative, the bijection induces a well-defined isomorphism between $I(\mathcal O_K, Nf)$ and $I(\mathcal O, Nf)$.
\[Lemma - relatively prime for ideals\] Let $\mathcal O$ be an order in an imaginary quadratic field and let $M, N$ be positive integers which are relatively prime. Then every ideal class in $\mathcal C_0(\mathcal O, N)$ contains a proper $\mathcal O$-ideal whose norm is relatively prime to $M$.
Since $\Phi$ is bijective, every ideal class in $\mathcal C_0(\mathcal O, N)$ is of the form $[\mathbb Z + \mathbb Z \tau_Q]$ for some $Q = ax^2 + bxy + cy^2 \in \mathrm{QF}(D)$ with $(a, N) = 1$, where $D$ is the discriminant of $\mathcal O$. Appealing to Lemma \[Lemma - relatively prime for quadratic forms\], we may assume that $a$ is also relatively prime to $M$. Since $[\mathbb Z + \mathbb Z \tau_Q] = [\mathbb Za + \mathbb Za \tau_Q]$, the desired assertion readily follows from the proof of Proposition \[Proposition - bijection\], which says that $\mathbb Za + \mathbb Za\tau_Q$ is a proper $\mathcal O$-ideal having norm $a$.
Let $\mathcal O$ be an order of conductor $f$ in an imaginary quadratic field $K$ and let $N$ denote a positive integer. Then the map $$\Psi : \mathcal C_0(\mathcal O_K, Nf) \longrightarrow \mathcal C_0(\mathcal O, N)$$ given by $[\mathfrak a \mathfrak b^{-1}] \longmapsto [(\mathfrak a \cap \mathcal O)(\mathfrak b \cap \mathcal O)^{-1}]$ is an isomorphism, where $\mathfrak a$ and $\mathfrak b$ are $\mathcal O_K$-ideals relatively prime to $Nf$.
We consider the homomorphism $$\Psi' : \mathcal I(\mathcal O_K, Nf) \longrightarrow \mathcal C_0(\mathcal O, N)$$ defined multiplicatively from the map taking $\mathfrak a$ to $[\mathfrak a \cap \mathcal O]$, where $\mathfrak a$ is an $\mathcal O_K$-ideal relatively prime to $Nf$. According to Proposition \[Proposition - norm-preserving isomorphism\], this map is well defined because $\mathfrak a \cap \mathcal O \in \mathcal I(\mathcal O, Nf) \subset \mathcal I(\mathcal O, N)$. The map is also surjective because Lemma \[Lemma - relatively prime for ideals\] says that every ideal class in $\mathcal C_0(\mathcal O, N)$ contains a proper $\mathcal O$-ideal $\mathfrak b$ which is relatively prime to $Nf$ and because we have $\mathfrak b \mathcal O_K \cap \mathcal O = \mathfrak b$ by applying Proposition \[Proposition - norm-preserving isomorphism\] again.
We claim that $\mathrm{ker}(\Psi') = \mathcal P_0(\mathcal O_K, Nf)$.
($\supset$) Let $\alpha = a + Nf \beta$ where $\beta \in \mathcal O_K$, $a \in \mathbb Z$, and $(a, Nf) = 1$. Since $f\beta \in \mathcal O$ and $(a, N) = 1$, we infer from Proposition \[Proposition - norm-preserving isomorphism\] that $$\alpha \mathcal O_K \cap \mathcal O \ = \ (\alpha \mathcal O)\mathcal O_K \cap \mathcal O \ = \ \alpha \mathcal O \ \in \ \mathcal P_0(\mathcal O, N).$$
($\subset$) Let $\mathfrak a \mathfrak b^{-1} \in \mathrm{ker}(\Psi')$ where $\mathfrak a$, $\mathfrak b$ are $\mathcal O_K$-ideals relatively prime to $Nf$. Since $\mathrm{N}(\mathfrak b) \mathcal O_K = \mathfrak b \bar{\mathfrak b}$, we have $\mathfrak a \mathfrak b^{-1} = \mathrm{N}(\mathfrak b)^{-1} \mathfrak a \bar{\mathfrak b}$ and hence we see $\mathfrak a \bar{\mathfrak b} \in \mathrm{ker}(\Psi')$ and it is sufficient to show that $\mathfrak a \bar{\mathfrak b} \in \mathcal P_0(\mathcal O_K, Nf)$. Set $\mathfrak a \bar{\mathfrak b} \cap \mathcal O = \alpha \mathcal O$ where $\alpha = a + N\beta$, $\beta \in \mathcal O$, $a \in \mathbb Z$, and $(a, N) = 1$. Since $\mathcal O = \mathbb Z + f \mathcal O_K$, we may put $\alpha = b + Nf \gamma$ where $\gamma \in \mathcal O_K$, $b \in \mathbb Z$, and $(b, N) = 1$. Then we deduce that $(b, Nf) = 1$ because $(\mathrm{N}(\mathfrak a \bar{\mathfrak b}), Nf) = 1$ and $$\mathrm{N}(\mathfrak a \bar{\mathfrak b}) \ = \ \mathrm{N}(\mathfrak a \bar{\mathfrak b} \cap \mathcal O) \ = \ \mathrm{N}(\alpha \mathcal O) \ = \ \mathrm{N}(\alpha) \ = \ b^2 + Nf(\gamma + \bar{\gamma}) + N^2 f^2 \gamma \bar{\gamma}.$$ Therefore we obtain by applying Proposition \[Proposition - norm-preserving isomorphism\] once more that $$\mathfrak a \bar{\mathfrak b} \ = \ (\mathfrak a \bar{\mathfrak b} \cap \mathcal O) \mathcal O_K \ = \ \alpha \mathcal O \mathcal O_K \ = \ \alpha \mathcal O_K \ \in \ \mathcal P_0(\mathcal O_K, Nf).$$
\(1) Let $f, N$ be positive integers and let $\mathcal O$ (respectively, $\mathcal O'$) be the order of conductor $f$ (respectively, $Nf$) in an imaginary quadratic field $K$. Then we have $$\mathcal C_0(\mathcal O, N) \ \cong \ \mathcal C(\mathcal O').$$
\(2) Let $N$ be a positive integer and let $D \equiv 0, 1 \pmod 4$ be a negative integer. Then we have $$\mathcal C(D, \Gamma_0(N)) \ \cong \ \mathcal C(DN^2).$$
It is also true that $\mathcal C(\mathcal O') \cong \mathcal C(\mathcal O', Nf)$ according to [@Cox Proposition 7.19].
\(1) By the above theorem, we readily see that $$\mathcal C_0(\mathcal O, N) \ \cong \ \mathcal C_0(\mathcal O_K, Nf) \ \cong \ \mathcal C_0(\mathcal O', 1) \ = \ \mathcal C(\mathcal O').$$
\(2) It follows from (1) and Theorem \[Theorem - isomorphism\].
The $N$-representations of integers by binary quadratic forms
=============================================================
Let $f(x, y) \in \mathbb Z[x, y]$ be a binary quadratic form and let $N \in \mathbb N$. An integer $m$ is said to be $N$-represented by $f(x, y)$ if $m = f(x_0, y_0)$ for some $x_0, y_0 \in \mathbb Z$ with $(x_0, N) = 1$, $y_0 \equiv 0 \ (\bmod \ N)$. If $(x_0, y_0) = 1$ is further satisfied, then we say that $m$ is properly $N$-represented by $f(x, y)$.
If $f(x, y)$ and $g(x, y)$ are $\Gamma_0(N)$-equivalent, then $$\{ f(x, y) \in \mathbb Z \, | \, (x, N) = 1, \, y \equiv 0 \, (\bmod \, N) \} \, = \, \{ g(x, y) \in \mathbb Z \, | \, (x, N) = 1, \, y \equiv 0 \, (\bmod \, N) \}.$$
An integer $m$ is properly $N$-represented by $f(x, y)$ if and only if there exist $b, c \in \mathbb Z$ such that $f(x, y)$ is $\Gamma_0(N)$-equivalent to $mx^2 + bxy + cy^2$.
($\Leftarrow$) By definition, there exists $(\begin{smallmatrix} a & b \\ c & d \end{smallmatrix}) \in \Gamma_0(N)$ such that $mx^2 + bxy + cy^2 = f(ax + by, cx + dy)$. Substituting $x = 1$ and $y = 0$, we see that $m = f(a, c)$. Since $(a, N) = 1$, $c \equiv 0 \ (\bmod \ N)$, and $(a, c) = 1$, $m$ is properly $N$-represented by $f(x, y)$.
($\Rightarrow$) By definition, there exist $x_0, y_0 \in \mathbb Z$ such that $m = f(x_0, y_0)$, $(x_0, N) = 1$, $y_0 \equiv 0 \ (\bmod \ N)$ and $(x_0, y_0) = 1$. Since $(x_0, y_0) = 1$, we see that $ux_0 + vy_0 = 1$ for some $u, v \in \mathbb Z$. Put $\gamma = (\begin{smallmatrix} x_0 & -v \\ y_0 & u \end{smallmatrix}) \in \Gamma_0(N)$. Then $f(x, y)$ is $\Gamma_0(N)$-equivalent to $f(x_0x - vy, y_0x + uy)$. Because the coefficient of $x^2$ in $f(x_0x - vy, y_0x + uy)$ is $f(x_0, y_0) = m$, we are done.
Let $D \equiv 0, 1 \ (\bmod \ 4)$ be an integer, $m$ an odd integer with $(m, D) = 1$, and $N \in \mathbb N$. Then $m$ is properly $N$-represented by a primitive form of discriminant $D$ if and only if $D$ is a quadratic residue modulo $m$.
($\Rightarrow$) By the preceding lemma, we may assume that $m$ is properly $N$-represented by a form $mx^2 + bxy + cy^2$ of discriminant $D$. Since $D = b^2 - 4mc$, $D$ is a quadratic residue modulo $m$.
($\Leftarrow$) Suppose $D \equiv b^2 \ (\bmod \ m)$ for some $b \in \mathbb Z$. Since $m$ is odd, we may assume that $D$ and $b$ have the same parity. From $D \equiv 0, 1 \ (\bmod \ 4)$, we infer that $D \equiv b^2 \ (\bmod \ 4m)$, i.e. that $D = b^2 - 4mc$ for some $c \in \mathbb Z$. Then $m$ is properly $N$-represented by the form $mx^2 + bxy + cy^2$ of discriminant $D$. Note that $mx^2 + bxy + cy^2$ is primitive because $(m, D) = 1$.
Given a positive integer $m$, let $p$ denote an odd prime not dividing $m$. Then for any $N \in \mathbb N$, we obtain that $(\frac{-m}{p}) = 1$ if and only if $p$ is $N$-represented by a primitive form of discriminant $-4m$.
This follows immediately from the preceding lemma because $-4m$ is a quadratic residue modulo $p$ if and only if $(\frac{-4m}{p}) = (\frac{-m}{p}) = 1$.
Given a positive integer $m$, let $p$ denote an odd prime not dividing $m$. Then for any $N \in \mathbb N$, we obtain that $(\frac{-m}{p}) = 1$ if and only if $p$ is $N$-represented by one of the $\Gamma_0(N)$-reduced forms of discriminant $-4m$.
The proof is clear.
Let $D \equiv 0, 1 \ (\bmod \ 4)$ be a negative integer, $N \in \mathbb N$, and let $\chi:(\mathbb Z / D\mathbb Z)^\times \rightarrow \{ \pm 1 \}$ be a homomorphism given by $\chi([m]) = (\frac{D}{m})$, where $(\frac{\cdot}{\cdot})$ denotes the Kronecker symbol. Then for any odd prime $p$ not dividing $D$, the residue class $[p] \in \mathrm{ker}(\chi)$ if and only if $p$ is $N$-represented by one of the $\Gamma_0(N)$-reduced forms of discriminant $D$.
Notice that $$\begin{aligned}
& & [p] \in \mathrm{ker}(\chi) \\
& \Leftrightarrow & (\frac{D}{p}) = 1 \\
& \Leftrightarrow & \mbox{$D$ is a quadratic residue modulo $p$} \\
& \Leftrightarrow & \mbox{$p$ is $N$-represented by a primitive form of discriminant $D$} \\
& \Leftrightarrow & \mbox{$p$ is $N$-represented by one of the $\Gamma_0(N)$-reduced forms of discriminant $D$.}\end{aligned}$$
Given a negative integer $D \equiv 0, 1 \ (\bmod \ 4)$, the principal form is defined to be $$\left\{
\begin{array}{ll}
x^2 - \frac{D}{4}y^2 & \mbox{if } D \equiv 0 \ (\bmod \ 4) \\
x^2 + xy + \frac{1 - D}{4}y^2 & \mbox{if } D \equiv 1 \ (\bmod \ 4).
\end{array}
\right.$$
Let $f(x, y) = ax^2 + bxy + cy^2$ be a primitive form and let $n, N \in \mathbb N$ with $(a, N) = 1$. Then there exists $m \in \mathbb N$ such that $(m, nN) = 1$ and $m$ is properly $N$-represented by $f(x, y)$.
Given a prime $p$ not dividing $N$, at least one of $f(1, 0)$, $f(1, N)$, $f(1 + N, N)$ is not divisible by $p$. Hence for each prime divisor $p_i$ ($i = 1, \ldots, \ell$) of $n$ not dividing $N$, we deduce by relabeling the indices if necessary that $p_i$ does not divide $f(x_i, y_i)$ for all $i$, where $$\begin{aligned}
x_i = 1 \mbox{ or } 1 + N, & & y_i = N \qquad (1 \leq i \leq k), \\
x_i = 1, & & y_i = 0 \qquad (k + 1 \leq i \leq \ell).\end{aligned}$$ By the Chinese remainder theorem, there exists $x \in \mathbb Z$ such that $$\left\{
\begin{array}{ll}
x \equiv x_i & (\bmod \ {p_i}) \qquad (1 \leq i \leq \ell), \\
x \equiv 1 & (\bmod \ N).
\end{array}
\right.$$ By Dirichlet’s theorem on arithmetic progressions, we can find a prime $q$ such that $(q, x) = 1$ and $p_{k + 1} \cdots p_\ell q \equiv 1 \ (\bmod \ {p_1 \cdots p_k})$.
Put $y = N p_{k + 1} \cdots p_\ell q$ and $m = f(x, y)$. Then $$\begin{aligned}
n \equiv f(x_i, y_i) \not \equiv 0 \ (\bmod \ {p_i}) & & (1 \leq i \leq \ell), \\
n \equiv f(1, 0) = a \ (\bmod \ N), & & \end{aligned}$$ whence $(m, nN) = 1$. Moreover, we have $x \equiv 1 \ (\bmod \ N)$, $y \equiv 0 \ (\bmod \ N)$, and $(x, y) = 1$ because $$(x, p_i) \ = \ (x_i, p_i) \ = \ 1 \qquad (k + 1 \leq i \leq \ell).$$
Given a negative integer $D \equiv 0, 1 \ (\bmod \ 4)$, let $\chi : (\mathbb Z / D\mathbb Z)^\times \rightarrow \{ \pm 1 \}$ be a homomorphism given by $\chi ([m]) = (\frac{D}{m})$. Let $N \in \mathbb N$ and let $f(x, y) = ax^2 + bxy + cy^2$ be a primitive form of discriminant $D$ with $(a, N) = 1$. Then
\(1) $H := \{ [m] \in (\mathbb Z / D\mathbb Z)^\times \, | \, \mbox{$m$ is $N$-represented by the principal form of discriminant $D$} \}$ is a subgroup of $\mathrm{ker}(\chi)$.
\(2) $\{ [m] \in (\mathbb Z / D\mathbb Z)^\times \, | \, \mbox{$m$ is $N$-represented by $f(x, y)$} \}$ is a coset of $H$ in $\mathrm{ker}(\chi)$.
We first show that $\{ [m] \in (\mathbb Z / D\mathbb Z)^\times \, | \, \mbox{$m$ is $N$-represented by $f(x, y)$} \}$ is contained in $\mathrm{ker}(\chi)$. Let $m \in \mathbb N$ be $N$-represented by the form $f(x, y)$. Then we can write $m = d^2 m'$, where $m'$ is properly $N$-represented by $f(x, y)$ and $\chi([m]) = \chi([d])^2 \chi ([m']) = \chi ([m'])$. Thus we may assume that $m$ is properly $N$-represented by $f(x, y)$. By the first lemma on a previous page, we deduce that $b^2 - 4mc = D$ for some $b, c \in \mathbb Z$. The properties of the Kronecker symbol imply that $$\chi([m]) \ = \ (\frac{D}{m}) \ = \ (\frac{b^2}{m}) \ = \ 1$$ and hence our claim is proved.
Put $d = (x, y)$. Then $m' := f(x/d, y/d)$ is properly represented by $f$ and $m' = \frac{m}{d^2}$ is relatively prime to $nN$.
\(1) By applying the preceding lemma with $n = D$ and $a = 1$, we see that $H \subset \mathrm{ker}(\chi)$ is nonempty. The identities $$\begin{aligned}
(x^2 + ay^2)(z^2 + aw^2) & = & (xz - ayw)^2 + a(xw + yz)^2, \\
(x^2 + xy + ay^2)(z^2 + zw + aw^2) & = & (xz - ayw)^2+ (xz - ayw)(xw + yz + yw) + a(xw + yz + yw)^2\end{aligned}$$ show that $H$ is closed under multiplication, and hence $H$ is a subgroup of $\mathrm{ker}(\chi)$.
\(2) If we apply the preceding lemma again with $n = D$, we may assume that $f(1, 0) = a$ is relatively prime to $ND$. Suppose that $D \equiv 0 \ (\bmod \ 4)$. Then $b$ is even and can be written as $2b'$, and then we have $$a f(x, y) \ = \ (ax + b'y)^2 - \frac{D}{4} y^2.$$ Note that $(ax + b'y, N) = 1$ if $(x, N) = 1$ and $y \equiv 0 \ (\bmod \ N)$. Since $(a, D) = 1$, it follows that the values in $(\mathbb Z / D\mathbb Z)^\times$ which are $N$-represented by $f(x, y)$ lie in the coset $[a]^{-1} H$.
Conversely, if $[c] \in [a]^{-1}H$, then $ac \equiv z^2 - \frac{D}{4}w^2 \ (\bmod \ D)$ for some $z, w \in \mathbb Z$ with $(z, N) = 1$, $w \equiv 0 \ (\bmod \ N)$. Using the above identity, it is easy to solve $f(x, y) \equiv c \ (\bmod \ D)$ with $(x, N) = 1$ and $y \equiv 0 \ (\bmod \ N)$. Thus the coset $[a]^{-1}H$ precisely consists of the values $N$-represented in $(\mathbb Z / D \mathbb Z)^\times$ by $f(x, y)$.
The case that $D \equiv 1 \ (\bmod \ 4)$ is similary by making use of the identity $$af(x, y) \ = \ (ax + \frac{b-1}{2}y)^2 + (ax + \frac{b - 1}{2}y)y + \frac{1 - D}{4}y^2.$$
With the notations as above, we define the $N$-genus of a coset of $H$ in $\mathrm{ker}(\chi)$ to be the set consisting of all primitive forms $ax^2 + bxy + cy^2$ with $(a, N) = 1$ of discriminant $D$ which $N$-represent the values of the given coset modulo $D$.
The preceding lemma immediately implies the following refinement of a previous theorem.
Let $D \equiv 0, 1 \ (\bmod \ 4)$ be negative and let $H \subset \mathrm{ker}(\chi)$ be as in the preceding lemma. If $[a]H$ is a coset of $H$ in $\mathrm{ker}(\chi)$ and $p$ is an odd prime not dividing $D$, then $[p]$ is contained in $[a]H$ if and only if $p$ is $N$-represented by a $\Gamma_0(N)$-reduced form of discriminant $D$ in the genus of $[a]H$.
The genus containing the principal form is called the principal genus.
\(1) Let $m, N \in \mathbb N$ and $p$ an odd prime not dividing $m$. Then $p$ is $N$-represented by a form of discriminant $-4m$ in the principal genus if and only if for some integer $x$ with $(x, N) = 1$, $$\left\{
\begin{array}{ll}
p \equiv x^2 \mbox{ or } x^2 + m \ (\bmod \ {4m}) & \mbox{if $N$ is odd}, \\
p \equiv x^2 \ (\bmod \ {4m}) & \mbox{if $N$ is even}.
\end{array}
\right.$$ (2) Let $m, N \in \mathbb N$ and $p$ an odd prime not dividing $1 - 4m$. Then $p$ is $N$-represented by a form of discriminant $1 - 4m$ in the principal genus if and only if for some integer $x$ with $(x, N) = 1$, $$p \equiv x^2 \ (\bmod \ {1 - 4m}).$$
Note that $$4(x^2 + xy + my^2) \equiv (2x + y)^2 \ (\bmod \ {1 - 4m}) \ \Rightarrow \ x^2 + xy + my^2 \equiv (x')^2 \ (\bmod \ {1 - 4m}),$$ where $x' := x + 2^{-1}y \in (\mathbb Z / (1 - 4m)\mathbb Z)^\times$.
Consider the case that $D = -28$ and $N = 2$. There are exactly two $\Gamma_0(2)$-reduced forms $x^2 + 7y^2$, $7x^2 + y^2$ in the $2$-genera for the discriminant $-28$. The values in $(\mathbb Z/ 28\mathbb Z)^\times$ which are $2$-represented by $x^2 + 7y^2$ are $1, 9, 25$. The values $2$-represented by $7x^2 + y^2$ in $(\mathbb Z/ 28\mathbb Z)^\times$ are $11$, $23$, $15$. Note that $\mathrm{ker}(\chi) = \{1, 9, 11, 15, 23, 25 \}$. Therefore we deduce from the preceding theorem that for any prime $p$ with $(p, 14) = 1$, $$\begin{aligned}
p \equiv 1, 9, 25 \ (\bmod \ {28}) & \Leftrightarrow & \mbox{$p$ is $2$-represented by $x^2 + 7y^2$}, \\
p \equiv 11, 15, 23 \ (\bmod \ {28}) & \Leftrightarrow & \mbox{$p$ is $2$-represented by $7x^2 + y^2$}.\end{aligned}$$ Note that $(7x^2 + y^2)(7z^2 + w^2) = (7xz - yw)^2 + 7(xw + yz)^2$ and $(7xz - yw, 2) = 1$, $xw + yz \equiv 0 \ (\bmod \ 2)$ if $(xz, 2) = 1$, $y \equiv w \equiv 0 \ (\bmod \ 2)$. From this, we further infer that for primes $p$, $q$ with $(pq, 14) = 1$, $$p, q \equiv 11, 15, 23 \ (\bmod \ {28}) \ \Rightarrow \ pq = x^2 + 7y^2, \ (x, 2) = 1, \ y \equiv 0 \ (\bmod \ 2).$$ In fact, there is a version of the above identity that holds for any form of discriminant $-4m$ $$(ax^2 + 2bxy + cy^2)(az^2 + 2bzw + cw^2) \ = \ (axz + bxw + byz + cyw)^2 + m(xw - yz)^2,$$ where $(axz + bxw + byz + cyw, N) = 1$ and $xw - yz \equiv 0 \ (\bmod \ N)$ if $(a, N) = 1$, $(xz, N) = 1$, and $y \equiv w \equiv 0 \ (\bmod \ N)$.
A positive integer $m$ is $N$-represented by a primitive form $f(x, y)$ of discriminant $D < 0$ if and only if there exists a proper $\mathcal O$-ideal $\mathfrak a \in \Phi([f(x, y)])$ such that $m = \mathrm{N}(\mathfrak a)$.
($\Rightarrow$) Let $m = f(x_0, y_0)$ for some $x_0, y_0 \in \mathbb Z$ with $(x_0, N) = 1$, $y_0 \equiv 0 \ (\bmod \ N)$. Put $d = (x_0, y_0)$. Then $a := m/d^2$ is properly $N$-represented by $f(x, y)$. Without loss of generality we may assume that $f(x, y) = ax^2 + bxy + cy^2$ for some $b, c \in \mathbb Z$ by a previous lemma. Since $\Phi([f(x, y)]) = [\mathbb Za + \mathbb Za\tau]$ and $\mathrm{N}(\mathbb Za + \mathbb Z a \tau) = a$ with $\tau = \frac{-b + \sqrt{D}}{2a}$, we see that $\mathrm{N}(\mathbb Zad + \mathbb Zad\tau) = ad^2 = m$ and $\mathbb Zad + \mathbb Zad\tau \in \Phi([f(x, y)])$.
($\Leftarrow$) Let $f(x, y) = ax^2 + bxy + cy^2$ be a primitive form of discriminant $D$ with $(a, N) = 1$ and let $\tau \in \mathbb H$ satisfy $f(\tau, 1) = 0$. From the proof of the surjectivity of $\Phi$, we see that $\mathfrak a = t(\mathbb Z + \mathbb Z \tau)$ for some $t \in \mathbb Z$ with $(t, N) = 1$. Then we have $\Phi([f(x, y)]) = [\mathbb Za + \mathbb Za\tau] = [\mathfrak a]$. We claim that $m$ is $N$-represented by $f(x, y)$. Because $t\tau \in \mathfrak a \subset \mathcal O = \mathbb Z + \mathbb Za\tau$, we put $t = as$ for some $s \in \mathbb Z$. Then we have $(s, N) = 1$ and $$m \ = \ \mathrm{N}(\mathfrak a) \ = \ \frac{t^2}{a} \ = \ as^2 \ = \ f(s, 0).$$
Let $\mathcal O$ be an order in an imaginary quadratic field and $N \in \mathbb N$. Given a positive integer $n$, every ideal class in $C_0(\mathcal O, N)$ contains a proper $\mathcal O$-ideal whose norm is relatively prime to $n$.
It follows directly from the preceding theorem and a previous lemma.
The fundamental region for $\Gamma_0(p)$
========================================
Throughout this section, let $p \geq 5$ denote a prime number. The purpose of this section is to construct the fundamental region $\mathfrak F_{\Gamma_0(p)}$ for $\Gamma_0(p)$ which satisfies the following properties:
1. $|\mathrm{Re}(\tau)| \leq 1/2$ for all $\tau \in \mathfrak F_{\Gamma_0(p)}$
2. the imaginary part of $\tau \in \mathfrak F_{\Gamma_0(p)}$ is maximal in the orbit $\Gamma_0(p)\tau$
3. the real part of $\tau \in \mathfrak F_{\Gamma_0(p)}$ is minimal among the points $\tau' \in \Gamma_0(p)\tau$ with $|\mathrm{Re}(\tau')| \leq 1/2$
4. for any $\tau \in \mathbb H$, there exists a unique point $\tau' \in \mathfrak F_{\Gamma_0(p)}$ such that $\tau' \in \Gamma_0(p) \tau$.
Given a congruence subgroup, we can explicitly construct its open fundamental region consisting of points whose imaginary part is maximal in its orbit. This can be done by applying the method given in Ferenbaugh’s paper [@Fer Section 3]. For a matrix $\gamma = (\begin{smallmatrix} a & b \\ c & d \end{smallmatrix})$, we write $a_\gamma := a$, $b_\gamma := b$, and so on. As in his paper, let $$\begin{aligned}
\mathrm{arc}(\gamma) & = & \{ \tau \in \mathbb H \, \big| \, |\tau - a_\gamma/c_\gamma| = 1/|c_\gamma| \}, \\
\mathrm{inside}(\gamma) & = & \{ \tau \in \mathbb H \, \big| \, |\tau - a_\gamma/c_\gamma| < 1/|c_\gamma| \}, \\
\mathrm{outside}(\gamma) & = & \{ \tau \in \mathbb H \, \big| \, |\tau - a_\gamma/c_\gamma| > 1/|c_\gamma| \}\end{aligned}$$ for any $\gamma \in \mathrm{SL}_2(\mathbb Z)$ with $c_\gamma \neq 0$. If $\gamma = \pm(\begin{smallmatrix} 1 & b \\ 0 & 1 \end{smallmatrix})$ with $b > 0$, then we define $$\begin{aligned}
\mathrm{arc}(\gamma) & = & \{ \tau \in \mathbb H \, \big| \, \mathrm{Re}(\tau) = b/2 \}, \\
\mathrm{inside}(\gamma) & = & \{ \tau \in \mathbb H \, \big| \, \mathrm{Re}(\tau) > b/2 \}, \\
\mathrm{outside}(\gamma) & = & \{ \tau \in \mathbb H \, \big| \, \mathrm{Re}(\tau) < b/2 \}.\end{aligned}$$ On the other hand, if $\gamma = \pm(\begin{smallmatrix} 1 & b \\ 0 & 1 \end{smallmatrix})$ with $b < 0$, then we define $\mathrm{arc}(\gamma)$ in the same way and reverse the definitions of $\mathrm{inside}(\gamma)$ and $\mathrm{outside}(\gamma)$.
An open fundamental region for a congruence subgroup $\Gamma$ of $\mathrm{SL}_2(\mathbb Z)$ is defined to be an open subset $R$ of $\mathbb H$ with the properties:
1. there do not exist $\gamma \in \Gamma$ and $\tau, \tau' \in R$ such that $\tau \neq \tau'$ and $\tau = \gamma(\tau')$
2. for any $\tau \in \mathbb H$, there exists $\gamma \in \Gamma$ such that $\gamma(\tau) \in \bar{R}$.
Let $\Gamma$ be a congruence subgroup containing $(\begin{smallmatrix} 1 & 1 \\ 0 & 1 \end{smallmatrix})$. Then $$R_\Gamma \ := \ \bigcap_{\gamma \in \Gamma - \{ \pm I \}} \mathrm{outside}(\gamma)$$ is an open fundamental region for $\Gamma$ such that
\(1) $|\mathrm{Re}(\tau)| \leq 1/2$ for all $\tau \in R_\Gamma$
\(2) the imaginary part of $\tau \in \overline{R_\Gamma}$ is maximal in the orbit $\Gamma \tau$.
See [@Fer Theorem 3.3] and its proof.
Thus the fundamental region $\mathfrak F_{\Gamma_0(p)}$ that we try to find lies between $R_{\Gamma_0(p)}$ and $\overline{R_{\Gamma_0(p)}}$.
$$R_{\Gamma_0(p)} \ = \ \{ \tau \in \mathbb H \, \big| \, |\mathrm{Re}(\tau)| < 1/2, \
|\tau - k/p| > 1/p \mbox{ for all } k = \pm 1, \pm 2, \ldots, \pm [p/2] \}.$$
Note that $$\bigcap_{\gamma \in \Gamma_0(p) - \{ \pm I \} \atop c_\gamma = 0} \mathrm{outside}(\gamma) \ = \ \{ \tau \in \mathbb H \, \big| \, |\mathrm{Re}(\tau)| < 1/2 \}$$ and that $$\bigcap_{\gamma \in \Gamma_0(p) - \{ \pm I \} \atop |c_\gamma| = p} \mathrm{outside}(\gamma) \ = \ \{ \tau \in \mathbb H \, \big| \, |\tau - k/p| > 1/p \mbox{ for all } k \in \mathbb Z \mbox{ with } (k, p) = 1 \}$$ because $$\gamma \in \Gamma_0(p), \ c_\gamma = \pm p \ \Longrightarrow \ \mathrm{outside}(\gamma) \ = \ \{ \tau \in \mathbb H \, \big| \, |\tau \mp a_\gamma / p| > 1/p \}.$$ Thus we see that $$\bigcap_{\gamma \in \Gamma_0(p) - \{ \pm I \} \atop |c_\gamma| \leq p} \mathrm{outside}(\gamma) \ = \ \{ \tau \in \mathbb H \, \big| \, |\mathrm{Re}(\tau)| < 1/2, \ |\tau - k/p| > 1/p \mbox{ for all } k = \pm 1, \pm 2, \ldots, \pm [p/2] \}.$$ Let $\tau$ be a point contained in the set right above. If $|\mathrm{Re}(\tau)| \leq 1 / 2p$, then it is clear that $\tau \in \mathrm{outside}(\gamma)$ for any $\gamma \in \Gamma_0(p)$ with $|c_\gamma| \geq 2p$. If $|\mathrm{Re}(\tau)| > 1 / 2p$, then we see that $\mathrm{Im}(\tau) > \sqrt{3}/2p$ and so we deduce that $\tau \in \mathrm{outside}(\gamma)$ for any $\gamma \in \Gamma_0(p)$ with $|c_\gamma| \geq 2p$ because every point in $\mathrm{inside}(\gamma)$ has its imaginary part less than $1 / 2p$.
Let $S_p = \{ \pm 1, \pm 2, \ldots, \pm \frac{p - 1}{2} \}$. For any $x \in \mathbb Z$ with $(x, p) = 1$, we denote by $x^{-1} \in S_p$ (respectively, $\langle x \rangle \in S_p$) the unique integer such that $xx^{-1} \equiv 1 \pmod p$ (respectively, $x \equiv \langle x \rangle \pmod p)$. For any $k \in S_p$, we define $$\gamma_k \ = \ \begin{pmatrix}
k & \frac{kk^{-1} - 1}{p} \\ p & k^{-1}
\end{pmatrix} \ \in \ \Gamma_0(p).$$ Then we have shown in the above lemma that $$R_{\Gamma_0(p)} \ = \ \{ \tau \in \mathbb H \, \big| \, |\mathrm{Re}(\tau)| < 1/2 \} \bigcap \ \big(
\bigcap_{k \in S_p} \mathrm{outside}(\gamma_k) \big).$$
Notice that for any $k \in S_p$, the matrix $\gamma_k$ maps $\mathrm{arc}(\gamma_{-k^{-1}})$ onto $\mathrm{arc}(\gamma_k)$ taking endpoints to endpoints as follows: $$\gamma_{k}\Big(\frac{-k^{-1} \pm 1}{p}\Big) \ = \ \frac{k \mp 1}{p}.$$ Observe also that an elliptic point of order $2$ can occur only at the top points of the arcs and that an elliptic point of order $3$ can occur only at the points where the lines and the arcs intersect, namely the points $\frac{k}{p} - \frac{1}{2p} + \frac{\sqrt{3}}{2p}i$ for $k \in S_p - \{ 1 \}$.
From these, we deduce that the top point of $\mathrm{arc}(\gamma_k)$ is an elliptic point of order $2$ if and only if $\mathrm{arc}(\gamma_{-k^{-1}})$ $=$ $\mathrm{arc}(\gamma_k)$ if and only if $k^2 \equiv -1 \pmod p$. So if we set $$E_p^{(2)} \ := \ \{ k \in S_p \, | \, k^2 \equiv -1 \pmod p \},$$ then the points $\frac{k}{p} + \frac{1}{p}i$ with $k \in E_p^{(2)}$ are all the inequivalent elliptic points of order $2$ for $\Gamma_0(p)$. We further observe that for such a $k$, the matrix $\gamma_k$ maps $\mathrm{arc}(\gamma_k)$ onto itself with the direction reversed. Hence, if $\tau \in \partial R_{\Gamma_0(p)} \cap \mathrm{arc}(\gamma_k)$ for some $k \in E_p^{(2)}$, then we need to choose $\tau$ such that $\mathrm{Re}(\tau) \leq \frac{k}{p}$ in order to construct $\mathfrak F_{\Gamma_0(p)}$ satisfying the prescribed property (3). On the other hand, if $\tau \in \partial R_{\Gamma_0(p)} \cap (\mathrm{arc}(\gamma_{-k^{-1}}) \cup \mathrm{arc}(\gamma_k))$ for some $k \in S_p - E_p^{(2)}$, then we need to choose $\tau$ such that $\tau \in \mathrm{arc}(\gamma_{k_{(2)}})$, where $$k_{(2)} \ := \ \min \{ k, \, -k^{-1} \}.$$ For example, if $k = 1$, then $k_{(2)} = -1$ and so we discard the points $\tau$ with $|\tau - 1/p| = 1/p$.
Now it remains to determine which points $\frac{k}{p} - \frac{1}{2p} + \frac{\sqrt{3}}{2p}i$ with $k \in S_p - \{ 1 \}$ are equivalent to each other under $\Gamma_0(p)$. For any $k \in S_p$ we see by a straightforward computation that $$\gamma_{k}\Big( \frac{1 - k^{-1}}{p} - \frac{1}{2p} + \frac{\sqrt{3}}{2p}i \Big) \ = \ \frac{k}{p} - \frac{1}{2p} + \frac{\sqrt{3}}{2p}i.$$ So the point $\frac{k}{p} - \frac{1}{2p} + \frac{\sqrt{3}}{2p}i$ is $\Gamma_0(p)$-equivalent to the point $\frac{1 - k^{-1}}{p} - \frac{1}{2p} + \frac{\sqrt{3}}{2p}i$ for any $k \in S_p$. Notice that the map $f: S_p - \{ 1 \} \longrightarrow S_p - \{ 1 \}$ taking $k$ to $\langle 1 - k^{-1} \rangle$ satisfies $f^3 = \mathrm{id}$, from which we infer that for any $k \in S_p - \{ 1 \}$ we have either $f(k) = k$ or ($k \neq f(k)$ and $f(k) \neq f^2(k)$ and $k \neq f^2(k)$). We see that $$\begin{aligned}
f(k) = k & \Longleftrightarrow & k + k^{-1} \equiv 1 \pmod p \\
& \Longleftrightarrow & k + k^{-1} = 1 \mbox{ or } 1 - p \\
& \Longleftrightarrow & k + k^{-1} = 1\end{aligned}$$ because $|k + k^{-1}| \leq p - 2$ by the assumption $p \geq 5$. Thus if $k^2 - k + 1 \equiv 0 \pmod p$, then $\gamma_k$ fixes the point $\frac{k}{p} - \frac{1}{2p} + \frac{\sqrt{3}}{2p}i$. On the other hand, if $k^2 - k + 1 \not\equiv 0 \pmod p$, then we see that the three distinct points $\frac{k}{p} - \frac{1}{2p} + \frac{\sqrt{3}}{2p}i$, $\frac{\langle 1 - k^{-1} \rangle}{p} - \frac{1}{2p} + \frac{\sqrt{3}}{2p}i$, $\frac{\langle 1 - (1 - k^{-1})^{-1} \rangle}{p} - \frac{1}{2p} + \frac{\sqrt{3}}{2p}i$ are equivalent to each other under $\Gamma_0(p)$. Therefore, if we set $$E_p^{(3)} \ := \ \{ k \in S_p \, | \, k^2 - k + 1 \equiv 0 \ (\bmod \ p) \},$$ then the points $\frac{k}{p} - \frac{1}{2p} + \frac{\sqrt{3}}{2p}i$ with $k \in E_p^{(3)}$ are all the inequivalent elliptic points of order $3$ for $\Gamma_0(p)$. Moreover, if $\tau$ is one of the three points above for some $k \in S_p - (\{ 1 \} \cup E_p^{(3)})$, then $\tau$ must be chosen as $\tau = \frac{k_{(3)}}{p} - \frac{1}{2p} + \frac{\sqrt{3}}{2p}i$, where $$k_{(3)} \ := \ \min \{ k, \, \langle 1 - k^{-1} \rangle, \, \langle 1 - (1 - k^{-1})^{-1} \rangle \}.$$ For example, if $k = -1$, then $k_{(3)} = -\frac{p-1}{2}$ and so the point $\tau = -\frac{1}{2} + \frac{\sqrt{3}}{2p}i$ must be chosen and the points $\tau = \pm\frac{3}{2p} + \frac{\sqrt{3}}{2p}i$ must be discarded.
Combining all of the above discussion, we conclude that $\mathfrak F_{\Gamma_0(p)}$ is the set of all points $\tau \in \mathbb H$ satisfying the following properties:
1. $|\mathrm{Re}(\tau)| \leq 1/2$
2. $|\tau - k/p| \geq 1/p$ for all $k \in S_p$
3. $|\mathrm{Re}(\tau)| = 1/2 \Longrightarrow \mathrm{Re}(\tau) = -1/2$
4. $|\tau - 1/p| \neq 1/p$
5. $|\tau - k/p| = 1/p$ for some $k \in E_p^{(2)}$ $\Longrightarrow$ $\mathrm{Re}(\tau) \leq k/p$
6. $|\tau - k/p| = 1/p$ for some $k \in S_p - (\{ \pm 1 \} \cup E_p^{(2)})$ $\Longrightarrow$ $\mathrm{Re}(\tau) \leq (2k_{(2)} + 1)/ 2p$
7. $\tau \neq (2k - 1) / 2p + i\sqrt{3}/2p$ for all $k \in S_p - (\{ 1 \} \cup E_p^{(3)})$ with $k \neq k_{(3)}$.
[10]{}
B. Cho, *Primes of the form $x^2 + ny^2$ with conditions $x \equiv 1 \bmod N$, $y \equiv 0 \bmod N$*, J. Number Theory, 130 (2010), 852–861.
B. Cho, *Integers of the form $x^2 + ny^2$*, Monatsh. Math., 174 (2014), 195–204.
B. Cho, *Representations of integers by the binary quadratic form $x^2 + xy + ny^2$*, J. Aust. Math. Soc., 100 (2016), 182–191.
H. Cohen, *A Course in Computational Algebraic Number Theory*, Graduate Texts in Mathematics (Volume 138), Springer, 2000.
D. A. Cox, *Primes of the form $x^2 + ny^2$*, second edition, Wiley, 2013.
C. R. Ferenbaugh, *The genus-zero problem for $n|h$-type groups*, Duke Mathematical Journal, 72 (1993), 31–63.
D. A. Marcus, *Number fields*, Universitext, Springer, 1995.
[^1]: The author was supported by NRF-2015R1C1A1A02037526 and the Dongguk University Research Fund of 2017.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We present the first fully relativistic longterm numerical evolutions of three equal-mass black holes in a system consisting of a third black hole in a close orbit about a black-hole binary. We find that these close-three-black-hole systems have very different merger dynamics from black-hole binaries. In particular, we see complex trajectories, a redistribution of energy that can impart substantial kicks to one of the holes, distinctive waveforms, and suppression of the emitted gravitational radiation. We evolve two such configurations and find very different behaviors. In one configuration the binary is quickly disrupted and the individual holes follow complicated trajectories and merge with the third hole in rapid succession, while in the other, the binary completes a half-orbit before the initial merger of one of the members with the third black hole, and the resulting two-black-hole system forms a highly elliptical, well separated binary that shows no significant inspiral for (at least) the first $t \sim 1000M$ of evolution.'
author:
- Manuela Campanelli
- 'Carlos O. Lousto'
- Yosef Zlochower
bibliography:
- '../../../Lazarus/bibtex/references.bib'
title: Close encounters of three black holes
---
[*Introduction:*]{} The recent dramatic breakthroughs in the numerical techniques to evolve black-hole-binary spacetimes [@Pretorius:2005gq; @Campanelli:2005dd; @Baker:2005vv] has led to rapid advancements in our understanding of black-hole physics. Notable among these advancements are developments in mathematical relativity, including systems of PDEs and gauge choices [@Lindblom:2005qh; @Gundlach:2006tw; @vanMeter:2006vi], the exploration of the cosmic censorship [@Campanelli:2006uy; @Campanelli:2006fg; @Campanelli:2006fy; @Rezzolla:2007rd; @Sperhake:2007gu], and the application of isolated horizon formulae [@Ashtekar:2000hw; @Dreyer:2002mx; @Schnetter:2006yt; @Campanelli:2006fg; @Campanelli:2006fy; @Krishnan:2007pu]. These breakthroughs have also influenced the development of data analysis techniques through the matching of post-Newtonian to fully-numerical waveforms [@Pan:2007nw; @Boyle:2007ft; @Hannam:2007ik]. Similarly, the recent discovery of very large merger recoil kicks [@Campanelli:2007ew; @Gonzalez:2007hi; @Campanelli:2007cga; @Herrmann:2007ac; @Koppitz:2007ev; @Herrmann:2006ks; @Baker:2006vn; @Gonzalez:2006md] has had a great impact in the astrophysical community, with several groups now seeking for observational traces of such high speed holes as the byproduct of galaxy collisions [@Bonning:2007vt; @HolleyBockelmann:2007eh]. In this letter, we continue our quest to discover new astrophysical consequences of black-hole interactions by simulating close encounters of three black holes to see the different behaviors introduced by the finite size of the holes, their nonlinear interactions, and the radiation of gravitational waves, as described by General Relativity. We find that the three-body relativistic problem shows far richer dynamics than the two-body problem, akin to the rich three-body dynamics in Newtonian gravity, but with added complexity due to mergers.
Three-body and four-body interactions are expected to be common in globular clusters [@Gultekin:2003xd; @Miller:2002pg], and in galactic cores hosting supermassive black holes (when stellar-mass-black-hole-binary systems interact with the Supermassive black hole). Hierarchical triplets of massive black holes might also be formed in galactic nuclei undergoing sequential mergers [@Makino:1990zy; @valtonen96]. The gravitational wave emission from such systems was recently estimated using post-Newtonian techniques [@Gultekin:2005fd].
[*Techniques:*]{} We evolve the three-black-hole data-sets using the [LazEv]{} [@Zlochower:2005bj] implementation of the ‘moving puncture approach’ [@Campanelli:2005dd; @Baker:2005vv]. We use the Carpet [@Schnetter-etal-03b] driver to provide a ‘moving boxes’ style mesh refinement. In this approach refined grids of fixed size are arranged about the coordinate centers of each hole. We use [AHFinderDirect]{} [@Thornburg2003:AH-finding] to locate apparent horizons. We extract the waveform on spheres centered about the origin and extrapolate the radiated energy/momentum to $r=\infty$ (the waveforms do no change qualitatively for $r>50\pm20M$).
[*Results:*]{} We chose one configuration (3BH1) with purely [*ad-hoc*]{} momentum parameters, which merged relatively quickly, to test the convergence and accuracy of our code. The initial data parameters for these configurations are summarized in Table \[table:ID\]. We evolved these configuration using 11 levels of refinement and a finest resolution of $h=M/80$. The outer boundaries were located at $640M$. In addition we evolved the 3BH1 configuration with grid-spacings rescaled by $5/6$ and $(5/6)^2$ to test convergence. In the table, the horizon mass is the Christodoulou mass, where ${m^H} = \sqrt{m_{\rm irr}^2 +
S^2/(4 m_{\rm irr}^2)},$ $S$ is the magnitude of the spin of the hole, and $m_{\rm irr}$ is the irreducible mass.
Config 3BH1 3BH101 3BH102
----------- ------------ ------------ ------------
$x_1/M$ -2.40856 -3.52462 -3.52238
$y_1/M$ 2.23413 2.58509 2.58509
$p^x_1/M$ -0.0460284 -0.0782693 0.0782693
$p^y_1/M$ -0.0126181 -0.0400799 -0.0433529
$m^p_1/M$ 0.315269 0.318143 0.317578
$m^H_1/M$ 0.335555 0.336201 0.335721
$x_2/M$ -2.40856 -3.52462 -3.52462
$y_2/M$ -2.10534 -2.58509 -2.58509
$p^x_2/M$ 0.130726 0.0782693 -0.0782693
$p^y_2/M$ -0.0126181 -0.0400799 -0.0433529
$m^p_2/M$ 0.315269 0.318143 0.317578
$m^H_2/M$ 0.3405205 0.336241 0.335767
$x_3/M$ 4.8735 7.04923 7.04476
$y_3/M$ 0.0643941 0 0
$p^x_3/M$ -0.0846974 0 0
$p^y_3/M$ 0.0252361 0.0801597 0.0867057
$m^p_3/M$ 0.315269 0.320815 0.318585
$m^H_3/M$ 0.332198 0.333115 0.331270
: Initial data parameters. $(x_i,y_i,0)$ and $(p^x_i, p^y_i,0)$ are the initial position and momentum of the puncture $i$, $m^p_i$ is the puncture mass parameter, and $m^H_i$ is the horizon mass.
\[table:ID\]
It is interesting to note that the same techniques used for black-hole binary evolutions work for configurations of three (and, according to a brief test by the authors, at least 22) black holes. We tested the convergence of our algorithm with three black holes by evolving configuration 3BH1 with three resolutions ($M/80, M/96, M/115.2$). We chose this configuration since it merges relatively quickly, thus reducing the computational expense. The resulting waveform converges to fourth-order. Note that late-time accumulation of errors can have a significant effect on the trajectories of 3-black-hole systems due to their inherent sensitivity to changes in configuration. We have confirmed, through the use of a 8th-order accurate code, that the trajectories presented here show the correct qualitative behavior.
We now show results for two similar initial configurations with qualitatively different outcomes. To aid in the discussion, we will denote the two holes in the binary with BH1 and BH2, and the third hole with BH3, where BH1 is initially located at $y>0$. We determined our initial data parameters by choosing a fiducial binary configuration with orbital frequency $M_B\Omega_B = 0.04$ and angular momentum $J_B/M_B^2 = 0.9104975$. We then treat the binary as a point particle of mass $M_B = 2/3 M$ and spin $a_B/M_B = 0.9105$ along with a non-spinning point particle of mass $M/3$, and choose position and momenta parameters such that this two particle system is in a quasi-circular orbit (up to 3 PN) at a separation equal to twice the binary’s separation. We set up the systems so that the binary’s orbital angular momentum is aligned with the total orbital angular momentum (configuration 3BH101), and anti-aligned (configuration 3BH102). Configuration 3BH101 has BH2 and BH3 merging after the binary completes nearly a half of an orbit. The result of this interaction is to significantly push BH1 away from the merger remnant, producing a new, highly elliptical, binary, with large orbital separation oscillating in time from $10.4M$ to $23.5M$ (see Fig. \[fig:cf101\_track\]).
![The horizon trajectories for configuration 3BH101. The three black holes are initially located at the points labeled by BH1, BH2, and BH3, respectively. BH1 and BH2 form a quasi-circular binary, which is disrupted by BH3. BH2 and BH3 merge at point M1. The BH1 and the BH2–BH3 merger remnant continue to orbit each other throughout the simulation. []{data-label="fig:cf101_track"}](cf101_track.eps){width="2.375in" height="1.9475in"}
The 3BH101 waveform (Fig. \[fig:cf101\_wave\]) shows a burst of radiation from the BH2–BH3 merger, as well as a small pulse at $t\sim
700M$ which corresponds to the point of closest approach of the BH2–BH3 merger remnant with BH1. We stopped the evolution at $t\sim1000M$ due to computational expense and boundary contamination.
![The $(l=2,m=2)$ mode of $\psi_4$ for 3BH101. The BH2–BH3 merger waveform is centered at $t\sim185M$, the small pulse at $t\sim700$ was produced by the close approach of the BH2–BH3 merger product to BH1.[]{data-label="fig:cf101_wave"}](cf101_wave.eps){width="3.0in" height="1.8in"}
Configuration 3BH102 displays very different behavior, as seen in Fig. \[fig:cf102\_track\]. Here the binary is disrupted almost immediately, and the individual holes follow complicated trajectories (note that the trajectories are similar to the Greek letters $\gamma$, $\tau$, and $\sigma$). BH3 and BH1 merge when BH3 almost completes 1.25 orbits. The BH3–BH1 merger product then quickly merges with BH2. The resulting waveform shows a double merger as seen in Fig. \[fig:cf102\_wave\].
![The horizon trajectories for configuration 3BH102 with ticks every $45M$ of evolution. The three black holes are initially located at the points labeled BH1, BH2, and BH3, respectively. BH1 and BH2 form a quasi-circular binary, which is almost immediately disrupted by BH3. BH1 and BH3 merge at point M1, and then merge with BH2 at M2. []{data-label="fig:cf102_track"}](cf102_track.eps){width="2.375in" height="2.3275in"}
![The $(\ell=2,m=2)$ mode of $\psi_4$ for 3BH102 shows two merger waveform signals arising from merger M1 and M2. []{data-label="fig:cf102_wave"}](cf102_wave.eps){width="3.0in" height="1.55in"}
It is important to note that the dramatic difference in dynamics between 3BH101 and 3BH102 is not strictly due to corotation versus anti-corotation of the binary, but rather is a result the precise configuration of the binary when the third black hole approaches.
Config 3BH1 3BH101$^*$ 3BH102
------------------------------------- --------------- ------------- -------------
$E_{\rm rad}/M (\times 1000)$ $6.06\pm0.02$ $4.4\pm0.8$ $6.1\pm0.7$
$J^z_{\rm rad}/M^2 (\times 100)$ $2.92\pm0.01$ $3.6\pm0.5$ $4.4\pm0.2$
$P^x_{\rm rad}/M({\rm km\,s^{-1}})$ $20.0\pm1.9$ $50\pm40$ $-1.3\pm14$
$P^y_{\rm rad}/M({\rm km\,s^{-1}})$ $-22.9\pm2.4$ $27\pm13$ $-15\pm13$
$M_{\rm H}/M$ 0.9835 \*\*\* 0.9885
$S^z_{\rm H}/M^2$ 0.532 \*\*\* 0.465
$t_{\rm M1}/M$ $\sim 27$ $\sim 115$ $\sim 218$
$t_{\rm M2}/M$ $\sim 40$ \*\*\* $\sim280$
: Total radiated energy, momentum, and angular momentum, as well as horizon mass spin and merger time, for 3BH102 and 3BH1. ($*$ 3BH101 did not merge in the time alloted and we only report values for the first merger).
\[table:Results\]
For all three configurations the radiated energy and angular momenta were a fraction of that for a quasi-circular equal-mass binary. This is due to the grazing type mergers, as also seen by the suppression of the radiated angular momentum. We expect that the full waveform for 3BH101 will show significantly more radiation as the system eventually ‘circularizes’ and merges.
We calculate the mass and spin of the first merger remnant and the final remnant using the fitted exponential decay rate and frequency of the quasi-normal behavior [@Echeverria89]. For 3BH102, we fit the real and imaginary parts of the $(\ell=2,m=2)$ mode separately and find $a_{H}/M_{H}=0.479, 0.378$ and $M_{H}=0.716, 0.678$ from fits of the real and imaginary parts of $\psi_4$ for the first merger product, and $a_{H}/M_{H}=0.478, 0.580$ and $M_{H}=0.994, 1.04$ from fits of the real and imaginary parts of $\psi_4$ for the final merger remnant. Note that the imaginary part of $\psi_4$ provides a better estimate (i.e. closer to the expected $0.66$) mass for the first remnant, but the real part provides a better estimate of both $a$ and $M_{H}$ for the final remnant. We note the quasi-normal frequencies were $\omega = 0.64, 0.46$ for the first and final merger remnants respectively. For 3BH101 we were only able to fit the real part of $\psi_4$ due to significant contamination from other modes in the imaginary part. We find a narrow region of width $10M$ where the waveform shows nearly exponential decay. A fit to this region yields $\omega= 0.63$, with a corresponding mass and spin of $M_{H} = 0.665$ and $a/M_{H} = 0.293$. The remnant masses, spins, and merger times are given in Table \[table:Results\] for the 3BH1 and 3BH102 configurations. The numbers quoted above should be taken only as indicative of the expected values of these parameters. Further runs, at higher resolutions, will be needed in order to establish the errors in these values.
[*Discussion:*]{} The relativistic study of quasi-circular orbits of a binary in the presence of a third comparable-mass hole, as an initial value problem, was studied in [@Campanelli:2005kr]. Here, by studying their dynamical evolution we find that the third hole perturbs the system to the extent that no true binary orbit is seen. We found that the close encounter of this third body can both trigger a quick merger of the three-body system, as well as impart a significant kick to one of the holes, producing a new long-lived, highly-elliptical binary. The generic effect of the third black hole is to reduce the gravitational radiation. This happens for two reasons. First, close-three-body interactions lead to grazing collisions, which emit far less radiation than quasi-circular mergers. Second, the resulting binary orbit will be elliptical, which is less efficient at emitting gravitational radiation than circular orbits at the final stages. Note however, that although we report radiated energies that are $1/{5^{\rm th}}$ that for a typical binary, here we scale the energy by the total mass. If we scale the radiated energy with the initial binary’s mass, then the rescaled radiated energy would be 3/2 times larger. The close-three-body systems also appear to be shorter lived than typical binaries.
The three-black-hole waveforms (See Figs. \[fig:cf101\_wave\] and \[fig:cf102\_wave\]) are distinct from the robust and simple form of the binary-black-hole waveform [@Baker:2001nu; @Baker:2002qf; @Campanelli:2006gf; @Baker:2006yw]. In addition, there seems to be a large exchange of energy among the components of the triple system, which occurs on a much shorter timescale than the radiation. It is important to note that these three-black-hole interactions provide a mechanism for producing highly-elliptical close-binaries, which would otherwise have circularized (due to emission of gravitational radiation during the inspiral).
Investigations of the Newtonian encounters of three bodies show that such encounters generically lead to the breakup of the system into a binary and the third body that escapes [@Monaghan76] in a ‘water-shed effect’. The distribution of the eccentricity of the remaining binary is bell shaped around $e=0.3$ for compact systems [@Duquennoy91; @Kroupa95]. Classical studies [@Hills75] show that the probability of exchange of the binary companion in a triple system is surprisingly high for all comparable masses, reaching near one for more massive $m_3$ [@SP:93]. The motion of the system can be chaotic, due to small denominators. The finite size of the black holes represents a natural regularization to the problem, and the dissipative effects of the gravitational radiation can prevent some configurations from becoming chaotic.
We have found that 3-black-hole systems exhibit complicated orbital dynamics analogous to the rich 3-body Newtonian dynamics, but with the added complexity introduced by mergers. Further study of this problem, including a comparison of Newtonian and relativistic dynamics, will be reported in a forthcoming paper, where we also examine configurations where the triple system is disrupted.
We thank Alessia Gualandris, David Merritt, and Hiroyki Nakano for valuable discussions. We gratefully acknowledge NSF for financial support from grant PHY-0722315, PHY-0701566, PHY 0714388, and PHY 0722703; and NASA for financial support from grant NASA 07-ATFP07-0158. We also thank Hans-Peter Bischof for producing 3-D visualizations of the three-black-hole configurations introduced in this paper. Computational resources were provided by the NewHorizons cluster at RIT and the Lonestar cluster at TACC.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: |
The Maximum Likelihood Estimator (MLE) is widely used in estimating information measures, and involves “plugging-in” the empirical distribution of the data to estimate a given functional of the unknown distribution. In this work we propose a general framework for analyzing the performance of the MLE in estimating functionals of discrete distributions, under the worst-case mean squared error criterion. We show that existing theory, which was developed to accommodate a fixed alphabet and a growing number of observations, is insufficient for analyzing the bias of the MLE, and apply the theory of approximation using positive linear operators to study this bias. The variance is controlled using the well-known tools from the literature on concentration inequalities.
Our techniques yield a characterization of the worst-case $L_2$ risk incurred by the MLE in estimating the Shannon entropy $H(P) = \sum_{i = 1}^S -p_i \ln p_i$, and $F_\alpha(P) = \sum_{i = 1}^S p_i^\alpha,\alpha>0$ up to a multiplicative constant, for any alphabet size $S\leq \infty$ and sample size $n$. As a corollary, for Shannon entropy estimation, we show that it is necessary and sufficient to have $n \gg S$ observations for the MLE to be consistent. In addition, we establish that it is necessary and sufficient to consider $n \gg S^{1/\alpha}$ samples for the MLE to consistently estimate $F_\alpha(P), 0<\alpha<1$. The minimax rate-optimal estimators for both problems require $S/\ln S$ and $S^{1/\alpha}/\ln S$ samples, which implies that the MLE has a strictly sub-optimal sample complexity. When $1<\alpha<3/2$, we show that the worst-case $L_2$ rate of convergence for the MLE is $n^{-2(\alpha-1)}$ for infinite alphabet size, while the minimax $L_2$ rate is $(n\ln n)^{-2(\alpha-1)}$. When $\alpha\geq 3/2$, the MLE achieves the minimax optimal $L_2$ convergence rate $n^{-1}$ regardless of the alphabet size.
We explicitly establish an equivalence between bias analysis of plug-in estimators for general functionals under arbitrary statistical models, and the theory of approximation using positive linear operators, a subarea of approximation theory with more than a century’s history. This equivalence between bias analysis and approximation theory is of relevance and consequence far beyond the specific problem setting in this paper.
author:
- 'Jiantao Jiao, , Kartik Venkat, , Yanjun Han, , and Tsachy Weissman, [^1]. [^2]'
bibliography:
- 'di.bib'
title: Maximum Likelihood Estimation of Functionals of Discrete Distributions
---
entropy estimation, maximum likelihood estimator, approximation theory, high dimensional statistics, Rényi entropy
Introduction
============
Entropy and related information measures arise in information theory, statistics, machine learning, biology, neuroscience, image processing, linguistics, secrecy, ecology, physics, and finance, among other fields. Numerous inferential tasks rely on data driven procedures to estimate these quantities (see, e.g. [@Olsen--Meyer--Bontempi2009impact; @Pluim--Maintz--Viergever2003mutual; @Viola--Wells1997alignment; @Batina--Gierlichs--Prouff--Rivain--Standaert--Veyrat2011mutual; @Hill1973diversity; @Franchini--Its--Korepin2008Renyi]). Consider the problem of estimating the Shannon entropy of an unknown discrete distribution $P$ based on $n$ i.i.d. samples. This problem has a rich history, of which we refer the reader to [@Jiao--Venkat--Han--Weissman2014minimax] for a review. One of the most widely used estimators for this purpose is the Maximum Likelihood Estimator (MLE), which is simply the empirical entropy. The empirical entropy is an instantiation of the plug-in principle in functional estimation, where a point estimate of the parameter (distribution $P$ in this case) is used to construct an estimator for a [*functional*]{} of the parameter. The idea of using the MLE for estimating information measures of interest (in this case entropy), is not only intuitive, but has sound justification: [*asymptotic efficiency*]{}.
The beautiful theory of Hájek and Le Cam [@Hajek1970characterization; @Hajek1972local; @LeCam1986asymptotic] shows that, as the number of observed samples grows without bound while the parameter dimension (e.g., alphabet size) remains fixed, the MLE performs optimally in estimating any differentiable functional complying with the benign LAN (Local Asymptotic Normality) condition [@LeCam1986asymptotic]. Thus, for finite dimensional problems, the problems of parameter and functional estimation are well understood in an asymptotic sense, and the MLE appears to be not only natural but also theoretically justified. But does it make sense to employ the MLE to estimate the entropy in most practical applications?
As it turns out, while asymptotically optimal in entropy estimation, the MLE is by no means sacrosanct in many real applications, especially in regimes where the alphabet size is comparable to, or even larger than the number of observations. It was shown that the MLE for entropy is strictly sub-optimal in the large alphabet regime [@Paninski2003; @Paninski2004]. Therefore, classical asymptotic theory does not satisfactorily address high dimensional settings, which are becoming increasingly important in the modern era of high dimensional statistics.
The main contribution of this paper is to provide novel tools for analysis of plug-in estimators such as the MLE, complementing the companion paper [@Jiao--Venkat--Han--Weissman2014minimax] which presents new estimators shown to be [*minimax rate-optimal*]{}. The papers [@Jiao--Venkat--Han--Weissman2014minimax; @Jiao--Venkat--Han--Weissman2014beyond] also demonstrate that employing the MLE in functional estimation can result in highly sub-optimal estimators in practical inferential applications.
Problem formulation
-------------------
To illustrate our methodology, we shall focus on the following problem. Suppose we observe $n$ independent samples from an unknown discrete probability distribution $P = (p_1, p_2, \ldots, p_S)$, with *unknown* support size $S$, and would like to estimate a functional of the distribution of the form: $$\label{eqn.generalf}
F(P) = \sum_{i=1}^{S} f(p_i),$$ where $f: [0,1] \to \mathbb{R}$ is a continuous function, and $f(0) = 0$. We shall focus on two concrete and well-motivated examples of functionals of this form. In particular, the Shannon entropy, $$\label{eqn.entropy}
H(P) \triangleq \sum_{i = 1}^S - p_i \ln p_i,$$ plays significant roles in information theory and serves as fundamental limits to various operational problems [@Shannon1948]. Another interesting class of functionals is $F_\alpha(P), \alpha > 0$: $$F_\alpha (P) \triangleq \sum_{i = 1}^S p_i^\alpha, \alpha>0.$$ The $\alpha$-moments of a distribution often emerge in various operational problems. In particular, the significance of $F_\alpha(P)$ can also be seen via the connection $H_\alpha(P) = \frac{\ln F_\alpha(P)}{1-\alpha}$, where $H_\alpha(P)$ is the Rényi entropy [@Renyi1961measures]. Like Shannon entropy, Rényi entropy is an important information measure emerging in an increasing variety of disciplines such as ecology, quantum information, information theory, and statistics, to name a few. Recently [@Acharya--Orlitsky--Suresh--Tyagi2014complexity] has studied the complexity of estimating the Rényi entropy.
In this work, we analyze, for the above two classes of functionals, the performance of MLE, widely considered the most natural plug-in rule.
The Maximum Likelihood Estimator (MLE) for the functional $F(P)$ in (\[eqn.generalf\]) is defined as: $$F(P_n) \triangleq \sum_{i = 1}^S f(P_n(i)),$$ where $P_n$ denotes the empirical distribution induced by $n$ observations.
We use the conventional decision theoretic framework, and present matching upper and lower bounds (up to multiplicative constants depending neither on $n$ nor $S$) on the maximum $L_2$ risk of the MLE, i.e. $$\sup_{P \in \mathcal{M}_S} \mathbb{E}_{P} (F(P) - F(P_n))^2,$$ where $\mathcal{M}_S$ denotes the set of discrete distributions of support size $S$. The subscript $P$ means that the expectation is taken with respect to the distribution $P$ that generates the i.i.d. observations. We omit the subscript for the expectation operator $\mathbb{E}$ if the meaning of the expectation is clear from the context. Understanding the benefits and limitations of the MLE serves two key purposes. First, the approach is a natural benchmark for comparing other more nuanced procedures for estimation of functionals. Second, performance analysis for the MLE reveals regimes where the problem is difficult, and motivates the development of improvements, such as those recently derived in [@Jiao--Venkat--Han--Weissman2014minimax; @Wu--Yang2014minimax; @Jiao--Venkat--Han--Weissman2014beyond].
Bias of the MLE: an approximation theoretic perspective
-------------------------------------------------------
In the entropy estimation literature, considerable effort has been devoted to understanding the non-asymptotic performance of the MLE $H(P_n)$ in estimating $H(P)$. One of the earliest investigations in this direction is due to Miller [@Miller1955] in 1955, who showed that, for any fixed distribution $P$, $$\label{eqn.miller}
\bE H(P_n) = H(P) - \frac{S-1}{2n} + O\left(\frac{1}{n^2}\right).$$ Equation (\[eqn.miller\]) was later refined by Harris [@Harris1975] using higher order Taylor series expansions to yield $$\label{eqn.harris}
\bE H(P_n) = H(P) - \frac{S-1}{2n} + \frac{1}{12 n^2} \left( 1- \sum_{i = 1}^S \frac{1}{p_i} \right) + O\left( \frac{1}{n^3} \right).$$
Harris’s result reveals an undesirable consequence of the Taylor expansion method: one cannot obtain uniform bounds on the bias of the MLE. Indeed, the term $\sum_{i = 1}^S \frac{1}{p_i}$ can be arbitrarily large for some distribution $P$. However, it is evident that both $H(P_n)$ and $H(P)$ are bounded above by $\ln S$, since the maximum entropy of any distribution supported on $S$ elements is $\ln S$. Conceivably, for such a distribution $P$ that would make $\sum_{i = 1}^S \frac{1}{p_i}$ very large, we need to compute even higher order Taylor expansions to obtain more accuracy, but even with such efforts we can never obtain a uniform bias bound for all $P$.
We gain one of our key insights into the bias of the MLE by relating it to the approximation error induced by the [*Bernstein polynomial approximation*]{} of the function $f$. To see this, we first compute the bias of $F(P_n)$.
\[lemma.biasgeneralf\] The bias of the estimator $F(P_n)$ is given by $$\begin{aligned}
\label{eqn.biasgeneralfequation}
\mathsf{Bias}(F(P_n)) &\triangleq \bE F(P_n) - F(P) \nonumber \\
&= \sum_{i = 1}^S \left( \sum_{j = 0}^n f \left( \frac{j}{n} \right) \binom{n}{j} p_i^j(1-p_i)^{n-j} - f(p_i) \right).\end{aligned}$$
The bias term in (\[eqn.biasgeneralfequation\]) can be equivalently expressed as[^3]. $$\mathsf{Bias}(F(P_n)) = \sum_{i = 1}^S \left( \sum_{j = 0}^n f\left( \frac{j}{n} \right) B_{j,n}(p_i) - f(p_i) \right) = \sum_{i = 1}^S \left( B_n[f](p_i) - f(p_i)\right),$$ where $B_{j,n}(x) \triangleq \binom{n}{j} x^j(1-x)^{n-j}$ is the well-known Bernstein polynomial basis, and $B_n[f](x)$ is the so-called Bernstein polynomial for function $f(x)$. Bernstein in 1912 [@Bernstein1958collected] provided an insightful constructive proof of the Weierstrass theorem on approximation of continuous functions using polynomials, by showing that the Bernstein polynomial of any continuous function converges uniformly to that function.
From a functional analysis viewpoint, the Bernstein polynomial is an operator that maps a continuous function $f\in C[0,1]$ to another continuous function $B_n[f] \in C[0,1]$. This operator is linear in $f$, and is *positive* because $B_n[f]$ is also pointwise non-negative if $f$ is pointwise non-negative. Evidently, bounding the approximation error incurred by the Bernstein polynomial is equivalent to bounding the bias of the MLE $f(X/n)$, where $X \sim \mathsf{Binomial}(n,x)$. As mentioned above, Taylor series expansions are not sufficient to satisfactorily analyze this bias. Fortunately, the theory of *approximation using positive linear operators* [@Paltanea2004] provides us with advanced tools that are very effective for the bias analysis our problem calls for. A century ago, probability theory served Bernstein in breaking new ground in function approximation. It is therefore very satisfying that advancements in the latter have come full circle to help us better understand probability theory and statistics. We briefly review the general theory of approximation using positive linear operators below.
Approximation theory using positive linear operators
----------------------------------------------------
Generally speaking, for any estimator $\hat{\theta}$ of a parametric model indexed by $\theta$, the expectation $f \mapsto \bE_\theta f(\hat{\theta})$ is a positive linear operator for $f$, and analyzing the bias $\bE_\theta f(\hat{\theta}) -f(\theta)$ is equivalent to analyzing the approximation properties of the positive linear operator $\bE_\theta f(\hat{\theta})$ in approximating $f(\theta)$. Hence, analyzing the bias of *any* plug-in estimator for functionals of parameters from *any* parametric families can be recast as a problem of approximation theory using positive linear operators [@Paltanea2004].
Why are approximation theorists interested in this theory? It turns out the most constructive proofs of the Weierstrass theorem concerning the approximation of continuous functions on a compact interval by polynomials use some sequences of positive linear operators. So do the classical operators of Gauss–Weierstrass, Landau, Vallée-Poissin, Jackson, and Bernstein. Concretely, an operator $L$ defined on a linear space of functions, $V$, is called *linear* if $$L(\alpha f + \beta g) = \alpha L(f) + \beta L(g), \quad \text{for all }f,g\in V, \alpha,\beta \in \mathbb{R},$$ and is called *positive*, if $$L(f)\geq 0\quad \text{for all }f\in V, f\geq 0.$$
Given that the most general problems of bias analysis about plug-in methods in functional estimation are a subset of the general theory of approximation using positive linear operators, can we show that for a natural large class of positive linear operators, there exist corresponding probabilistic models such that the bias analysis is exactly the same as approximation properties for the positive linear operator?
Surprising as it may sound, the answer is *yes*. Denote by $\mathcal{F}(I)$ the set of all functions with domain $I$, where $I\subset \mathbb{R}$ is an interval. We note that any positive linear operator $L: V \mapsto \mathcal{F}(I)$ can be expressed as a collection of linear positive functionals $(F_x)_{x\in I}$. Indeed, for any fixed point $x\in I$ we obtain a positive linear functional $F_x: V \mapsto \mathbb{R}$, defined by $F_x(f) \triangleq L(f)(x), f\in V$. Conversely, if we have a collection of positive linear functionals $(F_x)_{x\in I}, F_x: V\mapsto \mathbb{R}$, then we can construct a positive linear operator $L: V\mapsto \mathcal{F}(I)$, where $L(f)(x) \triangleq F_x(f), f\in V$.
Hence, the study of positive linear operators in this context can be reduced to that of positive linear functionals. It was shown in Paltanea [@Paltanea2004 Remark 1.1.2.] that if $I$ is an interval of $\mathbb{R}$, then for any positive linear functional $F: C(I)\to \mathbb{R}$, there exists a positive regular Borel measure $\mu$, such that we have the integral representation[^4]: $$\label{eqn.generalfriesz}
F(f) = \int_I f d\mu, \quad f\in C(I),$$ where $C(I)$ denotes the space of continuous functions on $I$. The key ingredient of this proof is the Riesz–Markov–Kakutani theorem, which can be invoked in more general settings in which $I$ is not a subset of $\mathbb{R}$.
Denote by $e_j,j\in \mathbb{N}_+ \cup \{0\}$, the monomial functions $e_j(y) = y^j,y\in I$. If we naturally require $F(e_0) = 1$, which means that the functional maps constant valued functions to the same constant, then the measure $\mu$ in (\[eqn.generalfriesz\]) is a probability measure, implying that $F(f)$ can be written as an expectation $$F(f) = \bE_\mu f(Z),\quad Z \sim \mu, \mathsf{supp}(\mu)\subset I.$$
If we go back to the positive linear operator setting, and with a slight abuse of notation still denote the operator by $F(f)$, then we can parametrize the measure $\mu$ to be $\mu_x, x\in I$, and represent any positive linear operator operating on $C(I)$ as $$F(f)(x) = \int_I f d\mu_x = \bE_{\mu_x} f(Z),\quad Z\sim \mu_x, x\in I, \mathsf{supp}(\mu_x)\subset I.$$
Thus, there is an equivalence between bias analysis of continuous plug-in estimators and approximation theory using positive linear operators. After more than a century’s active research on approximation using positive linear operators, where do we stand? Probably a proper description of the current theory is “highly developed for certain problems, but far from complete”. Indeed, we now have highly non-trivial tools for positive linear operators of functions on one dimensional compact sets, but the general theory for vector valued multivariate functions on non-compact sets is still far from complete [@Paltanea2004]. In the next subsection, we present a sample of existing results in approximation using positive linear operators, corollaries of which will be used to analyze the bias of the MLE for two examples: $F_\alpha(P)$ and $H(P)$. We note, however, that the problem we investigate is somewhat different from the approximation problem, since the parameter $P$ lies in a simplex, which is a strict subset of $\mathbb{R}^S$. Before embarking on discussing the main results of this paper in Section \[sec.mainresults\], we present a short introduction to bias bounding techniques via approximation theory.
Some general results in bias analysis {#sec.generalbiasanalysis}
-------------------------------------
First, some elementary approximation theoretic concepts need to be introduced in order to characterize the degree of *smoothness* of functions. For $I\subset \mathbb{R}$ an interval, the first-order modulus of smoothness $\omega^1(f,t),t\geq 0$ is defined as [@Paltanea2004] $$\omega^1(f,t) \triangleq \sup \{ |f(u) - f(v)|: u,v \in I, |u-v|\leq t \}.$$
The second-order modulus of smoothness $\omega^2(f,t),t\geq 0$ [@Paltanea2004] is defined as $$\omega^2(f,t) \triangleq \sup \left \{ \left| f(u) - 2f\left(\frac{u+v}{2}\right) + f(v) \right|: u, v\in I, |u-v|\leq 2t \right\}.$$
Ditzian and Totik [@Ditzian--Totik1987] introduced a class of modulus of smoothness, which proves to be extremely useful in characterizing the incurred approximation errors. For simplicity, for functions defined on $[0,1]$, $\varphi(x) = \sqrt{x(1-x)}$, the first-order Ditzian–Totik modulus of smoothness is defined as $$\omega^1_\varphi(f,t) \triangleq \sup \left \{ |f(u) - f(v)|: u,v\in [0,1], |u-v| \leq t \varphi\left( \frac{u+v}{2}\right) \right \},$$ and the second-order Ditzian–Totik modulus of smoothness is defined as $$\begin{aligned}
\omega_\varphi^2(f,t) & \triangleq \sup \left\{ \left | f(u) - 2f \left( \frac{u+v}{2}\right) +f(v) \right|: u, v\in [0,1], |u-v| \leq 2t \varphi\left( \frac{u+v}{2} \right) \right\} .\end{aligned}$$
Recall that we denote by $e_j,j\in \mathbb{N}_+ \cup \{0\}$, the monomial functions $e_j(y) = y^j,y\in I$. The first estimate for general positive linear operators, using modulus $\omega^2$ and with precise constants, was given by Gonska [@Gonska1979quantitative]. We rephrase Paltanea [@Paltanea2004 Cor. 2.2.1.] as follows. Note that notation $e_1 - x e_0$ denotes a continuous function on $I$ which is the difference of a linear function $y$ and a constant function with constant value $x$ over $I$. In other words, it is an abbreviation of $e_1(y) - x e_0(y),y\in I$, which is a function of $y$ rather than $x$.
[@Paltanea2004 Cor. 2.2.1.]\[lemma.paltaneacor221\] Let $F: C(I) \to \mathbb{R}$ be a positive linear functional, where $I \subset \mathbb{R}$ is an interval. Suppose that $F(e_0) = 1, t>0$. Then,
1. If $\mathrm{length}(I)\geq 2t, s\geq 2$, then $$|F(f) - f(x)| \leq |F(e_1 - xe_0)|\frac{\omega^1(f,t)}{t} + \left( 1 + \frac{F(|e_1 - xe_0|^s)}{2t^s} \right)\omega^2(f,t).$$
2. If we also have $F(e_1) = x$, $x$ is an interior point of $I$ and $\mathrm{length}(I)\leq 2t$, then $$|F(f) - f(x)| \leq \omega^2(f,t).$$
We remark that Lemma \[lemma.paltaneacor221\] can be applied to bound the bias of plug-in estimators in very general models. For example, consider an arbitrary statistical experiment $\{P_\theta, \theta \in I\}$, and we obtain $n$ i.i.d. samples $X_1,X_2,\ldots,X_n \sim P_\theta$. For any estimator $\hat{\theta}_n$, we would like to analyze the bias of the plug-in estimator $f(\hat{\theta}_n)$ for functional $f(\theta)$.
Suppose $\mathrm{length}(I)\geq 2t, s\geq 2$, then Lemma \[lemma.paltaneacor221\] implies that $$|\bE_\theta f(\hat{\theta}_n) - f(\theta)| \leq |\bE_\theta \hat{\theta}_n - \theta| \frac{\omega^1(f,t)}{t} + \left( 1+ \frac{\bE |\hat{\theta}_n - \theta|^s}{2t^s} \right)\omega^2(f,t).$$
If we further assume that $\hat{\theta}_n$ is an unbiased estimator for $\theta$, i.e., $\bE_\theta \hat{\theta}_n = \theta$ holds for all $\theta \in I$, then we have $$|\bE_\theta f(\hat{\theta}_n) -f(\theta)| \leq \left( 1+ \frac{\bE |\hat{\theta}_n - \theta|^s}{2t^s} \right)\omega^2(f,t).$$
Taking $s = 2$ and assuming $\mathsf{Var}(\hat{\theta}_n) \leq \mathrm{length}(I)/2$, we have $$|\bE_\theta f(\hat{\theta}_n) - f(\theta)| \leq \frac{3}{2} \omega^2(f, \sqrt{\mathsf{Var}(\hat{\theta}_n)}),$$ after we take $t = \sqrt{\bE |\hat{\theta}_n - \theta|^2}$.
We remark that Lemma \[lemma.paltaneacor221\] is only one way to analyze the bias, which is by no means always tight. For example, the following estimate using Ditzian–Totik modulus is significantly better than Lemma \[lemma.paltaneacor221\] for certain functionals such as the entropy.
[@Paltanea2004 Thm. 2.5.1.]\[lemma.paltaneathm251\] If $F: C[0,1] \to \mathbb{R}$ is a positive linear functional such that $F(e_0) = 1$, then $$|F(f) - f(x)| \leq \frac{|F(e_1 - xe_0)|}{2t\varphi(x)} \omega_\varphi^1(f,2t) + \left( 1 + \frac{3 F((e_1 - x e_0)^2)}{2 (t\varphi(x))^2} \right) \omega_\varphi^2(f,t),$$ for all $t\in (0,1/2]$.
Considering the same statistical experiment $\{P_\theta, \theta \in I\}$, and the plug-in estimator $f(\hat{\theta}_n)$ for $f(\theta)$, if $\hat{\theta}_n$ is unbiased for $\theta$ and $\mathsf{Var}(\hat{\theta}_n) \leq \frac{\varphi(\theta)^2}{4}$, then it follows from Lemma \[lemma.paltaneathm251\] that $$|\bE_\theta f(\hat{\theta}_n) - f(\theta)| \leq \frac{5}{2} \omega_\varphi^2 \left( f, \frac{\sqrt{\mathsf{Var}(\hat{\theta}_n)}}{\varphi(\theta)} \right),$$ after we take $t = \frac{\sqrt{\mathsf{Var}(\hat{\theta}_n)}}{\varphi(\theta)}$.
For certain functions $f(\theta)$ and statistical models Lemma \[lemma.paltaneathm251\] is stronger than Lemma \[lemma.paltaneacor221\]. For example, if $f(\theta) = -\theta \ln \theta, \theta \in [0,1]$, and we have $n\cdot\hat{\theta}_n \sim \mathsf{B}(n,\theta)$. We will show in Lemma \[lemma.dtmoduluscomputation\] that $\omega_\varphi^2(f,t) = \frac{t^2\ln 4}{1+t^2}$, and $\omega^2(f,t) = t\ln 4$. We also have $\mathsf{Var}(\hat{\theta}_n) = \frac{\theta(1-\theta)}{n}$. Hence, Lemma \[lemma.paltaneacor221\] gives the upper bound $$\label{eqn.pointwiseboundforentropyf}
|\bE_\theta f(\hat{\theta}_n) - f(\theta)| \leq \frac{3\ln 4}{2} \sqrt{\frac{\theta(1-\theta)}{n}},$$ whereas Lemma \[lemma.paltaneathm251\] gives $$\label{eqn.normboundentropyf}
|\bE_\theta f(\hat{\theta}_n) - f(\theta)| \leq \frac{5\ln 4}{2n} \cdot \frac{1}{1+1/n},$$ which is much stronger when $n$ is large and $\theta$ not too close to the endpoints of $[0,1]$.
There also exist various estimates for the bias when the parameter lies in sets other than an interval in $\mathbb{R}$. However, the bounds we presented are in general *not* optimal for specific functionals, thereby leaving ample room for future development. For example, the recent work [@Han--Jiao--Weissman2015Bayes] showed an improved result that gives strictly better estimates than Lemma \[lemma.paltaneathm251\] for all functionals satisfying the conditions. Note that (\[eqn.pointwiseboundforentropyf\]) is stronger than (\[eqn.normboundentropyf\]) when $\theta \leq 1/n$, but Han, Jiao, and Weissman [@Han--Jiao--Weissman2015adaptive] showed that even the pointwise bound in (\[eqn.pointwiseboundforentropyf\]) is strictly sub-optimal for certain functionals. Han, Jiao, and Weissman [@Han--Jiao--Weissman2015jensen] summarize existing techniques in bias analysis of plug-in estimators, with a general methodology for obtaining tight bounds.
Despite the extensive literature, obtaining the tightest bounds for specific functionals of interest remains a challenge. In the present paper we focus on the functionals $F_\alpha(P)$ and $H(P)$.
Main results {#sec.mainresults}
============
[*Notation:*]{} $a \wedge b$ denotes $\min\{a,b\}$, and $\mathcal{M}_S$ denotes the collection of discrete distributions with support size $S, S\leq \infty$. For two non-negative series $\{a_n\},\{b_n\}$, notation $a_n \lesssim b_n$ means that there exists a positive universal constant $C<\infty$ such that $\frac{a_n}{b_n}\leq C$, for all $n$. Notation $a_n \asymp b_n$ is equivalent to $a_n \lesssim b_n$ and $b_n \lesssim a_n$. Notation $a_n \gg b_n$ means that $\liminf_{n\to \infty} \frac{a_n}{b_n} =\infty$.
Below, we present non-asymptotic bounds on the maximum $L_2$ risk incurred by the MLE for estimating $H(P)$ and $F_\alpha(P)$.
\[thm.main\] For the functional $F_\alpha(P)$,
1. $\alpha\geq 2$: $$\begin{aligned}
\sup_{P \in \cM_S} \bE_P \left( F_\alpha(P_n) - F_\alpha(P) \right)^2 \leq \left( \frac{\alpha(\alpha-1)}{2n}\right)^2 + \frac{\alpha^2}{4n}.\end{aligned}$$
2. $1<\alpha<2$: $$\begin{aligned}
\sup_{P \in \cM_S} \bE_P \left( F_\alpha(P_n) - F_\alpha(P) \right)^2 \leq \left ( \frac{4}{n^{\alpha-1}} \wedge \frac{3S^{1-\alpha/2}}{n^{\alpha/2}} \wedge C_{\alpha,n} \frac{5S}{2n} \right)^2 + \frac{\alpha^2}{4n},\end{aligned}$$ where $C_{\alpha,n} \triangleq n \omega_\varphi^2(x^\alpha,n^{-1/2})>0$ satisfies $\limsup_{n\to \infty}C_{\alpha,n}<\infty$ for $1<\alpha<2$, and $\omega_\varphi^2$ is the second-order Ditzian–Totik modulus of smoothness.
3. $1/2 \leq \alpha<1$: $$\begin{aligned}
\sup_{P \in \cM_S} \bE_P \left( F_\alpha(P_n) - F_\alpha(P) \right)^2 \leq \left(\frac{3S^{1-\alpha/2}}{2n^{\alpha/2}} \wedge \frac{5S}{2n^{\alpha}} \right)^2 + \left( \frac{10S^{2-2\alpha}}{n}+ \frac{120}{\alpha^2}\left(\frac{S}{n^{2\alpha}}\wedge \frac{1}{n^{2\alpha-1}}\right) \right).\end{aligned}$$
4. $0<\alpha<1/2$: $$\begin{aligned}
\sup_{P \in \cM_S} \bE_P \left( F_\alpha(P_n) - F_\alpha(P) \right)^2 \leq \left(\frac{3S^{1-\alpha/2}}{2n^{\alpha/2}} \wedge \frac{5S}{2n^{\alpha}} \right)^2 + \left( \frac{10S}{n^{2\alpha}}+ \frac{120}{\alpha^2}\left(\frac{S}{n^{2\alpha}}\wedge \frac{1}{n^{2\alpha-1}}\right) \right).\end{aligned}$$
5. For the entropy $H(P)$, $$\sup_{P \in \cM_S} \bE_P \left( H(P_n) - H(P) \right)^2 \leq \left( \ln \left( 1+ \frac{S-1}{n} \right) \right)^2 + \left(\frac{(\ln n)^2}{n}\wedge \frac{2(\ln S + 2)^2}{n}\right).$$
Moreover, in all the bounds presented above, the first term bounds the square of the bias, and the second term bounds the variance.
There are several interesting implications of this result, highlighted in the following corollaries.
\[cor.remark.thmmain\] For the functional $F_\alpha(P), \alpha > 1$, if $n \gg 1$, MLE is consistent regardless of the alphabet size, i.e., $\sup_{P \in \mathcal{M}_\infty} E_P (F_\alpha(P_n) - F_\alpha(P))^2 \longrightarrow 0$ as $n \rightarrow \infty$, where $\mathcal{M}_\infty$ denotes the set of all discrete distributions.
Evidently, when the functional $F_\alpha(P)$ is differentiable everywhere, the number of samples required to make the maximum $L_2$ risk vanish has no dependence on the alphabet size $S$. Results of this form have appeared in the literature, for example, Antos and Kontoyiannis [@Antos--Kontoyiannis2001convergence] showed that it suffices to take $n \gg 1$ samples to consistently estimate $F_\alpha(P), \alpha\geq 2, \alpha \in \mathbb{Z}$. To our knowledge, Theorem \[thm.main\] gives the first non-asymptotic result for estimation of $F_\alpha(P)$, $\alpha$ non-integer. More refined than Corollary \[cor.remark.thmmain\] is the following:
\[cor.largerthanone\] There exist universal convergence rates for $F_\alpha(P), \alpha> 1$, any alphabet size $S$ (possibly infinite): $$\sup_S \sup_{P \in \cM_S} \bE_P \left( F_\alpha(P_n) - F_\alpha(P) \right)^2 \lesssim \begin{cases} n^{-2(\alpha-1)} & 1< \alpha<3/2 \\ n^{-1} & \alpha\geq 3/2 \end{cases}$$
Corollary \[cor.largerthanone\] implies that, when $\alpha\geq 3/2$, estimation of $F_\alpha(P)$ is extremely simple in terms of convergence rate: plug-in estimation achieves the best possible rate $n^{-1}$ (as shown in the theory of regular statistical experiments of classical asymptotic theory, see [@Ibragimov--Hasminskii1981 Chap. 1.7.]). However, when $1<\alpha<3/2$, the rate $n^{-2(\alpha-1)}$ is considerably slower. Are the bounds obtained in this regime loose? It turns out that the convergence rate $n^{-2(\alpha-1)}$ is in fact tight for the MLE when $1<\alpha<3/2$. Precisely, we have the following lower bound.
\[thm.loweralphalargerthanone\] If $S = cn, 1<\alpha<3/2$, $c>0$, then $$\liminf_{n\to \infty} n^{2(\alpha-1)} \cdot \sup_{P \in \mathcal{M}_S} \bE_P(F_\alpha(P_n) - F_\alpha(P))^2>0.$$
Interestingly, there exist estimators that demonstrate better convergence rates for estimating $F_\alpha(P), 1<\alpha<3/2$. The companion paper [@Jiao--Venkat--Han--Weissman2014minimax] shows that the minimax $L_2$ rate in estimating $F_\alpha(P), 1<\alpha<3/2$, is $(n\ln n)^{-2(\alpha-1)}$ as long as $S\gtrsim n\ln n$, which is achieved using the general methodology developed therein for constructing minimax rate-optimal estimators for nonsmooth functionals.
Let us now examine the case $0 < \alpha < 1$, another interesting regime that has not been characterized before. In this regime, we observe significant increase in the difficulty of the estimation problem. In particular, the relative scaling between the number of observations $n$ and the alphabet size $S$ for consistent estimation of $F_\alpha(P)$ exhibits a phase transition, encapsulated in the following.
\[cor.mlealphaless1\] For the functional $F_\alpha(P), 0<\alpha<1$, if $n \gg S^{\frac{1}{\alpha}}$, MLE is consistent.
In words, in the region $0 < \alpha < 1$, $S^{\frac{1}{\alpha}}$ samples are sufficient for estimating $F_\alpha(P)$. It turns out that this characterization of the performance of the MLE is in fact tight. Precisely, we have the following lower bound for the performance of the MLE in this region.
\[thm.lowerfalpha\] If $n \geq S$, then
1. $1/2\leq \alpha<1$: $$\begin{aligned}
& \sup_{P \in \mathcal{M}_S} \bE_P \left( F_\alpha(P_n) - F_\alpha(P) \right)^2 \nonumber \\
& \quad \geq \frac{\alpha^2(1-\alpha)^2}{72 n^{2\alpha}} (S-1)^2 \left(1-\frac{1}{n}\right)^2 \nonumber \\
& \quad \quad + \left(\frac{\alpha^2}{64en}\left[(2(S-1))^{1-\alpha}-2^{-\alpha} -\frac{1-\alpha}{4n}\left((2(S-1))^{1-\alpha}+2^{-\alpha}\right)\right]^2 - \frac{1}{2} e^{-n/4}S^{2(1-\alpha)} \right),\end{aligned}$$
2. $0<\alpha<1/2$: $$\begin{aligned}
\sup_{P \in \mathcal{M}_S} \bE_P \left( F_\alpha(P_n) - F_\alpha(P) \right)^2 \geq \frac{\alpha^2(1-\alpha)^2}{36 n^{2\alpha}} (S-1)^2 \left(1-\frac{1}{n}\right)^2.\end{aligned}$$
\[cor.l2ratesfalphalessone\] The maximum $L_2$ risk of the MLE $F_\alpha(P_n)$ in estimating $F_\alpha(P)$ is characterized as follows when $n\geq S$: $$\sup_{P \in \mathcal{M}_S} \bE_P \left( F_\alpha(P_n) - F_\alpha(P) \right)^2 \asymp \begin{cases} \frac{S^2}{n^{2\alpha}} + \frac{S^{2-2\alpha}}{n} & 1/2< \alpha<1 \\ \frac{S^2}{n^{2\alpha}} & 0<\alpha \leq 1/2 \end{cases}$$
Corollary \[cor.l2ratesfalphalessone\] follows directly from Theorem \[thm.main\] and Theorem \[thm.lowerfalpha\]. In particular, it implies that it is necessary and sufficient to take $n \gg S^{1/\alpha}$ samples to consistently estimate $F_\alpha(P), 0<\alpha<1$ using MLE. Thus, as one might expect, the scale of the number of measurements required for consistent estimation increases as $\alpha$ decreases. When $\alpha \to 0$, the number of samples required for the MLE grows super-polynomially in $S$. On the other hand, for $\alpha > 1$, it suffices to take $n \gg 1$ samples for the MLE to be consistent. We exhibit some of our findings by plotting the value required of $\ln n/\ln S$ for consistent estimation of $F_\alpha(P)$ using the MLE $F_\alpha(P_n)$, as a function of $\alpha$, in Figure \[fig.falphaphase\].
(0.6,2.8) – (0.6,0.6) – (2.8,0.6); (1.0316, 2.9171) – (1.0789, 2.6879) – (1.1263, 2.5000 ) – (1.1737, 2.3431 )– (1.2211, 2.2102) – (1.2684,2.0961 ) – (1.3158,1.9971 ) – (1.3632,1.9103 ) – (1.4105,1.8338 ) – (1.4579,1.7656 ) – (1.5053, 1.7047) – (1.5526,1.6497 ) – (1.6000, 1.6); (1.6, 0.6) – (2.8, 0.6); (1.6, 0.6) – (1.6,1.6); at (1.1,1) [not achievable via MLE]{}; at (1.1,1) [(Theorem \[thm.lowerfalpha\])]{}; at (0.6, 1.6) [1]{}; at (0.6,0.6) [0]{}; at (1.6, 0.6) [1]{}; at (2.6,0.6) [2]{}; at (2.8,0.6) [$\alpha$]{}; at (0.6,2.8) [$\frac{\ln n}{\ln S}$]{}; (1.1,1.6) – (1.4579,1.7656 ); at (1.1, 1.6) [$1/\alpha$]{}; at (2.4, 1.8) [achievable via MLE]{}; at (2.4, 1.8) [(Theorem \[thm.main\])]{};
\[fig.falphaphase\]
A natural question arising in light of Corollary \[cor.mlealphaless1\] and Theorem \[thm.lowerfalpha\] is whether one can construct estimators that are better than the MLE in terms of required sample complexity for consistent estimation. The answer turns out to be affirmative, as we show in the companion paper [@Jiao--Venkat--Han--Weissman2014minimax]. In particular, the scheme we introduce therein is a consistent estimator of $F_\alpha(P)$ in the regime $0 < \alpha < 1$ when $n\gg \frac{S^{\frac{1}{\alpha}}}{\ln S}$, which is a logarithmic improvement in the sample complexity over the MLE. It is also shown that the scheme of [@Jiao--Venkat--Han--Weissman2014minimax] is minimax rate-optimal.
Let us now shift our focus to the case of entropy $H(P)$, which may be roughly intuitively viewed as the functional $F_\alpha(P)$ when $\alpha \to 1^{-}$. The following corollary demonstrates that such intuition is indeed approximately correct as $H(P)$ coincides with $\alpha \to 1^{-}$ on the phase transition curve shown in Figure \[fig.falphaphase\]. Theorem \[thm.main\] implies the following corollary.
Case $H(P)$: The maximum $L_2$ risk of the MLE vanishes provided $n \gg S$.
As it turns out, $n \asymp S$ is the optimal scaling for consistency of $H(P_n)$ in estimating $H(P)$. Further, we show that the same is true of the Miller–Madow bias-corrected estimator [@Miller1955] defined as $$H^{\text{MM}}(P_n) = H(P_n) + \frac{S-1}{2n}.$$
\[thm.lowerentropy\] For the entropy $H(P)$, if $n \geq 15S$, then $$\sup_{P \in \mathcal{M}_S} \bE_P \left( H(P_n) - H(P) \right)^2 \geq \frac{1}{2}\left(\frac{S-1}{2n} + \frac{S^2}{20n^2} - \frac{1}{12n^2}\right)^2 + c \frac{\ln^2 S}{n}.$$ Moreover, if $n \geq 15S$, the Miller–Madow bias-corrected estimator satisfies $$\sup_{P \in \mathcal{M}_S} \bE_P \left( H^{\mathrm{MM}}(P_n) - H(P) \right)^2 \geq \frac{1}{2}\left(\frac{S^2}{20n^2} - \frac{1}{12n^2}\right)^2 + c \frac{\ln^2 S}{n},$$ where the positive constant $c>0$ in both expressions does not depend on $S$ or $n$.
Theorem \[thm.main\] and Theorem \[thm.lowerentropy\] imply the following corollary.
\[cor.entropyl2rate\] The maximum $L_2$ risk of the MLE $H(P_n)$ in estimating $H(P)$ is characterized as follows when $n\geq 15S$: $$\sup_{P \in \mathcal{M}_S} \bE_P \left( H(P_n) - H(P) \right)^2 \asymp \frac{S^2}{n^2} + \frac{\ln^2 S}{n}.$$
We can raise an analogous question for the estimation of $H(P)$ as we did for $F_\alpha(P), 0<\alpha<1$: does there exist an entropy estimator with vanishing maximum $L_2$ risk with sublinear $n\ll S$ number of samples? The answer is affirmative, as was shown in [@Valiant--Valiant2011], and more recently via a different scheme by the present authors in [@Jiao--Venkat--Han--Weissman2014minimax], that one can construct estimators for which $n = \omega\left( S/\ln S \right)$ samples suffice for consistent estimation. It was further shown in [@Valiant--Valiant2011; @Wu--Yang2014minimax] that this is the optimal order as the number of samples needed for any estimator to achieve vanishing maximum $L_2$ risk is at least of order $\omega(S/\ln S)$. As in the case of $F_\alpha(P),0<\alpha<1$, the MLE serves as a very good benchmark when we compare it with the minimax rate-optimal estimator.
To sum up our results regarding achievability and lower bounds on estimation of $F_\alpha(P)$ and $H(P)$, we have Table \[table.summary\], in which JVHW denotes [@Jiao--Venkat--Han--Weissman2014minimax], VV denotes [@Valiant--Valiant2011power; @Valiant--Valiant2011], and WY denotes [@Wu--Yang2014minimax].
The following table summarizes the minimax $L_2$ rates and the $L_2$ convergence rates of the MLE for estimating $F_\alpha(P),\alpha>0$ and $H(P)$.
Minimax $L_2$ rates $L_2$ rates of MLE
------------------------------------------ --------------------------------------------------------------------------------------------------------------------- -----------------------------------------------------------------------------------------------------------------------------------------
$H(P)$ $\frac{S^2}{(n \ln n)^2} + \frac{\ln^2 S}{n} \quad \left( n \gtrsim S/\ln S \right)$ (JVHW,WY,VV) $\frac{S^2}{n^2} + \frac{\ln^2 S}{n}\quad \left( n \gtrsim S \right)$ (Corollary \[cor.entropyl2rate\])
$F_\alpha(P), 0<\alpha \leq \frac{1}{2}$ $\frac{S^2}{(n \ln n)^{2\alpha}} \quad \left( n \gtrsim S^{1/\alpha}/\ln S, \ln n \lesssim \ln S \right)$ (JVHW) $\frac{S^2}{n^{2\alpha}}\quad \left( n \gtrsim S^{1/\alpha} \right)$ (Corollary \[cor.l2ratesfalphalessone\])
$F_\alpha(P), \frac{1}{2}< \alpha<1$ $\frac{S^2}{(n \ln n)^{2\alpha}} + \frac{S^{2-2\alpha}}{n}\quad \left( n \gtrsim S^{1/\alpha}/\ln S \right)$ (JVHW) $\frac{S^2}{n^{2\alpha}} + \frac{S^{2-2\alpha}}{n}\quad \left( n \gtrsim S^{1/\alpha} \right)$ (Corollary \[cor.l2ratesfalphalessone\])
$F_\alpha(P), 1< \alpha<\frac{3}{2}$ $(n \ln n)^{-2(\alpha-1)}\quad \left(S \gtrsim n\ln n \right)$ (JVHW) $n^{-2(\alpha-1)}\quad \left(S \gtrsim n \right)$ (Thm. \[thm.main\],\[thm.loweralphalargerthanone\])
$F_\alpha(P), \alpha\geq \frac{3}{2}$ $n^{-1}$ (Thm. \[thm.main\]) $n^{-1}$
: Summary of results in this paper and the companion [@Jiao--Venkat--Han--Weissman2014minimax][]{data-label="table.summary"}
Table \[table.summary\] demonstrates that the MLE cannot achieve the minimax risk for estimation of $H(P)$, and $F_\alpha(P)$ when $0 < \alpha < 3/2$. In these cases, there exist strictly better estimators whose performance with $n$ samples is roughly the same as that of the MLE with $n \ln n$ samples. In other words, the optimal estimators *enlarge* the effective sample size by a logarithmic factor.
To our knowledge, Paninski [@Paninski2003] was the first to have realized the connection between Bernstein polynomials and bias of MLE. In the same paper [@Paninski2003], Paninski demonstrated that if $n = cS$, where $c>0$ is a constant, the maximum squared bias of $H(P_n)$, and the Miller–Madow bias-corrected estimator $H^{\mathrm{MM}}(P_n)$, would be bounded from zero. Theorem \[thm.lowerentropy\] is the first to provide precise non-asymptotic constants in lower bounding the maximum $L_2$ risk of $H(P_n)$ and $H^{\mathrm{MM}}(P_n)$.
We now present the proofs of Theorem \[thm.main\], \[thm.loweralphalargerthanone\], \[thm.lowerfalpha\], and \[thm.lowerentropy\] in Section \[sec.outline\]. Proofs of lemmas are provided in the appendices.
Proofs of main results {#sec.outline}
======================
In the sequel, we denote the $n$ samples as $n$ i.i.d. random variables $\{Z_i\}_{1\leq i\leq n}$ taking values in $\mathcal{Z} = \{1,2,\ldots,S\}$ with probability $(p_1,p_2,\ldots,p_S)$. Defining $$X_i \triangleq \sum_{j = 1}^n \mathbbm{1}(Z_j = i),\quad 1\leq i \leq S,$$ we know that $(X_1,X_2,\ldots,X_S)$ follows a multinomial distribution with parameter $(n;p_1,p_2,\ldots,p_S)$. Denote $h_j \triangleq \sum_{i = 1}^S \mathbbm{1}(X_i = j),\quad 0\leq j \leq n$. Then it is evident that the MLE $F(P_n)$ can be alternatively represented as the following linear function of $(h_0,h_1,\ldots,h_n)$: $$F(P_n) = \sum_{j = 0}^n f\left( \frac{j}{n} \right) h_j.$$
Proof of Theorem \[thm.main\]
-----------------------------
### Bounding the bias
We first work to bound the bias. Lemma \[lemma.biasgeneralf\] shows that the bias of $F(P_n)$ could be represented as $$\mathsf{Bias}(F(P_n)) = \sum_{i = 1}^S \left( B_n[f](p_i) - f(p_i) \right),$$ where $B_n[f](x)$ is the Bernstein polynomial corresponding to $f(x)$. The following lemma summarizes some state-of-the-art bounds for approximation error of Bernstein polynomials. We remark that Lemma \[lemma.bernsteinerror\] can be derived easily from the general theory we presented in Section \[sec.generalbiasanalysis\].
\[lemma.bernsteinerror\] The following bounds are valid for function approximation error incurred by Bernstein polynomials:
1. [@Paltanea2004 Cor. 2.2.1][@Paltanea2008some] for all continuous functions $f$ on $[0,1]$, $$\label{eqn.bern1}
|f(x)-B_n[f](x)| \leq \frac{3}{2} \omega^2\Bigg(f, \sqrt{\frac{x(1-x)}{n}}\Bigg),$$ and the constant $3/2$ is shown by [@Paltanea2008some] to be the best constant;
2. [@Devore--Lorentz1993 Eqn. 10.3.4] for $f\in C^2[0,1]$, i.e., twice continuously differentiable, $$\label{eqn.bern2}
| f(x) - B_n[f](x) | \leq \| f'' \|_\infty \frac{x(1-x)}{2n};$$
3. [@Paltanea2004 Cor. 4.1.10] for $\varphi(x) = \sqrt{x(1-x)}$ and all continuous functions $f$ on $[0,1]$, we have $$\label{eqn.bern3}
\| B_n[f](x) - f(x) \|_\infty \leq \frac{5}{2}\omega_\varphi^2(f,n^{-1/2});$$
The first part of Lemma \[lemma.bernsteinerror\] follows from Lemma \[lemma.paltaneacor221\]. The third part of Lemma \[lemma.bernsteinerror\] follows from Lemma \[lemma.paltaneathm251\]. Regarding the second part, suppose random variable $X \sim \mathsf{B}(n,x)$. We have $$\begin{aligned}
|f(x) - B_n[f](x)| & = |\bE_x f(X/n) - f(x)| \\
& = |\bE_x (f(x) + f'(x)(X/n-x) + \frac{1}{2} f''(\xi_X) (X/n-x)^2) ) - f(x)| \\
& = \frac{1}{2} |\bE_x f''(\xi_X) (X/n-x)^2| \\
& \leq \frac{\| f''\|_\infty}{2} |\bE_x (X/n-x)^2| \\
& = \frac{\| f''\|_\infty}{2} \frac{x(1-x)}{n},\end{aligned}$$ where we used Taylor expansion for $f(X/n)$ at point $x$ with the Lagrange remainder. The proof is complete.
Note that although (\[eqn.bern3\]) is in the form of an upper bound, it has been shown to be a lower bound as well. Totik [@Totik1994approximation] showed the following equivalence property on the norm estimate of Bernstein approximation errors $$\| B_n[f](x) - f(x) \|_\infty \asymp \omega_\varphi^2(f,n^{-1/2}).$$
It is easy to calculate the second-order modulus of smoothness and the Ditzian–Totik second-order modulus of smoothness for functions $x^\alpha$ and $-x \ln x$. The results are presented in the following lemma.
\[lemma.dtmoduluscomputation\] We have
$x^\alpha, 0<\alpha<1$ $x^\alpha, 1<\alpha< 2$ $-x \ln x $
------------------------- ----------------------------------------------------- ------------------------- ---------------------------
$\omega^2(f,t)$ $|2-2^\alpha|t^\alpha$ $|2-2^\alpha|t^\alpha$ $t \ln 4$
$\omega^2_\varphi(f,t)$ $|2-2^\alpha| \frac{t^{2\alpha}}{(1+t^2)^{\alpha}}$ $ \asymp t^2$ $\frac{t^2 \ln 4}{1+t^2}$
where the second-order modulus results hold for $0<t\leq 1/2$, and the Ditizan–Totik second-order modulus results hold for $0<t\leq 1$.
We first bound the bias incurred by $F_{\alpha}(P_n)$.
1. $\alpha \geq 2$:
In this case, $f \in C^2[0,1]$, hence $$|f(x) - B_n[f](x)| \leq \frac{\alpha (\alpha-1) x(1-x)}{2n}.$$ Thus, we have $$|\mathsf{Bias}(F_{\alpha}(P_n))| \leq \sum_{i =1}^S \alpha (\alpha-1) \frac{p_i(1-p_i)}{2n} \leq \frac{\alpha (\alpha-1)}{2n}.$$
2. $1<\alpha<2$
The following lemma presents a bound on the bias of $F_\alpha(P_n)$, which does not depend on the alphabet size $S$.
\[lemma.biasalphaonetwo\] The bias of $F_\alpha(P_n)$ for estimating $F_\alpha(P), 1<\alpha<2$, is upper bounded by the following: $$|\mathsf{Bias}(F_{\alpha}(P_n))| \leq \frac{4}{n^{\alpha-1}}.$$
We also present two additional bounds involving the alphabet size $S$. Using the pointwise estimate, the bias term of the MLE is upper bounded as follows for all $0<\alpha<2, \alpha \neq 1$:
$$\begin{aligned}
\sum_{i = 1}^S \frac{3}{2}|2-2^\alpha| \left( \frac{p_i(1-p_i)}{n} \right)^{\alpha/2} & \leq \frac{3}{2}|2-2^\alpha|\frac{1}{n^{\alpha/2}} \sum_{i =1}^S p_i^{\alpha/2} \\
& \leq \frac{3}{2}|2-2^\alpha|\frac{1}{n^{\alpha/2}} S \frac{1}{S^{\alpha/2}} \\
& = \frac{3}{2}|2-2^\alpha|\frac{S^{1-\alpha/2}}{n^{\alpha/2}}. \label{eqn.biasoneandtwopoint}\end{aligned}$$
Using the norm estimate, when $1<\alpha<2$, the bias would be upper bounded by $C_{\alpha,n} \frac{5S}{2n}$, where $C_{\alpha,n} = n \omega_\varphi^2(x^\alpha,n^{-1/2})$ is a finite positive constant such that $\limsup_{n\to \infty} C_{\alpha,n}<\infty$ for $1<\alpha<2$. Combining Lemma \[lemma.biasalphaonetwo\], the pointwise estimate, and the norm estimate, we know that the bias of $F_\alpha(P_n)$ for $1<\alpha<2$ is upper bounded as $$|\mathsf{Bias}(F_\alpha(P_n))| \leq \frac{4}{n^{\alpha-1}} \wedge \frac{3}{2}|2-2^\alpha|\frac{S^{1-\alpha/2}}{n^{\alpha/2}} \wedge C_{\alpha,n}\frac{5S}{2n}.$$
3. $0<\alpha<1$:
The pointwise estimate is given by (\[eqn.biasoneandtwopoint\]). Using the norm estimate, the bias would be upper bounded by $|2-2^\alpha| \frac{5S}{2n^\alpha}$. Combining the pointwise estimate and the norm estimate, we know that the bias of $F_\alpha(P_n)$ for $0<\alpha<1$ is upper bounded as $$|\mathsf{Bias}(F_\alpha(P_n))| \leq \frac{3}{2}|2-2^\alpha|\frac{S^{1-\alpha/2}}{n^{\alpha/2}} \wedge |2-2^\alpha|\frac{5S}{2n^\alpha}.$$
4. Upper bound on $|\mathsf{Bias}(H(P_n))|$:
Using the norm estimate, we know $$|\mathsf{Bias}(H(P_n))| \leq \frac{5S \ln 4}{2n}.$$ Using the pointwise estimate, we obtain $$|\mathsf{Bias}(H(P_n))| \leq \frac{3}{2}\sqrt{\frac{S}{n}} \ln 4.$$
It was shown by [@Paninski2003 Prop. 1] that the squared bias of MLE $H(P_n)$ is upper bounded as $$\label{eqn.paninskientropy}
(\mathsf{Bias}(H(P_n)))^2 \leq \left( \ln \left( 1+ \frac{S-1}{n} \right) \right)^2,$$ which is better than the two bounds we obtained using Bernstein polynomial results. However, we remark that (\[eqn.paninskientropy\]) is obtained using special properties of the entropy function and connections between KL-divergence and $\chi^2$-divergence [@Tsybakov2008], which cannot be applied to general functions.
To the best of our knowledge, [@Strukov--Timan1977mathematical] was the first to analyze the non-asymptotic bias of the MLE in estimating $F_\alpha(P)$ and $H(P)$, but the techniques were specialized to these functionals.
### Bounding the variance
The Efron–Stein inequality gives a general upper bound on the variance.
\[lemma.esnew\][@Boucheron--Lugosi--Massart2013 Theorem 3.1] Let $Z_1,\ldots,Z_n$ be independent random variables and let $f(Z_1,Z_2,\ldots,Z_n)$ be a square integrable function. Moreover, if $Z_1',Z_2',\ldots,Z_n'$ are independent copies of $Z_1,Z_2,\ldots,Z_n$ and if we define, for every $i = 1,2,\ldots,n$, $$f_i' = f(Z_1,Z_2,\ldots,Z_{i-1},Z_i',Z_{i+1},\ldots,Z_n),$$ then $$\mathsf{Var}(f) \leq \frac{1}{2} \sum_{i = 1}^n \bE \left[ (f - f_i')^2 \right].$$
The following inequality, which is called the bounded differences inequality, is a useful corollary of the Efron–Stein inequality.
\[lemma.es\][@Boucheron--Lugosi--Massart2013 Corollary 3.2] If function $f: \mathcal{Z}^n \to \mathbb{R}$ has the *bounded differences property*, i.e. for some nonnegative constants $c_1,c_2,\ldots,c_n$, $$\sup_{z_1,\ldots,z_n,z_i'\in \mathcal{Z}} |f(z_1,\ldots,z_n) - f(z_1,\ldots,z_{i-1},z_i',z_{i+1},\ldots,z_n)| \leq c_i, 1\leq i \leq n,$$ then $$\mathsf{Var}(f(Z_1,Z_2,\ldots,Z_n)) \leq \frac{1}{4}\sum_{i=1}^n c_i^2,$$ given that $Z_1,Z_2,\ldots,Z_n$ are independent random variables.
\[cor.vargeneralf\] The variance of $F(P_n)$ satisfies the following upper bound: $$\mathsf{Var}(F(P_n)) \leq n\cdot\max_{0\leq j <n}(f((j+1)/n)-f(j/n))^2.$$ If $f(\cdot)$ is monotone, then we can strengthen the bound to be $$\mathsf{Var}(F(P_n)) \leq \frac{n}{4}\cdot\max_{0\leq j <n}(f((j+1)/n)-f(j/n))^2.$$
Now we bound the variance for $F_\alpha(P_n),\alpha>1$. We have $$\max_{0\leq j <n} (((j+1)/n)^\alpha - (j/n)^\alpha)^2 \leq \left( 1- \left( 1- \frac{1}{n} \right)^\alpha \right)^2 \leq\left( \frac{\alpha}{n} \right)^2,$$ where in the last step we used Bernoulli’s inequality: $(1+x)^r \geq 1+rx,\forall r \geq 1, x>-1, x\in \mathbb{R}$. Using Lemma \[lemma.es\], we know the variance is upper bounded by $$\mathsf{Var}(F_\alpha(P_n)) \leq \frac{\alpha^2}{4n}.$$
We bound the variance of $F_\alpha(P_n), 0<\alpha<1$ in the following lemma.
\[lemma.varianceboundfalphabetween01\] For $0<\alpha<1$, we have $$\begin{aligned}
\sup_{P\in\mathcal{M}_S}\mathsf{Var}(F_\alpha(P_n)) &\le \begin{cases} \frac{10S}{n^{2\alpha}}+ \left(\frac{3\alpha\cdot2^{3+2\alpha}+1}{8\alpha^2}\left(\frac{8\alpha}{e}\right)^{2\alpha}+4\right)\left(\frac{S}{n^{2\alpha}}\wedge \frac{1}{n^{2\alpha-1}}\right) & 0<\alpha<1/2 \\ \frac{10S^{2-2\alpha}}{n}+ \left(\frac{3\alpha\cdot2^{3+2\alpha}+1}{8\alpha^2}\left(\frac{8\alpha}{e}\right)^{2\alpha}+4\right)\left(\frac{S}{n^{2\alpha}}\wedge \frac{1}{n^{2\alpha-1}}\right) & 1/2\leq \alpha<1 \end{cases} \\
&\lesssim \begin{cases} \frac{S}{n^{2\alpha}} & \ 0<\alpha<1/2 \\ \frac{S^{2-2\alpha}}{n} + \left(\frac{S}{n^{2\alpha}}\wedge \frac{1}{n^{2\alpha-1}}\right) & \ 1/2\leq \alpha<1 \end{cases}
\end{aligned}$$
Further, one can show that for all $\alpha\in (0,1)$, $$\frac{3\alpha\cdot2^{3+2\alpha}+1}{8\alpha^2}\left(\frac{8\alpha}{e}\right)^{2\alpha}+4 \leq \frac{120}{\alpha^2},$$ which is used in Theorem \[thm.main\].
Regarding the variance of $H(P_n)$, we have
\[lemma.varianceboundentropymle\] $$\sup_{P \in \mathcal{M}_S} \mathsf{Var}(H(P_n)) \leq \frac{(\ln n)^2}{n}\wedge \frac{2(\ln S + 2)^2}{n} \lesssim \frac{(\ln S)^2 \wedge (\ln n)^2}{n}.$$
Proofs of Theorem \[thm.loweralphalargerthanone\], \[thm.lowerfalpha\] and \[thm.lowerentropy\]
------------------------------------------------------------------------------------------------
The following lemma was proved in [@Braess--Sauer2004].
\[lemma.qn1lower\] Let $k\geq 4$ be an even number. Suppose that the $k$-th derivative of $f$ satisfies $f^{(k)} \leq 0$ in $(0,1)$, $Q_{k-1}$ is the Taylor polynomial of order $k-1$ to $f$ at some $x_1$ in $(0,1)$. Then for $x\in [0,1]$, $$f(x) - B_n[f](x) \geq Q_{k-1} - B_n[Q_{k-1}](x).$$
### Proof of Theorem \[thm.loweralphalargerthanone\]: lower bounds for estimation of $F_\alpha(P)$ when $1<\alpha<3/2$
Consider $f_\alpha(x) = -x^\alpha, 1<\alpha<2, x\in [0,1]$. Applying Lemma \[lemma.qn1lower\] to $f_\alpha$, taking $k = 6$, we have the following result.
\[lemma.biasalphalarge\] Suppose $f_\alpha(x) = -x^\alpha, 1<\alpha<2$ on $[0,1]$. For all $x\in (0,1)$, we have $$\begin{aligned}
\hspace*{-4cm} & f_\alpha(x) -B_n[f_\alpha](x) \nonumber \\
\quad & \geq \max \left \{ \frac{\alpha(\alpha-1)x^{\alpha-2}(1-x)}{2n} \left( x + \frac{(2-\alpha)(3\alpha-1)x}{12n} + \frac{(2-\alpha)(5-3\alpha)}{12n} \right) + \frac{R_1(x)}{n^3} + \frac{R_2(x)}{n^4}, 0 \right \},\end{aligned}$$ where $$\begin{aligned}
R_1(x) & = \frac{\alpha(\alpha-1)(\alpha-2)(\alpha-3) x^{\alpha-3} (1-x)}{24} \left( 1+2(1-x)((5-2\alpha)x + \alpha-4) \right), \\
R_2(x) & = \frac{\alpha(\alpha-1)(\alpha-2)(\alpha-3)(\alpha-4) x^{\alpha-4} (1-x)(1-2x)(1-12x(1-x))}{120}.\end{aligned}$$
Note that we have assumed $S = cn,c>0$. If $c\leq 1$, we take a uniform distribution on $S$ elements $P = (1/S,1/S,\ldots,1/S)$, otherwise we take distribution $P = (1/n,1/n,\ldots,1/n,0,\ldots,0)$. We first analyze the $c\leq 1$ case. Applying Lemma \[lemma.biasalphalarge\], we have $$\begin{aligned}
& \sum_{i = 1}^S f_\alpha(1/S) - B_n[f_\alpha](1/S) \nonumber \\
& = \bE F_\alpha(P_n) - F_\alpha(P) \nonumber \\
& \geq S \cdot \Big( \frac{\alpha(\alpha-1)}{2 S^{\alpha-2}n} \left( \frac{1}{S} + \frac{(2-\alpha)(5-3\alpha)}{12n} \right) \nonumber \\
& \quad + \frac{\alpha(\alpha-1)(\alpha-2)(\alpha-3)}{24S^{\alpha-3}n^3}(1+2(\alpha-4)) + \frac{\alpha(\alpha-1)(\alpha-2)(\alpha-3)(\alpha-4)}{120S^{\alpha-4} n^4} + o(n^{-\alpha})\Big ) \nonumber \\
& = \frac{\alpha(\alpha-1)}{n^{\alpha-1}} \Big( \frac{1}{2c^{\alpha-3}}\left(\frac{1}{c}+\frac{(2-\alpha)(5-3\alpha)}{12}\right) + \frac{(\alpha-2)(\alpha-3)(1+2(\alpha-4))}{24 c^{\alpha-4}} \nonumber\\
& \quad + \frac{(\alpha-2)(\alpha-3)(\alpha-4)}{120c^{\alpha-5}} \Big) + o(n^{-(\alpha-1)}) \nonumber\\
& = \frac{\alpha(\alpha-1)c^{2-\alpha}}{n^{\alpha-1}} \Big(\frac{1}{2}+ \frac{(2-\alpha)(5-3\alpha)c}{24} + \frac{(\alpha-2)(\alpha-3)(1+2(\alpha-4))c^{2}}{24 } \nonumber\\
& \quad + \frac{(\alpha-2)(\alpha-3)(\alpha-4)c^{3}}{120} \Big) + o(n^{-(\alpha-1)}) \nonumber \\
& \geq \frac{\alpha c^{2-\alpha}(124-330\alpha + 285 \alpha^2-90 \alpha^3 + 11 \alpha^4)}{120 n^{\alpha-1}} + o(n^{-(\alpha-1)}), \nonumber\end{aligned}$$ where in the last step we have taken $c = 1$ in the following expression $$\frac{1}{2}+ \frac{(2-\alpha)(5-3\alpha)c}{24} + \frac{(\alpha-2)(\alpha-3)(1+2(\alpha-4))c^{2}}{24 } + \frac{(\alpha-2)(\alpha-3)(\alpha-4)c^{3}}{120},$$ and considered the fact that it is a monotonically decreasing function with respect to $c$ on $(0,1]$ for any $\alpha\in (1,3/2)$.
For cases when $c>1$, since we take $P = (1/n,1/n,\ldots,1/n,0,\ldots,0)$, the analysis is exactly the same as that above when we set $c=1$. One can verify that the function $\alpha(124-330\alpha + 285 \alpha^2-90 \alpha^3 + 11 \alpha^4)/120$ is positive on interval $(1,3/2)$. Defining $\sqrt{c_\alpha} = \alpha c^{2-\alpha}(124-330\alpha + 285 \alpha^2-90 \alpha^3 + 11 \alpha^4)/120>0$ when $c\leq 1$, and $\sqrt{c_\alpha} = \alpha (124-330\alpha + 285 \alpha^2-90 \alpha^3 + 11 \alpha^4)/120>0$ when $c>1$, Theorem \[thm.loweralphalargerthanone\] is proved.
### Proof of Theorem \[thm.lowerfalpha\]: lower bounds for estimation of $F_\alpha(P)$ when $0<\alpha<1$
Applying Lemma \[lemma.qn1lower\] to function $f_\alpha(x) = x^\alpha, \alpha \in (0,1)$, taking $k = 4$, we have the following result:
\[lemma.alphalower\] For $f_\alpha(x) = x^\alpha$ on $[0,1]$, $\alpha\in (0,1), x\in (0,1)$, we have $$f_\alpha(x) - B_n[f_\alpha](x) \geq \max \left\{ \frac{\alpha(1-\alpha)}{2n} x^{\alpha-2} (1-x) \left( x - \frac{2-\alpha}{3n} \right), 0\right\}$$
Suppose $\forall 1\leq i\leq S$, $p_i\geq \frac{1}{n}$. We have $n\geq S$. Applying Lemma \[lemma.alphalower\], $$\begin{aligned}
F_\alpha(P) - \bE F_\alpha(P_n) & \geq \sum_{i = 1}^S \frac{\alpha(1-\alpha)}{2n} p_i^{\alpha-2} (1-p_i) \frac{1}{3n} \\
& = \sum_{i =1}^S \frac{\alpha(1-\alpha)}{6n^2}p_i^{\alpha-2} (1-p_i)\end{aligned}$$
For distribution $W = (w_1,w_2,\ldots,w_S) \in \mathcal{M}_S$ such that $$1\leq i \leq S-1, w_i = \frac{1}{n}; \quad w_S = 1- \frac{S-1}{n},$$ we have $$\begin{aligned}
F_\alpha(W) - \bE_W F_\alpha(P_n) & \geq \sum_{i = 1}^{S-1} \frac{\alpha(1-\alpha)}{6n^2} \left(\frac{1}{n} \right)^{\alpha-2} \left( 1-\frac{1}{n} \right) \\
& = \frac{\alpha(1-\alpha)(S-1)}{6n^\alpha}\frac{n-1}{n}.\end{aligned}$$ Thus, we know for all $0<\alpha<1$, $$\sup_{P \in \mathcal{M}_S} \bE_P \left( F_\alpha(P) - F_\alpha(P_n) \right)^2 \geq \frac{\alpha^2(1-\alpha)^2 (S-1)^2}{36 n^{2\alpha}} \left( 1-\frac{1}{n}\right)^2.$$
It is shown in [@Jiao--Venkat--Han--Weissman2014minimax] that the following minimax lower bound holds for estimation of $F_\alpha(P), 1/2\leq \alpha<1$.
\[lemma.minimaxlowerfalpha12to1\] For $\frac{1}{2}\le\alpha<1$, we have $$\begin{aligned}
\inf_{\hat{F}}\sup_{P\in\mathcal{M}_S} \bE_P\left(\hat{F}-F_\alpha(P)\right)^2 &\ge \frac{\alpha^2}{32en}\left[(2(S-1))^{1-\alpha}-2^{-\alpha} -\frac{1-\alpha}{4n}\left((2(S-1))^{1-\alpha}+2^{-\alpha}\right)\right]^2 \nonumber \\
& \quad \quad - e^{-n/4}S^{2(1-\alpha)}\nonumber \\
&\gtrsim \frac{S^{2-2\alpha}}{n},
\end{aligned}$$ where the infimum is taken over all possible estimators.
Since this lower bound holds for all possible estimators, it also holds for the MLE $F_\alpha(P_n)$. In other words, $$\begin{aligned}
& \sup_{P \in \mathcal{M}_S} \bE_P \left( F_\alpha(P) - F_\alpha(P_n) \right)^2 \\
& \quad \geq \frac{\alpha^2}{32en}\left[(2(S-1))^{1-\alpha}-2^{-\alpha} -\frac{1-\alpha}{4n}\left((2(S-1))^{1-\alpha}+2^{-\alpha}\right)\right]^2 - e^{-n/4}S^{2(1-\alpha)}.\end{aligned}$$
Since $\max\{a,b\}\geq \frac{1}{2}(a+b)$, we have the desired bound in Theorem \[thm.lowerfalpha\].
### Proof of Theorem \[thm.lowerentropy\]: lower bounds for estimation of $H(P)$
Braess and Sauer [@Braess--Sauer2004] derived the following lower bound for the approximation error of Bernstein polynomials for the function $g(x) = -x \ln x$:
\[lemma.glower\] Define $g(x) = -x\ln x$ on $[0,1]$. For $x \geq \frac{15}{n}, x\in [0,1]$, we have $$g(x) - B_n[g](x) \geq \frac{1-x}{2n} + \frac{1}{20n^2 x} - \frac{x}{12 n^2}.$$
Applying Lemma \[lemma.glower\] to the estimation of $H(P)$, we know that if $\forall 1\leq i\leq S, p_i \geq \frac{15}{n}$, $$H(P) - \bE H(P_n) \geq \frac{S-1}{2n} + \frac{1}{20n^2} \left( \sum_{i = 1}^S \frac{1}{p_i} \right) - \frac{1}{12n^2}.$$ In order to guarantee that $\forall 1\leq i\leq S, p_i\geq \frac{15}{n}$, we need $$1 = \sum_{i = 1}^S p_i \geq \sum_{i =1}^S \frac{15}{n} = \frac{15S}{n},$$ which implies that $$n \geq 15S.$$ Since $$\sum_{i = 1}^S \frac{1}{p_i} \geq S^2,$$ we have $$\sup_{P\in \mathcal{M}_S} \left( H(P) - \bE H(P_n) \right) \geq \frac{S-1}{2n} + \frac{S^2}{20n^2} - \frac{1}{12n^2}.$$ Thus, when $n\geq 15S$, $$\sup_{P\in \mathcal{M}_S} \bE_P \left( H(P) - H(P_n) \right)^2 \geq \left( \frac{S-1}{2n} + \frac{S^2}{20n^2} - \frac{1}{12n^2} \right)^2.$$
It was shown in [@Wu--Yang2014minimax Prop. 1] that the following minimax lower bound holds.
\[lemma.minimaxlowerentropyvariance\] There exists a universal constant $c>0$ such that $$\inf_{\hat{H}} \sup_{P \in \mathcal{M}_S} \bE_P \left( \hat{H} - H(P) \right)^2 \geq c \frac{\ln^2 S}{n},$$ where the infimum is taken over all possible estimators $\hat{H}$.
Hence, we have $$\begin{aligned}
& \sup_{P\in \mathcal{M}_S} \bE_P \left( H(P) - H(P_n) \right)^2 \nonumber \\
& \quad \geq \max\left\{ \left( \frac{S-1}{2n} + \frac{S^2}{20n^2} - \frac{1}{12n^2} \right)^2 , c \frac{\ln^2 S}{n} \right\} \\
& \quad \geq \frac{1}{2}\left( \frac{S-1}{2n} + \frac{S^2}{20n^2} - \frac{1}{12n^2} \right)^2 + \frac{c}{2} \frac{\ln^2 S}{n}.\end{aligned}$$
Similar arguments can be applied to the Miller–Madow estimator.
Acknowledgments {#acknowledgments .unnumbered}
===============
We thank Dany Leviatan, Gancho Tachev, and Radu Paltanea for very helpful discussions regarding the literature on approximation theory using positive linear operators. We thank Jayadev Acharya, Alon Orlitsky, Ananda Theertha Suresh, and Himanshu Tyagi for communicating to us the independent discovery that it suffices to take $n\gg 1$ samples to consistently estimate $F_\alpha(P)$, when $\alpha>1$.
Auxiliary Lemmas
================
We begin with the definition of the negative association property, which allows us to upper bound the variance by treating each component of the empirical distribution $P_n(i)$ as “independent” random variables.
[@joag-dev1983 Def. 2.1] Random variables $X_1,X_2,\cdots,X_S$ are said to be negatively associated if for any pair of disjoint subsets $A_1,A_2$ of $\{1,2,\cdots,S\}$, and any component-wise increasing functions $f_1,f_2$, $$\begin{aligned}
\mathsf{Cov}\left(f_1(X_i,i\in A_1), f_2(X_j,j\in A_2)\right) \le 0.
\end{aligned}$$
To verify whether random variables $X_1,X_2,\cdots,X_S$ are negatively associated or not, the following lemma presents a useful criterion.
\[lem\_NA\] [@joag-dev1983 Thm. 2.9] Let $X_1,X_2,\cdots,X_S$ be $S$ independent random variables with log-concave densities. Then the joint conditional distribution of $X_1,X_2,\cdots,X_S$ given $\sum_{i=1}^S X_i$ is negatively associated.
In light of the preceding lemma, we can obtain the following corollary.
\[cor\_NA\_Multinomial\] For any discrete probability distribution vector $P\in\mathcal{M}_S$, the random variables $\mathbf{X}=(X_1,X_2,\cdots,X_S)$ drawn from the Multinomial distribution $\mathbf{X}\sim\mathsf{Multinomial}(n,P)$ are negatively associated.
Consider the Poissonized model $Y_i\sim \mathsf{Poi}(np_i), 1\le i\le S$ with all $Y_i$ independent, it is straightforward to verify that each $Y_i$ possesses a log-concave distribution. Then conditioning on $\sum_{i=1}^S Y_i=n$, we know that $(Y_1,Y_2,\cdots,Y_S)|(\sum_{i=1}^S Y_i=n) \sim \mathsf{Multinomial}(n,P)$, hence Lemma \[lem\_NA\] yields the desired result.
The following lemma gives some tail bounds for Poisson or Binomial random variables.
\[lem\_chernoff\] [@mitzenmacher2005probability Exercise 4.7] If $X\sim \mathsf{Poi}(\lambda)$ or $X\sim \mathsf{B}(n,\frac{\lambda}{n})$, then for any $\delta>0$, we have $$\begin{aligned}
\bP(X \geq (1+\delta) \lambda) & \leq \left( \frac{e^\delta}{(1+\delta)^{1+\delta}} \right)^\lambda, \\
\bP(X \leq (1-\delta) \lambda) & \leq \left( \frac{e^{-\delta}}{(1-\delta)^{1-\delta}} \right)^\lambda\leq e^{-\delta^2\lambda/2}.\end{aligned}$$
To establish the upper bound of the variance obtained by the plug-in estimator $F_\alpha(P_n)$, we split $p$ into two different regimes $p\le 1/n$ or $p>1/n$, and the following lemmas give the corresponding variance bounds.
\[lem\_variance\_p\_small\] For $nX\sim\mathsf{B}(n,p),p\le 1/n$, we have $$\begin{aligned}
\mathsf{Var}(X^\alpha) \le \frac{2}{n^{2\alpha}}\wedge \frac{2p}{n^{2\alpha-1}} \quad 0<\alpha<1 .
\end{aligned}$$
\[lem\_variance\_p\_big\] For $nX\sim\mathsf{B}(n,p),p> 1/n$, we have $$\begin{aligned}
\mathsf{Var}(X^\alpha) \le \frac{10p^{2\alpha-1}}{n} + \frac{3}{2\alpha}\left(\frac{16\alpha}{en}\right)^{2\alpha} + \frac{2}{n^{2\alpha}} + \frac{1}{8\alpha^2}\left(\frac{8\alpha}{en}\right)^{2\alpha} & \quad 0<\alpha<1\end{aligned}$$
Proofs of Main Lemmas
=====================
Proof of Lemma \[lemma.biasgeneralf\]
-------------------------------------
We compute the first moment of $F(P_n)$. $$\bE F(P_n) = \sum_{j = 0}^n f \left( \frac{j}{n} \right) \bE h_j,$$ and $$\bE h_j = \bE \sum_{i = 1}^S \mathbbm{1}(X_i = j) = \sum_{i = 1}^S \bP(X_i = j) = \sum_{i = 1}^S \binom{n}{j} p_i^j(1-p_i)^{n-j}.$$ Thus, we have $$\bE F(P_n) = \sum_{j = 0}^n f \left( \frac{j}{n} \right)\sum_{i = 1}^S \binom{n}{j} p_i^j(1-p_i)^{n-j} = \sum_{j = 0}^n \sum_{i = 1}^S f \left( \frac{j}{n} \right) \binom{n}{j} p_i^j(1-p_i)^{n-j}.$$
The bias of $F(P_n)$ is $$\mathsf{Bias}(F(P_n)) = \bE F(P_n) - F(P) = \sum_{i = 1}^S \left( \sum_{j = 0}^n f \left( \frac{j}{n} \right) \binom{n}{j} p_i^j(1-p_i)^{n-j} - f(p_i) \right).$$
Proof of Lemma \[lemma.biasalphaonetwo\]
----------------------------------------
The bias of $F_\alpha(P_n), 1<\alpha<2$ can be expressed as follows: $$\begin{aligned}
| \mathsf{Bias}(F_\alpha(P_n))| & = | \bE \sum_{i = 1}^S P_n^\alpha(i) - p_i^\alpha | \\
& \leq \left| \bE \sum_{i: p_i \leq \frac{1}{n}} P_n^\alpha(i) - p_i^\alpha \right| + \left| \bE \sum_{i: p_i > \frac{1}{n}} P_n^\alpha(i) - p_i^\alpha \right| \\
& \triangleq B_1 + B_2.\end{aligned}$$
Now we bound $B_1$ and $B_2$ separately. It follows from Jensen’s inequality that for any $i$, $$\bE P_n^\alpha(i) \geq p_i^\alpha, \quad 1\leq i \leq S.$$
Hence, we have $$\begin{aligned}
B_1 & = \bE \sum_{i: p_i \leq \frac{1}{n}} P_n^\alpha(i) - p_i^\alpha \\
& = \bE \sum_{i: p_i \leq \frac{1}{n}} P_n^\alpha(i) - \frac{(np_i)^\alpha}{n^\alpha} \\
& \leq \bE \sum_{i: p_i \leq \frac{1}{n}} P_n^\alpha(i) - \frac{(np_i)^2}{n^\alpha} \\
& = \bE \sum_{i: p_i \leq \frac{1}{n}} \frac{(n P_n(i))^\alpha}{n^\alpha}- \frac{(np_i)^2}{n^\alpha} \\
& \leq \bE \sum_{i: p_i \leq \frac{1}{n}} \frac{(n P_n(i))^2}{n^\alpha} - \frac{(np_i)^2}{n^\alpha} \\
& = \sum_{i: p_i \leq \frac{1}{n}} \frac{\bE(n P_n(i))^2}{n^\alpha}- \frac{(np_i)^2}{n^\alpha} \\
& = \sum_{i: p_i \leq \frac{1}{n}} \frac{(n p_i)^2 + np_i(1-p_i)}{n^\alpha}- \frac{(np_i)^2}{n^\alpha} \\
& \leq \sum_{i: p_i \leq \frac{1}{n}} \frac{np_i}{n^\alpha}\\
& = \frac{1}{n^{\alpha-1}},\end{aligned}$$ where we have used the fact that $nP_n(i) \geq 1$ for any $P_n(i)\neq 0$.
Regarding $B_2$, we have the following bounds: $$\begin{aligned}
B_2 & = \left| \bE \sum_{i: p_i >\frac{1}{n}} P_n^\alpha(i) - p_i^\alpha \right| \\
& \leq \sum_{i: p_i > \frac{1}{n}} \bE | P_n^\alpha(i) - p_i^\alpha | \\
& \leq \sum_{i: p_i > \frac{1}{n}} \frac{3}{2}|2-2^\alpha| \left( \frac{p_i(1-p_i)}{n} \right)^{\alpha/2} \\
& \leq \frac{3|2-2^\alpha|}{2n^{\alpha/2}} \sum_{i: p_i > \frac{1}{n}} p_i^{\alpha/2},\end{aligned}$$ where the second inequality follows from the pointwise bound on the Bernstein approximation error in Lemma \[lemma.bernsteinerror\].
Denoting $| \{ i: p_i > \frac{1}{n} \} | = K \leq n$, we know $$\sum_{i: p_i > \frac{1}{n}} p_i^{\alpha/2} \leq K^{1-\alpha/2} \leq n^{1-\alpha/2},$$ which implies that $$B_2 \leq \frac{3|2-2^\alpha|}{2n^{\alpha/2}}n^{1-\alpha/2} = \frac{3|2-2^\alpha|}{2n^{\alpha-1}}.$$
Therefore, we have $$\begin{aligned}
| \mathsf{Bias}(F_\alpha(P_n))| & \leq B_1 + B_2 \\
& \leq \frac{1}{n^{\alpha-1}}+ \frac{3|2-2^\alpha|}{2n^{\alpha-1}} \\
& \leq \frac{3|2-2^\alpha| + 2}{2n^{\alpha-1}} \\
& \leq \frac{4}{n^{\alpha-1}}.\end{aligned}$$
Proof of Lemma \[lemma.dtmoduluscomputation\]
---------------------------------------------
We first compute the second-order modulus. Fix $t, 0< t\leq 1/2$. Defining $M \triangleq \frac{u+v}{2}$, then the computation of second-order modulus is equivalent to maximization of $|f(M-t) - 2 f(M) + f(M+t)|$ over interval $M \in [t, 1-t]$, since all the functions we consider are strictly convex or concave over $[0,1]$.
For $f(x) = x^\alpha,0<\alpha<1$, $f(x)$ is strictly concave on $[0,1]$. It follows from Jensen’s inequality that $$g(M) = (M-t)^\alpha - 2M^\alpha + (M+t)^\alpha \leq 0,$$ and it suffices to minimize this function of $M$ in order to obtain the modulus. Taking derivative of $g(M)$, we have $$g'(M) = \alpha \left( (M-t)^{\alpha-1} - 2 M^{\alpha-1} + (M+t)^{\alpha-1} \right) \geq 0,$$ since $x^{\alpha-1}$ is a convex function on $[t,1-t]$. It implies that the function $g(M)$ is non-decreasing, and the minimum of $g(M)$ over $M\in [t,1-t]$ is attained at $M = t$, and the minimum value is $g(t) = (2^\alpha-2) t^\alpha$. Hence, the corresponding second-order modulus is $|2-2^\alpha| t^\alpha$.
Analogous procedures computes the second-order modulus for $x^\alpha,1<\alpha<2$ and $-x\ln x$.
Now we consider the computation of Ditzian–Totik second-order modulus. Fix $t, 0<t\leq 1$. Again denote $M \triangleq \frac{u+v}{2} \in [0,1]$. Then the optimization is over the regime $|u-v| \leq 2t\varphi(M) = 2t \sqrt{M(1-M)}$. Equivalently, it is the interval $[M-t\sqrt{M(1-M)},M+t\sqrt{M(1-M)}] \cap [0,1]$.
Since the function $f(x) = -x\ln x$ is strictly convex on $[0,1]$, the maximum of $\left|f(u)-2f\left(\frac{u+v}{2}\right)+f(v)\right|$ is definitely attained when $u$ and $v$ take the boundary values of the feasible interval $[M-t\sqrt{M(1-M)},M+t\sqrt{M(1-M)}] \cap [0,1]$.
Define $\Delta \triangleq t\sqrt{\frac{1-M}{M}}$. The feasible interval can be equivalently written as $[M-\Delta M, M+\Delta M] \cap [0,1]$. We have $$M - t\sqrt{M(1-M)} \geq 0 \Leftrightarrow M \geq \frac{t^2}{1+t^2},$$ as well as $$M + t\sqrt{M(1-M)} \leq 1 \Leftrightarrow M \leq \frac{1}{1+t^2}.$$
Hence, it is equivalent to maximize over three regimes:
1. Regime A:
$u = 0, v = 2M, 0\leq M \leq \frac{t^2}{1+t^2}$.
2. Regime B:
$u = M-\Delta M, v = M + \Delta M, M \in \left[ \frac{t^2}{1+t^2}, \frac{1}{1+t^2} \right]$
3. Regime C:
$u = 2M-1, v = 1, 1 \geq M \geq \frac{1}{1+t^2}$.
Over regime A, we have $$\left|f(u)-2f\left(\frac{u+v}{2}\right)+f(v)\right| = 2M \ln 2.$$ Maximizing over $0\leq M \leq \frac{t^2}{1+t^2}$, the maximum value is $\frac{t^2 \ln 4}{1+t^2}$, attained at $M = \frac{t^2}{1+t^2}$.
Over regime C, we have $$\left|f(u)-2f\left(\frac{u+v}{2}\right)+f(v)\right| = \left | 2M \ln M - (2M-1) \ln (2M-1) \right|.$$ Maximizing over $\frac{1}{1+t^2} \leq M \leq 1$, the maximum is attained at $M = \frac{1}{1+t^2}$, and the maximum value is no more than $\frac{t^2 \ln 4}{1+t^2}$.
Now we consider regime B. Since $M \in \left[ \frac{t^2}{1+t^2}, \frac{1}{1+t^2} \right]$ in regime B, we know $\Delta = t\sqrt{\frac{1-M}{M}} \in [t^2,1]$. We have
$$\left|f(u)-2f\left(\frac{u+v}{2}\right)+f(v)\right| = M\left | (1-\Delta)\ln(1-\Delta) + (1+\Delta)\ln(1+\Delta) \right|.$$
Since $\Delta = t\sqrt{\frac{1-M}{M}}$ implies $M = \frac{t^2}{t^2 + \Delta^2}$, we can recast the corresponding optimization problem as maximizing $$\frac{t^2}{\Delta^2 + t^2} \left | (1-\Delta)\ln(1-\Delta) + (1+\Delta)\ln(1+\Delta) \right|$$ subject to constraint $\Delta \in [t^2, 1]$. One can show that the maximum is always attained at $\Delta = 1$, with the maximum value $\frac{t^2 \ln 4}{1+t^2}$.
To sum up, we conclude that when $0< t\leq 1$, the maximum of the optimization problem defining $\omega^2_\varphi(-x\ln x,t)$ is always attained at $u = 0, v = \frac{2t^2}{1+t^2}$, with the resulting modulus $\frac{t^2 \ln 4}{1+t^2}$.
Analogous computation can also be done for function $x^\alpha,0<\alpha<1$. For the function $x^\alpha,1<\alpha<2$, it is hard to compute the modulus exactly, but it is easy to show that it is of order $t^2$.
Proof of Lemma \[lemma.varianceboundfalphabetween01\]
-----------------------------------------------------
In light of Lemma \[lem\_variance\_p\_small\] and \[lem\_variance\_p\_big\], we have $$\begin{aligned}
& \sum_{i=1}^S \mathsf{Var}(P_n(i)^\alpha) \nonumber \\
&= \sum_{i: p_i\le 1/n} \mathsf{Var}(P_n(i)^\alpha) + \sum_{i: p_i> 1/n} \mathsf{Var}(P_n(i)^\alpha)\\
&\le \sum_{i: p_i\le 1/n} \frac{2}{n^{2\alpha}}\wedge \frac{2p_i}{n^{2\alpha-1}}+ \sum_{i: p_i>1/n}\left(\frac{10p_i^{2\alpha-1}}{n} + \frac{3}{2\alpha}\left(\frac{16\alpha}{en}\right)^{2\alpha} + \frac{2}{n^{2\alpha}} + \frac{1}{8\alpha^2}\left(\frac{8\alpha}{en}\right)^{2\alpha}\right)\\
&\le \begin{cases} \left(\frac{2S}{n^{2\alpha}}\wedge \frac{2}{n^{2\alpha-1}}\right)+ \frac{10S}{n^{2\alpha}} + \left(\frac{3\alpha\cdot2^{3+2\alpha}+1}{8\alpha^2}\left(\frac{8\alpha}{e}\right)^{2\alpha}+2\right)\left(\frac{S}{n^{2\alpha}}\wedge \frac{1}{n^{2\alpha-1}}\right) & \ 0<\alpha<1/2 \\ \left(\frac{2S}{n^{2\alpha}}\wedge \frac{2}{n^{2\alpha-1}}\right)+ \frac{10S^{2-2\alpha}}{n} + \left(\frac{3\alpha\cdot2^{3+2\alpha}+1}{8\alpha^2}\left(\frac{8\alpha}{e}\right)^{2\alpha}+2\right)\left(\frac{S}{n^{2\alpha}}\wedge \frac{1}{n^{2\alpha-1}}\right) & 1/2\leq \alpha<1 \end{cases}\end{aligned}$$ where we have used the concavity of $x^{2\alpha-1}$ when $1/2\le \alpha<1$ in the last step.
Now we exploit the negative association property of all random variables $P_n(i),1\le i\le S$. Corollary \[cor\_NA\_Multinomial\] and monotonically increasing property of $x^\alpha$ yield $$\begin{aligned}
\mathsf{Var}(F_\alpha(P_n)) = \sum_{i=1}^S \mathsf{Var}(P_n(i)^\alpha) + 2\sum_{1\le i<j\le S}\mathsf{Cov}(P_n(i)^\alpha,P_n(j)^\alpha)\le \sum_{i=1}^S \mathsf{Var}(P_n(i)^\alpha),\end{aligned}$$ thereby finishing the proof of Lemma \[lemma.varianceboundfalphabetween01\].
Proof of Lemma \[lemma.varianceboundentropymle\]
------------------------------------------------
The upper bound $(\ln n)^2/n$ follows from Corollary \[cor.vargeneralf\]. We apply the Efron–Stein inequality (Lemma \[lemma.esnew\]) to obtain the other bound. Denote the $n$ i.i.d. samples from distribution $P$ as $Z_1,Z_2,\ldots,Z_n \in \mathcal{Z}$. Denoting the MLE $H(P_n)$ as $\hat{H}(Z_1,Z_2,\ldots,Z_n)$, since it is invariant to any permutation of $\{Z_1,Z_2,\ldots,Z_n\}$, we know that the Efron–Stein inequality implies $$\label{eqn.entropyesnew}
\mathsf{Var}(H(P_n)) \leq \frac{n}{2} \bE \left( \hat{H}(Z_1,Z_2,\ldots,Z_n) - \hat{H}(Z_1',Z_2,\ldots,Z_n) \right)^2,$$ where $Z_1'$ is an i.i.d. copy of $Z_1$.
Recall that $$X_i = \sum_{j = 1}^n \mathbbm{1}(Z_j = i),\quad 1\leq i \leq S.$$
For notional brevity, we denote the $S$-tuple $(X_1,X_2,\ldots,X_S)$ as $X_1^S$, and the $n$-tuple $(Z_1,Z_2,\ldots,Z_n)$ as $Z_1^n$. A specific realization of $(X_1,X_2,\ldots,X_S)$ is denoted by $x_1^S = (x_1,x_2,\ldots,x_S)$, and a specific realization of $(Z_1,Z_2,\ldots,Z_n)$ is denoted by $z_1^n = (z_1,z_2,\ldots,z_n)$.
In order to upper bound the right hand side of (\[eqn.entropyesnew\]), we first condition on $\{X_1,X_2,\ldots,X_S\}$. In other words, we use $$\begin{aligned}
& \bE \left( \hat{H}(Z_1,Z_2,\ldots,Z_n) - \hat{H}(Z_1',Z_2,\ldots,Z_n) \right)^2 \\
& \quad = \sum_{x_1^S} \bP(X_1^S = x_1^S) \bE \left[ \left( \hat{H}(Z_1,Z_2,\ldots,Z_n) - \hat{H}(Z_1',Z_2,\ldots,Z_n) \right)^2 \Bigg | X_1^S = x_1^S\right].\end{aligned}$$
The following lemma calculates the conditional distribution of $Z_1$ conditioned on $(X_1,X_2,\ldots,X_S)$.
\[lemma.conditionalpaninski\] The conditional distribution of $Z_1$ conditioned on $(X_1,X_2,\ldots,X_S)$ is given by the following discrete distribution: $$(X_1/n,X_2/n,\ldots,X_S/n).$$
By definition of conditional distribution, for any $k, 1\leq k \leq S$, we have $$\begin{aligned}
\bP(Z_1=k| X_1^S = x_1^S) & = \frac{\bP(Z_1 = k, X_1^S = x_1^S)}{\bP(X_1^S = x_1^S)} \\
& = \frac{\bP(Z_1 = k)\bP(X_1^S = x_1^S|Z_1 = k)}{\bP(X_1^S = x_1^S)} \\
& = \frac{p_k \binom{n-1}{x_1,x_2,\ldots,x_k-1,\ldots,x_S} p_k^{x_k-1} \prod_{i \neq k} p_i^{x_i}}{\binom{n}{x_1,x_2,\ldots,x_S} \prod_{1\leq i\leq S} p_i^{x_i}} \\
& = \frac{\binom{n-1}{x_1,x_2,\ldots,x_k-1,\ldots,x_S}}{\binom{n}{x_1,x_2,\ldots,x_S}} \\
& = \frac{x_k}{n},\end{aligned}$$ where the multinomial coefficient $\binom{n}{x_1,x_2,\ldots,x_S}$ is defined as $$\binom{n}{x_1,x_2,\ldots,x_S} = \frac{n!}{\prod_{i = 1}^S x_i!}.$$
Denoting $r(p) = -p\ln p$, we have $r(j/n) \triangleq\frac{-j}{n}\ln \frac{j}{n}$. We rewrite $$\hat{H}(Z_1',Z_2,\ldots,Z_n)-\hat{H}(Z_1,Z_2,\ldots,Z_n) = D_- + D_+,$$ where $$\begin{aligned}
D_- & = r\left( \frac{X_{Z_1}-1}{n} \right) - r\left(\frac{X_{Z_1}}{n}\right)\\
D_+ & = \begin{cases} r\left(\frac{X_{Z_1'+1}}{n} \right) - r\left(\frac{X_{Z_1'}}{n}\right) & Z_1 \neq Z_1'\\ r\left(\frac{X_{Z_1'}}{n} \right) - r\left(\frac{X_{Z_1'-1}}{n}\right) & Z_1 = Z_1' \end{cases}\end{aligned}$$
Here, $D_-$ is the change in $\hat{H}$ that occurs when $Z_1$ is removed according to the distribution specified in Lemma \[lemma.conditionalpaninski\], and $D_+$ is the change in $\hat{H}$ that occurs when $Z_1'$ is added back according to the true distribution $P$.
Now we compute $\bE [D_-^2|X_1^S]$ and $\bE [D_+^2| X_1^S]$. We have $$\begin{aligned}
\bE[D_-^2|X_1^S] & = \sum_{1\leq i\leq S} \frac{X_i}{n} \left( r\left(\frac{X_i-1}{n}\right) - r\left(\frac{X_i}{n}\right) \right)^2,\end{aligned}$$ and $$\begin{aligned}
& \bE[D_+^2|X_1^S] \nonumber \\
& \quad = \sum_{1\leq i\leq S} p_i \frac{X_i}{n} \left( r\left(\frac{X_i}{n}\right) -r\left(\frac{X_i-1}{n}\right) \right)^2 +\sum_{1\leq i\leq S} p_i\left(1-\frac{X_i}{n}\right) \left( r\left(\frac{X_i + 1}{n} \right)-r\left(\frac{X_i}{n}\right) \right)^2 .\end{aligned}$$
Note that we interpret $\frac{X_i}{n} \left( r\left(\frac{X_i}{n}\right) -r\left(\frac{X_i-1}{n}\right) \right)^2$ as $0$ when $X_i = 0$. Taking expectations of $\bE [D_-^2|X_1^S]$ and $\bE [D_+^2| X_1^S]$ with respect to $X_1^S$, we have $$\begin{aligned}
\bE [D_-^2] & = \sum_{1\leq i\leq S} \sum_{1\leq j\leq n} \frac{j}{n} \left( r\left(\frac{j-1}{n} \right) - r\left(\frac{j}{n}\right)\right)^2 \bP(\mathsf{B}(n,p_i) = j)\end{aligned}$$ and $$\begin{aligned}
& \bE[D_+^2] \nonumber \\
& \quad = \sum_{1\leq i\leq S} \sum_{0\leq j\leq n} \left( \frac{j}{n} \left(r\left(\frac{j-1}{n} \right)- r\left(\frac{j}{n}\right)\right)^2 + \left(1-\frac{j}{n}\right) \left(r\left(\frac{j}{n}\right) - r\left(\frac{j+1}{n}\right)\right)^2 \right) \nonumber \\
& \quad \quad \times p_i \bP(\mathsf{B}(n,p_i) = j).\end{aligned}$$
After some algebra, one can show that $\bE[D_+^2] = \bE[D_-^2]$. It then follows from (\[eqn.entropyesnew\]) that $$\begin{aligned}
\mathsf{Var}(H(P_n)) & \leq \frac{n}{2} \cdot \bE \left( \hat{H}(Z_1,Z_2,\ldots,Z_n) - \hat{H}(Z_1',Z_2,\ldots,Z_n) \right)^2 \\
& = \frac{n}{2} \cdot \bE \left( D_- + D_+ \right)^2 \\
& \leq n \cdot \bE (D_-^2 + D_+^2) \\
& \leq 2n \bE D_-^2 \\
& = 2n \cdot \sum_{1\leq i\leq S} \bE P_n(i) \left( r(P_n(i)) - r(P_n(i)-\frac{1}{n}) \right)^2 \label{eqn.espaninskifurther}\end{aligned}$$
The proof above is an elaborate version of that in [@Paninski2003 App. B.3]. Now we proceed to obtain non-asymptotic upper bounds of (\[eqn.espaninskifurther\]). For $x\geq 1/n$, it follows from Taylor expansion with integral form residue that $$\begin{aligned}
(x-\frac{1}{n}) \ln (x-\frac{1}{n}) & = x\ln x + (\ln x + 1) (-\frac{1}{n}) + \int_x^{x-\frac{1}{n}} (x-\frac{1}{n}-u) \frac{1}{u} du.\end{aligned}$$ Then, we have $$\begin{aligned}
\left|r(P_n(i)) - r\left(P_n(i) -\frac{1}{n}\right)\right| & \leq \frac{|\ln P_n(i) + 1|}{n} + \left| \int_{P_n(i)}^{P_n(i)-\frac{1}{n}} \frac{P_n(i)-\frac{1}{n}}{u}du \right| + \frac{1}{n} \\
& \leq \frac{|\ln P_n(i) + 1|+2}{n}.\end{aligned}$$
Hence, we have $$\mathsf{Var}(H(P_n)) \leq 2n\cdot\sum_{1\leq i\leq S} \bE P_n(i) \left( \frac{|\ln P_n(i) +1|+2}{n}\right)^2.$$ Noting that $\ln P_n(i) \leq 0$, we have $$\begin{aligned}
\mathsf{Var}(H(P_n)) & \leq \frac{2}{n}\cdot\sum_{1\leq i\leq S} \bE P_n(i) \left( \ln P_n(i)-2 \right)^2 \\
& \leq \frac{2}{n} \sum_{1\leq i\leq S} p_i(\ln p_i - 2)^2 \\
& \leq \frac{2}{n} S \cdot \frac{1}{S} (-\ln S - 2)^2 \\
& = \frac{2(\ln S + 2)^2}{n},\end{aligned}$$ where we have used the fact that $x(\ln x -2)^2$ is a concave function on $[0,1]$.
Proof of Lemma \[lemma.biasalphalarge\]
---------------------------------------
It is well known (see, e.g. [@Devore--Lorentz1993 Cor. 10.4.2]) that if $f$ is concave in $(0,1)$, then $$f(x) - B_n[f](x) \geq 0,\quad 0\leq x\leq 1.$$
Hence we focus on deriving the other bound. For concave function $f_\alpha(x) = -x^\alpha,\alpha\in (1,2) $, Taylor’s polynomial of degree $5$ at $x = x_0$ takes the form $$\begin{aligned}
Q_5(x) & = -\frac{\alpha(\alpha-1)}{2} x_0^{\alpha-2} (x-x_0)^2 - \frac{\alpha(\alpha-1)(\alpha-2)}{6} x_0^{\alpha-3} (x-x_0)^3 \\
& \quad -\frac{\alpha(\alpha-1)(\alpha-2)(\alpha-3)}{24} x_0^{\alpha-4}(x-x_0)^4 - \frac{\alpha(\alpha-1)(\alpha-2)(\alpha-3)(\alpha-4)}{120} x_0^{\alpha-5} (x-x_0)^5 \\
& \quad + \textrm{affine terms of }x\end{aligned}$$
We know that the Bernstein polynomial of any affine function on $[0,1]$ is the affine function itself, hence it suffices to consider the non-affine part of $Q_5(x)$. [@Braess--Sauer2004 Prop. 4] showed the following results for Bernstein polynomials:
\[lemma.brastaylor\] Let $0\leq x_0 \leq 1$. Then we have $$\begin{aligned}
B_n[(x-x_0)^2](x_0) & = \frac{x_0(1-x_0)}{n} \\
B_n[(x-x_0)^3](x_0) & = \frac{x_0(1-x_0)}{n^2}(1-2x_0) \\
B_n[(x-x_0)^4](x_0) & = 3 \frac{x_0^2(1-x_0)^2}{n^2} + \frac{x_0(1-x_0)}{n^3} [ 1-6x_0(1-x_0)] \\
B_n[(x-x_0)^5](x_0) & = \left( 10 \frac{x_0^2(1-x_0)^2}{n^3} + \frac{x_0(1-x_0)}{n^4}[1-12 x_0(1-x_0)] \right)(1-2x_0)\end{aligned}$$
Applying Lemma \[lemma.qn1lower\] and Lemma \[lemma.brastaylor\], taking $x_0 =x$, we have $$\begin{aligned}
& Q_5(x) - B_n[Q_5](x)\\
& = \frac{\alpha(\alpha-1)}{2}x^{\alpha-2} \frac{x(1-x)}{n} + \frac{\alpha(\alpha-1)(\alpha-2)}{6}x^{\alpha-3} \frac{x(1-x)}{n^2}(1-2x) \\
& \quad +\frac{\alpha(\alpha-1)(\alpha-2)(\alpha-3)}{24} x^{\alpha-4} \left( 3 \frac{x^2(1-x)^2}{n^2} + \frac{x(1-x)}{n^3} [ 1-6x(1-x)]\right) \\
& \quad + \frac{\alpha(\alpha-1)(\alpha-2)(\alpha-3)(\alpha-4)}{120} x^{\alpha-5} \left \{ \left( 10 \frac{x^2(1-x)^2}{n^3} + \frac{x(1-x)}{n^4}[1-12 x(1-x)] \right)(1-2x) \right \}\end{aligned}$$
After some algebra, we have $$\begin{aligned}
& Q_5(x) - B_n[Q_5](x)\nonumber \\
& \quad = \frac{\alpha(\alpha-1)x^{\alpha-2}(1-x)}{2n} \left( x + \frac{(2-\alpha)(3\alpha-1)x}{12n} + \frac{(2-\alpha)(5-3\alpha)}{12n} \right) + \frac{R_1(x)}{n^3} + \frac{R_2(x)}{n^4},\end{aligned}$$ where $$\begin{aligned}
R_1(x) & = \frac{\alpha(\alpha-1)(\alpha-2)(\alpha-3) x^{\alpha-3} (1-x)}{24} \left( 1+2(1-x)((5-2\alpha)x + \alpha-4) \right) \\
R_2(x) & = \frac{\alpha(\alpha-1)(\alpha-2)(\alpha-3)(\alpha-4) x^{\alpha-4} (1-x)(1-2x)(1-12x(1-x))}{120}\end{aligned}$$
Hence, we have $$\begin{aligned}
& f_\alpha(x) - B_n[f_\alpha](x) \nonumber\\
& \geq Q_5(x) - B_n[Q_5](x) \nonumber \\
& = \frac{\alpha(\alpha-1)x^{\alpha-2}(1-x)}{2n} \left( x + \frac{(2-\alpha)(3\alpha-1)x}{12n} + \frac{(2-\alpha)(5-3\alpha)}{12n} \right) + \frac{R_1(x)}{n^3} + \frac{R_2(x)}{n^4}. \nonumber\end{aligned}$$
Proof of Lemma \[lemma.alphalower\]
-----------------------------------
It is well known (see, e.g. [@Devore--Lorentz1993 Cor. 10.4.2]) that if $f$ is concave in $(0,1)$, then $$f(x) - B_n[f](x) \geq 0,\quad 0\leq x\leq 1.$$
Hence we focus on deriving the other bound. For function $f_\alpha(x) = x^\alpha$, Taylor’s polynomial of degree $3$ at $x = x_0$ takes the form $$Q_3(x) = \frac{\alpha(\alpha-1)}{2} x_0^{\alpha-2} (x-x_0)^2 + \frac{\alpha(\alpha-1)(\alpha-2)}{6} x_0^{\alpha-3} (x-x_0)^3 + \textrm{ affine terms of }x.$$
Applying Lemma \[lemma.qn1lower\] and Lemma \[lemma.brastaylor\], taking $x_0 =x$, we have $$\begin{aligned}
Q_3(x) - B_n[Q_3](x) & = - \frac{\alpha(\alpha-1)}{2}x^{\alpha-2} \frac{x(1-x)}{n} - \frac{\alpha(\alpha-1)(\alpha-2)}{6}x^{\alpha-3} \frac{x(1-x)}{n^2}(1-2x) \\
& = \frac{\alpha(1-\alpha)}{2n} x^{\alpha-2}(1-x) \left( x - \frac{2-\alpha}{3n} (1-2x) \right) \\
& \geq \frac{\alpha(1-\alpha)}{2n}x^{\alpha-2} (1-x) \left( x - \frac{2-\alpha}{3n} \right).\end{aligned}$$
Hence, we have $$f_\alpha(x) - B_n[f_\alpha](x) \geq Q_3(x) - B_n[Q_3](x) \geq \frac{\alpha(1-\alpha)}{2n}x^{\alpha-2} (1-x) \left( x - \frac{2-\alpha}{3n} \right).$$
Proof of Lemma \[lem\_variance\_p\_small\]
------------------------------------------
Since $nX$ is an integer, we have $(nX)^2\ge (nX)^{2\alpha},0<\alpha<1$. Hence, $$\begin{aligned}
\mathsf{Var}(X^\alpha) \le \bE X^{2\alpha} = \frac{\bE (nX)^{2\alpha}}{n^{2\alpha}} \le \frac{\bE (nX)^{2}}{n^{2\alpha}} = \frac{(np)^2+np(1-p)}{n^{2\alpha}}
\le \frac{2}{n^{2\alpha}}\wedge \frac{2p}{n^{2\alpha-1}},\end{aligned}$$ which completes the proof.
Proof of Lemma \[lem\_variance\_p\_big\]
----------------------------------------
Denoting $f(p)=p^\alpha,0<\alpha<1$, we have $$\begin{aligned}
\mathsf{Var}(f(X)) & = \bE f^2(X) - (\bE f(X))^2 \\
& = \bE f^2(X) - f^2(p) + f^2(p) - (\bE f(X))^2 \\
& \leq | \bE f^2(X) - f^2(p)|+ | f^2(p) - (\bE f(X)-f(p) + f(p))^2 |\\
& = | \bE f^2(X) - f^2(p)|+ |(\bE f(X)-f(p))^2 + 2f(p) (\bE f(X)-f(p))| \\
& \leq | \bE f^2(X) - f^2(p)| + |\bE f(X)-f(p)|^2 + 2f(p)|\bE f(X)-f(p)|. \label{eqn.varianceestimatelarge}\end{aligned}$$
Hence, it suffices to obtain bounds on $| \bE f^2(X) - f^2(p)|$ and $|\bE f(X) - f(p)|$. Denoting $r(x) = f^2(x)$, it follows from Taylor’s formula and the integral representation of the remainder term that $$r(X) = f^2(p) + r'(p)(X-p) + R_1(X;p),$$ $$R_1(X;p)=\int_p^X (X-u)r''(u)du = \frac{1}{2}r''(\eta_X)(X-p)^2, \quad \eta_X \in [\min\{X,p\},\max\{X,p\}].$$
Similarly, we have $$f(X) = f(p) + f'(p)(X-p) + R_2(X;p),$$ $$\quad R_2(X;p) = \int_p^X (X-u)f''(u)du = \frac{1}{2} f''(\nu_X)(X-p)^2,\quad \nu_X\in [\min\{X,p\},\max\{X,p\}].$$
Taking expectation on both sides with respect to $X$, where $nX \sim \mathsf{B}(n,p), p\geq \Delta$, we have $$\label{eqn.u2biaslarge}
|\bE f^2(X) - f^2(p)| = |\bE R_1(X;p) |.$$ Similarly, we have $$|\bE f(X) - f(p)| = | \bE R_2(X;p)|.$$
It is straightforward to show that $$\begin{aligned}
|r''(x)|&=2\alpha(2\alpha-1)x^{2\alpha-2} \le 2x^{2\alpha-2},\\
|f''(x)|&=\alpha(1-\alpha)x^{\alpha-2} \le \frac{1}{4}x^{\alpha-2}.\end{aligned}$$
Now we are in the position to bound $|\bE R_1(X;p)|$ and $|\bE R_2(X;p)|$. For $|\bE R_1(X;p)|$, we have $$\begin{aligned}
|\bE R_1(X;p)| & \leq \bE |R_1(X;p)| \\
& = \bE [|R_1(X;p)\mathbbm{1}(X\geq p/2)|] + \bE[R_1(X;p)\mathbbm{1}(X<p/2)] \\
& \leq \bE \left[ 2(p/2)^{2\alpha-2} (X-p)^2\right] + \bE[R_1(X;p)\mathbbm{1}(X<p/2)] \\
& \leq 8\frac{p^{2\alpha-1}}{n} + \sup_{x\leq p/2}|R_1(x;p)|\bP(nX<np/2) \\
& \leq 8\frac{p^{2\alpha-1}}{n} + \sup_{x\leq p/2}|R_1(x;p)| e^{-np/8},\end{aligned}$$ where in the last step we have used Lemma \[lem\_chernoff\]. Regarding $\sup_{x\leq p/2}|R_1(x;p)|$, for any $x\leq p/2$, we have $$\begin{aligned}
R_1(x;p) = \int_x^p (u-x)r''(u)du \leq \int_x^p (u-x) 2u^{2\alpha-2}du
\leq 2 \int_x^p u^{2\alpha-1}du
\leq 2 \int_0^p u^{2\alpha-1}du
= \frac{p^{2\alpha}}{\alpha}.\end{aligned}$$
Hence, we have $$|\bE R_1(X;p)| \leq \frac{8p^{2\alpha-1}}{n} + \frac{1}{\alpha}p^{2\alpha}e^{-np/8}.$$
Analogously, we obtain the following bound for $|\bE R_2(X;p)|$: $$|\bE R_2(X;p)|\leq \frac{p^{\alpha-1}}{n} + \frac{1}{4\alpha}p^\alpha e^{-np/8}.$$
Plugging these estimates of $|\bE R_1(X;p)|$ and $|\bE R_2(X;p)|$ into (\[eqn.varianceestimatelarge\]), we have for $p\geq 1/n$, $$\begin{aligned}
\mathsf{Var}(X^\alpha) &\leq \frac{8p^{2\alpha-1}}{n} + \frac{1}{\alpha}p^{2\alpha}e^{-np/8} + \left(\frac{p^{\alpha-1}}{n} + \frac{1}{4\alpha} p^\alpha e^{-np/8} \right)^2 +2f(p)\left(\frac{p^{\alpha-1}}{n} + \frac{1}{4\alpha}p^\alpha e^{-np/8} \right)\\
&\le \frac{8p^{2\alpha-1}}{n} + \frac{1}{\alpha}p^{2\alpha}e^{-np/8} + \frac{2p^{2(\alpha-1)}}{n^2} + \frac{1}{8\alpha^2}p^{2\alpha} e^{-np/4}+2p^\alpha\left(\frac{p^{\alpha-1}}{n} + \frac{1}{4\alpha}p^\alpha e^{-np/8} \right)\\
&\le \frac{10p^{2\alpha-1}}{n} + \frac{3}{2\alpha}p^{2\alpha}e^{-np/8} + \frac{2}{n^{2\alpha}} + \frac{1}{8\alpha^2}p^{2\alpha} e^{-np/4}\\
&\le \frac{10p^{2\alpha-1}}{n} + \frac{3}{2\alpha}\left(\frac{16\alpha}{en}\right)^{2\alpha} + \frac{2}{n^{2\alpha}} + \frac{1}{8\alpha^2}\left(\frac{8\alpha}{en}\right)^{2\alpha}.\end{aligned}$$ where we have used the following inequality in the last step: for $x\in(0,1)$ and any $c>0$, $$\begin{aligned}
x^{2\alpha}e^{-cnx} \le \left(\frac{2\alpha}{cen}\right)^{2\alpha}.\end{aligned}$$
Note that if $0<\alpha<1/2$, we can upper bound $\frac{10p^{2\alpha-1}}{n}$ by $\frac{10}{n^{2\alpha}}$, since we have constrained $p>\frac{1}{n}$.
Proof of Corollary \[cor.vargeneralf\]
--------------------------------------
In our case, apparently $F(P_n)$ is a function of $n$ independent random variables $\{Z_i\}_{1\leq i\leq n}$ taking values in $\mathcal{Z} = \{1,2,\ldots,S\}$. Changing one location of the sample would make some symbol with count $j$ to have count $j+1$, and another symbol with count $i$ to have count $i-1$. Then the total change in the functional estimator is $$f\left(\frac{j+1}{n}\right) - f \left( \frac{j}{n}\right) - f\left(\frac{i}{n}\right) + f\left(\frac{i-1}{n}\right).$$
If $f$ is monotone, then the total change would be upper bounded by $\max_{0\leq j<n} |f((j+1)/n) - f(j/n)|$. If $f$ is not monotone, the total change can be upper bounded by $2 \cdot\max_{0\leq j<n} |f((j+1)/n) - f(j/n)|$. Applying Lemma \[lemma.es\], we have the desired bounds.
[^1]: Jiantao Jiao, Kartik Venkat, and Tsachy Weissman are with the Department of Electrical Engineering, Stanford University, CA, USA. Email: {jiantao,kvenkat,tsachy}@stanford.edu
[^2]: Yanjun Han is with the Department of Electronic Engineering, Tsinghua University, Beijing, China. Email: [email protected]
[^3]: In the literature of combinatorics, the sum $\sum_{j = 0}^n a_{j,n}B_{j,n}(x)$ is called the Bernoulli sum, and various approaches have been proposed to evaluate its asymptotics [@Jacquet--Szpankowski1999entropy], [@Flajolet1999singularity], [@Cichon--Golkbiewski--Kardas--Klonowski]
[^4]: We remark that not every positive linear functional can be formulated in the form (\[eqn.generalfriesz\]). See [@Paltanea2004 Remark 1.1.3.].
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We consider nonlinear Choquard equation $$- \Delta u + V u = \bigl(I_\alpha \ast {\lvert u \rvert}^{\frac{\alpha}{N}+1}\bigr) {\lvert u \rvert}^{\frac{\alpha}{N}-1} u\quad\text{in \({{\mathbb R}}^N\)},$$ where $N \ge 3$, $V \in L^\infty ({{\mathbb R}}^N)$ is an external potential and $I_\alpha (x)$ is the Riesz potential of order $\alpha \in (0, N)$. The power $\frac{\alpha}{N}+1$ in the nonlocal part of the equation is critical with respect to the Hardy-Littlewood-Sobolev inequality. As a consequence, in the associated minimization problem a loss of compactness may occur. We prove that if $\liminf_{{\lvert x \rvert} \to \infty} \bigl(1 - V (x)\bigr){\lvert x \rvert}^2 > \frac{N^2 (N - 2)}{4 (N + 1)}$ then the equation has a nontrivial solution. We also discuss some necessary conditions for the existence of a solution. Our considerations are based on a concentration compactness argument and a nonlocal version of Brezis-Lieb lemma.'
address:
- |
Swansea University\
Department of Mathematics\
Singleton Park\
Swansea\
SA2 8PP\
Wales, United Kingdom
- |
Université Catholique de Louvain\
Institut de Recherche en Mathématique et Physique\
Chemin du Cyclotron 2 bte L7.01.01\
1348 Louvain-la-Neuve\
Belgium
author:
- Vitaly Moroz
- Jean Van Schaftingen
title: |
Groundstates of nonlinear Choquard equations:\
Hardy-Littlewood-Sobolev critical exponent
---
Introduction and results
========================
We consider a nonlinear Choquard type equation $$\tag{$\mathcal{P}$}
\label{equationNLChoquard-p}
- \Delta u + V u = \bigl(I_\alpha \ast {\lvert u \rvert}^p\bigr) {\lvert u \rvert}^{p-2} u\quad\text{in \({{\mathbb R}}^N\)},$$ where $N \in {{\mathbb N}}$, $\alpha\in(0,N)$, $p>1$, $I_\alpha : {{\mathbb R}}^N\setminus\{0\} \to {{\mathbb R}}$ is the Riesz potential of order $\alpha \in (0, N)$ defined for every $x \in {{\mathbb R}}^N \setminus \{0\}$ by $$I_\alpha(x)=\frac{\Gamma(\tfrac{N-\alpha}{2})}
{2^{\alpha}\pi^{N/2}\Gamma(\tfrac{\alpha}{2})\, {\lvert x \rvert}^{N-\alpha}},$$ and $V \in L^\infty ({{\mathbb R}}^N)$ is an external potential.
For $N = 3$, $\alpha = 2$ and $p = 2$ equation is the *Choquard-Pekar equation* which goes back to the 1954’s work by S.I.Pekar on quantum theory of a Polaron at rest and to 1976’s model of P.Choquard of an electron trapped in its own hole, in an approximation to Hartree-Fock theory of one-component plasma [@Lieb1977]. In the 1990’s the same equation reemerged as a model of self-gravitating matter and is known in that context as the *Schrödinger-Newton equation*.
Mathematically, the existence and qualitative properties of solutions of Choquard equation have been studied for a few decades by variational methods, see for earlier and for recent work on the problem and further references therein.
The following sharp characterisation of the existence and nonexistence of nontrivial solutions of in the case of constant potential $V$ can be found in [@MorozVanSchaftingen2013JFA].
\[Ground states of with constant potential [@MorozVanSchaftingen2013JFA]\*[theorems 1 and 2]{}\] Assume that $V\equiv 1$. Then has a nontrivial solution $u \in H^1({{\mathbb R}}^N)\cap L^\frac{2Np}{N-\alpha}({{\mathbb R}}^N)$ with $\nabla u\in H^1_{\mathrm{loc}}({{\mathbb R}}^N)\cap L^\frac{2Np}{N-\alpha}_{\mathrm{loc}} ({{\mathbb R}}^N)$ if and only if $p\in\big(\frac{\alpha}{N}+1,\frac{N + \alpha}{(N - 2)_+}\big)$.
If $p\in\big[\frac{\alpha}{N}+1,\frac{N + \alpha}{(N - 2)_+}\big]$ then $H^1({{\mathbb R}}^N)\subset L^\frac{2Np}{N-\alpha}({{\mathbb R}}^N)$ by the Sobolev inequality, and moreover, every $H^1$–solution of belongs to $W^{2,p}_{\mathrm{loc}}({{\mathbb R}}^N)$ for any $p\ge 1$ by a regularity result in [@MorozVanSchaftingenBL]\*[proposition 3.1]{}. This implies that the Choquard equation with a positive constant potential has no $H^1$–solutions at the end-points of the above existence interval.
In this note we are interested in the existence and nonexistence of solutions to with nonconstant potential $V$ at the [*lower critical exponent*]{} $p=\frac{\alpha}{N}+1$, that is, we consider the problem $$\tag{$\mathcal{P_*}$}
\label{equationNLChoquard}
- \Delta u + V u = \bigl(I_\alpha \ast {\lvert u \rvert}^{\frac{\alpha}{N}+1}\bigr) {\lvert u \rvert}^{\frac{\alpha}{N}-1} u\quad\text{in \({{\mathbb R}}^N\)}.$$ The exponent $\frac{\alpha}{N}+1$ is critical with respect to the Hardy-Littlewood-Sobolev inequality, which we recall here in a form of minimization problem $$c_\infty = \inf \Bigl\{ \int_{{{\mathbb R}}^N} {\lvert u \rvert}^2 {\;\vert\;}u \in L^2 ({{\mathbb R}}^N) \text{ and }\int_{{{\mathbb R}}^N} \bigl(I_\alpha \ast {\lvert u \rvert}^{\frac{\alpha}{N} + 1}\bigr) {\lvert u \rvert}^{\frac{\alpha}{N} + 1} = 1\Bigr\} > 0.$$
\[theoremHLS\] The infimum $c_\infty$ is achieved if and only if $$\label{HLSoptimal}
u(x) = C\left(\frac{\lambda}{\lambda^2 + {\lvert x - a \rvert}^2}\right)^{N/2},$$ where $C > 0$ is a fixed constant, $a \in {{\mathbb R}}^N$ and $\lambda \in (0, \infty)$ are parameters.
The form of minimizers in theorem \[theoremHLS\] suggests that a loss of compactness in may occur by translations and dilations.
In order to characterise the existence of nontrivial solutions for the lower critical Choquard equation we define the critical level $$c_* = \inf \Bigl\{ \int_{{{\mathbb R}}^N} {\lvert \nabla u \rvert}^2 + V {\lvert u \rvert}^2 {\;\vert\;}u \in H^1 ({{\mathbb R}}^N) \text{ and } \int_{{{\mathbb R}}^N} \bigl(I_\alpha \ast {\lvert u \rvert}^{\frac{\alpha}{N} + 1}\bigr) {\lvert u \rvert}^{\frac{\alpha}{N} + 1} = 1\Bigr\}.$$ It can be checked directly that if $u\in H^1({{\mathbb R}}^N)$ achieves the infimum $c_*$, then a multiple of the minimizer $u$ is a weak solution of Choquard equation .
Using a Brezis-Lieb type lemma for Riesz potentials [@MorozVanSchaftingen2013JFA]\*[lemma 2.4]{} and a concentration compactness argument (lemma \[lemmaWeakHilbert\]), we establish our main abstract result.
\[theoremExistence\] Assume that $V \in L^\infty ({{\mathbb R}}^N)$ and $$\label{eqliminfV}
\liminf_{{\lvert x \rvert} \to \infty} V (x) \ge 1.$$ If $c_* < c_\infty$ then the infimum $c_*$ is achieved and every minimizing sequence for $c_*$ up to a subsequence converges strongly in $H^1({{\mathbb R}}^N)$.
The inequality for the existence of minimizers is sharp, as shown by the following lemma for constant potentials.
\[lemmaV1Sharp\] If $V\equiv 1$, then $c_*=c_\infty$.
Since problem with $V\equiv 1$ has no $H^1$–solutions, this shows that the strict inequality $c_*<c_\infty$ is indeed essential for the existence of a minimizer for $c_*$.
In fact, the strict inequality $c_*<c_\infty$ is necessary at least for the strong convergence of [*all*]{} minimizing sequences.
\[propositionSharp\] Let $V \in L^\infty ({{\mathbb R}}^N)$. If $$\limsup_{{\lvert x \rvert} \to \infty} V (x) \le 1,$$ then $$c_* \le c_\infty.$$ In addition, if $$c_*=c_\infty,$$ then there exists a minimizing sequence for $c_*$ which converges weakly to $0$ in $H^1({{\mathbb R}}^N)$.
Using Hardy-Littlewood-Sobolev minimizers as a family of test functions for $c_*$, we establish a sufficient condition for the strict inequality $c_*<c_\infty$.
\[theoremSufficient-Hardy\] Let $V \in L^\infty ({{\mathbb R}}^N)$. If $$\label{eq:sufficient-Hardy}
\liminf_{{\lvert x \rvert} \to \infty} \bigl(1 - V (x)\bigr){\lvert x \rvert}^2 > \frac{N^2 (N - 2)_+}{4 (N + 1)},$$ then $c_*<c_\infty$ and hence the infimum $c_*$ is achieved.
In particular, if $N=1,2$ then condition reduces to $$\liminf_{{\lvert x \rvert} \to \infty} \bigl(1 - V (x)\bigr){\lvert x \rvert}^2 > 0,$$ that is, the potential $1-V$ should not decay to zero at infinity faster then the inverse square of $|x|$.
Employing a version of Pohožaev identity for Choquard equation (see proposition \[propositionPohozaev\] below), we show that a certain control on the potential $V$ is indeed necessary for the strict inequality $c_*<c_\infty$.
\[propositionPohozaev1\] Let $V \in C^1 ({{\mathbb R}}^N) \cap L^\infty ({{\mathbb R}}^N)$. If $$\label{equationSufficient1}
\sup \Bigl\{\int_{{{\mathbb R}}^N} \frac{1}{2}{( \nabla V (x) \vert x)} {\lvert \varphi (x) \rvert}^2{\,\mathrm{d}}x {\;\vert\;}\varphi \in C^1_c ({{\mathbb R}}^N) \text{ and } \int_{{{\mathbb R}}^N} {\lvert \nabla \varphi \rvert}^2 \le 1\Bigr\} < 1,$$ then Choquard equation does not have a nonzero solution $u\in H^1({{\mathbb R}}^N)\cap W^{2, 2}_\mathrm{loc}({{\mathbb R}}^N)$.
In particular, combining with Hardy’s inequality on ${{\mathbb R}}^N$, we obtain a simple nonexistence criterion.
\[propositionPohozaev2\] Let $N\ge 3$ and $V \in C^1 ({{\mathbb R}}^N) \cap L^\infty ({{\mathbb R}}^N)$. If for every $x \in {{\mathbb R}}^N$, $$\label{eq:PohozaevHardy}
\sup_{x \in {{\mathbb R}}^N} {\lvert x \rvert}^2 {( \nabla V (x) \vert x)} < \frac{(N - 2)^2}{2},$$ then Choquard equation does not have a nonzero solution $u\in H^1({{\mathbb R}}^N)\cap W^{2, 2}_\mathrm{loc}({{\mathbb R}}^N)$.
For example, for $N\ge 3$ and $\mu>0$, we consider a model equation $$\label{eq:model}
- \Delta u + \Big(1- \frac{\mu}{1+{\lvert x \rvert}^2}\Big) u = \bigl(I_\alpha \ast {\lvert u \rvert}^{\frac{\alpha}{N}+1}\bigr) {\lvert u \rvert}^{\frac{\alpha}{N}-1} u\quad\text{in \({{\mathbb R}}^N\)}.$$ Then proposition \[propositionPohozaev2\] implies that has no nontrivial solutions for $\mu < \frac{(N-2)^2}{4}$, while for $\mu >\frac{N^2 (N - 2)}{4 (N + 1)}$ assumption is satisfied and hence admits a groundstate. We note that $$\frac{\frac{(N - 2)^2}{4}}{\frac{N^2 (N - 2)}{4 (N + 1)}} = 1 - \frac{N - 2}{N^2},$$ so that the two bounds are asymptotically sharp when $N \to \infty$. We leave as an open question whether admits a ground state for $\mu \in \big[\frac{(N-2)^2}{4},\frac{N^2 (N - 2)}{4 (N + 1)}\big]$.
We emphasise that unlike the asymptotic sufficient existence condition , nonexistence condition is a global condition on the whole of ${{\mathbb R}}^N$. For example, a direct computation shows that for $a=0$ and every $\lambda>0$, a multiple of the Hardy-Littlewood-Sobolev minimizer solves the equation $$\label{eq:null}
- \Delta u + \Big(1+ \frac{N(2{\lvert x \rvert}^2-N\lambda^2)}{({\lvert x \rvert}^2+\lambda^2)^2}\Big) u = \bigl(I_\alpha \ast {\lvert u \rvert}^{\frac{\alpha}{N}+1}\bigr) {\lvert u \rvert}^{\frac{\alpha}{N}-1} u\quad\text{in \({{\mathbb R}}^N\)}.$$ Here fails on an annulus centered at the origin, while $V(x)>1$ and ${( \nabla V (x) \vert x)}<0$ for all ${\lvert x \rvert}$ sufficiently large. Moreover, $$\lim_{{\lvert x \rvert} \to \infty} (1 - V (x)) {\lvert x \rvert}^2 = - 2 N < 0 \le \frac{N^2 (N - 2)_+}{4 (N + 1)}.$$ Note that the constructed solution $u_\lambda$ satisfies $$\int_{{{\mathbb R}}^N} {\lvert \nabla u_\lambda \rvert}^2 + V {\lvert u_\lambda \rvert}^2=0.$$ In particular, we are unable to conclude that $c_*<c_\infty$. We do not know whether $u_\lambda$ is a groundstate of . However, if $u_\lambda$ was not a groundstate, then we would have $c_* < c_\infty$ and would then have a groundstate by theorem \[theoremExistence\].
Existence of minimizers under strict inequality: proof of theorem \[theoremExistence\]
======================================================================================
In order to prove theorem \[theoremExistence\] we will use a special case of the classical Brezis-Lieb lemma [@BrezisLieb1983] for Riesz potentials.
\[lemmaBrezisLiebriesz\] Let $N \in {{\mathbb N}}$, $\alpha \in (0, N)$, and $(u_n)_{n \in {{\mathbb N}}}$ be a bounded sequence in $L^2 ({{\mathbb R}}^N)$. If $u_n \to u$ almost everywhere on ${{\mathbb R}}^N$ as $n \to \infty$, then $$\begin{gathered}
\int_{{{\mathbb R}}^N} \bigl(I_\alpha \ast {\lvert u \rvert}^{\frac{\alpha}{N} + 1}\bigr) {\lvert u \rvert}^{\frac{\alpha}{N} + 1} = \lim_{n \to \infty} \int_{{{\mathbb R}}^N} \bigl(I_\alpha \ast {\lvert u_n \rvert}^{\frac{\alpha}{N} + 1}\bigr) {\lvert u_n \rvert}^{\frac{\alpha}{N} + 1}\\
- \int_{{{\mathbb R}}^N} \bigl(I_\alpha \ast {\lvert u_n - u \rvert}^{\frac{\alpha}{N} + 1}\bigr) {\lvert u_n - u \rvert}^{\frac{\alpha}{N} + 1}.\end{gathered}$$
Our second result is a concentration type lemma.
\[lemmaWeakHilbert\] Assume that $V \in L^\infty ({{\mathbb R}}^N)$ and $\liminf_{{\lvert x \rvert} \to \infty} V (x) \ge 1$. If the sequence $(u_n)_{n \in {{\mathbb N}}}$ is bounded in $L^2 ({{\mathbb R}}^N)$ and converges in $L^2_\mathrm{loc} ({{\mathbb R}}^N)$ to $u$ as $n \to \infty$, then $$\int_{{{\mathbb R}}^N} V {\lvert u \rvert}^2
\le \liminf_{n \to \infty} \int_{{{\mathbb R}}^N} V {\lvert u_n \rvert}^2 - \int_{{{\mathbb R}}^N} {\lvert u_n - u \rvert}^2.$$
Since the sequence $(u_n)_{n \in {{\mathbb N}}}$ is bounded in $L^2 ({{\mathbb R}}^N)$ and converges in measure to $u$, we deduce by the Brezis-Lieb lemma [@BrezisLieb1983] (see also [@LiebLoss2001]\*[theorem 1.9]{}) that $$\int_{{{\mathbb R}}^N} V {\lvert u \rvert}^2
= \lim_{n \to \infty} \int_{{{\mathbb R}}^N} V {\lvert u_n \rvert}^2 - \int_{{{\mathbb R}}^N} V {\lvert u_n - u \rvert}^2.$$ Now, we observe that for every $R > 0$ and every $n \in {{\mathbb N}}$, $$\int_{{{\mathbb R}}^N} (1 - V) {\lvert u_n - u \rvert}^2
\le \int_{B_R} (1 - V) {\lvert u_n - u \rvert}^2 + (1 - \inf_{{{\mathbb R}}^N \setminus B_R} V)_+ \int_{{{\mathbb R}}^N} {\lvert u_n - u \rvert}^2.$$ By the local $L^2_\mathrm{loc} ({{\mathbb R}}^N)$ convergence, we note that $$\lim_{n \to \infty} \int_{B_R} (1 - V) {\lvert u_n - u \rvert}^2 = 0.$$ Since $\lim_{R \to \infty} (1 - \inf_{{{\mathbb R}}^N \setminus B_R} v)_+ = 0$ and $(u_n - u)_{n \in {{\mathbb N}}}$ is bounded in $L^2 ({{\mathbb R}}^N)$, we conclude that $$\limsup_{n \to \infty} \int_{{{\mathbb R}}^N} (1 - V) {\lvert u_n - u \rvert}^2 \le 0;$$ the conclusion follows.
Let $(u_n)_{n \in {{\mathbb N}}}\subset H^1 ({{\mathbb R}}^N)$ be a minimizing sequence for $c_*$, that is $$\int_{{{\mathbb R}}^N} \bigl(I_\alpha \ast {\lvert u_n \rvert}^{\frac{\alpha}{N} + 1}\bigr){\lvert u_n \rvert}^{\frac{\alpha}{N} + 1} = 1$$ and $$\lim_{n \to \infty} \int_{{{\mathbb R}}^N} {\lvert \nabla u_n \rvert}^2 + V {\lvert u_n \rvert}^2 \to c_*.$$ In view of our assumption we observe that the sequence $(u_n)_{n \in {{\mathbb N}}}$ is bounded in $H^1 ({{\mathbb R}}^N)$. So, there exists $u \in H^1 ({{\mathbb R}}^N)$ such that, up to a subsequence, the sequence $(u_n)_{n \in {{\mathbb N}}}$ converges to $u$ weakly in $H^1 ({{\mathbb R}}^N)$ and, by the classical Rellich-Kondrachov compactness theorem, strongly in $L^2_\mathrm{loc} ({{\mathbb R}}^N)$. By the lower semi-continuity of the norm under weak convergence, $$\int_{{{\mathbb R}}^N} {\lvert \nabla u \rvert}^2 + V {\lvert u \rvert}^2\le \lim_{n \to \infty} \int_{{{\mathbb R}}^N} {\lvert \nabla u_n \rvert}^2 + V {\lvert u_n \rvert}^2 = c_*.$$ and by Fatou’s lemma $$\int_{{{\mathbb R}}^N} \bigl(I_\alpha \ast {\lvert u_n \rvert}^{\frac{\alpha}{N} + 1}\bigr){\lvert u_n \rvert}^{\frac{\alpha}{N} + 1} \le 1.$$ In order to conclude, it suffices to prove that equality is achieved in the latter inequality.
We observe that by lemma \[lemmaBrezisLiebriesz\], $$\lim_{n \to \infty} \int_{{{\mathbb R}}^N} \bigl(I_\alpha \ast {\lvert u_n - u \rvert}^{\frac{\alpha}{N} + 1}\bigr){\lvert u_n - u \rvert}^{\frac{\alpha}{N} + 1}
= 1 - \int_{{{\mathbb R}}^N} \bigl(I_\alpha \ast {\lvert u \rvert}^{\frac{\alpha}{N} + 1}\bigr){\lvert u \rvert}^{\frac{\alpha}{N} + 1}$$ while by lemma \[lemmaWeakHilbert\] and by the lower-semicontinuity of the norm under weak convergence, $$\begin{split}\label{eqBL-2}
\int_{{{\mathbb R}}^N} {\lvert \nabla u \rvert}^2 + V {\lvert u \rvert}^2
&\le\liminf_{n \to \infty} \int_{{{\mathbb R}}^N} {\lvert \nabla u_n \rvert}^2 +
\liminf_{n \to \infty} \int_{{{\mathbb R}}^N} V {\lvert u_n \rvert}^2 - {\lvert u_n - u \rvert}^2\\
&\le \liminf_{n \to \infty} \int_{{{\mathbb R}}^N} {\lvert \nabla u_n \rvert}^2 + V {\lvert u_n \rvert}^2 - {\lvert u_n - u \rvert}^2\\
&= c_* - \limsup_{n \to \infty} \int_{{{\mathbb R}}^N}{\lvert u_n - u \rvert}^2.
\end{split}$$ By definition of $c_\infty$, we have $$\int_{{{\mathbb R}}^N} {\lvert u_n - u \rvert}^2
\ge c_\infty \Bigl(\int_{{{\mathbb R}}^N} \bigl(I_\alpha \ast {\lvert u_n - u \rvert}^{\frac{\alpha}{N} + 1}\bigr){\lvert u_n - u \rvert}^{\frac{\alpha}{N} + 1} \Bigr)^\frac{N}{N + \alpha}.$$ Therefore, we conclude that $$\int_{{{\mathbb R}}^N} {\lvert \nabla u \rvert}^2 + V {\lvert u \rvert}^2
\le c_* - c_\infty\Bigl( 1 - \int_{{{\mathbb R}}^N} \bigl(I_\alpha \ast {\lvert u \rvert}^{\frac{\alpha}{N} + 1}\bigr){\lvert u \rvert}^{\frac{\alpha}{N} + 1}\Bigr)^\frac{N}{N + \alpha}.$$ In view of the definition of $c_*$ this implies that $$c_* \ge c_\infty \Bigl( 1 - \int_{{{\mathbb R}}^N} \bigl(I_\alpha \ast {\lvert u \rvert}^{\frac{\alpha}{N} + 1}\bigr){\lvert u \rvert}^{\frac{\alpha}{N} + 1}\Bigr)^\frac{N}{N + \alpha} + c_* \Bigl( \int_{{{\mathbb R}}^N} \bigl(I_\alpha \ast {\lvert u \rvert}^{\frac{\alpha}{N} + 1}\bigr){\lvert u \rvert}^{\frac{\alpha}{N} + 1}\Bigr)^\frac{N}{N + \alpha}.$$ Since by assumption $c_* < c_\infty$, we conclude that $$\int_{{{\mathbb R}}^N} \bigl(I_\alpha \ast {\lvert u \rvert}^{\frac{\alpha}{N} + 1}\bigr){\lvert u \rvert}^{\frac{\alpha}{N} + 1}=1,$$ and hence, by definition of $c_*$, $$\int_{{{\mathbb R}}^N} {\lvert \nabla u \rvert}^2 + V {\lvert u \rvert}^2 = c_*,$$ that is the infimum $c_*$ is achieved at $u$. Moreover, from we conclude that $u_n\to u$ in $L^2({{\mathbb R}}^N)$. Since $V\in L^\infty({{\mathbb R}}^N)$, this implies that $Vu_n\to Vu$ in $L^2({{\mathbb R}}^N)$. Using again, we conclude that $$\int_{{{\mathbb R}}^N} {\lvert \nabla u \rvert}^2=\lim_{n \to \infty} \int_{{{\mathbb R}}^N} {\lvert \nabla u_n \rvert}^2.$$ Since $(u_n)_{n \in {{\mathbb N}}}$ converges to $u$ weakly in $H^1 ({{\mathbb R}}^N)$, this implies that $(u_n)_{n \in {{\mathbb N}}}$ also converges to $u$ strongly in $H^1 ({{\mathbb R}}^N)$.
Optimality of the strict inequality
===================================
In this section we prove lemma \[lemmaV1Sharp\] and proposition \[propositionSharp\].
Let us denote by $\Tilde{c}_\infty$ the infimum on the right-hand side. By density of the space $H^1 ({{\mathbb R}}^N)$ in $L^2 ({{\mathbb R}}^N)$ and by continuity in $L^2$ of the integral functionals involved in the definition of $c_\infty$, it is clear that $\Tilde{c}_\infty \ge c_\infty$. We choose now $u \in H^1 ({{\mathbb R}}^N)$ and define for $\lambda > 0$ the function $u_\lambda \in H^1 ({{\mathbb R}}^N)$ for every $x \in {{\mathbb R}}^N$ by $$u_\lambda (x) = \lambda^{N/2}u (\lambda x).$$ We compute for every $\lambda > 0$ that $$\int_{{{\mathbb R}}^N} \bigl(I_\alpha \ast {\lvert u_\lambda \rvert}^{\frac{\alpha}{N} + 1}\bigr) {\lvert u_\lambda \rvert}^{\frac{\alpha}{N} + 1}=\int_{{{\mathbb R}}^N} \bigl(I_\alpha \ast {\lvert u \rvert}^{\frac{\alpha}{N} + 1}\bigr) {\lvert u \rvert}^{\frac{\alpha}{N} + 1}$$ and $$\int_{{{\mathbb R}}^N} {\lvert \nabla u_\lambda \rvert}^2 + {\lvert u_\lambda \rvert}^2 = \frac{1}{\lambda^2} \int_{{{\mathbb R}}^N} {\lvert \nabla u \rvert}^2 + \int_{{{\mathbb R}}^N} {\lvert u \rvert}^2.$$ Hence, $$\inf_{\lambda > 0} \int_{{{\mathbb R}}^N} {\lvert \nabla u_\lambda \rvert}^2 + {\lvert u_\lambda \rvert}^2 = \int_{{{\mathbb R}}^N} {\lvert u \rvert}^2,$$ and we conclude that $\Tilde{c}_\infty \le c_\infty$.
For $\lambda>0$, let $$u_\lambda (x) = C\left(\frac{\lambda}{\lambda^2 + {\lvert x \rvert}^2}\right)^\frac{N}{2} = \lambda^{-\frac{N}{2}}u_1\Big(\frac{x}{\lambda}\Big)$$ be a family of minimizers for $c_\infty$ given in . We observe that $$\int_{{{\mathbb R}}^N} \bigl(I_\alpha \ast {\lvert u_\lambda \rvert}^{\frac{\alpha}{N} + 1}\bigr) {\lvert u_\lambda \rvert}^{\frac{\alpha}{N} + 1}= 1,$$ whereas by a change of variables, $$\int_{{{\mathbb R}}^N} {\lvert \nabla u_\lambda \rvert}^2 + V {\lvert u_\lambda \rvert}^2
= \frac{1}{\lambda^2} \int_{{{\mathbb R}}^N} {\lvert \nabla u_1 \rvert}^2
+ \int_{{{\mathbb R}}^N} V \Big(\frac{y}{\lambda}\Big) \frac{C^2}{1 + {\lvert y \rvert}^2}{\,\mathrm{d}}y.$$ By Lebesgue’s dominated convergence theorem $$\limsup_{\lambda \to 0} \int_{{{\mathbb R}}^N} V \Big(\frac{y}{\lambda}\Big) \frac{C^2}{1 + {\lvert y \rvert}^2}{\,\mathrm{d}}y
\le \int_{{{\mathbb R}}^N} \frac{C^2}{1 + {\lvert y \rvert}^2}{\,\mathrm{d}}y = c_\infty,$$ so we conclude that $c_*\le c_\infty$. If, in addition, $c_*=c_\infty$ then for any $\lambda_n\to 0$, $(u_{\lambda_n})_{n\in{{\mathbb N}}}$ is a minimizing sequence for $c_*$, and the conclusion follows.
Sufficient conditions for the strict inequality: proof of theorem \[theoremSufficient-Hardy\]
=============================================================================================
For $a\in{{\mathbb R}}^N$ and $\lambda>0$, let $$u_\lambda (x) = C\left(\frac{\lambda}{\lambda^2 + {\lvert x-a \rvert}^2}\right)^{N/2}$$ be a family of minimizers for $c_\infty$ as in . Then $$\int_{{{\mathbb R}}^N} {\lvert \nabla u_\lambda \rvert}^2 + V {\lvert u_\lambda \rvert}^2
= c_\infty + \int_{{{\mathbb R}}^N} {\lvert \nabla u_\lambda \rvert}^2 + \int_{{{\mathbb R}}^N} (V - 1) {\lvert u_\lambda \rvert}^2.$$ Denote $$\mathcal I_V(a,\lambda):=\lambda^2\int_{{{\mathbb R}}^N} {\lvert \nabla u_\lambda \rvert}^2 + \lambda^2\int_{{{\mathbb R}}^N} (V - 1) {\lvert u_\lambda \rvert}^2 < 0.$$ To obtain a sufficient conditions for $c_*<c_\infty$ it is enough to show that for some $a\in{{\mathbb R}}^N$, $$\label{eq:presufficient}
\inf_{\lambda>0}\mathcal{I}_V (a, \lambda)<0,$$
If $N \le 2$, then by there exists $\mu > 0$ such that $$\liminf_{{\lvert x \rvert} \to \infty} (1 - V (x)){\lvert x \rvert}^2 \ge \mu.$$ Therefore $$\lim_{\lambda \to \infty} \lambda^2 \int_{{{\mathbb R}}^N} (1 - V) {\lvert u_\lambda \rvert}^2
= \lim_{\lambda \to \infty} \int_{{{\mathbb R}}^N} \frac{\lambda^2(1 - V (\lambda x))}{(1 + {\lvert x \rvert}^2)^N} {\,\mathrm{d}}x \ge \int_{{{\mathbb R}}^N} \frac{\mu}{{\lvert x \rvert}^2 (1 + {\lvert x \rvert}^2)^N} {\,\mathrm{d}}x = \infty.$$ Since for every $\lambda > 0$, $$\lambda^2 \int_{{{\mathbb R}}^N} {\lvert \nabla u_\lambda \rvert}^2 = \int_{{{\mathbb R}}^N} {\lvert \nabla u_1 \rvert}^2 < \infty,$$ the condition is satisfied.
If $N\ge 3$, we observe that for every $\lambda > 0$, $$\int_{{{\mathbb R}}^N} {\lvert \nabla u_\lambda \rvert}^2 = \frac{N^2 (N - 2)}{4 (N + 1)} \int_{{{\mathbb R}}^N} \frac{{\lvert u_\lambda (x) \rvert}^2}{{\lvert x \rvert}^2}{\,\mathrm{d}}x.$$ This follows from the fact that $$\int_{{{\mathbb R}}^N} \frac{{\lvert x \rvert}^2}{(1 + {\lvert x \rvert}^2)^{N + 2}}{\,\mathrm{d}}x = \frac{N - 2}{4 (N + 1)} \int_{{{\mathbb R}}^N} \frac{1}{{\lvert x \rvert}^2 (1 + {\lvert x \rvert}^2)^{N}}{\,\mathrm{d}}x,$$ which can be proved by two successive integrations by parts. Then, after a transformation $x=\lambda y+a$, $$\mathcal I_V(a,\lambda)=
\int_{{{\mathbb R}}^N} \Big(\frac{\tfrac{N^2 (N - 2)}{4 (N + 1)}}{{\lvert y \rvert}^2} - \lambda^2 (1 - V (a+\lambda y)) \Big)\frac{C^2}{(1 + {\lvert y \rvert}^2)^N} {\,\mathrm{d}}y,$$ and in view of , sufficient condition is is satisfied for $a=0$, so we conclude that $c_*<c_\infty$.
Note that if the function $\lambda \mapsto \lambda^2(1 - V (a+\lambda y))$ is nondecreasing for every $y\in{{\mathbb R}}^N$, then $\lambda \mapsto \mathcal{I}_V (a, \lambda)$ is nonincreasing. Therefore $\mathcal{I}_V (a, \lambda)$ admits negative values if and only if it has a negative limit as $\lambda\to\infty$. The latter is ensured in theorem \[theoremSufficient-Hardy\] via asymptotic condition . This explains that if the function $\lambda \mapsto \lambda^2(1 - V (a+\lambda y))$ is nondecreasing, like for instance, in the special case $$V(x)=1- \frac{\mu}{1+{\lvert x \rvert}^2},$$ then integral sufficient condition is in fact equivalent to the asymptotic sufficient condition .
Pohožaev identity and necessary conditions for the existence
============================================================
We establish a Pohožaev type identity, which extends the identities obtained previously for constant potentials $V$ .
\[propositionPohozaev\] Let $N\ge 3$ and $V \in C^1 ({{\mathbb R}}^N) \cap L^\infty ({{\mathbb R}}^N)$ and $u \in W^{1, 2} ({{\mathbb R}}^N)$. If $$\sup_{x \in {{\mathbb R}}^N} {\lvert {( \nabla V (x) \vert )}x \rvert} < \infty,$$ and $u \in W^{2, 2}_\mathrm{loc}({{\mathbb R}}^N)$ satisfies Choquard equation then $$\int_{{{\mathbb R}}^N} {\lvert \nabla u \rvert}^2 = \frac{1}{2} \int_{{{\mathbb R}}^N} {( \nabla V (x) \vert x)}\, {\lvert u (x) \rvert}^2{\,\mathrm{d}}x.$$
We fix a cut-off function $\varphi \in C^1_c({{\mathbb R}}^N)$ such that $\varphi = 1$ on $B_1$ and we test for $\lambda \in (0, \infty)$ the equation against the function $v_\lambda \in W^{1, 2} ({{\mathbb R}}^N)$ defined for every $x \in {{\mathbb R}}^N$ by $$v_\lambda (x) = \varphi(\lambda x) {( \nabla u (x) \vert x)}$$ to obtain the identity $$\int_{{{\mathbb R}}^N} {( \nabla u \vert \nabla v_\lambda)} + \int_{{{\mathbb R}}^N} V u v_\lambda
= \int_{{{\mathbb R}}^N} (I_\alpha \ast {\lvert u \rvert}^{\frac{\alpha}{N} + 1}) {\lvert u \rvert}^{\frac{\alpha}{N} - 1} u v_\lambda.$$ We compute for every $\lambda > 0$, by definition of $v_\lambda$, the chain rule and by the Gauss integral formula, $$\begin{split}
\int_{{{\mathbb R}}^N} V u v_\lambda &= \int_{{{\mathbb R}}^N} V (x) u(x) \varphi (\lambda x){( x \vert \nabla u(x))} {\,\mathrm{d}}x\\
&= \int_{{{\mathbb R}}^N} V (x) \varphi (\lambda x){( x \vert \nabla \bigl(\tfrac{{\lvert u \rvert}^2}{2}\bigr) (x))} {\,\mathrm{d}}x\\
&= - \int_{{{\mathbb R}}^N} \bigl( (N V (x) + {( \nabla V (x) \vert x)}) \varphi (\lambda x)
+ V (x) \lambda {( x \vert \nabla \varphi (\lambda x))} \bigr)
\frac{{\lvert u(x) \rvert}^2}{2}
{\,\mathrm{d}}x.
\end{split}$$ Since $\sup_{x \in {{\mathbb R}}^N} {( \nabla V (x) \vert x)} < \infty$, by Lebesgue’s dominated convergence theorem it holds $$\lim_{\lambda \to 0} \int_{{{\mathbb R}}^N} V u v_\lambda = - \frac{N}{2} \int_{{{\mathbb R}}^N} V {\lvert u \rvert}^2 - \frac{1}{2} \int_{{{\mathbb R}}^N} {( \nabla V (x) \vert x)} {\lvert u \rvert}^2.$$ By Lebesgue’s dominated convergence again, since $u \in W^{1, 2} ({{\mathbb R}}^N)$, we have (see [@MorozVanSchaftingen2013JFA]\*[proof of proposition 3.1]{} for the details) $$\lim_{\lambda \to 0} \int_{{{\mathbb R}}^N} {( \nabla u \vert \nabla v_\lambda)} = - \frac{N - 2}{2} \int_{{{\mathbb R}}^N} {\lvert \nabla u \rvert}^2.$$ and $$\lim_{\lambda \to 0} \int_{{{\mathbb R}}^N} (I_\alpha \ast {\lvert u \rvert}^{\frac{\alpha}{N} + 1}) {\lvert u \rvert}^{\frac{\alpha}{N} - 1} u\, v_\lambda
= - \frac{N}{2} \int_{{{\mathbb R}}^N} (I_\alpha \ast {\lvert u \rvert}^{\frac{\alpha}{N} + 1}) {\lvert u \rvert}^{\frac{\alpha}{N} + 1}.$$ We have thus proved the Pohožaev type identity $$\begin{gathered}
\label{eqPohozhaev}
\frac{N - 2}{2} \int_{{{\mathbb R}}^N} {\lvert \nabla u \rvert}^2 + \frac{N}{2} \int_{{{\mathbb R}}^N} V {\lvert u \rvert}^2
+ \frac{1}{2} \int_{{{\mathbb R}}^N} {( \nabla V (x) \vert x)} \, {\lvert u (x) \rvert}^2 {\,\mathrm{d}}x\\ = \frac{N}{2} \int_{{{\mathbb R}}^N} (I_\alpha \ast {\lvert u \rvert}^{\frac{\alpha}{N} + 1}) {\lvert u \rvert}^{\frac{\alpha}{N} + 1}.\end{gathered}$$ If we test the equation against $u$, we obtain the identity $$\int_{{{\mathbb R}}^N} {\lvert \nabla u \rvert}^2 + \int_{{{\mathbb R}}^N} V {\lvert u \rvert}^2 = \int_{{{\mathbb R}}^N} (I_\alpha \ast {\lvert u \rvert}^{\frac{\alpha}{N} + 1}) {\lvert u \rvert}^{\frac{\alpha}{N} + 1};$$ the combination of those two identities yields the conclusion.
Proposition \[propositionPohozaev1\] is a direct consequence of proposition \[propositionPohozaev\], while proposition \[propositionPohozaev2\] follows from proposition \[propositionPohozaev\] and the classical optimal Hardy inequality on ${{\mathbb R}}^N$, $$\frac{(N - 2)^2}{4}\int_{{{\mathbb R}}^N} \frac{{\lvert u(x) \rvert}^2}{{\lvert x \rvert}^2}{\,\mathrm{d}}x
\le \int_{{{\mathbb R}}^N} {\lvert \nabla u \rvert}^2$$ which is valid for all $u\in H^1({{\mathbb R}}^N)$ (see for example [@Willem2013]\*[theorem 6.4.10 and exercise 6.8]{}).
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'I discuss the problem of phase separation in cuprates from the point of view of the Landau theory of Fermi liquids. I calculate the rate of growth of unstable regions for the hydrodymanics and collisionless limit and, in presence of long range Coulomb interactions, the size of these regions. These are analytic results valid for any strength of the Landau parameters.'
address: |
Institute for Theoretical Physics\
University of California at Santa Barbara\
Santa Barbara, CA, 93106-4030
author:
- 'A. H. Castro Neto'
title: The Landau theory of phase separation in cuprates
---
200
There is experimental evidence that in the cuprates there exists a phase separation between hole rich and hole poor regions [@emery]. Also numerical studies in the $t-J$ model (which is believed to describe the low temperature physics of these materials) show clearly signs of phase separation [@lin]. It is well established that at zero doping and temperatures higher than the critical temperature the cuprates can be well described as a Heisenberg antiferromagnet. This has been confirmed theoretically from the studies of non-linear sigma models (which reproduces the long wavelength behavior of the Heisenberg model) and the scaling laws of many measurable quantities [@chakravarty; @sokol]. At finite doping holes start to move in the antiferromagnetic background and interact by exchanging spin waves. At the level of the $t-J$ model the physics at long wavelengths is described by the Shraiman-Siggia model [@shraiman]. In the last few years Emery and Kivelson [@emery] have also studied the problems of holes in an antiferromagnetic background and show that the phenomena of phase separation is quite natural.
In this paper I study the physics of phase separation using the tools of the Landau theory of Fermi liquids. This work is inspired by the problem of growth of instabilities in three dimensional Fermi liquids studied by Pethick and Ravenhall [@pethick]. Thus, I will make the assumption of a Fermi liquid behavior for the holes. This is actually not a strong assumption because I am going to use a part of the Landau theory which depends only on conservation laws for the distortions of the Fermi surface. Recently it was shown that this part of the Landau theory can also be applied in Luttinger liquids in integrable systems [@carmelo]. In any event, analytic studies of the ferromagnetic region of the Hubbard model at strong coupling (close to the antiferromagnetic instability) have shown that the holes form a Fermi liquid [@ioffe].
I will consider a two dimensional system of holes (spinless) with density $n=\frac{p_F^2}{4 \pi}$ where $p_F$ is the Fermi momentum ($n$ is the doping density). Since I am interested in the low doping regime I will assume that the Fermi surface is spherical and I will not take into account lattice effects. The density of states at the Fermi level is given by $N(0)=\frac{m^*}{2 \pi}$ where $m^*$ is the effective mass of the quasi-holes and $m$ is the bare mass.
I will consider angle depend distortions of the Fermi surface (in this case a line) and parametrize the distortions by an angle $\theta$, $$\delta n(\theta) = u_0 + 2 \sum_{n=1}^{\infty} u_n \cos(n \theta),$$ where I expanded the deviations in Fourier series with coefficients $u_n$. Also, it is assumed that the quasi-holes interact through the Landau parameters $f_{\vec{p},\vec{p}'}$ which in principle can be obtained from the experiment or can be calculated from some theoretical model (for instance, the $t-J$ model). It is also useful to consider the dimensionless Landau parameters $F_{\vec{p},\vec{p}'} = N(0) f_{\vec{p},\vec{p}'}$.
Since the interaction between the particles at the Fermi surface only depend on the angle between the momenta of the quasi-holes we can also expand the Landau parameters in Fourier series, $$F_{\vec{p},\vec{p}'} = F_0 + 2 \sum_{n=1}^{\infty} F_n \cos(n
(\theta-\theta'))$$ where $F_n$ are the Fourier components.
At large densities a fermionic liquid is stable against phase separation and density fluctuations which results in a positive compressibility. As we decrease the density it can enter a metastable region where it separates into a liquid-gas phase but it is still stable against density fluctuations of long wavelengths. In this phase the compressibility is positive. As the compressibility approaches zero (the so-called spinodal line) the liquid becomes unstable to long wavelength density fluctuations as well. In the context of the Fermi liquid theory this result can be understood from the expression for the sound velocity, $c_S$, in a Fermi liquid which is given by [@baym], $$c_S^2= \frac{n (1+F_0)}{m N(0)} = \frac{m^*}{m} \frac{1+F_0}{2} v_F^2,$$ where $v_F =\frac{p_F}{m^*}$ is the Fermi velocity. Observe that if $F_0 < -1$ then $c_S^2 < 0$ and an unstable collective mode grows in the system and there is a phase separation. For the moment I am considering as a starting point a homogeneous system and local interactions. It means that the instability grows in the system as a whole. If we include the Coulomb interaction, as I will show later, then the instability has a finite size. For the moment we consider local interactions since the inclusion of Coulomb interactions is trivial.
It is also well established that in the cuprates the effective mass of the holes can be many times the bare mass. In the context of the Fermi liquid theory this is due to the a non-isotropic interaction on the Fermi surface. Indeed, the relation between the effective mass and the bare mass is given by [@baym], $$\begin{aligned}
\frac{m^*}{m} = 1 - \sum_{\vec{p}'} f_{\vec{p},\vec{p}'} \frac{\partial
n^0(p')}{\partial \epsilon(p')} \frac{\vec{p} \cdot \vec{p}'}{p^2}
\nonumber\end{aligned}$$ where $n^0(p)=\Theta(\mu - \epsilon(p))$ is the zero temperature distribution for the holes and $\mu$ is the chemical potential. Substituting (2) and using the orthogonality relation between the Fourier components one finds, $$\frac{m^*}{m} = 1 + F_1 .$$ Therefore, the change in the effective mass depends only on the second Fourier component. From now on we will keep only $F_0$ (related to the instability) and $F_1$ (related to the increase in the mass) since these are the main ingredients for the physics of the system.
Observe that for $F_0 < -1$ the hydrodynamical rate of growth of the instability in the system (which determines the time scale for the growth, $\Gamma^{-1}$) is obtained straight from (3), $$\Gamma(q) = \left(\frac{(1+F_1) |1+F_0|}{2}\right)^{1/2} q v_F.$$ This result is valid when the collisions dominate the physics, that is, the mean free path is short. However, since the density of holes is small the mean free path can be quite large and the collisionless limit is more interesting.
The collisionless limit can be obtained from the Landau equation for sound waves in the absence of a collision integral (which can also be included [@next]). It reads [@baym], $$\frac{\partial \delta n_{\vec{p}}}{\partial t} + \vec{v}_{\vec{p}}
\cdot \nabla \delta n_{\vec{p}} - \frac{\partial n^0(p)}{\partial \epsilon(p)}
\vec{v}_{\vec{p}} \cdot \nabla \left( U + \sum_{\vec{p}'} f_{\vec{p},\vec{p}'}
\delta n_{\vec{p}}\right) = 0,$$ where $U$ is an external field and in terms of the deviations (1) we have $ \delta n_{\vec{p}} = \frac{\delta n(\theta)}{N(0)} \left( - \frac{\partial
n^0(p)}{\partial \epsilon(p)}\right)$.
Assuming an harmonic evolution in time, that is, $\delta n(\theta) \sim e^{i(\vec{q} \cdot \vec{r}- \omega t)}$, the Landau equation has the form, $$\left(q v_F \cos(\theta)-\omega \right) \delta n(\theta) + q v_F \cos(\theta)
\left(N(0) U + \int_{0}^{2 \pi} \frac{d \theta'}{2 \pi} F(\theta-\theta')
\delta n(\theta')\right) = 0,$$ which can be simplified by our assumption of keeping only the two first Fourier components, $$\left(q v_F \cos(\theta)-\omega \right) \delta n(\theta) + q v_F \cos(\theta)
\
\left(N(0) U + F_0 u_0 + 2 F_1 u_1 \cos(\theta)\right) = 0.$$
Integrating (8) over $\theta$ one finds, $$u_1 = \frac{\omega u_0}{q v_F (1+F_1)},$$ which is essentially the conservation of the number of particles.
Solving (8) for $\delta n(\theta)$ one gets, $$\delta n(\theta) = \frac{q v_F \cos(\theta)}{\omega-q v_F \cos(\theta)} \left(
N(0) U + F_0 + 2 F_1 u_1 \cos(\theta)\right),$$ and an integration over $\theta$ leads to, $$u_0 = \chi_{RPA}(s) \left(N(0) U + F_0 u_0 + 2 s F_1 u_1\right)$$ where $s = \frac{\omega}{q v_F}$ and $$\chi_{RPA}(s) = \int_{0}^{2 \pi} \frac{d \theta}{2 \pi}
\frac{\cos(\theta)}{s-\cos(\theta)} = \frac{s}{\sqrt{s^2-1}}-1$$ is the RPA density-density correlation function [@baym].
Since we are interested in the growth of collective modes we calculate the density-density correlation function for the problem, $$\chi(s)= \frac{u_0}{N(0) U} =
\frac{ \chi_{RPA}(s)}{1- \chi_{RPA}(s) \left(F_0 + s^2 A_1\right)}$$ where, $$A_1 = \frac{2 F_1}{1+F_1}.$$
As it is well known the collective modes are given by the poles of the correlation function (13), that is, the points where $\chi^{-1}=0$. Unlike the three dimensional case [@pethick], which leads to a transcendental equation that can be solved only numerically, in two dimensions we can solve this problem analytically for any strength of the interaction. The equation has two roots which read, $$s_{\pm}^2 = \frac{-1+2 A_1-2 F_0+2 A_1 F_0 \pm
\sqrt{1+4 A_1+4 F_0+4 F_0^2 + 4 A_1 F_0}}{2 A_1 (2-A_1)}.$$ It is easy to conclude that $s_+^2 \geq 0$ and therefore is related to a true collective mode in the system and $s_-^2 \leq 0$ which means that the rate of growth of the instability in given by, $$\Gamma(q) = s_- q v_F.$$
The inclusion of Coulomb interaction, $V(r) = \frac{e^2}{ \epsilon r}$ (where $\epsilon$ is the dielectric constant), is rather trivial since the interaction is isotropic. The Fourier transform of this interactions in two dimensions is given by $V(q) = \frac{2 \pi e^2}{\epsilon q}$ and following the prescription of the Landau theory [@baym] the only change in our calculation would be to make the substitution, $F_0 \to F_0 + \frac{2 \pi e^2 N(0)}{\epsilon q} $. Observe that this modification would lead to a momentum depend relation for the build up of the instability ($F_0 + \frac{2 \pi e^2 N(0)}{\epsilon q}
<-1$) namely, $$q \geq \frac{e^2 m (1+F_1)}{\epsilon |1+F_0|}.$$ This relation means that the instability can be probed only for wavelengths smaller than $\frac{\epsilon |1+F_0|}{e^2 m (1+F_1)}$ which established the minimum size for the unstable regions in the plane, that is, the size of the role rich regions. Of course, in the absence of Coulomb forces ($e=0$) the instability region has the size of the system.
In conclusion, in this paper I use the Landau theory of Fermi liquids in order to calculate analytically the rate of growth of phase separation in a two dimensional systems of holes in cuprates for any strength of the Landau parameters. I also estimate the size of the unstable regions due to the Coulomb repulsion between the holes.
I would like to acknowledge S.A. Kivelson, H.Q. Lin, E. Miranda and H. Monien for many illuminating comments. This research was supported in part by the National Science Foundation under the Grant No. PHY89-04035.
V.J. Emery and S.A. Kivelson, Physica C [**209**]{}, 597, (1993) and references there in. V.J.Emery, S.A.Kivelson and H.Q. Lin, Phys.Rev.Lett. [**64**]{}, 475, (1990); Phys.Rev.B [**42**]{}, 6523, (1990). S.Chakravarty, B.I. Halperin and D.R. Nelson, Phys.Rev.B [**39**]{}, 2344, (1989). A.J. Millis and H. Monien, Phys.Rev.Lett. [**70**]{}, 2810, (1993); [**71**]{}, 210, (1993); A. Sokol, R.L. Glenister and R.R.P. Singh, Phys.Rev.Lett.[**72**]{}, 1549, (1994). B. Shraiman and E. Siggia, Phys.Rev.Lett. [**62**]{}, 1564, (1989). C.J.Pethick and D.G. Ravenhall, Ann.Phys. [**183**]{}, 131, (1988) and references there in. J.M.P. Carmelo, A.H.Castro Neto and D.K. Campbell, Phys.Rev.B [**50**]{}, 3667, (1994); Phys.Rev.B [**50**]{}, 3683, (1994). L.B.Ioffe and A.I.Larkin, Phys.Rev.B [**37**]{}, 5730, (1988). G.Baym and C.J. Pethick, [*Landau Fermi-Liquid Theory*]{}, (John Wiley, New York, 1991). A.H.Castro Neto, unpublished.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'The deformation of the Fermi surface induced by Coulomb interactions is investigated in the $t-t''$–Hubbard model. The interplay of the local $U$ and extended $V$ interactions is analyzed. It is found that exchange interactions $V$ enhance small anisotropies producing deformations of the Fermi surface which break the point group symmetry of the square lattice at the Van Hove filling . This Pomeranchuck instability competes with ferromagnetism and is suppressed at a critical value $U(V)$. The interaction $V$ renormalizes the $t''$ parameter to smaller values what favors nesting. It also induces changes on the topology of hole of the Fermi surface which can go from hole to electron-like what may explain recent ARPES experiments.'
address: |
$^1$Unidad Asociada ICMM-UC3M, Instituto de Ciencia de Materiales, Consejo Superior de Investigaciones Cient[í]{}ficas. Cantoblanco, 28049 Madrid. Spain.\
$^2$Unidad Asociada ICMM-UC3M, Departamento de Matemáticas, Universidad Carlos III de Madrid, Avda. de la Universidad 30, 28911 Leganés, Madrid, Spain.
author:
- 'B. Valenzuela $^1$ and M. A. H. Vozmediano $^2$'
title: Deformation of the Fermi surface in the extended Hubbard model
---
epsf.sty
The shape of the Fermi surface plays a crucial role in determining the instabilities of interacting Fermi systems. The importance of Fermi surface deformations by interaction in the non-spherical case was already recognized in the early papers on the foundation of the Fermi liquid concept [@KL1; @KL2], and remains a central question both under a fundamental point of view as well as for the phenomenological implications.
The physics of the cuprates and the recent ARPES experiments has renewed interest on the anisotropic Fermi surfaces. Main questions are how close is the meassured Fermi surface from the non–interacting?, how much can the Fermi surface be deformed before it breaks down driving the system to a broken symmetry phase?, does the deformation always respect the point group symmetry of the lattice?
In this paper we address the former questions in a very simplified model of the cuprates. We study the Fermi surface deformation in the $t-t'$–Hubbard model on the squared two–dimensional lattice in the presence of week to moderate Coulomb interactions. We include an on-site coupling $U$ to take account of the magnetic interactions and an extended exchange interaction $V$ which can be as large as $t$ [@auer] and is widely used recently to account for the stripe feature and the possible phase separated phase in the cuprates. It will be seen that the $V$ interaction plays a crucial role when the shape of the Fermi surface does not have special features favoring magnetic instabilities.
The deformation of the Fermi surface in the Hubbard and $t-J$ models has been studied recently [@HM1; @morita; @himeda; @nojiri] with a variety of methods. In most of these approaches attention is centered on the influence of magnetic interactions so the exchange interaction $V$ is not taken into account. Our results show that, even at the mean field level, this interaction has a strong influence on the Fermi surface renormalization.
In a very recent paper [@HM4] it was argued that the Fermi surface of the $t-t'$– Hubbard model may undergo a deformation breaking the point symmetry of the underlying lattice (Pomeranchuck instability) driven by the renormalization of the forward channel. A deformation of this type seems to be observed in recent ARPES experiments [@asensio]. Ferromagnetic instabilities which would directly compete with the Pomeranchuck instability were not taken into account. We investigate the conditions for the formation of Pomeranchuck instabilities and the competition with ferromagnetism. We argue that the instability is a fine tuning effect which only takes place at the Van Hove filling and is suppressed by ferromagnetism.
Our analysis provides also interesting results concerning the renormalization of the $t'$ parameter which is shown to decrease with $V$.
The hamiltonian that we shall use is $$H=-\sum_{i,j,\sigma} t_{ij} c^+_{i,\sigma}
c_{j,\sigma}
\;+\;U\sum_i n_{i,\uparrow} n_{i,\downarrow}\;+V\sum_{<ij>,ss'}n_{is}n_{js'}
\;\;\;,
\label{ham}$$ where $t_{ij}=t$ for nearest neighbors, $t_{ij}=t'$ for next nearest neighbors, and it is zero otherwise, and where we have included an extended Coulomb interaction $V$.
The non-interacting band structure of (\[ham\]) in the squared lattice is $$\varepsilon^0({\bf k})=
-2t\cos k_x-2t\cos k_y+4t'\cos k_x\cos k_y
\label{disp}$$ what originates the constant energy contours shown in fig. 1.
The shape of the interacting Fermi surface is encoded in the two–point Green’s function of the interacting system $$G(\omega,{\bf k})=\frac{1}
{\omega-(\varepsilon^0({\bf k})-\mu)-\Sigma(\omega,{\bf k})} \;\;,$$ where $\Sigma(\omega,{\bf k})$ is the fermion self–energy. The Fermi surface of the interacting system is defined by the equation [@KL2] $$\mu-\varepsilon^0({\bf k})-\Re\Sigma(\omega,{\bf k})=0\;\;,
\label{mu}$$ where $\Re \Sigma$ stands for the real part of the electron self–energy. A mean field decoupling of the interacting term in (\[ham\]) gives rise to an exchange interaction $$V({\bf k}_x,{\bf k}_y)=
2V\sum_{k'_x,k'_y}[\cos (k_x-k'_x)+ \cos (k_y-k'_y)]n({\bf k'}) \;\;\;,
\label{V}$$ where $n({\bf k})$ is the occupation of the site ${\bf k}$ in momentum space. The ${\bf k}$ dependence of the interacion $V$ in (\[V\]) is the key point of the results.
The diagrams contributing to the electron self–energy in perturbation theory are shown in fig. \[graphs\].
Local interactions $U$ do not change the shape of the Fermi surface at the one loop level. Diagrams in fig. 2a) and fig. 2b) give rise to a constant that is absorbed in the chemical potential so as to keep the system at constant density. The first contribution to the Fermi surface deformation in the local case comes from the diagram in fig. 2c) and has been computed in the literature [@HM1; @nojiri] and found to be very small. The interaction $V$ contributes to the fermion self–energy at the one loop level through diagram in fig. 2b).
We compute numerically the contribution of diagram in fig. 2b) to the Fermi surface deformation in eq. (\[mu\]) by solving self-consistently the equation $$\begin{aligned}
\varepsilon({\bf k}) &= & \varepsilon_{kin}({\bf k}) +\varepsilon_{exch}({\bf k})-\mu
\nonumber \\
& = & -2t\cos k_x-2t\cos k_y+4t'\cos k_x\cos k_y -
2V\sum_{k'_x,k'_y}[\cos (k_x-k'_x)+ \cos (k_y-k'_y)]n({\bf k'}) -\mu \;\;\;,
\label{nospin}\end{aligned}$$ where $\mu$ is adjusted so that the total number of electrons remains constant and spin interactions are not taken into account.
Ferromagnetism is included by starting with two different Fermi surfaces for electrons with spin up and down and solving the equations $$\begin{aligned}
\varepsilon_\uparrow ({\bf k})& = & \varepsilon_\uparrow^0 ({\bf k})
+\varepsilon_\uparrow^{exch} ({\bf k})+U\frac{n_\downarrow}{N}-\mu \nonumber \\
\varepsilon_\downarrow ({\bf k})& = & \varepsilon_\downarrow^0 ({\bf k})
+\varepsilon_\downarrow^{exch} ({\bf k})+U\frac{n_\uparrow}{N}-\mu \;\;,
\label{ferro}\end{aligned}$$ where the total number of electrons $N=n_\uparrow +n_\downarrow$ remains constant.
From eqs. (\[nospin\]) and (\[ferro\]), it is clear that in the absence of feromagnetism, i.e starting with two identical Fermi surfaces in (\[ferro\]), the interaction $U$ contributes to the total energy with a global constant that is absorbed in the chemical potential. The deformations induced by $V$ are in general similar to the ones obtained in the literature when computing diagram in fig. 2c) with a local $U$ interaction. The reason is that the electron susceptibility inserted in the loop acts as a ${\bf k}$–dependent effective interation. The global sign is also the same (attractive). In the second order perturbation theory because of the closed fermion loop in the diagram, and in $V$ due to the exchange. The present analysis has the relevance of acting already at the one loop level.
Our results can be summarized as follows. The deformation of the Fermi surface depends on the filling. The general effect of $V$ is to smoothen out anisotropies in the curvature of the Fermi surface. This has a number of interesting consequences. In particular for closed surfaces (electron–like), the inflection points present with a finite $t'$ which were related in [@us] to pairing instabilities with an extended-s order parameter, dissapear at a value of $V\sim 1$. In the context of ref. [@us], the presence of $V$ favors a d-wave pairing. This result seems at odds with the claims in [@himeda] which find spontaneous generation of $t'$ in a Monte Carlo analysis of the $t-J$ model. We think that this is due to the sign of the interaction. The effect of a positive $V$ is the reversed one: it reinforces the curvature of the Fermi surface leading to bigger values of $t'$.
The general agreement of our computation with other very different approaches [@morita] suggests that there will be a general tendency towards nesting in the $t-t'$–Hubbard model with an extended attractive interaction of any origin.
A related topic of experimental relevance [@exp1] concerns the possible change in Fermi surface topology induced by interactions. The presence of a $t'$ parameter breaking the electron–hole symmetry of the Hubbard model is crucial in this analysis. We find no changes in the Fermi surface topology with $t'=0$ in agreement with [@HM1]. In the presence of finite $t'$ and $V$, the Fermi surface does undergo changes in topology from open (hole–like) to closed (electron–like) when the initial Fermi energy lies close to – and slightly over – the Van Hove filling. This effect can explain recent messurements of electron–like surfaces in BISCO [@exp1; @exp2]. The reversed change would be observed by reversing the sign of $V$.
It is interesting to note that the interacting Fermi surface lies at the level of the Van Hove singularities for a variety of initial fillings in the hole–like regime. The Van Hove filling is never reached if starting with an overdoped situation (closed Fermi surface). This last result is very insensitive to the value of $t'$ and mimics the result about the pinning of the Fermi surface obtained in [@nos].
We have searched for Pomeranchuck instabilities by investigating the response of the system to very small perturbations of the free Fermi surface. The result is extremely sensitive to the value of the chemical potential of the final system. The effect of the interactions is, in general, to restore the lattice symmetry for generic values of the final $\mu$. Even in the case of very strong initial perturbation (of the order of a 4 per cent), the final Fermi surface is four-fold symmetric. The exception arises when the chemical potential of the interacting system coincides with the Van Hove singularity . We have checked that, in the absence of ferromagnetism, the Van Hove filling is unstable towards Pomeranchuck deformations. This result agrees with the general considerations exposed in [@and] that the Fermi level of electron systems tend to avoid peaks in the density of states.
Fig. \[vh\] shows the deformation caused by $V=2$ in a paramagnetic situation when a point is removed by hand in the initial Fermi surface of a 61 x 61 lattice (corresponding to an anisotropy of less than $10^{-3}$).
The solid line represents the free Fermi surface, the dotted line is the interacting Fermi surface in the absence of the perturbation, and the dashed line shows the final Fermi surface obtained after perturbing the free Fermi surface. Similar results are obtained with different values of V(n). This result is independent of $t'$ and occurs also at $t'=0$. It is nevertheless extremely sensitive to the value of the final chemical potential which has to be fine tuned by selecting the initial number of particles.
Magnetic interactions do substantially change the former picture. We have centered our attention on the effect of ferromagnetism as the predominant competing interaction in the forward channel [@us2]. Our results show that ferromagnetism is enhanced by $U$ and $t'$ as expected, but it is disfavored by $V$. This can be understood on the light of our previous analysis showing that $V$ renormalizes the $t'$ parameter downwards. For small to moderate values of $U$, the free ferromagnetic system evolves to an interacting paramagnetic state. There is, however, a critical value of $U(V)$ where the final state corresponds to a fully polarized ferromagnetic system. This supression of ferromagnetism by exchange interactions might explain the absence of ferromagnetic phases in real systems which are otherwise well described by the $t-t'$–Hubbard model as the one in [@us2].
The Pomeranchuck instabilities described above compete with ferromagnetism. A schematic phase diagram showing the evolution of the ferromagnetic behavior of the system as a function of $U$ and $V$ is despicted in fig. \[phased\].
Pomeranchuck instabilities are found systematically under the appropriate conditions in the paramagnetic zone of the phase diagram. By increasing $U$ and keeping $V$ constant, they enter into the phase denoted by FM$\downarrow$ in fig. \[phased\] changing it into a paramagnetic phase. Finally there is a critical value of $U$ for which ferromagnetism prevails and the deformation dissapears.
In conclusion we have shown that exchange interactions renormalize $t'$ to smaller values. An inmediate consequence of this is that, in our search for microscopic models of the cuprates, one should start the computations with a bigger value of $t'$ to account for a phenomenologically fitted value of $t'$. The effect is doping–dependent and should be stronger in the underdoped situation where screening is less effective. In this regime it will increase the nesting of the Fermi surface and enhance antiferromagnetic instabilities. The renormalization of the $t'$ parameter may also explain the absence of ferromagnetic phases in some real systems described by the $t-t'$–Hubbard model, and recent meassurements of electron–like Fermi surfaces in some cuprates.
Pomeranchuck instabilities are found as the response of the system to small impurities when the Fermi surface lies at the Van Hove filling and can be regarded as an instability of the Van Hove level. They compete with ferromagnetism and dissapear at a critical value U(V).
Although the results presented in this paper are mean field results , we believe that the main features will survive the test with more accurate methods and that exchange interactions should be included – or checked to be negligible – in Hubbard related studies.
[**Acknowledgements.**]{} We are very grateful to F. Guinea for valuable discussions. Financial support from MEC (Spain) through grant PB96/0875 and CAM (Madrid) through grant 07/0045/98 is acknowledged.
W. Kohn and J. M. Luttinger, Phys. Rev. [**118**]{}, 41 (1960).
W. Kohn and J. M. Luttinger, Phys. Rev. [**119**]{}, 1153 (1960).
A. Auerbach, “Interacting electrons and quantum magnetism”, Springer (1998).
C. J. Halboth and W. Metzner, Z. Phys. B [**102**]{}, 501 (1975).
K. Morita and K. Miyake, “Deformation of the Fermi surface due to antiferromagnetic correlation”, preprint (cond-mat/9912292).
A. Himeda and M. Ogata, “Spontaneous deformation of the Fermi surface due to strong correlation in the two–dimensional $t-J$ model”, preprint (cond-mat/0003278).
H. Nojiri, J. Phys. Soc. Jpn. [**68**]{}, 548 (1999).
C. J. Halboth and W. Metzner, “d-wave superconductivity and Pomeranchuck instability in the two dimensional Hubbard model”, preprint (cond-mat/0003349).
M. C. Asensio, private communication.
J. González, F. Guinea, and M.A.H. Vozmediano, Phys. Rev. Lett. [**79**]{}, 3514 (1997).
A. D. Gromko, Y. -D. Chuang, D. S. Dessau, K. Nakamura, and Y. Ando, “Fermi surface topology of $Bi_2Sr_2CaCu_2O_{8+\delta}$ at $h\nu=33 eV$: hole or electron like?”, preprint (cond-mat/0003017).
P. V. Bogdanov, A. Lanzara, X. J. Zhou, S. A. Kellar, D. L. Feng, E. D. Lu, J. -I. Shimoyama, K. Kishio, Z. Hussain, and Z. X. Shen, “ARPES study of Pb doped $Bi_2Sr_2CaCu_2O_8$ - an unambiguous case for an electron–like Fermi surface”, preprint (cond-mat/0005394).
J. González, F. Guinea, and M.A.H. Vozmediano, Europhys. Lett. [**34**]{}, 711 (1996); J. González, F. Guinea, and M.A.H. Vozmediano, Nucl. Phys. [**B 485**]{} (1997) 694.
P. W. Anderson, “Van Hove singularities as a source of anomalies in the cuprates?", Princeton preprint; and P. W. Anderson, “The theory of high–$T_c$ superconductivity", Princeton University Press, (1997).
A. Abanov and A. V. Chubukov, “Spin-fermion model near the quantum critical point: one-loop renormalization group results”, preprint (cond-mat/0002122).
J. V. Alvarez, J. González, F. Guinea, and M.A.H. Vozmediano, J. Phys. Soc. pn. [**67**]{}, 1868 (1998).
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We show that exposing TiO$_2$ anatase (001) surface to hydrogen plasma causes a slight structural modification of the surface. On the modified surface the topmost TiO$_2$ layer is intact, but out of registry with the bottom layers. Nevertheless, the modified surface has significantly improved ability to split water under exposure to sunlight. First, we show by explicit calculation of the water splitting reaction that the energy barrier that exists on a pristine surface isn’t present on the modified surface. Second, we show that the valence band maximum of the surface is raised relative to the pristine surface, which is a favorable way of adjusting the band gap in TiO$_2$ to the solar spectrum.'
author:
- Ming Lei
- Sinisa Coh
bibliography:
- 'pap.bib'
title: 'Prediction of hydrogen plasma induced modified anatase TiO$_2$ (001) surface with desirable water splitting performance'
---
Introduction
============
Anatase TiO$_{2}$ is a promising and well studied material for photocatalytic water splitting. TiO$_2$ is cheap, environment-friendly, and stable. As is well known, due to its large band gap, TiO$_2$, can’t utilize the sunlight efficiently. Furthermore, to provide enough driving force for the photo oxidation of water, the valence band maximum (VBM) of TiO$_2$ should be increased and positioned closer to, but lower than, the potential of H$_2$O/O$_{2}$ pair.[@valdes2008oxidation] Up to now, plenty of research has been conducted to narrow its band gap and to elevate the VBM.
Doping is one common strategy to reduce the band gap of TiO$_{2}$. Doping can be done with either a metal ion (La,[@liqiang2004preparation] V,[@klosek2001visible] Co,[@chen2007carbon] Nb,[@lu2010improved] Ag[@sobana2006nano]) or a nonmetal ion (S,[@ohno2003photocatalytic] B,[@in2007effective] C,[@irie2003carbon; @park2006novel] N[@di2005characterization]). However, some dopants, especially the metal ions, can lead to severe carrier recombination which reduces the overall quantum efficiency.[@choi1994role] Moreover, the metal ion dopants are often polluting.
An alternative approach for narrowing the band gap, and raising the VBM, is to modify the surface structure of anatase TiO$_{2}$ by hydrogenation.[@chen2011increasing] This treatment changes the color of anatase TiO$_{2}$ from white to black,[@chen2015black; @chen2011increasing; @liu2013hydrogenation] and produces highly rough and amorphous surface of TiO$_{2}$, both in the case of nanoparticles [@teng2014preparation; @yan2013] and nanotubes.[@xu2014integrated; @wu2013enhanced] Furthermore, according to the X-ray photoelectron spectroscopy (XPS) measurement, the band gap narrowing of black TiO$_2$ is achieved by raising the VBM without changing the CBM (conduction band minimum)[@chen2011increasing] which is a favorable band alignment for photocatalytic water splitting. While it remains unknown exactly which structural modification leads to the favorable band alignment of black TiO$_2$, there is evidence that likely the increase in VBM is due to the surface modification of TiO$_2$.[@chen2011increasing; @liu2013hydrogenation] On the other hand, XPS study on a similar system (nanowires instead of nanoparticles) by Wang et al.[@wang2011hydrogen] found no shift in the valence band. They assigned the dark color of nanowires to the formation of defect states or impurities. Furthermore, Alberto and co-workers[@naldoni2012effect] synthesized black TiO$_{2}$ nanoparticles with crystalline core and disordered shell morphology, and found that in addition to the surface modification, the presence of oxygen vacancies could also contribute to the visible light absorption of TiO$_2$.
In this work we explore possible structural changes to the surface of TiO$_2$ induced by absorption of hydrogen atoms to the TiO$_2$ surface. Our calculations show that the structure of pristine (001) surface of TiO$_2$ anatase is modified when exposed to a hydrogen atom pressure of 0.3–30 Pa, which can be achieved in the laboratory.[@nakamura2000role] Furthermore, by doing an explicit calculation of the water splitting process, we find that the structural modification induced by hydrogen atom removes the rate-limiting step for the water splitting process. More specifically, while the pristine anatase TiO$_{2}$ (001) surface has a barrier of 0.55 eV in the Gibbs free energy profile of the oxygen evolution reaction, there is no such barrier on the modified surface. Therefore, our calculations show that the hydrogen-plasma treated anatase TiO$_{2}$ (001) surface is a suitable candidate for water splitting applications.
This paper is organized as follows: in Sec. \[sec:details\] we describe the calculation details. In Sec. \[sec:surface\] and Sec. \[sec:water\] we present and discuss our results. We give an outlook and conclude in Sec. \[sec:conclusion\].
Calculation details {#sec:details}
===================
For our calculations we use density functional theory as implemented in the Quantum Espresso package.[@giannozzi2009quantum] We use the Generalized Gradient Approximation (GGA) of Perdew, Burke and Ernzerhof (PBE)[@perdew1996] along with the ultra-soft pseudopotentials from the GBRV database.[@garrity2014pseudopotentials] These pseudopotentials describe the valence electrons 3s3p3d4s in Ti, 2s2p in O and 1s in H. In order to obtain sufficient precision, we cutoff the plane wave basis for the wavefunction at 40 Ry and 400 Ry for the density. We use 15 Å of vacuum to avoid the interaction between neighboring slabs. In cases when the surfaces of the slab are different, for example when molecules are absorbed on one of the sides of the slab, we use the dipole correction in the direction perpendicular to the surface. All surface energies in the paper are reported per one side of the slab. We sample the electron’s Brillouin zone on a 6$\times$6$\times$1 Monkhorst-Pack grid.
To model a slab of TiO$_2$ we set the in-plane lattice constant of the slab equals to the in-plane bulk lattice constant. Each time we fully relax the slab with an only constraint that the in-plane lattice constant remains unchanged. We use slab with a thickness of 8 layers of TiO$_{2}$. For most calculations we use minimal in-plane unit cell, but for the water adsorption and oxygen evolution reaction processes, we use a in-plane supercell that is doubled along one of the in-plane lattice vectors.
We calculate the surface energy density $\Delta \gamma$ as $$\begin{aligned}
\Delta \gamma & = \frac{1}{A} \left[ E_{\rm slab}-E_{\rm clean}-N_{\rm H} \mu_{\rm H}(T,p) \right]\end{aligned}$$ Here $E_{\rm slab}$ is the energy of the slab with $N_{\rm H}$ hydrogen atoms adsorbed on the surface. $E_{\rm clean}$ is the energy of the slab without adsorbed hydrogen atoms. The surface area is $A$. $\mu_{\rm H}(T,p)$ is the chemical potential of H atom as a function of temperature ($T$) and pressure ($p$), $$\begin{aligned}
\mu_{\rm H} (T, p) & = \Delta\mu_{\rm H} (T,p) + \frac{1}{2} E_{{\rm H}_2}\label{mu}, \\
\Delta \mu_{\rm H} (T, p) & = \frac{1}{2} \left[ \tilde\mu_{{\rm H}_2} (T,p^0)+ k_{\rm B} T \ln (p_{{\rm H}_2}/p^0) \right]\label{delta}.\end{aligned}$$ Here, $p^0$ is the pressure of a reference state, and the temperature dependence of the chemical potential $\tilde\mu_{{\rm H}_2}(T,p^0)$ is tabulated in the thermochemical reference tables.[@stull1971janaf]
Hydrogen induced surface modification {#sec:surface}
=====================================
Now we present our results on the hydrogen-atom induced structural modification of the TiO$_2$ anatase $(001)$ surface. In the next section we will study its water splitting performance.
The conventional unit cell of bulk anatase TiO$_{2}$ is shown in Fig. \[fig:bulk\]a. Anatase TiO$_2$ crystallizes in space group I$4_1/$amd (space group 141). Titanium atoms are at Wyckoff orbit $b$ while oxygen atoms are at Wyckoff orbit $e$. Therefore, all titanium atoms in the crystal structure are equivalent to each other, and all oxygen atoms are equivalent to each other. Our calculated relaxed lattice parameters of bulk anatase TiO$_{2}$ are $a$ = $b$ = 3.804 Å and $c$ = 9.695 Å, which is close to the experimental result $a$ = $b$ = 3.804 Å and $c$ = 9.614 Å (0.84% deviation).[@horn1972refinement] The structural unit of TiO$_2$ anatase is a TiO$_6$ octahedron with Ti atom in the center of the octahedron, and O atoms in the corners of the octahedron. These octahedra are connected to each other and are forming an edge sharing network. Each O atom is bonded to three Ti atoms.
Based on our calculation of bulk anatase TiO$_{2}$ we constructed a model of pristine $(001)$ surface. This surface is shown in Fig. \[fig:bulk\]b. While in the bulk TiO$_2$ all Ti atoms are six-fold coordinated, this is clearly not the case on the $(001)$ surface. Here, the breaking of the Ti–O bond perpendicular to the surface reduces the coordination of the topmost Ti atom from six to five. Furthermore, there are now two symmetry inequivalent oxygen atoms at the surface: only one of which is nominally saturated, as it is surrounded by three Ti atoms. Another oxygen atom is unsaturated, as it is surrounded by only two Ti atoms. Therefore, all Ti and half of the O atoms at the topmost layer of TiO$_2$ (001) surface are nominally unsaturated.
![\[fig:bulk\]The conventional unit cell of bulk anatase TiO$_{2}$ (a). The pristine (b) and modified (c) surfaces of anatase TiO$_2$ differ only by translation of the top-most layer of TiO$_2$ by vector $(a/2,a/2)$. Ti, gray; O, red.](Fig_1.jpg){width="3.4in"}
Hydrogen adsorbed on a pristine surface
---------------------------------------
We now discuss the hydrogen passivization of the (001) surface. We consider all the Ti and O atoms on both the first and second layer of the surface as potential hydrogen atom adsorption sites. Relative energies of sites with strongest (optimal) absorption are shown in Fig. \[fig:stability\] as a function of number of adsorbed H atoms.
When only one H atom is adsorbed per surface unit cell we find that the H atom prefers to absorb horizontally on the surface two-coordinated O atom. The adsorption energy is $-2.23$ eV. The negative sign for the adsorption energy means that it is energetically favorable for H atom to absorb on the surface. This binding energy is specified relative to a single isolated H atom. With two adsorbed H atoms, one prefers to adsorb again to the topmost two-coordinated O atom, while the second adsorbs on the surface Ti atom. The total adsorption energy for these two H atoms taken together is $-4.06$ eV. Therefore, adsorption energy per atom is now decreased from $-2.23$ eV to $-4.06/2=-2.03$ eV. With three adsorbed H atoms we find that first two adsorb as before, while the third one prefers to adsorb on the O atom in the second layer. In this case the total adsorption energy for all three atoms is $-5.33$ eV. Finally, we find that with four adsorbed H atoms, two are adsorbed in the first layer, and another two on the second layer. The total adsorption energy for these four H atoms is $-6.85$ eV. The adsorption energy per hydrogen atom in this case is therefore reduced to $-6.85/4=-1.71$ eV.
We find that the pristine (001) surface can’t absorb more than four H atoms per cell. If we try adding the fifth H atom to the surface, we find that the added H atom combines with another H atom on the surface to generate a H$_2$ molecule and moves away from the surface. Therefore, we conclude that the unsaturated pristine surface of (001) TiO$_2$ anatase can be saturated by adsorption of at most four H atoms per surface unit cell.
In all of these cases we fully relaxed the TiO$_2$ surface in the presence of hydrogen atoms. The surface relaxation is significant, with the maximal atomic displacements on the order of 0.4 Å. As one would expect, surface relaxation is smallest with one adsorbed hydrogen atom and largest with four adsorbed hydrogen atoms.
![\[fig:stability\]Relative stability of pristine and modified surfaces after adsorption of hydrogen atoms.](Fig_2.pdf){width="3.4in"}
Hydrogen adsorbed on a modified surface
---------------------------------------
Now we consider the possibility that presence of adsorbed hydrogen atoms will lead (001) surface to reconstruct in a distinct basin of energy, with structure that is significantly different from the pristine (001) surface.
Indeed, our calculations show that when four hydrogen atoms are adsorbed per a single surface unit cell that the surface of TiO$_2$ (001) anatase is reconstructed. This reconstruction is similar to the bulk structure discussed in Ref. . The structure can be approximately described as translation of Ti atoms on the surface by $(a/2,a/2)$ relative to the pristine surface. This translation vector is shown in Fig. \[fig:bulk\]. Note that if the topmost layer was isolated from the other layers, that the pristine and reconstructed surface would be indistinguishable from each other. This can easily be seen from Fig. \[fig:bulk\], or by realizing that in a surface primitive unit cell oxygen atoms are at coordinates $(0,a/2)$ and $(a/2,0)$ while titanium atom is at the origin, $(0,0)$. Therefore, if we translate titanium atom by $(a/2,a/2)$ we get a crystal structure equivalent to the one where we translate both all atoms by $(a/2,a/2)$, as oxygen atoms get mapped into periodic images of each other. Therefore, this modified structure is in some sense minimally perturbed relative to the pristine surface, as the only difference of the modified surface is that the topmost layer is out of registry with the rest.
We studied adsorption of hydrogen atoms to the modified surface following steps analogous to those used for the pristine surface. The calculated relative energy of modified surface adsorbed with different number of H atoms is indicated in Fig. \[fig:stability\] with red color. We took the energy of the pristine surface without H atom adsorption as a reference state with zero energy.
We find that without H atom adsorption, the energy of the modified surface is 1.4 eV higher than that of pristine surface. However, once H atoms are adsorbed the energy difference between modified and pristine surface diminishes. Eventually, with four adsorbed H atoms the modified surface becomes energetically favorable compared to the pristine surface.
More specifically, we find that the absorption energy of the first hydrogen atom is $-2.32$ eV. While this surface absorbs hydrogen atom slightly more strongly than the pristine surface ($-2.32$ eV compared to $-2.23$ eV on a pristine surface), the difference is not large enough to compensate for the increased surface energy of the modified surface relative to the pristine surface (1.4 eV). However, if we increase the number of hydrogen atoms to two per surface unit cell, the total adsorption energy increases to $-4.58$ eV (compared to $-4.06$ eV in the pristine case). With three hydrogen atoms, it is $-6.52$ eV, and finally, it is $-8.34$ eV with four H atoms. Therefore, when four H atoms are adsorbed, the modified surface becomes favorable relative to the pristine surface, as the stronger preference of H-adsorption on the modified surface ($-8.34$ eV versus $-6.85$ eV on the pristine surface, the difference is $-6.85+8.34=1.49$ eV) is large enough to compensate for the difference in the surface energy between modified and pristine surface (1.4 eV). However the energy difference between pristine and modified surfaces with four adsorbed H atoms is very small, only 0.09 eV, and it is likely comparable to the error of the GGA approximation used in our calculation. Nevertheless, our calculation clearly shows that the modified surface has tendency to absorb more hydrogen atoms and will thus be energetically more and more favorable at high enough pressure of hydrogen atoms.
Required hydrogen pressure
--------------------------
Based on the surface adsorption energies of H atoms we will now determine the required pressure of hydrogen atoms needed to modify the pristine TiO$_2$ surface. Figure \[fig:potential\] shows the surface energy $\Delta \gamma$ of the pristine and modified surfaces as a function of the hydrogen chemical potential $\rm \mu_H$. As can be seen from the figure, surface without hydrogen atom adsorption $N_{\rm H}=0$ is favorable at hydrogen chemical potential below $-2.25$ eV. As hydrogen chemical potential is increased, preferred surface becomes the pristine surface with $N_{\rm H}=1$ hydrogen atom adsorption. At $\mu_{\rm H} = -1.83$ eV the preferred surface becomes the pristine surface with $N_{\rm H}=2$. Finally, when the hydrogen chemical potential is larger than $-1.44$ eV, the modified surface with $N_{\rm H}=4$ becomes the most favorable surface. Since a small difference in the chemical potential can cause a large change of the pressure, based on equations (\[mu\]) and (\[delta\]), we report the needed pressure of hydrogen atoms as a range 0.3–30 Pa. This range assumes a 10% error in the calculated hydrogen chemical potential.
![\[fig:potential\]The surface energy density $\Delta \gamma $ of the pristine (solid) and modified (dashed) surfaces with $N_{\rm H}=0$ through 4. Above $\mu_{\rm H}=1.44$ eV modified surface (dashed green) becomes the most favorable surface.](Fig_3.pdf){width="3.4in"}
Energy barrier between pristine and modified surface
----------------------------------------------------
Now we turn to calculating the barrier between the pristine and modified surface. We calculated the barrier using the nudged elastic band (NEB) approach. As shown in Fig. \[fig:barrier\] the barrier is very high when there are no hydrogen atoms adsorbed on the surfaces (it is 2.3 eV, per primitive surface unit cell). However, once four hydrogen atoms are adsorbed on the surface, the barrier is reduced to only 0.5 eV per primitive surface unit cell. The reduced energy pathway between pristine and modified surface is indicated with a dash line in Fig. \[fig:barrier\].
As discussed earlier, hydrogen atoms do not absorb at the equivalent locations in the case of pristine and modified surface. In particular, on the pristine surface one of the hydrogen atoms is optimally bonded below the O atom in the second layer, while it is bonded above on the modified surface. However, both configurations, on either of the surfaces, are nearly degenerate in energy (they differ by 0.18 eV). Therefore, for the sake of comparison, we performed the NEB calculation only between the structures in which hydrogen atoms are adsorbed at equivalent locations. This path is indicated with points 1 through 6 in Fig. \[fig:barrier\]. Point 7 indicates energy of a modified structure with the optimal configuration of adsorbed hydrogen atoms.
![\[fig:barrier\]Calculated energy profiles for the structure transformation from the pristine to the modified structure, both without (solid line) and with (dash line) adsorbed hydrogen.](Fig_4.pdf){width="3.4in"}
Other potential structural modifications
----------------------------------------
So far we only discussed one reconstruction of TiO$_2$ anatase (001) surface in the presence of hydrogen atoms. In what follows we consider some other possible reconstructions compatible with the minimal 1$\times$1 surface unit cell, and show that they are all energetically less stable. We leave to future studies the possibility of other reconstructions, especially those that might increase the size of the primitive surface unit cell, as that would make the calculation more computationally demanding.
Modified structure of TiO$_2$ we discussed earlier can be constructed by translating the top most Ti atom by $(a/2,a/2)$. This translation breaks one of the Ti-O bonds, the one that is perpendicular to the surface, which is why the energy of the surface increases by 1.4 eV. However, as discussed earlier, this energy difference is compensated by the fact that the surface with broken Ti-O bond can absorb more hydrogen atoms. Motivated by this finding, we will now consider different ways to break Ti-O bonds on the surface and check whether they can also be compensated energetically by absorbing additional hydrogen atoms.
The first alternative way to break Ti-O bond we considered was to simply increase the vertical distance between Ti and O atoms. If we try inserting additional H atom between the bond-breaking Ti and O, we find that instead of formation of Ti-H or O-H bond, hydrogen atoms bind together and form a H$_2$ molecule inside the slab. The energy of this structure is 0.45 eV higher than the total energy of the pristine surface adsorbed with four H atoms (plus one isolated H atom, to keep total number of H atoms the same). Therefore, we conclude that hydrogen atoms can’t stabilize breaking of the vertical Ti-O bond, unless one translates Ti atom by $(a/2,a/2)$, as in the modified structure.
The second structure we tried has a broken Ti-O bond that is parallel to the surface. As in the previous case, we broke the bond simply by increasing the distance between the Ti and O atoms in the bond. If we try adding two additional H atoms between Ti and O, we find that two H$_2$ molecules are generated during the structural relaxation. The energy of this configuration is 0.52 eV higher than the total energy of the pristine surface adsorbed with four H atoms (plus energy of two isolated H atoms). Therefore, we conclude that in this scenario hydrogen atoms can’t break the in-plane Ti-O bond.
The third structure we tried contains oxygen vacancy at the surface. To form a vacancy we removed one surface two-coordinated O atom. Next we tried to stabilize this surface by putting one additional H atom at the vacancy site. We again find that during the relaxation process one H$_2$ molecule is generated. The formation energy of this oxygen vacancy is very high (it is 3.74 eV), so this structure can’t be stabilized by addition of hydrogen atoms. We quantified the formation energy of the vacancy by taking a difference between the total energy of the slab with a vacancy (plus one isolated O atom) and the total energy of the pristine surface adsorbed with four H atoms (plus one isolated H atom).
![\[fig:doswater\]Density of states of (a) pristine (001) surface and (b) modified (001) surface with one water molecule on the surface. The conduction band shown is scissor shifted by 1.39 eV following Ref. . Blue line shows oxygen 2p-like states projected onto two topmost layers of TiO$_2$.](Fig_5.pdf){width="3.4in"}
Water splitting performance of the modified surface {#sec:water}
===================================================
After showing that the modified surface becomes favorable at high hydrogen plasma pressure, now we turn to the study of the water splitting performance of the modified surface.
Water adsorption
----------------
When anatase TiO$_2$ is used to catalyze water splitting, the catalyst is immersed in the aqueous environment. Therefore, we will first compare the adsorption of water molecule on pristine and modified surfaces. We find that after full structural relaxation the water molecule binds to the pristine surface by forming a bond between the Ti atom on the surface and O atom of the water molecule (bond length is 2.339 Å). One of the H atom in the water molecule interacts with the surface O atom and generates a bond with length of 1.620 Å. We expect that a very similar binding geometry will occur on the modified surface, as the top layer is nearly the same as in the pristine case. This is precisely what we find, as the Ti-O bond length on the modified surface is 2.343 Å while the H-O bond length is 1.636 Å.
Despite similarities in the structure, the energy level alignments are not the same on two surfaces. Figure \[fig:doswater\] shows calculated density of states (DOS) of the pristine and modified surfaces with adsorbed water molecule. The energy in that figure is aligned relative to the vacuum level above the surfaces. While we find that the conduction band minimum (CBM) is nearly the same in the pristine and the modified surface, the valence band maximum (VBM) is higher by 1 eV on the modified surface. Furthermore, as indicated in the figure, this increase in the VBM originates from the oxygen 2p states in the first and second layer (blue line) on the TiO$_2$ surface and not from the oxygen 2p states in the water (yellow line).
Reaction path
-------------
Since the modified surface has a higher VBM than the pristine surface, the reduced band gap of the modified surface should contribute to better water splitting performance. To test this hypothesis, we studied the explicit chemical reaction on the surfaces. Several groups have already studied theoretically the mechanism of photocatalytic water splitting on the pristine anatase TiO$_2$ surface.[@li2010mechanism; @zhao2012] Since pristine and modified surface differ only in the registry between the two top-most layers of TiO$_2$, we suspect that the reaction pathways might be similar. After all, the topmost layer of TiO$_2$ that is exposed to the water molecules is structurally nearly the same in two cases. However, the magnitude of energy barriers needs not be the same, as electronic structures are different.
As a reference path for the pristine TiO$_2$ surface we took the reaction path proposed by Liu et al. in Ref. . We didn’t implement the solvation effects in our calculations as solvation effects have a small effect on the relative energy of each state. For example, we find that the rate-limiting barrier without solvation effects is 0.55 eV while Ref. reports that the same barrier is 0.61 eV with solvation effects.
![\[fig:path\]Gibbs free energy profiles of oxygen evolution reaction on pristine (001) and modified (001) surface (1.93 V versus the standard hydrogen electrode).](Fig_6.pdf){width="3.4in"}
Figure \[fig:path\] shows calculated Gibbs free energy profiles of oxygen evolution reaction on the pristine (black line) and modified surface (red line). State 1 on the figure represents the surface adsorbed with one dissociated water molecule where one H$^+$ ion is adsorbed on the surface O atom and one OH$^-$ ion adsorbed on the surface Ti atom. This state 1 is the initial state of the water splitting reaction. From state 1 to state 2, one H$^+$ ion is extracted away from the surface, $$\rm H_2O/TiO_2 + h^+ \to OH/TiO_2 + H^+$$ (Adsorption of H$_2$O or OH on the surface we denoted with H$_2$O/TiO$_2$ and OH/TiO$_2$. The hole is denoted by $\rm h^+$.) In step $2 \to 3$ the second H$^+$ ion is extracted. Next, another water molecule absorbs on the surface (state 4), and the third H$^+$ ion is extracted away from the surface (state 5). From state $5 \to 8$ the fourth H$^+$ ion is extracted away from the surface and one O$_2$ molecule is generated during the process. State 9 is the final state of water splitting reaction without any adsorption. State 9 will go back to state 1 after chemical adsorption of one water molecule. States 1 through 9 shown in Fig. \[fig:path\] are equivalent to states 9 through 17 in Ref. . The Gibbs energy difference between the beginning state 1 and state 9 is the reaction energy ($\Delta$G) for process $$2\rm H_2O + 4h^+ \to 4H^+ + O_2.$$ The Gibbs energy difference between state 9 and the last state 1 is the chemisorption energy of one H$_2$O molecule on the surface.
As already found in Ref. the rate-controlling step on the pristine surface is the first proton removal step $1 \to 2$. However, this barrier does not exist on the modified surface, as $1 \to 2$ on the modified surface is exothermic by $-0.49$ eV. The rate-controlling step on the modified surface is the fourth proton removal step $7 \to 8$, with a relatively small barrier of only 0.13 eV. Therefore, we find that the modified (001) surface has much better water splitting performance than the pristine surface.
Origin of the reduced energy barrier
------------------------------------
Now we will discuss possible origin of the reduced energy barriers on the modified TiO$_2$ surface. Fig. \[fig:structure\] shows the optimized structures of states 1 and 2 in the reaction path. Comparing Fig. \[fig:structure\] (b) and (d), we find that the Ti–O bond lengths in state 2 are shorted on the modified state 2 compared to the pristine surface. We can rationalize this by noting that the surface Ti atom is less saturated on the modified than on the pristine surface. Therefore, the Ti atoms binds more firmly to O atom and OH$^-$. As a result, this relatively lower energy of the modified state 2 shown in Fig. \[fig:path\] contributes to the reduced energy barrier on the modified TiO$_2$ surface.
Furthermore, as shown in Fig. \[fig:doswater\], the VBM of modified TiO$_2$ surface after adsorbing one H$_2$O is higher than that of pristine TiO$_2$ surface. Which is well known to contribute to the improved chemical activity on the surface.[@xu2000]
![\[fig:structure\]Optimized structures of the intermediate states in the reaction path: (a) pristine state 1; (b) pristine state 2; (c) modified state 1; (d) modified state 2. (The bond lengths are given in Å.)](Fig_7.pdf){width="3.4in"}
\[sec:conclusion\]Outlook and conclusion
========================================
Our calculations show that the modified (001) surface of TiO$_2$ anatase has electronic structure that is favorable for water photocatalysis. There are several ways to synthesize TiO$_{2}$ (001) films in experiment. One is by adding hydrofluoric acid into the TiO$_2$ precursor, since the hydrofluoric acid can act as a shape controlling agent and makes (001) energetically preferable to (101) facet.[@yang2008anatase; @yang2009solvothermal; @liu2011anatase; @chen2011formation] Another is to grow the (001)-oriented anatase TiO$_{2}$ (001) film on a seed layer substrate, such as RbLaNb$_{2}$O$_{7}$,[@nakajima2011001] Ca$_{2}$Nb$_{3}$O$_{10}$,[@shibata2010fabrication] and amine functionalized glasses.[@wang2006surface]
According to our calculations, the (001) surface of TiO$_2$ anatase can be modified when exposed to partial pressure of hydrogen atoms about 0.3–30 Pa. Such pressures can be obtained in the experiments. For example, Nakamura[@nakamura2000role] synthesized the plasma-treated TiO$_2$ photocatalyst. In his experiment, the chamber pressure was about 270 Pa. Dobele[@abdel2006determination] measured the H$_2$ dissociation degree in processing plasma, and found the dissociation ratio is about 4.85%. According to this H$_2$ dissociation ratio, the H atom partial pressure is roughly 13 Pa in Nakamura’s experiment. Therefore, the hydrogen atom pressure range of 0.3–30 Pa is within the experimental range of achievable hydrogen plasma conditions.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Safety and decline of road traffic accidents remain important issues of autonomous driving. Statistics show that unintended lane departure is a leading cause of worldwide motor vehicle collisions, making lane detection the most promising and challenge task for self-driving. Today, numerous groups are combining deep learning techniques with computer vision problems to solve self-driving problems. In this paper, a Global Convolution Networks (GCN) model is used to address both classification and localization issues for semantic segmentation of lane. We are using color-based segmentation is presented and the usability of the model is evaluated. A residual-based boundary refinement and Adam optimization is also used to achieve state-of-art performance. As normal cars could not afford GPUs on the car, and training session for a particular road could be shared by several cars. We propose a framework to get it work in real world. We build a real time video transfer system to get video from the car, get the model trained in edge server (which is equipped with GPUs), and send the trained model back to the car.'
author:
- 'Tejas Mahale Chaoran Chen and Wenhui Zhang \* equal contribution'
bibliography:
- 'sample-bibliography.bib'
title: 'End to End Video Segmentation for Driving : Lane Detection For Autonomous Car'
---
Introduction
============
Statistics show that unintended lane departure is a leading cause of worldwide vehicle collisions, which leads to substantial financial costs to both society and the individuals involved \[1\]. To reduce the number of traffic accidents and to improve safety, research on Driver Assistance System (DAS) have been conducted worldwide for many years and even expanded to autonomous cars. Autonomous car is always equipped with intelligent vehicle (IV) system, and lane detection is an important information for this system. In this system, some tasks including road following, keeping within the correct lane, maintaining a safe distance between vehicles, controlling the speed of a vehicle according to traffic conditions and road conditions, moving across lanes in order to overtake vehicles and avoid obstacles, searching for the correct and shortest route to a destination have to be completed \[2\]. The premise of all the tasks are high related to the lanes. Only when we know the accurate lanes will we finish above tasks better. Furthermore, The distance between the lane boundaries and obstacle is also important because it makes sure that the vehicle is in a safe distance from another vehicle or obstacles to avoid any possible collisions\[2\]. To obtain the information of the distance to the lane boundaries, we need to know the lane boundaries first. That is why lane detection is an important part of intelligent vehicle system.\
Lane detection mainly focus on locating both the center line and the edge of each lane in a road image. lane markings help differentiate the lanes from other characteristic objects on the road such as other vehicles, pedestrian, and animals running into the road. But there are some challenges we have to face. Firstly, does markings really mark the correct lane? As we know, there are various road conditions and environment conditions that lanes maybe not visible in a image. Sometimes the quality of the image can also be the reason. Apart from that, road splitting or merging and the interference from roadside objects or shadows could be another problem which will worsen the detection.\
However, the great variety of road environments necessitates the use of complex vision algorithms that not only requires expensive hardware to implement but also relies on many adjustable parameters that are typically determined from experience \[3\]. With the development of the deep learning techniques, numerous groups have applied a variety of deep learning techniques to computer vision problems \[4\].\
In this paper, we gather original data from Carla simulator In Implementation Details section, we do some pre-processing work to clean the data, which is a preparation for using a image segmentation techniques to deal with the lane detection problem. The method is based on color to differentiate different objects. Then in Experiments and Evaluation section, we conduct a Global Convolution Networks (GCN) model using color-based segmentation and use Adam as an optimizer to optimize the results. After obtaining the results(using test data to get training image), we compare the image we get with the actual image to evaluate the model.
In summary, we highlight our potential contributions below:
- This report introduces a fine grained architecture of GCN model;
- We implements a GCN network with optimization method of Residual-based boundary refinement and Adam optimization;
- This paper discusses and conducts performance diagnostics of lane detection tasks under various weather conditions.
- We proposed a framework to get the model trained and tested in real world.
The rest of this report is structured as follows. Section \[sec:related-works\] gives out overview for related works. Section \[sec:formular\] explains and gives out an overview on algorithms and math background of our model. In Section \[sec:implementation\] , we present our detailed implementation for data pre-processing pipeline and data augmentation. In Section \[sec:evaluation\] , we present our proposal for evaluating performance analysis for our lane detection model. It also explains performance of our model on train, validation and test data.
Related Work {#sec:related-works}
============
Image Segmentation Techniques
-----------------------------
Image segmentation is the operation of partitioning an image into a collection of connected sets of pixels. Main methods of region segmentation includes, region growing, clustering and split and merge. Region growing techniques start with one pixel of a potential region and try to grow it by adding adjacent pixels till the pixels being compared are too far from similar. The first pixel selected can be just the first unlabeled pixel in the image or a set of seed pixels can be chosen from the image. Usually a statistical test is used to decide which pixels can be added to a region. \[6\] Clustering methods includes K-means Clustering and Variants, Isodata Clustering, Histogram-Based Clustering and Recursive Variant and Global-Theoretic Clustering. \[7\] In some image sets, lines, curves, and circular arcs are more useful than regions or helpful in addition to regions. Thus people use lines and arcs segmentation as basis of image segmentation. Basic idea is looking for a neighborhood with strong signs of change, and detect them as edge. Edge detectors are based on differential operators. Differential operators attempt to approximate the gradient at a pixel via masks. And then threshold the gradient to select the edge pixels. One widely used edge detector used in auto driving lane detection in earlier days is Canny Edge Detector (CED). CED firstly smooth the image with a Gaussian filter, then compute gradient magnitude and direction at each pixel of the smoothed image. It zeros out any pixel response smaller than the two neighboring pixels on either side of it, along the direction of the gradient, and track high-magnitude contours. Afterwards it keeps only pixels along these contours, so weak little segments go away. \[8\]
These classical image segmentation methods are good, however they do not work well with raining or snowing road situations. It is hard for these classical image segmentation methods to handle various scenarios. Thus higher order image segmentation methods, such as deep learning’s convolution neural network (CNN) is a must in auto driving.
Structured Deep Learning
------------------------
Classical machine learning on structured datasets methods includes kernel-based methods and graph-based regularization techniques \[1, 2\]. However these machine learning requires a lot of feature engineering before certain tasks such as image classification. Furthermore, most of the time these features require domain knowledge, creativity and a lot of trial and error. Structured Deep Learning (SDL) is a fast, no domain knowledge requiring, and high performing machine learning method.
In the last couple of years, a number of papers re-visited machine learning on structured datasets, like graphs. Some works has done in generalizing neural networks to work on arbitrarily structured graphs \[4, 5\]. Some of them achieve promising results in domains that have previously been dominated by classical machine learning methods mentioned above. Kipf \[9\] came up with a graph convolutions method, which is good for image segmentation. Global convolutions are generalization of convolutions, and easiest to define in spectral domain. General Fourier transform scales poorly with size of data so we need relaxations. In this paper they use first order approximation in Fourier-domain to obtain an efficient linear-time graph-CNNs. We adapt this approach in our paper. We illustrate here what this first-order approximation amounts to on a 2D lattice one would normally use for image processing, where actual spatial convolutions are easy to compute in this application the modelling power of the proposed graph conv-networks is severely impoverished, due to the first-order and other approximations made. It uses concept of a convolution filter for image pixels or a linear array of signals.
Optimization Techniques on Gradients
------------------------------------
Optimization algorithms used to accelerate convergence includes stochastic gradient methods and stochastic momentum methods. Stochastic gradient methods can generally be written $$\label{eq:desc}
w_{k+1} = w_k - \alpha_k\, \tilde \nabla f(w_k),$$ where $\tilde \nabla f(w_k) := \nabla f(w_k; x_{i_k})$ is the gradient of some loss function $f$ computed on a batch of data $x_{i_k}$.
Stochastic momentum methods have been used to accelerate training. These methods could be written as $$\begin{aligned}
\label{eq:gen-mom}
w_{k+1} &= w_k - \alpha_k \, \tilde \nabla f(w_k + \gamma_k (w_k-w_{k-1})) + \beta_k(w_{k} - w_{k-1}).\end{aligned}$$ Sequence of iterates includes Polyak’s heavy-ball method (HB) with $\gamma_k = 0$, and Nesterov’s Accelerated Gradient method (NAG) with $\gamma_k = \beta_k$.
However there are some exceptions as well, such as and . These are adaptive gradient and adaptive momentum methods. These methods construct the entire sequence of iterates, say $(w_1, \cdots, w_k)$ as a local distance measure. AdaGrad \[10\], RMSProp \[11\], and Adam \[12\] can generally be written as $$\begin{aligned}
\label{eq:adap-mom}
w_{k+1} &= w_k - \alpha_k \mathrm{H}_k^{-1}\tilde \nabla f' +\beta_k\mathrm{H}_k^{-1} \mathrm{H}_{k-1} (w_k - w_{k-1}),\end{aligned}$$ where $\tilde \nabla f'$ is $\tilde \nabla f(w_k + \gamma_k(w_k-w_{k-1}))$, and $\mathrm{H}_k:= H(w_1, \cdots, w_k)$ is a positive definite matrix. $\mathrm{H}_k$ is a diagonal matrix. Its entries are defined as square roots of a linear combination of squares of past gradient components.
In deep learning acceleration of convergence, specific settings of the parameters are stated in Table \[table:update\]. In this table, ${D}_k$ = $({g}_k\circ{g}_k)$ and ${G}_k$ = ${H}_k\circ{H}_k$.
As we could see from Table \[table:update\], adaptive methods change and adapt to geometry of the data. However, stochastic gradient descent and related variants use the $\ell_2$ geometry inherent. This is considered as equivalent to making$\mathrm{H}_k=\mathrm{I}$ in adaptive methods. Performance is defined as loss function than the function $f$ used in training. Thus, it seems like Adam is more favorable in our GCN project.
SGD HB NAG AdaGrad RMSProp Adam
---------------- -------------- -------------- -------------- ------------------------------------ --------------------------------------------------------- --------------------------------------------------------------------------------------------------
$\mathrm{G}_k$ $\mathrm{I}$ $\mathrm{I}$ $\mathrm{I}$ $\mathrm{G_{k-1}} + \mathrm{D}_k$ $\beta_2 \mathrm{G_{k-1}} + (1 - \beta_2) \mathrm{D}_k$ $\frac{\beta_2}{1- \beta_2^k} \mathrm{G_{k-1}} + \frac{(1 - \beta_2)}{1- \beta_2^k}\mathrm{D}_k$
$\alpha_k$ $\alpha$ $\alpha$ $\alpha$ $\alpha$ $\alpha$ $\alpha \frac{1-\beta_1}{1- \beta_1^k}$
$\beta_k$ 0 $\beta$ $\beta$ 0 0 $\frac{\beta_1(1 - \beta_1^{k-1})}{1 - \beta_1^k}$
$\gamma$ 0 0 $\beta$ 0 0 0
: Parameter Settings of Optimization Techniques used in Deep Learning. []{data-label="table:update"}
Formulation {#sec:formular}
===========
Lane detection involves following steps. Firstly, compute the camera calibration matrix and distortion coefficients given a set of chessboard images. Then, apply a distortion correction to raw images, and use color transforms, gradients, etc., and apply deep learning for classification analysis to create a threshold binary image. Afterwards, from a birds-eye view, apply a perspective transform to rectify binary image. Then, detect lane pixels and fit to find the lane boundary, and determine the curvature of the lane and vehicle position with respect to center. The last step is to warp the detected lane boundaries back onto the original image. Last but not least, put visual display of the lane boundaries and numerical estimation of lane curvature and vehicle position.
![Details of BR[]{data-label="fig:BR"}](GCN.png){width="100.00000%"}
![Details of BR[]{data-label="fig:BR"}](BR.png){width="100.00000%"}
In this paper, the GCN utilizes connectivity structure of graph as the filter to perform neighborhood mixing. Its architecture may be elegantly summarized as: $$H^{(l+1)}=\sigma(\hat{A}H^{(l)}W^{(l)}),$$ where $\hat{A}$ is normalization of the graph adjacency matrix, $H^{(l)}$ is row-wise embedding of graph vertices in the $l$th layer, $W^{(l)}$ is a parameter matrix, and $\sigma$ is non-linearity.
in this paper we consider a building block defined as: $y= F(x, W_{i}) + x.$ Here $x$ and $y$ are the input and output vectors of a particular layer. The function $F(x, W_{i})$ represents learning residual mapping. $F=W_{2} * \sigma (W_{1}x)$ in which $\sigma$ denotes ReLU and the biases are omitted for simplifying notations. The operation $F+x$ is performed by a shortcut connection and element-wise addition. We adopt the second non-linearity after the addition.
{width="1\linewidth"}
In this section, we first propose a novel GCN to address the contradictory aspects — classification and localization in semantic segmentation. Then using GCN \[fig:GCN\] we design a fully-convolutional framework for semantic segmentation task. Comparing with other many semantic segmentation models, our model has faster training speeds with smaller model size, while keeping similar accuracy. In features extraction stage, we use the max pooling module following by conv-layers and an up-sampling layer. After these two modules, feature maps has same size as input image. When the enlargement factor is large, it brings noise and make boundary pixels of two areas difficult to classify. And that is why we add refinement module at the end of the model to refine the segmentation area \[fig:BR\] In refinement module, we use combination of convolution layers and pooling layers, with a conv-net with ReLU, and then process it with convolution of size 3x3x21x21.
In convolution neural network training, network weights are adjusted by loss function evaluation. As images vary from class to class, influence of each class to the loss is different. In each iteration, we calculate the weights based on current input batch. Weights are different in each iteration. The weights is calculated as in Eq.\[equ1\]. $$w_i = \left\{
\begin{array}{lc}
1, & n_i = 0 \\
\beta, & w_i < \beta\\
\dfrac{N}{2*c*n_i}, & \beta < w_i < \alpha\\
\alpha, & w_i > \alpha
\end{array}
\right.
\label{equ1}$$
Where, $w_i$ is weight of class $i$, $c$ is class number, and value of $i$ is from 0 to $c$. $\beta$ and $\alpha$ are lower and upper threshold of $w_i$, we set threshold to avoid excessive weights differences. $N$ is the total pixel number of this batch, $n_i$ is the pixel number of class $i$, when $n_i$ = 0, it means that the class $i$ does not appear in this batch, we set the weight to 1. Because we need to increase the effect of small pixel number class on loss, so the smaller the $n_i$, the larger the $w_i$ is. $N$ and $c$ are constant, $w_i$ is just changed by $n_i$. When the $n_i$ is the average number, $w_i$ is calculated to be $\dfrac{1}{2}$, the multiplicative coefficient of $\dfrac{1}{2}$ is also used to decrease the $w_i$ of large pixel number of class. The loss function is shown in Eq.\[equ2\], where $x_{ij}$, $y_{ij}$ are prediction class and label in pixel (i, j), $w$ is the loss weights.
$$LOSS = \sum_{i = 1}\sum_{j = 1}w\left \| x_{ij} - y_{ij} \right \|^2
\label{equ2}$$
And the whole architecture could be seen as shown in Fig \[fig:whole-pipeline\].
Encode - Decoder
----------------
To decodify our depth first we get the int24. $R + G*256 + B*256*256$. Then normalize it in the range $[0, 1]$. $Ans / ( 256*256*256 - 1 )$. And finally multiply for the units that we want to get. We have set the far plane at 1000 metres. Ans \* far. Encoded stage includes encoding image to segmented image. Every pixel of label image is red channel class label. We use retained class labels with road and vehicle and converted them to green and blue respectively. Encoder network consist of convolution layers, batch normalization and max pooling for down-sampling\[13\]. For classification problems, multiple convolution layers with decrease in size of kernel and increase in number of filters with each layer pooling is advised\[14\]. On other hand decoder block consists of strided convolution and up-sampler which will re-size image into original shape. When we transfer the trained image back to color map to be project to videos, we just use $blue = value/(256*256)$, $ green = (value - blue*256*256)/256$ and $red = value - green*256 + blue*256*256$.
GCN and BR
----------
There are two tasks in segmentation, classification and localization. Conical CNN architectures are good for classifications. In case of object segmentation, large kernel size plays important role in term of object localization\[15\]. But usage of large kernels is heavy on weights of model as kernel size increases, number of parameters also increases. To achieve global convolution (GCN) performance with less weights, GCN was divided into two one dimensional kernels. Effect of both kernels added at end to get actual GCN output. Boundary refinement block(BR) is residual block which helps to get boundary structure of shape properly.
Adam
----
In this paper, we use Adam. Similar to momentum, it keeps exponentially decaying of past gradients $m_t$. It compute bias-corrected first and second moment estimates:
$$\begin{aligned}
\begin{split}
\hat{m}_t &= \frac{m_t}{1 - \beta^t_1}\\
\hat{v}_t &= \frac{v_t}{1 - \beta^t_2}
\end{split}\end{aligned}$$
$m_t$ and $v_t$ are estimates of the first moment and the second moment of the gradients respectively. Parameters are updated using: $$\theta_{t+1} = \theta_{t} - \frac{\eta}{\sqrt{\hat{v}_t} + \epsilon} \hat{m}_t$$
In this paper, we use default values of $0.9$ for $\beta_1$, $0.999$ for $\beta_2$, and $10^{-8}$ for $\epsilon$.
Implementation Details {#sec:implementation}
======================
Dataset
-------
First, we gathered our 3000 road images from a platform named Carla Simulator. It is an open-source simulator especially for autonomous driving research. It is a good tool since it provides various different images in a variety of circumstances including different types of weather conditions and different time of day. Through driving a car in urban conditions, we could get train images(shown in Figure \[original image\]) and corresponding encoded label images(shown in Figure \[converted image\]). 3,000 images along with semantic labels with resolution 600 x 800 x 3 are collected as our dataset.
![converted image[]{data-label="converted image"}](Picture1.png){width="100.00000%"}
![converted image[]{data-label="converted image"}](Picture2.png){width="100.00000%"}
Data Pre-processing
-------------------
To reduce the workload of computer and accelerate the running speed, car hood and sky portion are cropped in each image for 240\*800 pixels(From Figure \[600\*800\*3\] to Figure \[360\*800\*3\]). That leads to a dataset consisting of 360\*800 images of the red, green, and blue color channels.
![360\*800\*3[]{data-label="360*800*3"}](Picture3.png){width="100.00000%"}
![360\*800\*3[]{data-label="360*800*3"}](Picture4.png){width="100.00000%"}
The server provides an image with the tag information encoded in the red channel. A pixel with a red value of x displays an object with tag x. The following tags are displayed (shown in Figure \[fig:tags\]). For training purposes, encoded image(shown in Figure \[fig:encoded image\]) should be transformed to segmented image(shown in Figure \[fig:segmemted image\]). We retained class labels with road and vehicle and converted them to green and blue respectively.
![Tags[]{data-label="fig:tags"}](Picture0.png){width="0.4\linewidth"}
![encoded image to Segmented image(1)[]{data-label="fig:encoded image"}](Picture5.png){width="0.7\linewidth"}
![encoded image to Segmented image(2)[]{data-label="fig:segmemted image"}](Picture6.png){width="0.7\linewidth"}
Data Augmentation
-----------------
To ensure the diversity of dataset, we made some data augmentation. As is shown in Figure \[fig:rotation0\] and Figure \[fig:rotation1\], Rotation (0, 30 degree) like that is given to the original dataset to get another 3000 images, while in Figure \[fig:shifting0\] and Figure \[fig:shifting1\], shifting (Width = (0, 0.2), height = (0, 0.1)) like that is given to the original dataset as well to get other 3000 images. In total, we obtained 9000 images dataset at present, which is of course a large amount of images and also make sure the diversity of the dataset.
![Shifting[]{data-label="fig:shifting0"}](Picture8.png){width="100.00000%"}
![Shifting[]{data-label="fig:shifting0"}](Picture10.png){width="100.00000%"}
![shifting[]{data-label="fig:shifting1"}](Picture9.png){width="100.00000%"}
![shifting[]{data-label="fig:shifting1"}](Picture11.png){width="100.00000%"}
Evaluation and Discussion {#sec:evaluation}
=========================
Metric
------
In this paper, evaluation on both Global Convolution Network and Boundary refinement are conducted. We evaluate our GCN based lane detection model in terms of minimum loss performance with help of minimum square error and mean average error. Our experiment setup is composed of one class of server. For training and experiments performed on Tesla P100-PCIE gpu memoryclock rate 1.3285 with 11.95 GB video memory on Linux distribution.Conda environment is used with tensorflow gpu v1.10, Keras gpu v2.2.2, python version 3.5.
Our implementation for training is as follows. We re-size images with its shorter side randomly sampled in 600 x 800 x 3 for scale augmentation 360 x 800 x 3 resolution. The standard color augmentation mentioned in previous section is used. Batch normalization is adopted right after each convolution layer before activation. We initialize the weights as per Xavier initializer, and train encoder nets from scratch. We use Adam with a global-batch size of 64. The learning rate starts from 0.001 and the models are trained for up to 40 iterations. We use a weight decay of 0.9 and a momentum of 0.999.
In order to judge whether a lane segmentation is successfully detected, we view lane markings as two classified value and calculate the Minimum Square Error (MSE) and Mean Absolute Error (MAE) between the ground truth and the prediction. MSE is average of square of pixel to pixel difference between two images where as MAE is pixel to pixel absolute difference between images.
MSE = $\frac{1}{N}\sum_{n=1} ^ {N} (Y_{true} - Y_{predicted})^ {2}$\
MAE = $\frac{1}{N}\sum_{n=1}^{N} |(Y_{true} - Y_{predicted})|$
Hyper-parameter tuning
----------------------
Hyper-parameter tuning is important process of deep learning algorithms. In this project, we have implemented encoder-decoder architecture which has parameters like number of filters in each encoder layer, filter size, learning rate of optimizer, number of iterations to get minimum loss.\
Parameter Value
----------------------------- ---------------
Optimizer Adam
Learning rate 0.001
Number of iterations 40
Number of Filters per layer 8, 16, 20, 32
Table 5.1
As explained earlier, we implemented conical structure for encoder blocks which consist of convolution layer, batch normalization and max pooling to half the image resolution at every step along with gradual increase in number of filters. We trained our model for 5-10 iterations of various combination of hyper-parameter and observed the decay in mean square loss. Finally we came with hyper-parameter settings that give minimum loss at 40th iteration.
Training and Validation Results
-------------------------------
We divided our image dataset into train, validation and test categories. We trained our model on train data and tuned hyper-parameter on validation data.
Minimum Square Error (MSE)
Iterations Training (MSE) Validation (MSE)
---------------- ---------------- ------------------
0 Initial loss 1592.1951 1443.7032
10 Iterations 192.1637 337.5560
20 Iterations 102.3213 119.4761
30 Iterations 74.0532 89.0769
40 Iterations 42.7771 61.4360
Table 5.2
Mean Average Error (MAE)
Iterations Training (MAE) Validation (MAE)
---------------- ---------------- ------------------
0 Initial loss 13.3800 13.3908
10 Iterations 3.1369 6.8221
20 Iterations 2.9696 3.7389
30 Iterations 2.0202 2.8999
40 Iterations 1.7011 2.4999
\
Table 5.3
After 40 iterations, train loss was further decreasing but on same time validation loss started increasing, so we had to stop training process. We can see final iteration values of MAE and MSE and conclude that we have got optimized parameters with minimum loss and less over-fitting.
![MSE for Train and Validation []{data-label="fig:1"}](Lane_Curve.PNG){width="190.00000%"}
Test Results
------------
Once we concurred best hyper-parameters for minimum loss on given model with help of MSE and MAE values, we tested our model on reserved test dataset to get unbiased evaluation of our model.
Metric Test set evaluation
---------------------- ---------------------
Minimum Square Error 57.5875
Mean Absolute Error 2.2104
\
Table 5.4
From table 5.2 and 5.4, we can compare MSE values on train, validation and test set. There is not much difference in train and validation MSE. Also, validation and train values are almost in range. It implies that model is not facing over-fitting problem.
![Test Image 2[]{data-label="fig:t2"}](Og1.png){width="110.00000%"}
![Test Image 2[]{data-label="fig:t2"}](Og2.png){width="110.00000%"}
![Ground truth of Image 2[]{data-label="fig:g2"}](Lb1.PNG){width="110.00000%"}
![Ground truth of Image 2[]{data-label="fig:g2"}](lb2.PNG){width="110.00000%"}
![Model Prediction image 2[]{data-label="fig:m2"}](00.jpg){width="110.00000%"}
MSE = 48.8590
![Model Prediction image 2[]{data-label="fig:m2"}](11.jpg){width="110.00000%"}
MSE = 63.3626
We calculated minimum square error for test image1 and test image2 from test dataset. Both images were able to capture shape of vehicle as well as lane boundaries. From ground truth label image and model predicted image, we compared MSE score of both images.
Method MSE
----------------------------- -------------
FCN 8s 65.3
DPN 59.1
CRFasRNN 62.5
Scale invariant $CNN + CRF$ 66.3
Dilation10 67.1
DeepLabv2-CRF 70.4
Adelaide\_context 71.6
LRR-4x 71.8
Large-Batch 76.9
**Our approach** **57.5875**
\
Table 5.5 Experimental results on test set
Conclusion {#sec:conclusion}
==========
In this paper, a Global Convolution Networks (GCN) model is used to address both classification and localization issues in lane detection. A residual-based boundary refinement and Adam optimization is also used to achieve 57.5875 performance. We also build a picamera on RaspberryPi and put it on a car, this will send real time video to our edge server. On edge server side, we will have training scripts running and sending trained GCN model back to the car. The code could be found here:
$https://github.com/wenhuizhang/auto\_driving\_car$
References
==========
\[1\] Narote, S. P., Bhujbal, P. N., Narote, A. S., & Dhane, D. M. (2018). A review of recent advances in lane detection and departure warning system. Pattern Recognition, 73, 216-234.
\[2\] Chiu, K. Y., Lin, S. F. (2005, June). Lane detection using color-based segmentation. In Intelligent Vehicles Symposium, 2005. Proceedings. IEEE (pp. 706-711).
\[3\] Yim, Y. U., Oh, S. Y. (2003). Three-feature based automatic lane detection algorithm (TFALDA) for autonomous driving. IEEE Transactions on Intelligent Transportation Systems, 4(4), 219-225.
\[4\]Huval, B., Wang, T., Tandon, S., Kiske, J., Song, W., Pazhayampallil, J., ... & Mujica, F. (2015). An empirical evaluation of deep learning on highway driving. arXiv preprint arXiv:1504.01716.
\[3\] Martin Abadi et al. TensorFlow: Large-scale machine learning on heterogeneous systems, 2015.
\[4\] Bruna, J., Zaremba, W., Szlam, A., & LeCun, Y. (2013). Spectral networks and locally connected networks on graphs. arXiv preprint arXiv:1312.6203.
\[5\] Henaff, M., Bruna, J., & LeCun, Y. (2015). Deep convolutional networks on graph-structured data. arXiv preprint arXiv:1506.05163.
\[6\] Al-Hujazi, E., & Sood, A. (1990). Range image segmentation combining edge-detection and region-growing techniques with applications sto robot bin-picking using vacuum gripper. IEEE transactions on systems, man, and cybernetics, 20(6), 1313-1325.
\[7\] Lucchese, L., & Mitra, S. K. (2001). Colour image segmentation: a state-of-the-art survey. Proceedings-Indian National Science Academy Part A, 67(2), 207-222.
\[8\] Ding, L., & Goshtasby, A. (2001). On the Canny edge detector. Pattern Recognition, 34(3), 721-725.
\[9\] Kipf, T. N., & Welling, M. (2016). Semi-supervised classification with graph convolutional networks. arXiv preprint arXiv:1609.02907.
\[10\] Duchi, J., Hazan, E., & Singer, Y. (2011). Adaptive subgradient methods for online learning and stochastic optimization. Journal of Machine Learning Research, 12(Jul), 2121-2159.
\[11\] Tieleman, T., & Hinton, G. (2012). Rmsprop: Divide the gradient by a running average of its recent magnitude. coursera: Neural networks for machine learning. COURSERA Neural Networks Mach. Learn.
\[12\] Kingma, D. P., & Ba, J. (2014). Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980.
\[13\] “Autoencoders, Unsupervised Learning, and Deep Architectures”, Pierre Baldi, 2012
\[14\] “Very Deep Convolutional Networks for Large-Scale Image Recognition”, Karen Simonyan, Andrew Zisserman, September 2014
\[15\] “Large Kernel Matters – Improve Semantic Segmentation by Global Convolutional Network”, Chao Peng, Xiangyu Zhang, Gang Yu, Guiming Luo, Jian Sun, March 17
| {
"pile_set_name": "ArXiv"
} |
2em
[**Yang-Mills solutions on de Sitter space of any dimension**]{}
5em
Olaf Lechtenfeld$^{*\dagger}$ [and]{} Gönül Ünal$^\dagger$
2em
*$^*$ Institut für Theoretische Physik*
*$^\dagger$ Riemann Center for Geometry and Physics*
0.5em
*Leibniz Universität Hannover, Appelstra[ß]{}e 2, 30167 Hannover, Germany*
4em
> [**Abstract**]{}
>
> For gauge groups SO$(n{+}1)$, SU$(m{+}1)$ and Sp$(\ell{+}1)$, we construct equivariant Yang–Mills solutions on de Sitter space in $n{+}1$, $2(m{+}1)$ and $4(\ell{+}1)$ spacetime dimensions. The latter is conformally mapped to a finite cylinder over a coset space realizing an appropriate unit sphere. The equivariance condition reduces the Yang–Mills system to an analog Newtonian particle in one or two dimensions subject to a time-dependent friction and a particular potential. We analyze some properties of the solutions such as their action and energy and display all analytic ones. Beyond dS$_4$ all such configurations have finite energy but infinite action.
Introduction and summary
========================
Analytic solutions of the Yang–Mills equations are not easy to come by in any spacetime manifold of dimension larger than two. Yet, they play a central role in the analysis of semiclassical behavior, vacuum structure and admissible string backgrounds, to name a few. Imposing a sufficient amount of symmetry on Yang–Mills solutions, however, reduces the coupled system of nonlinear partial differential equations to ordinary (albeit cubic) matrix differential equations, which is amenable to solving explicitly. The simplest situation is that of maximally symmetric spacetimes and the gauge group being equal to its maximal compact symmetry subgroup $G$. Restricting attention to $G$-equivariant Yang–Mills configurations then further simplifies the matrix equations of motion to scalar equations, which can be interpreted as Newtonian dynamics of a particle in some Euclidean space with a particular quartic potential. The latter system sometimes allows for consistent one-dimensional double-well subsystems, yielding special explicit solutions.
A lot is known about analytic Yang–Mills solutions in Minkowski space $\mathbb{R}^{1,n}$ [@Actor]. Typically, however, these are singular or have infinite energy, unless a Higgs field is added. The situation is better on de Sitter space in various spacetime dimensions $n{+}1$, because equal-time slices may be chosen compact and the metric regulates the action integral at early and late times. The closed slicing of dS$_{n+1}$ describes it as a cosh-cylinder over the $n$-sphere $S^n$, with a $G$-action on $S^n$. Representing the unit sphere as a coset space $G/H$ via $$\label{spherecosets}
S^n \= \sfrac{\textrm{SO}(n{+}1)}{\textrm{SO}(n)} \qquad\textrm{or}\qquad
S^{2m+1} \= \sfrac{\textrm{SU}(m{+}1)}{\textrm{SU}(m)} \qquad\textrm{or}\qquad
S^{4\ell+3} \= \sfrac{\textrm{Sp}(\ell{+}1)}{\textrm{Sp}(\ell)}$$ suggests to look for Yang–Mills solutions with gauge groups SO$(n{+}1)$, SU$(m{+}1)$ or Sp$(\ell{+}1)$ on the spaces dS$_{n+1}$, dS$_{2m+2}$ or dS$_{4\ell+4}$, respectively. Indeed, classical finite-energy and finite-action SU(2) pure Yang–Mills fields have been found on dS$_4$ [@IvLePo1; @IvLePo2], employing the isomorphy of $S^3$ to SU(2) and a conformal map from dS$_4$ to ${\cal I}\times S^3$, where ${\cal I}$ is a finite time interval of length $\pi$.
It is thus natural to extend this analysis to higher-dimensional de Sitter spaces in the orthogonal, unitary and symplectic incarnation based on (\[spherecosets\]). This is the subject of the paper. After a geometric description of de Sitter space and the various spherical cosets in the following section, we perform the equivariant reduction of the Yang–Mills equations on the orthogonal, unitary and symplectic cosets, arriving at a one- or two-dimensional Newtonian system with a particular quartic potential and a peculiar friction term. The latter arises away from four spacetime dimensions due to the non-invariance of the Yang–Mills equations under the conformal map $$\textrm{dS}_{n+1} \ \to\ {\cal I}\times S^n \qquad\textrm{with}\quad{\cal I}=\bigl(-\sfrac{\pi}{2},\sfrac{\pi}{2}\bigr) \ .$$ For all three cases, we compute the equations of motion, Newtonian potential, action and energy of the analog-particle system and write down the Yang–Mills fields in terms of the particle trajectory. The final section discusses properties of generic and special solutions. As expected, above four spacetime dimensions all solutions have infinite action but finite energy. The friction is detrimental below four dimensions as it renders the field strengths singular at the temporal boundary. However, purely color-magnetic solutions avoid the friction, and there exists an analytic one in every dimension. Corresponding to the unique local maximum of the Newtonian potential, it is unstable and given by the canonical $H$-connection.
In the critical spacetime dimension four, one may use either SU(2) or $\frac{\textrm{SO}(4)}{\textrm{SO}(3)}$ to describe the three-sphere. The absence of the friction term leads to a family of Abelian and of non-Abelian solutions, where the latter are given in terms of elliptic functions which solve the mechanical double-well problem. We reproduce the known SU(2) results (where the color-magnetic solution is [*half*]{} the canonical SU(2) connection) [@AFF; @Luscher; @Gibbons] and complement them with corresponding SO(4) solutions, which however turn out to be equivalent.
We have this constructed new explicit finite-energy (but infinite-action) Yang–Mills configurations on de Sitter space of any dimension. The story may be repeated for other coset representations, such as $S^6=\frac{G_2}{\textrm{SU}(3)}$. Finally, our spacetime background is non-dynamical; we do not consider its gravitational backreaction. The result of this effect is known for the four-dimensional Yang–Mills solutions (see e.g. Section 8 of [@Volkov] and references therein) and may be generalized to higher dimensions.
Description of de Sitter space
==============================
de Sitter space dS$_{n+1}$ as a cylinder over $S^n$
---------------------------------------------------
It is well-known that $(n+1)$-dimensional de Sitter space dS$_{n+1}$ can be embedded into $(n+2)$-dimensional Minkowski space $\mathbb{R}^{n+1,1}$ as $$\delta_{\alpha\beta} y^\alpha y^\beta-(y^{n+2})^2\=R^2 \quad \text{where} \quad \alpha,\beta=1,\dots, n{+}1\,,$$ with the metric $$\dd s^2\=\dd y^\alpha\dd y^\beta-\dd y^{n+2}\dd y^{n+2}\,.
\label{flatmetric}$$ Topologically de Sitter space dS$_{n+1}$ is a cylinder over $S^n$, which is easily seen from the parametrization $\{\tau,\omega^\alpha\}$, $$y^\alpha=R\,\omega^\alpha \cosh \tau \und y^{n+2}=R\,\sinh \tau \quad\with \omega^\alpha\omega^\alpha=1\,,$$ where ${-}\infty{<}\tau{<}\infty$ and $\omega^\alpha$ embed the unit $n$-sphere into $\mathbb{R}^{n+1}$. An explicit form of the $\omega^\alpha$ can be found in terms of the global coordinates $\{\varphi,\theta_1,\ldots,\theta_{n-1}\}$ as $$\omega^1=\rhọ_{n-1} \sin \varphi\,,\quad \omega^2=\rhọ_{n-1} \cos \varphi\,, \quad \omega^k=\rhọ_{n-k+1} \cos \theta_{k-2}
\for 3\leq k\leq n+1$$ where $$\rho_\ell=\prod_{m=1}^\ell \sin \theta_{n-m}\,,\quad \rho_0=1\,,\quad 0\leq \varphi\leq 2\pi\,, \quad 0\leq \theta_1,\ldots,\theta_{n-1}<\pi\,.$$ With the help of this parametrization, the flat metric (\[flatmetric\]) induces on de Sitter space the metric $$\dd s^2\=R^2\bigl(-\dd\tau^2+\cosh^2\!\tau\,\dd\Omega^2_{n}\bigr)
\label{formalmetric}$$ where $\dd\Omega^2_{n}$ is the metric on the unit $n$-sphere. From the coordinates $\{\tau,\varphi,\theta_1,\ldots,\theta_{n-1}\}$ we pass to conformal coordinates $\{t,\varphi,\theta_a\}$ by the time reparametrization [@HawkingEllis] $$t=\arctan(\sinh \tau)=2\arctan(\tanh\sfrac{\tau}{2})\qquad \textrm{with} \quad \dv{\tau}{t}=\cosh\tau=\frac{1}{\cos t}\,,
\label{timereparametrization}$$ so that $$t \in (-\sfrac{\pi}{2}, \sfrac{\pi}{2}) \= \mathcal{I}$$ ranges over the finite interval $\mathcal{I}$, and the metric becomes $$\dd s^2\=\frac{R^2}{\cos^2\!t}\bigl(-\dd t^2+\dd\Omega^2_{n}\bigr)\,.
\label{metric}$$ We see that de Sitter space dS$_{n+1}$ is conformally equivalent to a cylinder $\mathcal{I}\times S^n$ over the $n$-sphere, with a conformal factor of $\frac{R^2}{\cos^2\!t}$.
$S^n$ as coset $\textrm{SO}(n{+}1)/\textrm{SO}(n)$
--------------------------------------------------
In this subsection, we briefly summarize the well-known geometric properties of $S^n$ realized as the coset space $\frac{SO(n{+}1)}{SO(n)}$. We denote by $\lbrace I_A\rbrace$ with $A=1,\ldots,\frac{n(n+1)}{2}$ the $SO(n{+}1)$ generators satisfying $$[I_A, I_B]={f_{AB}}^C I_C\,.
\label{commutationG}$$ The structure constants ${f_{AB}}^C$ give rise to the Cartan–Killing metric $g$ on the Lie algebra $so(n{+}1)$, $$g_{AB}\={f_{AD}}^C{f_{CB}}^D\=-\tr_{\textrm{adj}}\bigl(I_A I_B\bigr)\=\varkappa_n\,\delta_{AB}\,,
\label{metricG}$$ where the generators are taken in the adjoint representation and $\varkappa_n$ is some constant depending on their normalization.
Let us decompose $$so(n{+}1) = so(n) \oplus \mathfrak{m}$$ so that $\mathfrak{m}$ is the orthogonal complement of the Lie subalgebra $so(n)$ inside $so(n{+}1)$. Then, the set of $SO(n{+}1)$ generators splits into two subsets, $$\lbrace I_A\rbrace=\lbrace I_a\rbrace \cup \lbrace I_i\rbrace \quad\where
I_a\in\mathfrak{m} \and I_i\in so(n)$$ with the index ranges $a=1,\ldots,n$ and $i=n{+}1,\ldots,\sfrac{n(n{+}1)}{2}$. Since this coset is a symmetric space, ${f_{ab}}^c=0$, and the commutation relations (\[commutationG\]) decompose as $$[I_i, I_j]={f_{ij}}^kI_k\,,\qquad [I_i, I_a]={f_{ia}}^bI_b\,,\qquad [I_a, I_b]={f_{ab}}^i I_i$$ and the components of the Cartan–Killing metric (\[metricG\]) read $$g_{ij}={f_{ik}}^l{f_{lj}}^k+{f_{ia}}^b{f_{bj}}^a=\varkappa_n\,\delta_{ij}\,,\qquad
g_{ab}=2{f_{ad}}^i{f_{ib}}^d=\varkappa_n\,\delta_{ab}\,,\qquad g_{ia}=0\,.
\label{Killing-2}$$ We note that $S^1=\textrm{SO}(2)$ is special since the index $i$ takes no values, hence most terms are absent.
$S^{2m{+}1}$ as coset $\textrm{SU}(m{+}1)/\textrm{SU}(m)$
---------------------------------------------------------
Odd-dimensional spheres $S^{2m+1}$ can also be realized as (non-symmetric) coset spaces $\frac{\textrm{SU}(m{+}1)}{\textrm{SU}(m)}$. Let $\{\tilde{I}_{A}\}$ with $A=1,\ldots,m(m{+}2)$ be the generators of $\textrm{SU}(m{+}1)$ subject to $$[\tilde{I}_{A},\tilde{I}_{B}]=\tilde {f}_{AB}^{\hspace{2.5ex}C}\tilde{I}_{C}\,.
\label{commutationGtilde}$$ Analogous to the previous subsection, we decompose $$su(m{+}1) = su(m) \oplus \tilde{\mathfrak{m}}$$ and divide $$\{\tilde{I}_A\}=\{\tilde{I}_a\} \cup \{\tilde{I}_i\} \quad\where
\tilde{I}_a\in\tilde{\mathfrak{m}} \and \tilde{I}_i\in su(m)$$ with $a=1,\ldots,2m{+}1$ and $i=2m{+}2,\ldots,m(m{+}2)$. In this case, the commutation relations (\[commutationGtilde\]) read $$[\tilde{I}_i, \tilde{I}_j]=\tilde{f}_{ij}^{\hspace{1.5ex}k}\tilde{I}_k\,,\qquad
[\tilde{I}_i, \tilde{I}_a]=\tilde{f}_{ia}^{\hspace{1.5ex}b}\tilde{I}_b\,,\qquad
[\tilde{I}_a, \tilde{I}_b]=\tilde{f}_{ab}^{\hspace{1.5ex}i} \tilde{I}_i+\tilde{f}_{ab}^{\hspace{1.5ex}c}\tilde{I}_c\,.
\label{splitcom}$$ The normalization of the Cartan–Killing metric $$\tilde{g}_{AB}\=\tilde{f}_{AD}^{\hspace{2.5ex}C}\tilde{f}_{CB}^{\hspace{2.5ex}D}\=-\tr_{\textrm{adj}}\bigl(\tilde{I}_A\tilde{I}_B\bigr)\=\tilde{\varkappa}_m\,\delta_{AB}
\label{metricG2}$$ implies that $$\tilde{g}_{ij}=\tilde{f}_{ik}^{\hspace{1.5ex}l}\tilde{f}_{lj}^{\hspace{1.5ex}k}+\tilde{f}_{ia}^{\hspace{1.5ex}b}\tilde{f}_{bj}^{\hspace{1.5ex}a}=\tilde{\varkappa}_m\,\delta_{ij}\,,\qquad
\tilde{g}_{ab}=2\tilde{f}_{ad}^{\hspace{1.5ex}i}\tilde{f}_{ib}^{\hspace{1.5ex}d}+\tilde{f}_{ad}^{\hspace{1.5ex}c}\tilde{f}_{cb}^{\hspace{1.5ex}d}=\tilde{\varkappa}_m\,\delta_{ab}\,,\qquad
\tilde{g}_{ia}=0\,.
\label{Killing-3}$$ We remark that $S^3=\textrm{SU}(2)$ is particular since the index $i$ takes no values, so that many terms are vanishing.
$S^{4\ell+3}$ as coset $\textrm{Sp}(\ell{+}1)/\textrm{Sp}(\ell)$
----------------------------------------------------------------
Finally, there exists a symplectic coset realization of $S^{4\ell+3}$ as $\frac{\textrm{Sp}(\ell{+}1)}{\textrm{Sp}(\ell)}$. The $\textrm{Sp}(\ell{+}1)$ generators $\hat{I}_A$ with $A=1,\ldots,(\ell{+}1)(2\ell{+}3)$ obey $$[\hat{I}_A,\hat{I}_B]=\hat {f}_{AB}^{\hspace{2.5ex}C}\hat{I}_{C}\,.
\label{commutationGhat}$$ Splitting $$sp(\ell{+}1) = sp(\ell) \oplus \hat{\mathfrak{m}} \und
\{\hat{I}_a\} = \{\hat{I}_a\}\cup\{\hat{I}_i\} \quad\where
\hat{I}_a\in\hat{\mathfrak{m}} \and \hat{I}_i\in sp(\ell)$$ with $a=1,\ldots,4\ell{+}3$ and $i=4\ell{+}4,\ldots,(\ell{+}1)(2\ell{+}3)$, the commutation relations (\[commutationGhat\]) decompose as in (\[splitcom\]). Analogously, relations (\[metricG2\]) and (\[Killing-3\]) hold, with tildes replaced by hats and a normalization constant $\hat{\varkappa}_\ell$.
Connections on spheres
----------------------
To define a connection on $S^n$, we introduce orthonormal frames as follows: Let $\{{\hat E}_A\}$ be the left-invariant vector fields on $\textrm{SO}(n{+}1)$ satisfying the same commutation relations as the generators $I_A$, and let $\{{\hat e}^A\}$ be the left-invariant one-forms on $\textrm{SO}(n{+}1)$ dual to the vector fields ${\hat E}_a$. The SO$(n{+}1)$ group multiplication induces a natural map $$\alpha :\quad \textrm{SO}(n{+}1)\ \to\ S^n\equiv\sfrac{\textrm{SO}(n{+}1)}{\textrm{SO}(n)} \quad\with\quad
g\ \mapsto\ g\cdot\textrm{SO}(n)\,.$$ On any open subset $U\subset S^n$, one can invert this map by $\beta:U\rightarrow\textrm{SO}(n{+}1)$ with $\alpha\,\circ\,\beta=\textrm{id}$. In other words, $\beta$ is a local section of a principal bundle $\textrm{SO}(n{+}1)\rightarrow S^n$. This allows us to pull back the one-forms ${\hat e}^A$ to left-invariant one-forms $e^A=\beta^*{\hat e}^A$ on $S^n$. Splitting $\{e^A\}=\{e^a\}\cup\{e^i\}$, the set $\{e^a\}$ forms a basis on $S^n$, and the remaining one-forms $e^i$ are dependent, i.e.$e^i={e^i}_ae^a$ with real functions ${e^i}_a$. Using the group action, we can extend $e^a$ from $U$ to everywhere on $S^n$.
The normalization of the $e^A$ is related to that of the structure constants via the Maurer–Cartan equations $$\dd e^A+\sfrac12{f_{BC}}^A\,e^B\wedge e^C=0 \,,
\label{Maurer-Cartan0}$$ and we choose it in such a way that $$\delta_{ab}\,e^a e^b \= \dd\Omega_{n}^2$$ is the metric on the unit sphere, i.e. on $S^n$ with radius one. This fixes the proportionality constant $\varkappa_n$ in (\[metricG\]), as is derived in the Appendix. The metric (\[metric\]) on the cylinder then takes the conformally flat form $$\dd s^2\=\frac{R^2}{\cos^2\!t}\bigl(-\dd t^2+\delta_{ab}e^a e^b\bigr)\,.
\label{metric3}$$
The connection one-forms ${\omega^a}_b$ on $S^n$ may be found from Cartan’s structure equations $$\dd e^a+{\omega^a}_b\wedge e^b=0$$ by comparing with the Maurer–Cartan equations: $$0\=\dd e^a+{f_{bi}}^a\,e^b\wedge e^i\=\dd e^a+{f_{bi}}^a{e^i}_c\,e^b\wedge e^c \qquad\Rightarrow\qquad
{\omega^a}_b\={\omega^a}_{cb}\,e^c\={f_{ib}}^a {e^i}_c\,e^c\,.
\label{Maurer-Cartan1}$$
The connection on the non-symmetric coset for the odd-dimensional sphere $S^{2m+1}$ can be obtained analogously. Left-invariant one-forms $\tilde{e}^A$ on SU$(m{+}1)$ give rise to left-invariant one-forms $\tilde{e}^a$ on $S^{2m+1}$, which we normalize (fixing $\tilde{\varkappa}_m$) in such a way that the metric on the unit sphere reads in terms of which the metric reads $$\dd\Omega_{2m+1}^2\=\delta_{ab}\,\tilde{e}^a \tilde{e}^b
\qquad\Rightarrow\qquad
\dd s^2\=\frac{R^2}{\cos^2\!t}\bigl(-\dd t^2+\delta_{ab}\tilde{e}^a \tilde{e}^b\bigr)\,.
\label{metric4}$$ From the Maurer–Cartan equations and $\tilde{e}^i=\tilde{e}^i_{\hspace{0.5ex}a}\tilde{e}^a$ we read off the connection one-forms ${\tilde{\omega}^a}_{\hspace{1ex}b}$ on $S^{2m+1}$: $$0\=\dd \tilde{e}^a+\tilde{f}_{bi}^{\hspace{1.5ex}a}\,\tilde{e}^b\wedge \tilde{e}^i
+\sfrac{1}{2}\tilde{f}_{bc}^{\hspace{1.5ex}a}\,\tilde{e}^b\wedge \tilde{e}^c \qquad\Rightarrow\qquad
{\tilde{\omega}^a}_{\hspace{1ex}b} \= {\tilde{\omega}^a}_{\hspace{1ex}cb}\,\tilde{e}^c
\=\bigl(\tilde{f}_{ib}^{\hspace{1ex}a}\tilde{e}^i_{\hspace{0.5ex}c}+\sfrac{1}{2}\tilde{f}_{cb}^{\hspace{1ex}a}\bigr)\tilde{e}^c\,.
\label{Maurer-Cartan2}$$
The connection on the symplectic coset for $S^{4\ell+3}$ takes the same form, with left-invariant one-forms $\hat{e}^A$ on $\textrm{Sp}(\ell{+}1)$ descending to $\hat{e}^a$ on $S^{4\ell+3}$. Again we normalize by fixing $\hat{\varkappa}_\ell$ so that $$\dd\Omega_{4\ell+3}^2\=\delta_{ab}\,\hat{e}^a \hat{e}^b
\qquad\Rightarrow\qquad
\dd s^2\=\frac{R^2}{\cos^2\!t}\bigl(-\dd t^2+\delta_{ab}\hat{e}^a \hat{e}^b\bigr)\,.
\label{metric5}$$ The connection one-forms ${\hat{\omega}^a}_{\hspace{1ex}b}$ on $S^{4\ell+3}$ are obtained as in (\[Maurer-Cartan1\]).
Equivariant reduction of the Yang–Mills equations
=================================================
Yang–Mills equations on dS$_{n+1}$
----------------------------------
Since the Yang–Mills equations transform in a simple fashion under a conformal rescaling of the metric, solutions on de Sitter space may be obtained from transformed Yang–Mills equations on the cylinder $\mathcal{I}\times S^n$ with the flat metric $$\dd\bar{s}^2 \= -\dd t^2+\delta_{ab}\,\bar{e}^a\bar{e}^b\=-e^0 e^0+\delta_{ab}\,\bar{e}^a\bar{e}^b \quad \with e^0=\dd t
\and \bar{e}^a=\Bigl\{\begin{smallmatrix} e^a \\[2pt] \tilde{e}^a \\[2pt] \hat{e}^a \end{smallmatrix}\Bigr.\,,$$ depending on the case (\[metric3\]) or (\[metric4\]) or (\[metric5\]), respectively. Therefore, we may raise and lower spacetime indices $\mu\in\{0,a\}$ with the Minkowski metric $(\eta_{\mu\nu})=\textrm{diag}({-}1,1,\ldots,1)$. Structure constants with all indices down are completely antisymmetric. In this subsection, we treat the three cases in parallel, denoting the connection coefficients by ${\bar\omega^{\cdot}}_{\cdot\cdot}$.
We expand the gauge potential $\mathcal{A}$ and gauge field $\mathcal{F}$ in terms of the one-form basis $\lbrace \bar{e}^0, \bar{e}^a\rbrace$, $$\mathcal{A}=\mathcal{A}_0\,\bar{e}^0+\mathcal{A}_a\,\bar{e}^a \und
\mathcal{F}=\mathcal{F}_{0a}\,\bar{e}^0\wedge \bar{e}^a+\sfrac{1}{2}\mathcal{F}_{ab}\,\bar{e}^a\wedge \bar{e}^b\,,$$ and take the temporal gauge $\mathcal{A}_0=\mathcal{A}_t=0$. Then, the Yang–Mills equations on $\mathcal{I}\times S^n$ read $$\begin{aligned}
&\!\!\!\!\! E_a\mathcal{F}^{a0}+{\bar\omega^a}_{ac}\mathcal{F}^{c0}+{\bar\omega^0}_{a0} \mathcal{F}^{a0}+{\bar\omega^0}_{ab}\mathcal{F}^{ab}
+[\mathcal{A}_a, \mathcal{F}^{a0}]\=0 \und \\[4pt]
&\!\!\!\!\! E_0\mathcal{F}^{0b}+E_a\mathcal{F}^{ab}+{\bar\omega^0}_{00}\mathcal{F}^{0b}+{\bar\omega^a}_{a0}\mathcal{F}^{0b}
+{\bar\omega^0}_{0a}\mathcal{F}^{ab}+{\bar\omega^c}_{ca}\mathcal{F}^{ab}+{\bar\omega^b}_{ca}\mathcal{F}^{ca}+[\mathcal{A}_a,\mathcal{F}^{ab}]\=0\,,
\label{YM-1}
\end{aligned}$$ where $E_0=\dv{}{t}$ and $E_a$ are the vector fields satisfying $E_a\bar{e}^b=\delta_a^b$. We employ the flat metric $\dd\bar{s}^2$ on the cylinder, so Latin indices may be raised and lowered freely. The only effect of passing from $\dd\bar{s}^2$ to $\dd s^2$ is the occurrence of additional connection coefficients $${\bar\omega^0}_{00}=\tan t\,,\quad {\bar\omega^0}_{aa}={\bar\omega^a}_{0a}=n\tan t\quad \textrm{(sum over $a$)}\,,\quad
{\bar\omega^0}_{0b}={\bar\omega^a}_{0b}={\bar\omega^0}_{ab}=0$$ due to the conformal factor $\frac{R^2}{\cos^2\!t}$. Inserting these into (\[YM-1\]), the Yang-Mills equations take the form $$\begin{aligned}
&E_a\mathcal{F}^{a0}+{\bar\omega^a}_{ac}\mathcal{F}^{c0}+[\mathcal{A}_a, \mathcal{F}^{a0}]\=0\,,\\[4pt]
&E_0\mathcal{F}^{0b}+E_a\mathcal{F}^{ab}+(n{+}1)\tan t\, \mathcal{F}^{0b}+{\bar\omega^c}_{ca}\mathcal{F}^{ab}
+{\bar\omega^b}_{ca}\mathcal{F}^{ca}+[\mathcal{A}_a,\mathcal{F}^{ab}]\=0\,.
\label{YM-2}
\end{aligned}$$ Our aim is to find SO$(n{+}1)$, SU$(m{+}1)$ or Sp$(\ell{+}1)$ equivariant solutions to these equations.
Equivariant reduction on $\frac{\textrm{SO}(n{+}1)}{\textrm{SO}(n)}$
--------------------------------------------------------------------
At this point we restrict ourselves to the canonical choice of SO$(n{+}1)$ for the gauge group, so the gauge potential $\mathcal{A}$ and the field strength $\mathcal{F}$ live in the adjoint representation of $so(n{+}1)$. Expanding in the generators $\{I_a,I_i\}$ and the one-forms $\{e^b\}$, $$\mathcal{A} \= I_i\,{X^i}_b\,e^b + I_a\,{X^a}_b\,e^b\,,
\label{expandA}$$ introduces a set $\{{X^i}_b,{X^a}_b\}$ of functions on $\mathcal{I}\times S^n$. In order to find explicit solutions to the equations (\[YM-2\]), we impose $\textrm{SO}(n{+}1)$ equivariance on the gauge potential $\mathcal{A}$, which enforces $${X^i}_b = {e^i}_b \und {X^a}_b = {X^a}_b(t) \with {f_{ia}}^c\,{X^a}_b = {f_{ib}}^a\,{X^c}_a\,.$$ The first equation turns the first term in (\[expandA\]) into the canonical SO$(n)$ connection $I_i\,e^i$, while the last equation is more succinctly written in terms of $n$ matrix functions $X_a(t)\in\mathfrak{m}$, $$[I_i,X_a]={f_{ia}}^b X_b \quad\for X_b := I_a\,{X^a}_b\,.
\label{constraint}$$ Hence, the equivariant gauge potential takes the form [@Zoupanos; @IvLe08; @IvLePoRa09; @BaIvLeLu10] $$\mathcal{A}\=I_i\,e^i+X_a(t)\,e^a \qquad\Leftrightarrow\qquad \mathcal{A}_a=I_i\,{e^i}_a+X_a(t)\,.
\label{ansatz}$$ The curvature of this connection is readily computed with the help of (\[constraint\]), and the equivariant gauge field reads $$\mathcal{F}\=\dd\mathcal{A}+\mathcal{A}\wedge \mathcal{A}
\=\dot{X}_b\,e^0\wedge e^b-\sfrac{1}{2}\bigl({f_{bc}}^iI_i-[X_b, X_c]\bigr)\,e^b\wedge e^c$$ with the components $$\mathcal{F}_{0b}\=\dot{X}_b \und \mathcal{F}_{ab}\=-{f_{ab}}^iI_i+[X_a, X_b]\,.
\label{ansatz-2}$$ Here, overdot denotes a derivative with respect to $t$, i.e. $\dot{X}_a:=\dv{X_a}{t}$.
Making use of (\[ansatz\]) and (\[Maurer-Cartan1\]), the first Yang–Mills equation in (\[YM-2\]) reduces to $$[X_a,\dot{X}^a]\=0
\label{DE1}$$ while the second equation becomes $$-\ddot{X}_a-(n{-}3)\tan t\,\dot{X}_a+\sfrac12\varkappa_n\,X_a+\bigl[X_b,[X_b,X_a]\bigr]\=0\,,
\label{DE2}$$ where we have exploited the Jacobi identity and the normalization ${f_{ba}}^j{f_{jb}}^c=\sfrac12\varkappa_n\,{\delta_{a}}^c$ of the Killing metric (\[Killing-2\]). We still have to solve the equivariance condition (\[constraint\]). Due to the decomposition $$\textrm{adj}(so(n{+}1))\quad\longrightarrow\quad \textrm{adj}(so(n)) \oplus \bm{n}$$ with $\bm{n}$ denoting the $so(n)$ vector representation, there is only one free parameter in the general solution $$X_a(t)\=\phi(t)\,I_a\,,
\label{choice1}$$ where $\phi(t)$ is a real function of $t$. With this, (\[DE1\]) is automatically satisfied, and the ordinary matrix differential equation (\[DE2\]) reduces to $$\ddot{\phi}+(n{-}3)\tan t\,\dot{\phi}-\sfrac12\varkappa_n(1{-}\phi^2)\,\phi\=0\,.
\label{eom-1}$$ Any solution $\phi(t)$ gives rise to a Yang–Mills field $$\mathcal{F}_{0b}\= \dot{\phi}\,I_b \und
\mathcal{F}_{ab}\= \bigl(\phi^2{-}1\bigr) {f_{ab}}^i\,I_i \,.$$
Alternatively, one may evaluate the action functional on dS$_{n+1}$, $$S\=\sfrac{1}{4}\int_{\textrm{dS}_{n+1}} {\tr}(\mathcal{F}\wedge *\mathcal{F})
\=\sfrac{1}{8}\int_{\mathcal{I}\times S^n} \tr(\mathcal{F}_{\mu\nu}\mathcal{F}^{\mu\nu})\,
\bigl(\sfrac{R}{\cos t}\bigr)^{n+1}\,e^0\wedge e^1\wedge\cdots\wedge e^n\,,$$ where the factor $\bigl(\sfrac{R}{\cos t}\bigr)^{n+1}$ arises from the square root of the determinant of the metric (\[metric3\]). The trace is taken in the adjoint representation. Substituting the components (\[ansatz-2\]) of the equivariant gauge field with $\mathcal{F}^{\mu\nu}{=}\,\bigl(\sfrac{R^2}{\cos^2\!t}\bigr)^{-2}\mathcal{F}_{\mu\nu}$, we obtain ($\tr=\tr_{\textrm{adj}}$) $$\begin{aligned}
S&\=\sfrac{1}{8}\int \bigl(\sfrac{R}{\cos t}\bigr)^{n-3}\,\tr\bigl(-2\mathcal{F}_{0a}\mathcal{F}_{0a}+\mathcal{F}_{ab}\mathcal{F}_{ab}\bigr)\,e^0\wedge e^1\wedge\cdots\wedge e^n
\\&\=\sfrac{1}{8}\int\!\dd{t}\dd^n{\Omega}_{n}\,\bigl(\sfrac{R}{\cos t}\bigr)^{n-3}\,\bigl(-2\,\dot{\phi}^2\,\tr(I_aI_a)+(\phi^2{-}1)^2{f_{bc}}^i{f_{bc}}^j\,\tr(I_iI_j)\bigr)
\\&\=\sfrac{1}{8}\varkappa_n\textrm{vol}(S^n)\int_{-\pi/2}^{\pi/2}\!\dd{t}\,\bigl(\sfrac{R}{\cos t}\bigr)^{n-3}\,\bigl(2n\,\dot{\phi}^2-(\phi^2{-}1)^2{f_{bc}}^i{f_{bc}}^i\bigr)
\\&\=\sfrac12 n\varkappa_n\textrm{vol}(S^n)\int_{-\pi/2}^{\pi/2}\!\dd{t}\,\bigl(\sfrac{R}{\cos t}\bigr)^{n-3}\,\bigl(\sfrac12\dot{\phi}^2-\sfrac18\varkappa_n(\phi^2{-}1)^2\bigr)\,,
\label{Action-1}
\end{aligned}$$ where we have used ${f_{bc}}^i{f_{bc}}^i=\sfrac12\varkappa_n\delta_{bb}=\sfrac12\varkappa_n n$ from (\[Killing-2\]). It is straightforward to see that the variation with respect to $\phi$ reproduces the equation of motion (\[eom-1\]). To summarize, SO$(n{+}1)$ equivariance turns the Yang–Mills equations to the Newtonian dynamics of a particle on $\mathbb{R}$ with time-dependent friction $\sim(n{-}3)\tan t$ and subject to the double-well potential $$V(\phi)\=\sfrac{1}{8}\varkappa_n(\phi^2-1)^2\,.
\label{pot1}$$
Finally, we compute the energy $\mathcal{E}$ of our classical Yang–Mills configurations on dS$_{n+1}$, dual to de Sitter time $\tau$. Mapping again to $\mathcal{I}\times S^n$ and fixing $t(\tau)$, the result is $$\begin{aligned}
\mathcal{E}&\=-\sfrac{1}{4}\bigl(\sfrac{R}{\cos t}\bigr)^{n-4}\int_{S^n}\!\dd^n\Omega_n\,
{\tr}\bigl(\mathcal{F}_{0a}\mathcal{F}_{0a}+\sfrac{1}{2}\mathcal{F}_{ab}\mathcal{F}_{ab}\bigr) \\[4pt]
&\=\sfrac12 n\varkappa_n \textrm{vol}(S^n)\,(R\,\cosh\tau)^{n-4}\,
\bigl(\sfrac12\dot{\phi}^2+\sfrac18\varkappa_n(1{-}\phi^2)^2\bigr)\big|_{t=t(\tau)}\,.
\end{aligned}$$
For the Abelian case of $n{=}1$ the potential vanishes because $\varkappa_1=0$. The general solution to (\[eom-1\]) is $$\phi(t)=c\tan t+d \qquad\Rightarrow\qquad \mathcal{F}_{01}=\frac{c}{\cos^2\!t}\,I_1\,,$$ which yields a singular action integral except for the trivial constant solution.
Equivariant reduction on $\frac{\textrm{SU}(m{+}1)}{\textrm{SU}(m)}$
--------------------------------------------------------------------
In this subsection, we turn our attention to SU$(m{+}1)$-equivariant gauge fields on odd-dimensional spheres and fix the gauge group to be SU$(m{+}1)$. We pass from the orthogonal to the unitary coset by putting tildes over most symbols. The SU$(m{+}1)$ equivariant connection takes the form $$\mathcal{A}\=\tilde{I}_i\,\tilde{e}^i+\tilde{X}_a(t)\,\tilde{e}^a \qquad\Leftrightarrow\qquad
\mathcal{A}_a\=\tilde{I}_i\,{\tilde{e}^i}_{\hspace{0.5ex} a}+\tilde{X}_a(t)\,,
\label{ansatz2}$$ where the $2m{+}1$ matrix functions $\tilde{X}_a(t)$ are subject to the equivariance constraint $$[\tilde{I}_i,\tilde{X}_a]\={\tilde{f}_{ia}}^{\hspace{1.5ex}b} \tilde{X}_b\,.
\label{constraint2}$$ The connection (\[ansatz2\]) gives rise to the curvature $$\mathcal{F}\=\dot{\tilde{X}}_b\,\tilde{e}^0\wedge \tilde{e}^b
-\sfrac{1}{2}\bigl(\tilde{f}_{bc}^{\hspace{1.5ex}i}\tilde{I}_i+\tilde{f}_{bc}^{\hspace{1.5ex}a}\tilde{X}_a-[\tilde{X}_b, \tilde{X}_c]\bigr)\tilde{e}^b\wedge \tilde{e}^c$$ with the components $$\mathcal{F}_{0b}\=\dot{\tilde{X}}_b \und
\mathcal{F}_{ab}\=-\tilde{f}_{ab}^{\hspace{1.5ex}i}\tilde{I}_i-\tilde{f}_{ab}^{\hspace{1.5ex}c}\tilde{X}_c+[\tilde{X}_a, \tilde{X}_b]\,.
\label{ansatz-3}$$
While the first Yang–Mills equation in (\[YM-2\]) yields $$[\tilde{X}_a,\dot{\tilde{X}}^a]\=0
\label{DE1.1}$$ the second equation descends to $$-\ddot{\tilde{X}}_a-2(m{-}1)\tan t\, \dot{\tilde{X}}_a
+\bigl(\tilde{f}_{da}^{\hspace{1.5ex}i}\tilde{f}_{id}^{\hspace{1.5ex}b}-\sfrac{1}{2}\tilde{f}_{dc}^{\hspace{1.5ex}a}\tilde{f}_{dc}^{\hspace{1.5ex}b}\bigr)\tilde{X}_b
+\sfrac{3}{2}\tilde{f}_{bc}^{\hspace{1.5ex}a}[\tilde{X}_b,\tilde{X}_c]+\bigl[\tilde{X}_b,[\tilde{X}_b,\tilde{X}_a]\bigr]\=0\,,
\label{YM2-2}$$ again with the help of the Jacobi identity on the structure constants. In contrast to the previous subsection, our coset is no longer symmetric, which is manifested by $\tilde{f}_{ab}^{\hspace{1.5ex}c}\neq0$. So to evaluate the coefficient of $\tilde{X}_b$ in (\[YM2-2\]), we need more information on the SU$(m{+}1)$ structure constants. Since the coset $\frac{\textrm{SU}(m{+}1)}{\textrm{SU}(m)}$ is an $\alpha$-Sasakian manifold (with $\alpha{=}{-}\sfrac12\tilde{\varkappa}_m$ in our normalization), the contact one-form defines a preferred direction which we associate with $a{=}1$, and we may single out the corresponding $\tilde{\mathfrak{m}}$ generator as special, thus subdividing $$\{\tilde{I}_a\} \= \{\tilde{I}_1\} \cup \{\tilde{I}_{a^{\prime}}\} \quad\with a'=2,\ldots,2m{+}1\,.$$ The structure constants can be chosen such that $$\tilde{f}_{a^{\prime}b^{\prime}}^{\hspace{2.5ex}i}\neq0 \and
\tilde{f}_{a^{\prime}b^{\prime}}^{\hspace{2.5ex}1}\neq0
\qquad\textrm{but}\qquad
\tilde{f}_{i\,b^{\prime}}^{\hspace{2ex}1}=\tilde{f}_{a^{\prime}b^{\prime}}^{\hspace{2.5ex}c^{\prime}}=0 \,.
\label{structureconstant1}$$ The middle equation in (\[Killing-3\]) then splits into three parts, $$\tilde{f}_{c^{\prime}d^{\prime}}^{\hspace{2.5ex}1}\tilde{f}_{c^{\prime}d^{\prime}}^{\hspace{2.5ex}1}=\tilde{\varkappa}_m \qquad\Rightarrow\qquad
\tilde{f}_{c^{\prime}1}^{\hspace{2ex}a^{\prime}}\tilde{f}_{c^{\prime}1}^{\hspace{2ex}b^{\prime}}=\sfrac{1}{2m}\tilde{\varkappa}_m\,\delta^{a^\prime b^\prime} \und
\tilde{f}_{c^{\prime}i}^{\hspace{2ex}a^{\prime}} \tilde{f}_{c^{\prime}i}^{\hspace{2ex}b^{\prime}}=\sfrac{m-1}{2m}\tilde{\varkappa}_m\delta^{a^\prime b^\prime} \,.
\label{structureconstant2}$$ As a consequence, (\[YM2-2\]) simplifies to two different forms, $$\begin{aligned}
-\ddot{\tilde{X}}_{a'}-2(m{-}1)\tan t\, \dot{\tilde{X}}_{a'}
+\sfrac{m-2}{2m}\tilde{\varkappa}_m\tilde{X}_{a'}
+3\tilde{f}_{1\,b'}^{\hspace{2.5ex}a'}[\tilde{X}_1,\tilde{X}_{b'}]
+\bigl[\tilde{X}_{b'},[\tilde{X}_{b'},\tilde{X}_{a'}]\bigr]
+\bigl[\tilde{X}_1,[\tilde{X}_1,\tilde{X}_{a'}]\bigr]&=0\,,\\[4pt]
-\ddot{\tilde{X}}_1-2(m{-}1)\tan t\, \dot{\tilde{X}}_1
-\sfrac12\tilde{\varkappa}_m\tilde{X}_1
+\sfrac{3}{2}\tilde{f}_{b'c'}^{\hspace{2.5ex}1}[\tilde{X}_{b'},\tilde{X}_{c'}]
+\bigl[\tilde{X}_{b'},[\tilde{X}_{b'},\tilde{X}_1]\bigr]&=0\,.
\end{aligned}
\label{2YM}$$
This is consistent with the decomposition $$\textrm{adj}(su(m{+}1))\quad\longrightarrow\quad \textrm{adj}(su(m)) \oplus \bm{m} \oplus {\bm{\bar{m}}} \oplus \bm{1}\,,$$ which implies that the general solution to the equivariance constraint (\[constraint2\]) contains two free parameters, one for the fundamental su(m) representation $\bm{m}$ and its complex conjugate $\bm{\bar{m}}$ ($\mathcal{A}$ is anti-Hermitian), indexed by $a'$, and one for the singlet $\bm{1}$, indexed by $1$. Hence, the matrix functions $X_a(t)$ take the form [^1] $$\tilde{X}_{a^\prime}(t)\=\varphi(t)\,\tilde{I}_{a^\prime} \und \tilde{X}_{1}(t)\=\sqrt{2m}\,\psi(t)\,\tilde{I}_{1}
\label{choice2}$$ where $\varphi(t)$ and $\psi(t)$ are real functions of $t$. This form automatically fulfills (\[DE1\]), and the two ordinary matrix differential equations (\[2YM\]) become $$\begin{aligned}
-\ddot{\varphi}-2(m{-}1)\tan t\,\dot{\varphi}
+\sfrac{m-2}{2m}\tilde{\varkappa}_m\,\varphi
+\sfrac{3}{\sqrt{2m}}\tilde{\varkappa}_m\,\varphi\,\psi
-\sfrac12\tilde{\varkappa}_m\,\varphi^3
-\tilde{\varkappa}_m\,\varphi\,\psi^2&\=0\,, \\[4pt]
-\ddot{\psi}-2(m{-}1)\tan t\,\dot{\psi}
-\sfrac12\tilde{\varkappa}_m\,\psi
+\sfrac{3/2}{\sqrt{2m}}\tilde{\varkappa}_m\,\varphi^2
-\tilde{\varkappa}_m\,\psi\,\varphi^2&\=0\,.
\end{aligned}
\label{eom-2}$$ For any solution $\bigl(\varphi(t),\psi(t)\bigr)$ we gain a Yang–Mills configuration $$\begin{aligned}
\mathcal{F}_{0\,b'} &\= \dot{\varphi}\,\tilde{I}_{b'} \qquad\und
\mathcal{F}_{a'b'} \= \bigl(\varphi^2{-}1\bigr) \tilde{f}_{a^{\prime}b^{\prime}}^{\hspace{2.5ex}i}\,\tilde{I}_i
+\bigl(\varphi^2{-}\sqrt{2m}\psi\bigr) \tilde{f}_{a^{\prime}b^{\prime}}^{\hspace{2.5ex}1}\,\tilde{I}_1 \,,\\[4pt]
\mathcal{F}_{0\,1} &\= \sqrt{2m}\dot{\psi}\,\tilde{I}_1 \und
\mathcal{F}_{a'1} \,\= \varphi\bigl(\sqrt{2m}\psi{-}1\bigr) \tilde{f}_{a^{\prime}1}^{\hspace{2.5ex}b'}\,\tilde{I}_{b^\prime} \,.
\end{aligned}$$
Alternatively, the Yang–Mills action on dS$_{2m+2}$ with the input (\[choice2\]) can be computed as ($\tr=\tr_{\textrm{adj}}$) $$\begin{aligned}
S&\=\sfrac{1}{4}\int_{\mathcal{I}\times S^{2m+1}} {\tr}(F\wedge *F) \\[4pt]
&\=\sfrac{1}{8}\int\!\dd{t}\dd^{2m+1}{\Omega}_{2m+1} \bigl(\sfrac{R}{\cos t}\bigr)^{2m-2}
\Bigl\{(-2\,\dot{\varphi}^2\,\tr(\tilde{I}_{a^\prime} \tilde{I}_{a^\prime})-4m\,\dot{\psi}^2\,\tr(\tilde{I}_1 \tilde{I}_1)
+(1{-}\varphi^2)^2{\tilde{f}_{b^\prime c^\prime}}^{\hspace{2ex}i}{\tilde{f}_{b^\prime c^\prime}}^{\hspace{2ex}j}\,\tr(\tilde{I}_i\tilde{I}_j) \Bigr.\\
&\Bigl.\qquad\qquad\qquad\qquad\qquad
+\,(\sqrt{2m}\psi{-}\varphi^2)^2{\tilde{f}_{b^\prime c^\prime}}^{\hspace{2ex}1}{\tilde{f}_{b^\prime c^\prime}}^{\hspace{2ex}1}\,\tr(\tilde{I}_1\tilde{I}_1)
+2\,\varphi^2(1{-}\sqrt{2m}\psi)^2{\tilde{f}_{1\,c^\prime}}^{\hspace{2.5ex}a^{\prime}}{\tilde{f}_{1\,c^\prime}}^{\hspace{2.5ex}d^{\prime}}
\tr(\tilde{I}_{a^{\prime}}\tilde{I}_{d^{\prime}})\Bigr\}\\[4pt]
&\ \propto\ \int\!\dd{t}\,\Bigl(\sfrac{R}{\cos t}\bigr)^{2m-2}\,
\bigl\{4m\,\dot{\varphi}^2+4m\,\dot{\psi}^2-(m{-}1)\tilde{\varkappa}_m(1{-}\varphi^2)^2-\tilde{\varkappa}_m(\sqrt{2m}\psi{-}\varphi^2)^2
-2\tilde{\varkappa}_m\varphi^2(1{-}\sqrt{2m}\psi)^2\Bigr\}\,.
\label{Action-2}
\end{aligned}$$ Variation with respect to $\varphi$ and $\psi$ properly yields the equations of motion (\[eom-2\]). This time, we find a two-dimensional Newtonian dynamics with time-dependent friction. Dividing by $8m$, the external potential is extracted as $$\begin{aligned}
\tilde{\varkappa}_m^{-1}\,V(\varphi,\psi)
&\= \sfrac{m-1}{8m}\bigl(1-\varphi^2\bigr)^2
+\sfrac{1}{4m}\,\varphi^2\bigl(1-\sqrt{2m}\psi\bigr)^2
+\sfrac{1}{8m}\bigl(\sqrt{2m}\psi-\varphi^2\bigr)^2 \\[4pt]
&\=\sfrac{m-1}{8m}-\sfrac{m-2}{4m}\,\varphi^2+\sfrac14\,\psi^2
-\sfrac{3/2}{\sqrt{2m}}\,\varphi^2\psi+\sfrac18\,\varphi^4+\sfrac12\,\varphi^2\psi^2\ .
\end{aligned}
\label{pot2}$$ Contours plot for $m{=}1$ and $m{=}2$ are displayed in Fig. 1 at the end. Finally, the de Sitter energy read $$\mathcal{E} \= m\tilde{\varkappa}_m \textrm{vol}(S^{2m+1})\,(R\,\cosh\tau)^{2m-3}\,
\bigl(\sfrac12\dot{\varphi}^2+\sfrac12\dot{\psi}^2 + V(\varphi,\psi)\bigr)\big|_{t=t(\tau)}\ .$$
Equivariant reduction on $\frac{\textrm{Sp}(\ell{+}1)}{\textrm{Sp}(\ell)}$
--------------------------------------------------------------------------
Finally, we look at equivariant solutions for the symplectic gauge group $\textrm{Sp}(\ell{+}1)$. Equations (\[ansatz2\])–(\[DE1.1\]) of the previous subsection carry over with tildes exchanged for hats, and the matrix differential equations read $$-\ddot{\hat{X}}_a-4\ell\,\tan t\, \dot{\hat{X}}_a
+\bigl(\hat{f}_{da}^{\hspace{1.5ex}i}\hat{f}_{id}^{\hspace{1.5ex}b}-\sfrac{1}{2}\hat{f}_{dc}^{\hspace{1.5ex}a}\hat{f}_{dc}^{\hspace{1.5ex}b}\bigr)\hat{X}_b
+\sfrac{3}{2}\hat{f}_{bc}^{\hspace{1.5ex}a}[\hat{X}_b,\hat{X}_c]+\bigl[\hat{X}_b,[\hat{X}_b,\hat{X}_a]\bigr]\=0\,.
\label{YM3-2}$$ At this stage we exploit the fact that $\frac{\textrm{Sp}(\ell{+}1)}{\textrm{Sp}(\ell)}$ is $\alpha$-tri-Sasakian, which defines a special subalgebra in $\hat{\mathfrak{m}}$ that we span with the generators $\hat{I}_1$, $\hat{I}_2$ and $\hat{I}_3$, thus splitting $$\{\hat{I}_a\} \= \{\hat{I}_{a''}\} \cup \{\hat{I}_{a^{\prime}}\} \quad\with
a''=1,2,3 \and a'=4,\ldots,4\ell{+}3\,.$$ The structure constants can be arranged such that $$\hat{f}_{a^{\prime}b^{\prime}}^{\hspace{2.5ex}i}\neq0 \and
\hat{f}_{a^{\prime}b^{\prime}}^{\hspace{2.5ex}c''}\neq0\and
\hat{f}_{1\,2}^{\hspace{2ex}3}\neq0
\qquad\textrm{but}\qquad
\hat{f}_{i\,b^{\prime}}^{\hspace{2ex}c''}=
\hat{f}_{i\,b^{\prime\prime}}^{\hspace{2ex}c''}=
\hat{f}_{a'b''}^{\hspace{2.5ex}c''}=\hat{f}_{a^{\prime}b^{\prime}}^{\hspace{2.5ex}c^{\prime}}=0 \,.
\label{structureconstant3}$$ We normalize to unit radius the $S^3\simeq\textrm{Sp}(1)$ generated by $\{\hat{I}_{a''}\}$ by taking $$\diff\Omega_3^2=\delta_{a''b''}\,\hat{e}^{a''}\hat{e}^{b''} \quad\ \Leftrightarrow\ \quad
\hat{f}_{a''d''}^{\hspace{3ex}c''}\hat{f}_{c''b''}^{\hspace{3ex}d''}=\hat{\varkappa}_0\,\delta_{a''b''}
\quad\textrm{with}\quad\hat{\varkappa}_0=8 \quad\ \Leftrightarrow\ \quad
\hat{f}_{a''b''}^{\hspace{3ex}c''}=2\,\epsilon_{a''b''}^{\hspace{3ex}c''}\ .$$ The middle equation in (\[Killing-3\]) with hats instead of tildes then determines the following partial sums: $$\begin{aligned}
\hat{f}_{c^{\prime\prime}d^{\prime\prime}}^{\hspace{3ex}a^{\prime\prime}} \hat{f}_{c^{\prime\prime}d^{\prime\prime}}^{\hspace{3ex}b^{\prime\prime}}
&\=\frac{2\,\hat{\varkappa}_\ell}{(\ell{+}2)}\delta^{a^{\prime\prime} b^{\prime\prime}}\ ,\qquad\qquad
\hat{f}_{c^{\prime}d^{\prime}}^{\hspace{2.5ex}a^{\prime\prime}}\hat{f}_{c^{\prime}d^{\prime}}^{\hspace{2.5ex}b^{\prime\prime}}
\=\frac{\ell\,\hat{\varkappa}_\ell}{(\ell{+}2)}\delta^{a^{\prime\prime} b^{\prime\prime}}\ ,\\[4pt]
\hat{f}_{c^{\prime}d^{\prime\prime}}^{\hspace{2.5ex}a^{\prime}}\hat{f}_{c^{\prime}d^{\prime\prime}}^{\hspace{2.5ex}b^{\prime}}
&\=\frac{3\,\hat{\varkappa}_\ell}{4(\ell{+}2)}\,\delta^{a^\prime b^\prime}\ ,\qquad\qquad\ \
\hat{f}_{c^{\prime}i}^{\hspace{2ex}a^{\prime}} \hat{f}_{c^{\prime}i}^{\hspace{2ex}b^{\prime}}
\=\frac{(2\ell{+}1)\hat{\varkappa}_\ell}{4(\ell{+}2)}\,\delta^{a^\prime b^\prime}\ .
\end{aligned}$$ After this splitting, the matrix differential equations (\[YM3-2\]) read $$\begin{aligned}
-\ddot{\hat{X}}_{a'}-4\ell\tan t\, \dot{\hat{X}}_{a'}
+\sfrac{(\ell-1)}{2(\ell+2)}\hat{\varkappa}_\ell\hat{X}_{a'}
+3\hat{f}_{b'c^{\prime\prime}}^{\hspace{2.5ex}a'}[\hat{X}_{b'},\hat{X}_{c^{\prime\prime}}]
+\bigl[\hat{X}_{b'},[\hat{X}_{b'},\hat{X}_{a'}]\bigr]
+\bigl[\hat{X}_{b^{\prime\prime}},[\hat{X}_{b^{\prime\prime}},\hat{X}_{a'}]\bigr]&\=0\,,\\[4pt]
-\ddot{\hat{X}}_{a^{\prime\prime}}-4\ell\tan t\, \dot{\hat{X}}_{a^{\prime\prime}}
-\sfrac{1}{2}\hat{\varkappa}_\ell\hat{X}_{a^{\prime\prime}}
+\sfrac{3}{2}\hat{f}_{b\,c}^{\hspace{1.5ex}a^{\prime\prime}}[\hat{X}_{b},\hat{X}_{c}]
+\bigl[\hat{X}_{b},[\hat{X}_{b},\hat{X}_{a^{\prime\prime}}]\bigr]&\=0\,,
\end{aligned}
\label{YM3-3}$$ where in the last two terms $\{b,c\}=\{b',c'\}\cup\{b'',c''\}$, of course. When we consider the decomposition $$\textrm{adj}\bigl(sp(\ell{+}1)\bigr)\quad\longrightarrow\quad
\left(\textrm{adj}\bigl(sp(\ell)\bigr),\bm{1}\right) \oplus \left(\bm{2\ell},\bm{2}\right) \oplus \left(\bm{1},\textrm{adj}\bigl(sp(1)\bigr)\right)\ ,$$ of $sp(\ell{+}1)$ irreps into $sp(\ell)\otimes sp(1)$ ones, we see that again there are two free parameters in the equivariance condition (\[constraint2\]), one for $4\ell$-dimensional fundamental representation indexed by $a'$, and one for three-dimensional adjoint representation of $sp(1)$, indexed by $a^{\prime\prime}$. Bearing this in mind, the equivariant form of $\hat{X}_{a}(t)$ is [^2] $$\hat{X}_{a'}(t)\=\sqrt{3}\,\varphi(t)\,\hat{I}_{a^\prime}\und \hat{X}_{a^{\prime\prime}}(t)\=\sqrt{4\ell}\,\psi(t)\,\hat{I}_{a^{\prime\prime}}\ .
\label{ansatz3}$$ It automatically satisfies the condition (\[DE1.1\]) and reduces the matrix differential equations (\[YM3-3\]) to two coupled differential equations, $$\begin{aligned}
-\ddot{\varphi}-4\ell\tan t\,\dot{\varphi}+\sfrac{\ell-1}{2(\ell+2)}\hat{\varkappa}_\ell\,\varphi
+\sfrac{9\sqrt{\ell}}{2(\ell+2)}\hat{\varkappa}_\ell\,\varphi\,\psi-\sfrac{3}{2}\hat{\varkappa}_\ell\,\varphi^3
-\sfrac{3\ell}{\ell+2}\hat{\varkappa}_\ell\,\varphi\,\psi^2&\=0\ ,\\[4pt]
-\ddot{\psi}-4\ell\tan t\, \dot{\psi}-\sfrac{1}{2}\hat{\varkappa}_\ell\,\psi
+\sfrac{6\sqrt{\ell}}{\ell+2}\hat{\varkappa}_\ell\,\psi^2+\sfrac{9\sqrt{\ell}}{4(\ell+2)}\hat{\varkappa}_\ell\,\varphi^2
-\sfrac{8\ell}{\ell+2}\hat{\varkappa}_\ell\,\psi^3-\sfrac{3\ell}{\ell+2}\hat{\varkappa}_\ell\,\psi\,\varphi^2&\=0\ .
\label{eom-3}
\end{aligned}$$ For any solution $(\varphi,\psi)$ to these equations, we obtain a Yang–Mills configuration $$\begin{aligned}
\mathcal{F}_{0\,b'} &\= \sqrt{3}\dot{\varphi}\,\hat{I}_{b'} \und
\mathcal{F}_{a'b'} \= \bigl(3\varphi^2{-}1\bigr) \hat{f}_{a^{\prime}b^{\prime}}^{\hspace{2.5ex}i}\,\hat{I}_i
+\bigl(3\varphi^2{-}\sqrt{4\ell}\psi\bigr) \hat{f}_{a^{\prime}b^{\prime}}^{\hspace{2.5ex}a^{\prime\prime}}\,\hat{I}_{a^{\prime\prime}} \ ,\\[4pt]
\mathcal{F}_{0\,a^{\prime\prime}} &\= \sqrt{4\ell}\dot{\psi}\,\hat{I}_{a^{\prime\prime}} \ ,\quad
\mathcal{F}_{a'b^{\prime\prime}} \= \sqrt{3}\varphi\bigl(\sqrt{4\ell}\psi{-}1\bigr)
\hat{f}_{a'b^{\prime\prime}}^{\hspace{2.5ex}c'}\,\hat{I}_{c^\prime} \ ,\quad
\mathcal{F}_{a^{\prime\prime}b^{\prime\prime}} \= \sqrt{4\ell}\psi\bigl(\sqrt{4\ell}\psi{-}1\bigr)
\hat{f}_{a^{\prime\prime}b^{\prime\prime}}^{\hspace{2.5ex}c^{\prime\prime}}\,\hat{I}_{c^{\prime\prime}} \,.
\end{aligned}$$ The action functional on de Sitter space dS$_{4\ell+4}$ with the ansatz (\[ansatz3\]) can be computed as $$\begin{aligned}
S&\=\sfrac{1}{4}\int_{\mathcal{I}\times S^{4\ell+3}} {\tr}(F\wedge *F) \\[4pt]
&\=\sfrac{1}{8}\int\!\dd{t}\dd^{4\ell+3}{\Omega}_{4\ell+3} \bigl(\sfrac{R}{\cos t}\bigr)^{4\ell}
\Bigl\{(-6\,\dot{\varphi}^2\,\tr(\hat{I}_{a^\prime} \hat{I}_{a^\prime})-8\ell\,\dot{\psi}^2\,\tr(\hat{I}_{a^{\prime\prime}} \hat{I}_{a^{\prime\prime}})
+(1{-}3\varphi^2)^2{\hat{f}_{b^\prime c^\prime}}^{\hspace{2ex}i}{\hat{f}_{b^\prime c^\prime}}^{\hspace{2ex}j}\,\tr(\hat{I}_i\hat{I}_j) \Bigr.\\
&\Bigl.\qquad\qquad\qquad\qquad
+\,(\sqrt{4\ell}\psi{-}3\varphi^2)^2{\hat{f}_{b^\prime c^\prime}}^{\hspace{2ex}a^{\prime\prime}}
{\hat{f}_{b^\prime c^\prime}}^{\hspace{2ex}b^{\prime\prime}}\, \tr(\hat{I}_{a^{\prime\prime}}\hat{I}_{b^{\prime\prime}})
+6\varphi^2(1{-}\sqrt{4\ell}\psi)^2{\hat{f}_{b^{\prime\prime}c^\prime}}^{\hspace{2.5ex}a^{\prime}}
{\hat{f}_{b^{\prime\prime}c^\prime}}^{\hspace{2.5ex}d^{\prime}} \tr(\hat{I}_{a^{\prime}}\hat{I}_{d^{\prime}}) \\
&\qquad\qquad\qquad\qquad\!\!
+\,4\ell\psi^2(1{-}\sqrt{4\ell}\psi)^2{\hat{f}_{b^{\prime\prime}c^{\prime\prime}}}^{\hspace{2.5ex}a^{\prime\prime}}
{\hat{f}_{b^{\prime\prime}c^{\prime\prime}}}^{\hspace{2.5ex}d^{\prime\prime}} \tr(\hat{I}_{a^{\prime\prime}}\hat{I}_{d^{\prime\prime}})\Bigr\}\\[4pt]
\propto\!\!\int\!\!\dd{t}&\bigl(\sfrac{R}{\cos t}\bigr)^{4\ell}
\bigl\{24\ell\,\dot{\varphi}^2+24\ell\,\dot{\psi}^2-\sfrac{\ell(2\ell+1)}{\ell+2}\hat{\varkappa}_\ell(1{-}3\varphi^2)^2
-\sfrac{3\ell}{\ell+2}\hat{\varkappa}_\ell(\sqrt{4\ell}\psi{-}3\varphi^2)^2
-\sfrac{6\ell}{\ell+2}\hat{\varkappa}_\ell\,(3\varphi^2{+}4\psi^2)(1{-}\sqrt{4\ell}\psi)^2\Bigr\}
\end{aligned}
\label{Action-3}$$ where again $\tr=\tr_{\textrm{adj}}$. It is possible to verify the equations of motions (\[eom-3\]) by varying this action with respect to $\varphi$ and $\psi$. Dividing by $48\ell$, the corresponding external potential is read off as $$\begin{aligned}
\sfrac{1}{\hat{\varkappa}_\ell}V(\varphi,\psi)&\=
\sfrac{2\ell+1}{48(\ell+2)}(1-3\varphi^2)^2+\sfrac{1}{16(\ell+2)}(\sqrt{4\ell}\psi-3\varphi^2)^2
+\sfrac{1}{8(\ell+2)}(3\varphi^2+4\psi^2)(1-\sqrt{4\ell}\psi)^2 \\
&\=\sfrac{2\ell+1}{48(\ell+2)}-\sfrac{\ell-1}{4(\ell+2)}\varphi^2+\sfrac{1}{4}\psi^2
-\sfrac{9\sqrt{\ell}}{4(\ell+2)}\varphi^2\psi-\sfrac{2\sqrt{\ell}}{\ell+2}\psi^3
+\sfrac{3}{8}\varphi^4+\sfrac{3\ell}{2(\ell+2)}\varphi^2\psi^2+\sfrac{2\ell}{\ell+2}\psi^4\ .
\end{aligned}$$ Fig. 2 at the end of the paper shows contour plots for $\ell{=}1$ and $\ell{=}2$. The energy of these Yang–Mills configurations on de Sitter space dS$_{4\ell+4}$ computes as $$\mathcal{E} \= 6\ell\hat{\varkappa}_\ell \textrm{vol}(S^{4\ell+3})\,(R\,\cosh\tau)^{4\ell-1}\,
\bigl(\sfrac12\dot{\varphi}^2+\sfrac12\dot{\psi}^2 + V(\varphi,\psi)\bigr)\big|_{t=t(\tau)}\ .$$
Finite-action Yang–Mills solutions on dS$_{n+1}$
================================================
The conformal factor and the friction
-------------------------------------
In the previous section, we have shown that equivariance with respect to the isometry group of the coset reduces the Yang–Mills equations on dS$_{n+1}$ to Newtonian mechanics of a particle in a particular potential in one or two space dimensions with an additional friction term $\sim\tan t\,\dot{\phi}$ or $\sim\tan t\,(\dot{\varphi},\dot{\psi})$, respectively. Away from $n{=}3$, the general solution to the nonlinear differential equations (\[eom-1\]), (\[eom-2\]) or (\[eom-3\]) is not known, and the only analytic solutions there are the constant ones, with the analog particle sitting at a local extremum of the potential. For nonconstant solutions, it is possible to figure out the behavior of $\phi$ (or $(\varphi,\psi)$) and of the action numerically and asymptotically near the boundary $t=\pm\frac{\pi}{2}$. One finds that $$\phi(\sfrac{\pi}{2}{-}\epsilon) \ \buildrel{\epsilon\to0}\over{\sim}\
{\small\begin{cases} \epsilon^{-1} &\for n=1 \\[-4pt] \log\epsilon &\for n=2 \\[-4pt] \textrm{constant}+\epsilon^{n-2} &\for n\ge3 \end{cases}}\,,$$ and likewise for $(\varphi,\psi)$. Hence, the friction term leads to a blowing-up for $n<3$ (negative friction) and to a freezing for $n>3$ (positive friction). For $n>3$, the friction is harmless but the action (\[Action-1\]) is infinite due to the conformal factor $(\frac{R}{\cos t})^{n-3}$, unless we deal just with the vacuum solution. For $n<3$ the conformal factor is benign but the blowing-up at the boundary $t{=}{\pm}\frac{\pi}{2}$ renders the action (\[Action-1\]) divergent, unless $\phi$ or $(\varphi,\psi)$ is constant. The $n{=}3$ case is very special because the friction term always vanishes, and the particle solutions in a double-well potential are known explicitly in terms of elliptic functions. Here, both action and energy are finite.
Vacuum and purely magnetic solutions
------------------------------------
As is exemplified in Figs. 1 and 2 below, the vacuum solutions ($V{=}0$) in the three cases are given by $$\phi_{\textrm{vac}} = \pm 1 \qquad\textrm{or}\qquad
\bigl(\varphi,\psi\bigr)_{\textrm{vac}} = \bigl(\pm1,\sfrac1{\sqrt{2m}}\bigr) \qquad\textrm{or}\qquad
\bigl(\varphi,\psi\bigr)_{\textrm{vac}} = \bigl(\pm\sfrac1{\sqrt{3}},\sfrac1{\sqrt{4\ell}}\bigr)\ ,$$ and we shall not consider them further because they correspond to the pure-gauge configuration $$\mathcal{A} = \bar{I}_A\,\bar{e}^A \qquad\Rightarrow\qquad \mathcal{F} \equiv 0\,.$$
The only nontrivial analytic solution to (\[eom-1\]), (\[eom-2\]) or (\[eom-3\]) in any dimension is the constant solution $$\phi_{\textrm{mag}}=0 \qquad\textrm{or}\qquad (\varphi,\psi)_{\textrm{mag}}=(0,0)\,,$$ respectively. It corresponds to the analog particle sitting on a local maximum or on a saddle point (for $m{\ge}2$) of the potential, respectively (see Figs. 1 and 2). Since $\mathcal{F}_{0b}=0$ the ensuing gauge field is purely magnetic, with $$\begin{aligned}
\mathcal{A}_a= I_i\,e^i_{\ a} \quad&\Rightarrow\quad \mathcal{F}_{ab} = -{f_{ab}}^i\,I_i \ , \\[2pt]
\mathcal{A}_a=\tilde{I}_i\,\tilde{e}^i_{\ a} \quad&\Rightarrow\quad
\mathcal{F}_{a'b'} = -\tilde{f}_{a^{\prime}b^{\prime}}^{\hspace{2.5ex}i}\,\tilde{I}_i \and \mathcal{F}_{a''1}= 0 \ , \\[2pt]
\mathcal{A}_a=\hat{I}_i\,\hat{e}^i_{\ a} \quad&\Rightarrow\quad
\mathcal{F}_{a'b'} = -\hat{f}_{a^{\prime}b^{\prime}}^{\hspace{2.5ex}i}\,\hat{I}_i \and \mathcal{F}_{a''b'}= 0 =\mathcal{F}_{a''b''}\ .
\end{aligned}
\label{magnetic}$$ Note that $\mathcal{F}$ is valued in the stabilizer subgroup. Its action (\[Action-1\]), (\[Action-2\]) or (\[Action-3\]) becomes $$\begin{aligned}
S&\=-\sfrac{1}{8}\sfrac{n}{2}\varkappa_n^2\,\textrm{vol}(S^n)\!\smallint_{-\pi/2}^{\pi/2}\!\!\!\dd{t} \bigl(\sfrac{R}{\cos t}\bigr)^{n-3} \ ,\\
S&\=-\sfrac{1}{8}(m{-}1)\tilde{\varkappa}_m^2\,\textrm{vol}(S^{2m+1})\!\smallint_{-\pi/2}^{\pi/2}\!\!\!\dd{t} \bigl(\sfrac{R}{\cos t}\bigr)^{2m-2}\ ,\\
S&\=-\sfrac{1}{8}\sfrac{\ell(2\ell+1)}{\ell+2}\hat{\varkappa}_\ell^2\,\textrm{vol}(S^{4\ell+3})\!\smallint_{-\pi/2}^{\pi/2}\!\!\!\dd{t} \bigl(\sfrac{R}{\cos t}\bigr)^{4\ell}\ ,
\end{aligned}
\label{Action-1.1}$$ respectively, which is finite only for $n{\leq}3$ or $m{\le}1$. Finally, the de Sitter energy is given by $$\begin{aligned}
\mathcal{E} &\= \sfrac{1}{8}\sfrac{n}{2}\varkappa_n^2\,\textrm{vol}(S^n)\,(R\cosh\tau)^{n-4}\ ,\\[4pt]
\mathcal{E} &\= \sfrac{1}{8}(m{-}1)\tilde{\varkappa}_m^2\,\textrm{vol}(S^{2m+1})\,(R\cosh\tau)^{2m-3}\ ,\\[4pt]
\mathcal{E} &\= \sfrac{1}{8}\sfrac{\ell(2\ell+1)}{\ell+2}\hat{\varkappa}_\ell^2\,\textrm{vol}(S^{4\ell+3})\,(R\cosh\tau)^{4\ell-1}\ .
\end{aligned}
\label{energy}$$
The exception is $S^3=\textrm{SU}(2)$ (the case $m{=}1$), where $(\varphi,\psi)=(0,0)$ is again a vacuum solution. Since for $m{=}1$ the index $i$ takes no values, the field strength completely vanishes, as do action and energy. However, as can be seen in the top half of Fig. 1 below, in this case potential admits two saddle points half-way between the vacuum solutions, $$(\varphi,\psi)_{\textrm{mag}}^{m=1}=\bigl(\pm\sfrac12,\sfrac{1}{2\sqrt{2}}\bigr)\,,$$ which yield the configurations $$\mathcal{A}_{a'} = \pm\sfrac12\,\tilde{I}_{a'} \and \mathcal{A}_1 = \sfrac12\,\tilde{I}_1
\qquad\Rightarrow\qquad
\mathcal{A} = \sfrac12\,\tilde{I}_a\,\tilde{e}^a
\qquad\Rightarrow\qquad
\mathcal{F} = -\sfrac18 \tilde{f}_{ab}^{\hspace{1.5ex}c}\,\tilde{I}_c\,\tilde{e}^a\wedge\tilde{e}^b \,,$$ i.e. just one-half of the canonical SU(2) connection. For comparison, the orthogonal coset representation of the same space (the case $n{=}3$) yields precisely the canonical SO(3) connection inside SO(4) as in (\[magnetic\]).
With the values $$\label{varkappasmall}
\varkappa_2=2 \quad\Leftrightarrow\quad {f_{ab}}^3=\epsilon_{ab} \ ,\qquad
\varkappa_3=4 \quad\Leftrightarrow\quad {f_{ab}}^i={\epsilon_{ab}}^{i-3} \ ,\qquad
\tilde{\varkappa}_1=8 \quad\Leftrightarrow\quad \tilde{f}_{ab}^{\hspace{1.5ex}c}=2\,{\epsilon_{ab}}^c\,,$$ the values for the action and the energy of the magnetic Yang–Mills solutions on dS$_3$ and dS$_4$ (the orthogonal and the unitary coset) are displayed in the table below.
space value of the action $S$ value of the energy $\mathcal{E}$
---------------------------------- ------------------------------------------------------------------ -----------------------------------
$\vphantom{\bigg|}$ dS$_{3}$ $-\frac{2\cdot 2^2}{16}\,\textrm{vol}(S^2)\,\frac{2}{R}=-4\pi/R$ $2\pi\,(R\cosh\tau)^{-2}$
\[4pt\] dS$_{4}^{\textrm{orth}}$ $-\frac{3\cdot 4^2}{16} \,\textrm{vol}(S^3)\,\pi=-6\pi^3$ $6\pi^2(R\cosh\tau)^{-1}$
\[4pt\] dS$_{4}^{\textrm{unit}}$ $-\frac{3\cdot 8^2}{128}\,\textrm{vol}(S^3)\,\pi=-3\pi^3$ $3\pi^2(R\cosh\tau)^{-1}$
General equivariant dS$_4$ solutions
------------------------------------
As already mentioned, for dS$_4$ all equivariant Yang–Mills solutions enjoy finite action and energy, for gauge group SO(4) (case $n{=}3$) as well as for gauge group SU(2) (case $m{=}1$). We have the SO(4)- and SU(2)-equivariant connections $$\mathcal{A} \= I_i\,e^i + \phi\,I_a\,e^a \und
\mathcal{A} \= \varphi\,\tilde{I}_{a'}\,\tilde{e}^{a'} + \sqrt{2}\psi\,\tilde{I}_1\,\tilde{e}^1$$ which with (\[varkappasmall\]) produce the fields $$\begin{aligned}
\mathcal{F} &\= \dot{\phi}\,I_a\,e^0{\wedge}e^a +
\sfrac{1}{2}\bigl(\phi^2{-}1\bigr)\,{\epsilon_{ab}}^{i-3}\,I_i\,e^a{\wedge}e^b \und \\[4pt]
\mathcal{F} &\= \dot{\varphi}\,\tilde{I}_{a'}\,e^0{\wedge}\tilde{e}^{a'}
+ \sqrt{2}\dot{\psi}\,\tilde{I}_1\,e^0{\wedge}\tilde{e}^1
+\bigl(\varphi^2{-}\sqrt{2}\psi\bigr)\,{\epsilon_{a'b'}}^1\,
\tilde{I}_1\,\tilde{e}^{a'}{\wedge}\tilde{e}^{b'}
+2\varphi\bigl(1{-}\sqrt{2}\psi\bigr)\,{\epsilon_{1\,a'}}^{b'}\,
\tilde{I}_{b'}\,\tilde{e}^{a'}{\wedge}\tilde{e}^1 \,,
\end{aligned}$$ respectively. The Yang–Mills equations boil down to $$\begin{aligned}
&\ddot{\phi} \= 2\phi-2\phi^3\= -\sfrac{\partial V}{\partial\phi}
\qquad\qquad\qquad\qquad\qquad\quad\ \, \with V\= \sfrac12(\phi^2{-}1) \und \\[4pt]
\biggl\{ & \begin{matrix}
\ddot{\varphi} \= -4\varphi+12\sqrt{2}\varphi\,\psi-4\varphi^3-8\varphi\,\psi^2
\= -\sfrac{\partial V}{\partial\varphi} \\[4pt]
\ddot{\psi} \= -4\psi+6\sqrt{2}\varphi^2-8\psi\,\varphi^2
\qquad\qquad\!\= -\sfrac{\partial V}{\partial\psi}
\end{matrix} \biggr\} \quad\with
V\= 2\varphi^2(\sqrt{2}\psi{-}1)^2 + (\sqrt{2}\psi{-}\varphi^2)^2 \,,
\end{aligned}
\label{eom4d}$$ respectively.
Let us first look at the unitary case. Besides the constant solutions discussed above, one finds two types of non-constant analytic solutions easily visible in the top contour plot of Fig. 1, $$\begin{aligned}
\textrm{Abelian:} \quad \bigl(\varphi,\psi\bigr) &\= \bigl(0\,,\,c\cos 2(t{-}t_0)\bigr)
\qquad\quad\with c,t_0\in\mathbb{R}\,, \\[4pt]
\textrm{Non-Abelian:} \quad \bigl(\varphi,\psi\bigr)
&\= \bigl(\sfrac12(1{+}\phi)\,,\,\sfrac{1}{2\sqrt{2}}(1{+}\phi)\bigr)
\quad\ \with\ddot{\phi}=2\phi-2\phi^3\,.
\end{aligned}$$ Both types of solutions and the related Yang–Mills fields have been discussed in [@IvLePo1; @IvLePo2]. Interestingly, the “Newton equation” for $\phi(t)$ above is the same as in the orthogonal case, presented in the first line of (\[eom4d\]). Its general solution is given in terms of elliptic functions, except for the bounce solution $$\phi(t) \= \sqrt{2}\,\textrm{sech}\bigl(\sqrt{2}(t{-}t_0)\bigr)\,,$$ which approaches the purely magnetic solution $\phi{=}0$ for $t\to\pm\infty$. For any solution $\phi(t)$, the su(2) Yang–Mills potential and field take the form $$\mathcal{A} \= \sfrac12(1{+}\phi)\,\tilde{I}_a\,\tilde{e}^a \und
\mathcal{F} \= \sfrac12\dot{\phi}\,\tilde{I}_a\,e^0{\wedge}\tilde{e}^a
+\sfrac14(\phi^2{-}1)\,{\epsilon_{ab}}^c\,\tilde{I}_c\,\tilde{e}^a{\wedge}\tilde{e}^b$$ Let us compare this to the orthogonal case. The so(4)-valued configurations $$\mathcal{A} \= \bigl( I_i\,{e^i}_a + \phi\,I_a\bigr) e^a \und
\mathcal{F} \= \dot{\phi}\,I_a\,e^0{\wedge}e^a +
\sfrac{1}{2}\bigl(\phi^2{-}1\bigr)\,{\epsilon_{ab}}^{i-3}\,I_i\,e^a{\wedge}e^b$$ merely embed the su(2) Yang–Mills solution into so(4) and represent nothing new.
Acknowledgements {#acknowledgements .unnumbered}
================
O.L. thanks Lutz Habermann for help with the Appendix. G.Ü. is grateful for a Riemann Fellowship at Leibniz Universität Hannover, where the project was initiated. This work was partially supported by the Deutsche Forschungsgemeinschaft under grant LE 838/13. This article is based upon work from COST Action MP1405 QSPACE, supported by COST (European Cooperation in Science and Technology).
Appendix: Structure constant normalization for unit spheres
===========================================================
How are the structure constants to be normalized so that the corresponding coset manifold is a sphere of radius one? To answer this question, we embed the unit spheres as $$S^n \hookrightarrow \mathbb{R}^{n+1}\ ,\qquad
S^{2m+1} \hookrightarrow \mathbb{C}^{m+1}\ ,\qquad
S^{4\ell+3} \hookrightarrow \mathbb{H}^{\ell+1}$$ where $\mathbb{H}$ denotes the quaternionic number field. The three spaces $\mathbb{R}^{n+1}$, $\mathbb{C}^{m+1}$ and $\mathbb{H}^{\ell+1}$ carry the defining unitary representation of SO$(n{+}1)$, SU$(m{+}1)$ and Sp$(\ell{+}1)$, respectively, With elementary matrices $E_{\alpha\beta}$ defined by matrix elements $$\bigl(E_{\alpha\beta}\bigr)_{\gamma\delta} = \delta_{\alpha\gamma} \delta_{\beta\delta}
\qquad\textrm{with}\quad \alpha,\beta,\gamma,\delta=1,2,\ldots,n{+}1
\ \textrm{or} \ m{+}1 \ \textrm{or} \ \ell{+}1\ ,$$ some of the one-parameter subgroups in all three cases have the form $$\textrm{U}_{\alpha\beta}(t) \=
\bigl(E_{\alpha\alpha}{+}E_{\beta\beta}\bigr)\cos t +
\bigl(E_{\alpha\beta}{-}E_{\beta\alpha}\bigr)\sin t \=
\exp\bigl\{ t\,(E_{\alpha\beta}{-}E_{\beta\alpha}) \bigr\}
\qquad\textrm{with}\quad t\in[0,2\pi)\ .$$ For a suitable choice of the subgroup SO$(n)$, SU$(m)$ or Sp$(\ell)$, the circle $\textrm{U}_{\alpha\beta}(t)$ realizes a great circle in the coset space, i.e. the unit sphere, with an obvious circumference of $2\pi$. In other words, we may identify $E_{\alpha\beta}{-}E_{\beta\alpha}$ with a generator $\bar{I}_a$ of $\mathfrak{m}$ in the defining representation.
To connect this to the structure constants, we compute the trace of the square of this generator in the defining as well as in the adjoint representation of SO$(n{+}1)$, SU$(m{+}1)$ or Sp$(\ell{+}1)$: $$\tr_{\textrm{def}}\bigl(\bar{I}_a^2\bigr) \=
\tr\bigl((E_{\alpha\beta}{-}E_{\beta\alpha})^2\bigr) \= -2 \und
\tr_{\textrm{adj}}(\bar{I}_a \bar{I}_b) \=
\bar{f}_{aC}^{\hspace{2ex}D}\bar{f}_{bD}^{\hspace{2ex}C} \=
-\bar{g}_{ab} \= -\bar{\varkappa}\,\delta_{ab}\ .$$ On the other hand, the two traces are related by the dual Coxeter number $h^\vee$, $$\tr_{\textrm{adj}}(XY) \= \begin{cases}
h^\vee\,\tr_{\textrm{def}} (XY)
& \textrm{for}\quad X,Y \in so(n{+}1) \quad\ \textrm{with}\quad h^\vee=n{-}1 \\
2\,h^\vee\,\tr_{\textrm{def}} (XY)
& \textrm{for}\quad X,Y \in su(m{+}1) \quad\textrm{with}\quad h^\vee=m{+}1 \\
h^\vee\,\tr_{\textrm{def}} (XY{+}YX)
& \textrm{for}\quad X,Y \in sp(\ell{+}1) \quad\ \,\textrm{with}\quad h^\vee=\ell{+}2
\end{cases}\ ,$$ We conclude that (including even $n=2$) $$\varkappa_n = 2(n{-}1) \ ,\qquad
\tilde{\varkappa}_m = 4(m{+}1) \ ,\qquad
\hat{\varkappa}_\ell = 4(\ell{+}2)$$ for the orthogonal, unitary and symplectic case, respectively.
A rescaling of the structure constants effects the following changes, $$\bar{f}_{\cdot\cdot}^{\ \cdot} \mapsto \sfrac{1}{\rho}\,\bar{f}_{\cdot\cdot}^{\ \cdot}\ ,\quad
\bar{I}_\cdot \mapsto \sfrac{1}{\rho}\,\bar{I}_\cdot\ ,\quad
\bar{e}^\cdot \mapsto \rho\,\bar{e}^\cdot\ ,\quad
\diff\Omega_\cdot^2 \mapsto \rho^2\,\diff\Omega_\cdot^2\ ,\quad
\bar{g}_{\cdot\cdot} \mapsto \rho^2\, \bar{g}_{\cdot\cdot}\ ,\quad
\varkappa \mapsto \rho^2\,\varkappa \ ,$$ leading to a sphere of radius $\rho$.
[99]{} A. Actor,\
“Classical solutions of SU(2) Yang-Mills theories,”\
[Rev. Mod. Phys. [**51**]{} (1979) 461]{}.
T.A. Ivanova, O. Lechtenfeld and A.D. Popov,\
“Solutions to Yang-Mills equations on four-dimensional de Sitter space,”\
[Phys. Rev. Lett. [**119**]{} (2017) 061601]{}, [[arXiv:1704.07456 \[hep-th\]]{}]{}.
T.A. Ivanova, O. Lechtenfeld and A.D. Popov,\
“Finite-action solutions of Yang-Mills equations on de Sitter dS$_4$ and anti-de Sitter AdS$_4$ spaces,”\
[JHEP [**11**]{} (2017) 017]{}, [[arXiv:1708.06361 \[hep-th\]]{}]{}.
V. De Alfaro, S. Fubini and G. Furlan,\
“A new classical solution of the Yang-Mills field equations,”\
[Phys. Lett. B [**65**]{} (1976) 163]{}.
M. Lüscher,\
“SO(4) symmetric solutions of Minkowskian Yang-Mills field equations,”\
[Phys. Lett. B [**70**]{} (1977) 321]{}.
G.W. Gibbons and A.R. Steif,\
“Sphalerons and conformally compactified Minkowski spacetime,”\
[Phys. Lett. B [**346**]{} (1995) 255]{}, [[arXiv:hep-ph/9412210]{}]{}.
M.S. Volkov and D.V. Gal’tsov,\
“Gravitating non-Abelian solitons and black holes with Yang-Mills fields,”\
[Phys. Rept. [**319**]{} (1999) 1]{}, [[arXiv:hep-th/9810070]{}]{}.
S.W. Hawking and G.F.R. Ellis,\
[*The large scale structure of space-time,*]{}\
Cambridge University Press, Cambridge, 1975.
D. Kapetanakis and G. Zoupanos,\
“Coset space dimensional reduction of gauge theories,”\
[Phys. Rept. [**219**]{} (1992) 1]{}.
T.A. Ivanova and O. Lechtenfeld,\
“Yang-Mills instantons and dyons on group manifolds,”\
[Phys. Lett. B [**670**]{} (2008) 91]{}, [[arXiv:0806.0394 \[hep-th\]]{}]{}.
T.A. Ivanova, O. Lechtenfeld, A.D. Popov and T. Rahn,\
“Instantons and Yang-Mills flows on coset spaces,”\
[Lett. Math. Phys. [**89**]{} (2009) 231]{}, [[arXiv:0904.0654 \[hep-th\]]{}]{}.
I. Bauer, T.A. Ivanova, O. Lechtenfeld and F. Lubbe,\
“Yang-Mills instantons and dyons on homogeneous $G_2$-manifolds,”\
[JHEP [**10**]{} (2010) 044]{}, [[arXiv:1006.2388 \[hep-th\]]{}]{}.
![Newtonian potential contours for SU(2) (top) and $\sfrac{\textrm{SU}(3)}{\textrm{SU}(2)}$ (below).[]{data-label="fig1"}](v_u1.eps)
![Newtonian potential contours for SU(2) (top) and $\sfrac{\textrm{SU}(3)}{\textrm{SU}(2)}$ (below).[]{data-label="fig1"}](v_u2.eps)
![Newtonian potential contours for $\sfrac{\textrm{Sp}(2)}{\textrm{Sp}(1)}$ (top) and $\sfrac{\textrm{Sp}(3)}{\textrm{Sp}(2)}$ (below).[]{data-label="fig2"}](v_s1.eps)
![Newtonian potential contours for $\sfrac{\textrm{Sp}(2)}{\textrm{Sp}(1)}$ (top) and $\sfrac{\textrm{Sp}(3)}{\textrm{Sp}(2)}$ (below).[]{data-label="fig2"}](v_s2.eps)
[^1]: The prefactor $\sqrt{2m}$ is chosen for later convenience.
[^2]: The functions $\varphi(t)$ and $\psi(t)$ in this subsection are different from those in the previous one. We choose the prefactors for later convenience.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We propose a simple pedagogical way of introducing the Euler-MacLaurin summation formula in an undergraduate course on statistical mechanics. We put forward two alternative routes: the first one is the simplest and yields the first two terms of the expansion. The second one is somewhat more elaborate and takes into account all the correction terms. We apply both to the calculation of the simplest one-particle canonical partition functions for the translational, vibrational and rotational degrees of freedom.'
author:
- 'Francisco M. Fernández'
title: A simple way of approximating the canonical partition functions in statistical mechanics
---
Introduction {#sec:intro}
============
In statistical mechanics the thermodynamic functions are given in terms of the logarithm of the partition function and its derivatives with respect to volume, temperature, etc. The mathematical expression of the canonical partition function is an infinite sum over all the states of the system or, under some simplifying assumptions, the states of the individual particles.[@M73] Some of these infinite sums are commonly calculated approximately by means of the Euler-MacLaurin summation formula[@AS72] and most textbooks simply show how to apply it to the cases of interest.[@M73] A rigorous derivation of the summation formula may be rather too demanding for inexperienced students and one avoids it in introductory courses on statistical thermodynamics or statistical mechanics.
There is a relatively simple way of deriving the summation formula by means of operator methods[@FC96] but it also requires some kind of mathematical expertise that students of introductory courses may not posses. However, the students may feel more confident and confortable if they are shown how to carry out such calculations by means of mathematical methods that they have already learned in a first course on mathematical analysis.
In an introductory undergraduate course we show how to obtain the first two terms of the Euler-MacLaurin summation formula in an extremely simple way that only requires the students to be familiar with the Taylor expansion. The aim of this paper is to put forward this approach that we deem suitable for pedagogical purposes. In section \[sec:simple\_appro\] we outline the method and apply it to some simple examples in section \[sec:examples\]. If we decide to show the students how to obtain terms of higher order this simple approach results to be rather cumbersome. For that reason, in section \[sec:systematic\] we show how to derive the full summation formula by a relatively minor change of strategy that requires some additional mathematical skills. The reader may choose one or another approach depending on the level of the course. Finally, in section \[sec:conclusions\] we summarize the results and draw conclusions.
The simplest approximation {#sec:simple_appro}
==========================
The simplest canonical partition functions in statistical mechanics can be expressed as sums of the form $$S=\sum_{n=0}^{\infty }f(n), \label{eq:S}$$ where $f(n)$ should tend to zero sufficiently fast when $n\rightarrow \infty$ because, otherwise, the sum does not converge. In most cases one cannot obtain this sum in closed form and therefore resorts to some kind of approximation valid under certain conditions; for example, sufficiently high temperature. In order to derive such an approach we define $$S(x)=\sum_{n=0}^{\infty }f(n+x), \label{eq:S(x)}$$ that satisfies $$\begin{aligned}
S(0) &=&S \nonumber \\
\lim\limits_{x\rightarrow \infty }S(x) &=&0, \label{eq:S(x)_prop1}\end{aligned}$$ and $$S(x)-S(x+1)=f(x). \label{eq:S(x)_prop2}$$
Substituting the Taylor expansion $S(x+h)=S(x)+S^{\prime }(x)h+\frac{1}{2}S^{\prime \prime }(x)h^{2}+\ldots $ for $h=1$ into (\[eq:S(x)\_prop2\])
we obtain, after some rearrangement, $$-S^{\prime }(x)=f(x)+\frac{1}{2}S^{\prime \prime }(x)+\ldots .$$ If we integrate this expression between $x$ and $\infty $ and take into account that $S(x)$ and all its derivatives vanish at the upper limit we have $$S(x)=\int_{x}^{\infty }f(t)\,dt-\frac{1}{2}S^{\prime }(x)+\ldots ,
\label{eq:S(x)_iter}$$ that can be solved iteratively. In the first step we omit the derivatives of $S(x)$ so that this function is approximately given by $S(x)$ $\approx \int_{x}^{\infty }f(t)\,dt$. If we substitute this approximate result into the right-hand side of (\[eq:S(x)\_iter\]) we have $$S(x)\approx \int_{x}^{\infty }f(t)\,dt+\frac{1}{2}f(x).
\label{eq:S(x)_approx1}$$ Thus, for $x=0$ we obtain a simple approximation to the sum (\[eq:S\]): $$S(0)=S\approx \int_{0}^{\infty }f(t)\,dt+\frac{1}{2}f(0).
\label{eq:S_approx_simp}$$
The simple method just outlined is suitable for introductory courses because it only requires basic knowlegde in mathematical analysis. This procedure is not suitable for the systematic calculation of the corrections of higher order to the basic formula (\[eq:S\_approx\_simp\]) because it soon becomes rather cumbersome. However, this result is suitable for most porposes in an introductory course on statistical mechanics because several partitions functions for atomic and simple molecular systems can be easily derived from it.[@M73]
Some simple examples {#sec:examples}
====================
Under some simplifying assumptions that we will not discuss in this paper statistical mechanics tell us how to express the thermodynamic functions of atomic and molecular systems in terms of one-particle partition functions of the form[@M73] $$q(V,T)=\sum_{n}g_{n}e^{-\epsilon _{n}/(k_{B}T)}, \label{eq:q(V,T)}$$ where $V$ is the volume of the container, $T$ the absolute temperature, $\epsilon _{n}$ the $n$-th energy level (assumed to be $g_{n}$-fold degenerate) of the particle and $k_{B}$ the Boltzman constant.
Let us first consider a particle of mass $m$ in a one-dimensional box of length $L$ with impenetrable walls which is the starting point for the calculation of the thermodynamic properties of an ideal monoatomic gas.[@M73] If the particle has no internal structure its spectrum is only given by the translational degree of freedom: $$\epsilon _{n}^{t}=\frac{\hbar^{2}\pi^2
n^{2}}{2mL^{2}},\;n=1,2,\ldots , \label{eq:e_n_PB}$$ where $\hbar$ is the Planck constant $h$ divided $2\pi$ and $g_{n}=1$. Upon defining $\alpha =h^{2}/($ $8mk_{B}TL^{2})$ the partition function $q_t$ reads $$\begin{aligned}
q_{t} &=&S_{t}-1 \nonumber \\
S_{t} &=&\sum_{n=0}^{\infty }e^{-\alpha n^{2}}, \label{eq:S_PB}\end{aligned}$$ and the approximation (\[eq:S\_approx\_simp\]) yields $$S_{t}\approx \frac{\sqrt{\pi }}{2\sqrt{\alpha }}+\frac{1}{2}.
\label{eq:S_PB_appro}$$ This approximate expression is accurate for sufficiently small values of $\alpha $ or, equivalently, when the thermal de Broglie wavelength $\Lambda =h/\sqrt{2\pi mk_{B}T}$ is much smaller than the box length $L$.[@M73] The addition of corrections of higher order commonly improves the result of the approximate partition function, but in this case all of them vanish.[@FC96] Figure \[fig:PB\] shows that this approximate expression becomes increasingly more accurate as $\alpha $ decreases.
The spectrum of a one-dimensional harmonic oscillator is given by $$\epsilon _{n}^{v}=\left( n+\frac{1}{2}\right) h\nu ,\;n=0,1,\ldots ,
\label{eq:e_n_HO}$$ where $\nu $ is the frequency of the oscillation and $g_{n}=1$. The partition function for this degree of freedom is given by $$\begin{aligned}
q_{v} &=&e^{-\alpha /2}S_{v}, \nonumber \\
S_{v} &=&\sum_{n=0}^{\infty }e^{-\alpha n}, \label{eq:S_HO}\end{aligned}$$ where $\alpha =h\nu /(k_{B}T)$. In this case the sum $S_{v}$ is a geometric series that can be calculated exactly and therefore we can derive its small-$\alpha $ expansion in a straightforward way: $$S_{v}=\frac{1}{1-e^{-\alpha }}=\frac{1}{\alpha }+\frac{1}{2}+\frac{\alpha }{12}-\frac{\alpha ^{3}}{720}+\ldots . \label{eq:S_exact_exp}$$ If we define the vibrational temperature $\theta _{v}=h\nu /h_{B}$ then $\alpha =\theta _{v}/T$ and the Euler-MacLaurin formula is a good approximation when $\theta _{v}\ll T$. The approximate expression (\[eq:S\_approx\_simp\]) yields the first two terms of this series exactly.
The last example is the partition function for a rigid rotor with moment of inertia $I$. In this case, every energy level $$\epsilon _{J}^{r}=\frac{\hbar ^{2}}{2I}J(J+1),\;J=0,1,\ldots ,
\label{eq:e_J_RR}$$ is $(2J+1)$-fold degenerate and the partition function for the rotational degree of freedom is given by $$q_{r}=S_{r}=\sum_{J=0}^{\infty }(2J+1)e^{-\alpha J(J+1)}, \label{eq:S_RR}$$ where $\alpha =\hbar ^{2}/(2Ik_{B}T)$. The approximate expression (\[eq:S\_approx\_simp\]) yields $$S_{r}\approx \frac{1}{\alpha }+\frac{1}{2}, \label{eq:S_RR_appro1}$$ that is accurate enough if $\alpha $ is sufficently small as shown in figure \[fig:RR\]. It is custommary to define the rotational temperature $\theta
_{r}=\hbar ^{2}/(2Ik_{B})$[@M73] so that $\alpha =\theta _{r}/T\ll 1$ when $\theta _{r}\ll T$. It is worth mentioning that the $\alpha $-independent term in equation (\[eq:S\_RR\_appro1\]) is not exact as shown in the following section.
Systematic approach {#sec:systematic}
===================
If we decide to show the students how to obtain expression of higher order it is convenient to proceed in a different way. If we write the Taylor expansion discussed in section \[sec:simple\_appro\] as $$S(x+1)=\sum_{j=0}^{\infty }\frac{1}{j!}S^{(j)}(x), \label{eq:S(x+1)_Taylor}$$ where $S^{(j)}(x)$ is the $j$-th derivative of $S(x)$ with respect to $x$, then equation (\[eq:S(x)\_iter\]) reads $$S(x)=F(x)-\sum_{j=1}^{\infty }\frac{1}{(j+1)!}S^{(j)}(x),
\label{eq:S(x)_iter2}$$ where $F(x)=\int_{x}^{\infty }f(t)\,dt$.
Instead of trying to solve equation (\[eq:S(x)\_iter2\]) iteratively we propose a solution of the form $$S(x)=\sum_{k=0}^{\infty }a_{k}F^{(k)}(x), \label{eq:S(x)_ansatz}$$ where $F^{(k)}(x)=-f^{(k-1)}(x)$. In order to obtain the coefficients $a_{k}$ we substitute (\[eq:S(x)\_ansatz\]) into (\[eq:S(x)\_iter2\]) and compare the coefficients of $F^{(n)}(x)$ in the left- and right-hand sides; the result is $$\begin{aligned}
a_{n} &=&-\sum_{j=1}^{n}\frac{a_{n-j}}{(j+1)!},\;n=1,2,\ldots , \nonumber \\
a_{0} &=&1. \label{eq:a_n_rec}\end{aligned}$$ We thus have $$S(x)=F(x)-\sum_{k=0}^{\infty }a_{k+1}f^{(k)}(x), \label{eq:S(x)_appro_gen}$$ and $$S=\int_{0}^{\infty }f(t)\,dt-\sum_{k=0}^{\infty }a_{k+1}f^{(k)}(0).
\label{eq:S_appro_gen1}$$
Straightforward inspection of the first coefficients $a_{n}$ $$\begin{aligned}
a_{1} &=&-\frac{1}{2},\,a_{2}=\frac{1}{12},\,a_{3}=0,\,a_{4}=-\frac{1}{720},\,a_{5}=0,\,a_{6}=\frac{1}{30240},\,a_{7}=0,\, \nonumber \\
a_{8} &=&-\frac{1}{1209600},\,a_{9}=0,\,a_{10}=\frac{1}{47900160},
\label{eq:a_n}\end{aligned}$$ suggests that $a_{2j+1}$ vanish for all $j>0$. In order to prove this conjecture we consider the function $$u(x)=\frac{x}{e^{x}-1}, \label{eq:u(x)}$$ that satisfies $$u(x)-u(-x)=-x. \label{eq:u(x)-u(-x)}$$ If we substitute the Taylor expansion $$u(x)=\sum_{j=0}^{\infty }u_{j}x^{j}, \label{eq:u(x)_series}$$ into equation (\[eq:u(x)-u(-x)\]) we conclude that $u_{2j+1}=0$ for all $j>0$. If we now expand $(e^{x}-1)u(x)=x$ in a Taylor series about $x=0$ and compare the coefficients of $x^{n}$ in the left- and right-hand sides of the resulting equation we obtain a recurrence relation for the coefficients $u_{n}$ that is identical with equation (\[eq:a\_n\_rec\]) for the coefficients $a_{n}$. We thus conclude that $u_{n}=a_{n}$ for all $n$. Therefore, equation (\[eq:S\_appro\_gen1\]) reduces to $$S=\int_{0}^{\infty }f(t)\,dt+\frac{1}{2}f(0)-\sum_{k=1}^{\infty
}a_{2k}f^{(2k-1)}(0). \label{eq:S_appro_gen2}$$ It is worth noting that the coefficients $a_{j}$ are related to the Brillouin numbers $B_{j}$[@AS72] in the following way: $a_{j}=B_{j}/j!$.
The operator method leads to this result in a more straightforward way and is also more convenient for the discussion of the radius of convergence of the series.[@FC96] However, we do not discuss it here because it requires the introduction of functions of operators that may not be suitable for an undergraduate course.
Let us apply this more accurate approach to the examples discussed in section \[sec:examples\]. As pointed out in that section nothing can be done with the particle in a box because all the corrections of higher order vanish. This surprising result can be explained very easily and applies to all the sums in which $f(-x)=f(x)$ because $f^{(2k-1)}(-x)=-f^{(2k-1)}(x)$ and $f^{(2k-1)}(0)=0$. A more rigorous analysis of such problems, based on the Poisson summation formula, can be found eslewhere.[@FC96]
The harmonic oscillator is a suitable simple example for testing the summation formula (\[eq:S\_appro\_gen2\]) because we can calculate the exact expansion as shown in equation (\[eq:S\_exact\_exp\]). In this case every new term added to the summation formula (\[eq:S\_appro\_gen2\]) yields one more term of the small-$\alpha $ series (\[eq:S\_exact\_exp\]) as one may easily verify.
The application of the summation formula (\[eq:S\_appro\_gen2\]) to the partition function for the rigid rotor should be carried out with care. The reason is that $f^{(2k+1)}(0)=\alpha
^{k}P_{k+1}(\alpha )$, where $P_{k+1}(\alpha )$ is a polynomial function of $\alpha $ of degree $k+1$ and $P_{k+1}(0)\neq 0$ . Therefore, the summation formula that includes all the terms through $f^{(2k+1)}(0)$ only yields the small-$\alpha $ series correctly through degree $k$. For this reason the simple expression (\[eq:S\_approx\_simp\]) does not yield the correct $\alpha $-independent term that receives contributions from $f(0)$ and $f^{\prime }(0)$. The expansion accurate to $\alpha ^{4}$ is easily shown to be $$S_{r}=\frac{1}{\alpha }+\frac{1}{3}+\frac{\alpha }{15}+\frac{4\alpha ^{2}}{315}+\frac{\alpha ^{3}}{315}+\frac{4\alpha ^{4}}{3465}+\ldots ,
\label{eq:S_RR_EM}$$ if we add all the terms trough $f^{(9)}(0)$ in the summation formula (\[eq:S\_appro\_gen2\]) as argued above. Since this series does not converge[@FC96] one should truncate it before the terms start to increase.
Conclusions {#sec:conclusions}
===========
As stated in the introduction we show the students the simple method developed in section \[sec:simple\_appro\] because it is sufficient for the purposes of our course. However, if a motivated student wants to learn how to obtain the corrections of greater order that appear in some of the available texbooks on statistical mechanics then one can suggest him or her to try the systematic approach of section \[sec:systematic\] or even the operator method.[@FC96] In our opinion the present way of deriving the Euler-MacLaurin summation formula for statistical mechanics applications is more convenient for pedagogical purposes than the traditional one that appears in most textbooks on mathematics or numerical analysis.[@A99] Such traditional approaches are certainly more rigorous and general but require the students to be more experienced in mathematics.
In passing, it is worth mentioning that the summation formula given by equations (\[eq:a\_n\_rec\]) and (\[eq:S\_appro\_gen2\]) is suitable for motivating the students to resort to a computer algebra system. In this way they bypass the tedious algebraic manipulation of the equations that is necessary for the calculation of contributions of large order like those in equation (\[eq:S\_RR\_EM\]) and practise programming in any of such useful languages.
[9]{} D. A. McQuarrie, *Statistical Mechanics*, (Harper & Row, New York, 1973).
M. Abramowitz and I. A. Stegun, *Handbook of Mathematical Functions*, Ninth ed. (Dover, New York, 1972).
F. M. Fernández and E. A. Castro, *Algebraic Methods in Quantum Chemistry and Physics*, Mathematical Chemistry Series, (CRC, Boca Raton, New York, London, Tokyo, 1996).
T. M. Apostol, “An elementary view of Euler’s summation formula”, Amer. Math. Month. **106**, 409-418 (1999).
![$S_t$ calculated by means of the sum (\[eq:S\_PB\]) (points) and the simple Euler-MacLaurin expression (\[eq:S\_PB\_appro\]) (solid line)[]{data-label="fig:PB"}](PB.eps){width="9cm"}
![$S_r$ calculated by means of the sum (\[eq:S\_RR\]) (points) and the simple Euler-MacLaurin expression (\[eq:S\_RR\_appro1\]) (solid line)[]{data-label="fig:RR"}](RR.eps){width="9cm"}
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'In the last time the cold dark matter (CDM) model has suggested more and more that it is not able to describe all the properties of nearby galaxies that can be observed in great detail as well as that it has some problems in the mechanism by which matter is more rapidly gathered into large-scale structure such as galaxies and clusters of galaxies. In this work we revisit an alternative model, the scalar field dark matter (SFDM) model, which proposes that the galactic haloes form by condensation of a scalar field (SF) very early in the Universe, i.e., in this model the haloes of galaxies are astronomical Bose-Einstein Condensate drops of SF. On the other hand, large-scale structures like clusters or superclusters of galaxies form similar to the $\Lambda$CDM model, by hierarchy, thus all the predictions of the $\Lambda$CDM model at cosmological scales are reproduced by SFDM. This model predicts that all galaxy haloes must be very similar and exist for higher redshifts than in the $\Lambda$CDM model. In the first part of this review we revisit the cosmological evolution of SFDM model with a scalar potential $m^2\Phi^2/2+\lambda\Phi^4/4$ with two different frameworks: the field and fluid approach. We derive the evolution equations of the SF in the linear regime of perturbations as well. The scalar fluctuations have an oscillating growing mode and therefore, this kind of dark matter could lead to the early formation of gravitational structures in the Universe. We also revisit how BEC dark matter haloes exhibit a natural cut of the mass power spectrum. In the last part, we study the core central density profiles of BEC dark matter haloes and fit high-resolution rotation curves, we show a sample of some low surface brightness galaxies. The mean value of the logarithmic inner density slopes is $\alpha = - $0.27 $\pm$ 0.18. Using a model independent new definition of the core in the BEC density profile, we show that the recent observation of the constant dark matter central surface density can be reproduced. We conclude that in light of the difficulties that the $\Lambda$CDM model is currently facing the SFDM model can be a worthy alternative to keep exploring further.'
author:
- 'Juan Magaña[^1]'
- 'Tonatiuh Matos$^{*}$'
- 'Victor H. Robles$^{*}$'
- 'Abril Suárez$^{*}$'
title: A brief Review of the Scalar Field Dark Matter model
---
Introduction
============
Nowadays the most accepted model in cosmology which explains the evolution of the Universe is known as $\Lambda$CDM. In this model 4 per cent of the total content of the Universe is baryonic matter, 22 per cent is non-baryonic dark matter (DM) and the rest is in some form of cosmological constant. $\Lambda$CDM has achieved several observations with outstanding success, like for example the fact that the cosmic microwave background radiation can be explained in great detail and that it provides a framework within one can understand the large-scale isotropy of the Universe and important characteristics on the origin, nature and evolution of density fluctuations which are believed to give rise to galaxies and other cosmic structures. There remain, however, certain conflicts at galactic scales, like the cusp profile of central densities in galactic halos, the overpopulation of substructures predicted by $N$-body numerical simulations which are an order of magnitude larger than what has been observed, etc (see for example [@clowe], [@klypin], [@moore], and [@penny]).
There are other important issues of the $\Lambda$CDM model which require further considerations. Observations point out to a better understanding of the theory beginning with the Local Void, which contains far fewer galaxies than the expected. Another problem arises for the so called pure disk galaxies, which do not appear in numerical simulations of structure formation in the standard model. These problems would be solved, if the structure grew faster than it does in the standard paradigm [@peebles]. On the other hand, @lee_komatzu also found that the collision velocity of 3000km/s at R200 for the Bullet Cluster is very unlikely and might no be explained by the standard model. A final example of inconsistencies can be seen in a paper by @Shaun, who found anomalies in the mass power spectrum obtained by SDSS and the one obtained with the CDM model. With these and other results it seems necessary to change the $\Lambda$CDM paradigm to try and explain the formation of structure in the Universe.
The incorporation of a new kind of DM, different from the one proposed by the $\Lambda$CDM model into the big bang theory holds out the possibility of resolving some of these issues. Several authors have proposed interesting alternatives in where they try to solve the latter explained difficulties that the CDM scenario seem not solve. Some of these models consider warm dark matter ([@colin], [@villa]), self interacting dark matter ([@loeb],[@spergel]), and other exotic scenarios as braneworld models [@miguel]. In fact, there are models that do not include DM but instead modify the Newtonian force law [@milgrom; @sanders]. Other alternative scenario that has received much attention in the last years is the scalar field dark matter (SFDM) model. The main idea is simple (see @guzmanII), the nature of the DM is completely determined by a fundamental scalar field $\Phi$. The SFDM model proposes that galactic haloes form by Bose-Einstein condensation of a scalar field (SF) whose boson has an ultra-light mass of the order of $m\sim 10^{-22}$eV. From this mass it follows that the critical temperature of condensation $T_c\sim1/m^{5/3}\sim$TeV is very high, therefore, they form Bose-Einstein Condensates (BEC) drops very early in the Universe. In addition, the Compton length $\lambda_{c}=2 \pi \hbar/m$ associated to this boson is about $\sim$ kpc that corresponds to the dark halo-size of typical galaxies in the Universe. Thus, it has been proposed that these drops are the haloes of galaxies (see @further), $i.e.$, that haloes are huge drops of SF. On the other hand, big structures form like in the $\Lambda$CDM model, by hierarchy [@further; @suarez], thus, all successful predictions of the standard model at large scales are well reproduced by SFDM. In other words, in the SFDM model the haloes of galaxies do not form hierarchically, they are formed at the same time and in the same way when the Universe reaches the critical temperature of condensation of the SF. From this it follows that all galaxies must be very similar because they formed in the same manner and at the same moment [@leelim]. Therefore, from this paradigm we have to expect that there exist well formed galaxy haloes at higher redshifts than in the $\Lambda$CDM model. In this model the scalar particles with that ultra light mass are such that their wave properties avoid the cusp problem and reduce the high number of small satellites by the quantum uncertainty principle [@hu; @lundgren] which is another problem that is still present in the $\Lambda$CDM model [@harko_core; @klypin; @victor]. Summarizing, it is remarkable that with only one free parameter, the ultra-light scalar field mass ($m \sim 10^{-22}$eV), the SFDM model fits:
1. The evolution of the cosmological densities [@phi2].
2. The acoustic peaks of the cosmic microwave background [@ivan].
3. The scalar field has a natural cut off, thus the substructures in clusters of galaxies is avoided naturally. With a scalar field mass of $m_\phi\sim10^{-22}$eV the amount of substructure is compatible with the observed one [@further; @hu; @suarez].
4. We expect that SFDM forms galaxies earlier than the CDM model, because they form BECs at a critical temperature $T_c >> $TeV. So if SFDM is right, we have to see big galaxies at high redshifts with similar features [@phi2; @suarez].
5. The rotation curves of big galaxies and LSB galaxies [@Lesgourgues:2002hk; @Arbey:2003sj; @harkomnras11; @argelia; @victor].
6. With this mass, the critical mass of collapse for a real scalar field is just $10^{12}\,M_{\odot}$, i.e., the one observed in galaxies haloes [@alcubierre].
7. The observed properties of dwarf galaxies, i.e., the minimum length scale, the minimum mass scale, and their independence from the brightness [@leelim].
8. And recently it has been demonstrated that if the scalar mass is $m\sim10^{-22}$eV, the SFDM haloes would have cores large enough to explain the longevity of the cold clump in Ursa Minor and the wide distribution of globular clusters in Fornax (see @lora).
The idea was first considered by @sin [@ji_sin] and independently by @L9 [@guzmanI; @guzmanII; @lee_koh] suggested bosonic dark matter as a model for galactic halos [@dehnen]. In the BEC model, DM haloes can be described, in the non-relativistic regime, as Newtonian gravitational condensates made up of ultra light bosons condensated in a single macroscopic wave function. Thus, BECs haloes can be described as a coherent scalar field $\Phi$.
Several authors have introduced a dynamic scalar field with a certain potential $V(\Phi)$ as a candidate to dark matter, although there is not yet an agreement for the correct form of the potential of the field. One interesting work pointing this way was done by [@mat_ure2000] and independently by [@sahni] where they used a potential of the form $V(\Phi)=V_{0}\left[\cosh\left( \xi \Phi \right)-1\right]$ to explain the core density problem for disc galaxy halos in the $\Lambda$CDM model (see also [@Wetterich], [@Ratra] and [@Copeland]). @guzmanII presented a model for the DM in spiral galaxies, in which they supposed that DM is an arbitrary SF endowed with a scalar potential. Other scalar potential widely used to describe DM is $V(\Phi)=m^{2}\Phi^{2}/2$ [@turner; @phi2]. This potential is very interesting because it can mimic the cosmological evolution of the Universe predicted by the $\Lambda$CDM model. If we consider a SF self-interaction, we need to add a quartic term to the SF potential [@Arbey:2001qi; @Arbey:2005zc; @Arbey:2006it; @Arbey:2008um; @fabio], in this case the equation of state of the SF is that of a polytope of index n=1 (see [@suarez; @harkomnras11; @harko_cosmo; @Arbey:2001jj; @Lesgourgues:2002hk]).
Different issues of the cosmological behavior of the SFDM/BEC model have been studied (see for example [@further; @Arbey:2001qi; @hu; @kainb; @luis_bose; @lundgren; @marsh; @ivan; @ivan2; @suarez; @harko_cosmo; @harkomnras11; @phi2; @Arbey:2005zc; @matos_open; @chavanisIII; @velter; @woo]). @hu proposed fuzzy dark matter that is composed of ultra light scalar particles who are initially in the form of a BEC. In their work @woo used a bosonic dark matter model to explain the structure formation via high-resolution simulations, @urenaAIP [@luis_bose] reviewed the key properties that may arise from the bosonic nature of SFDM models. Recently @harko_cosmo [@harkomnras11] developed a further analysis of the cosmological dynamics of SFDM/BEC as well as of the evolution of their fluctuations (see also [@chavanisIII]). In the same direction, @suarez studied, the growth of scalar fluctuations and large-scale structure formation with the fluid approach of the SFDM/BEC model.
On the other hand, SFDM/BEC model has provided to be a good candidate for DM haloes of galaxies in the Universe because it can explain many aspects where the standard model fails ([@argelia; @boharko; @lee2; @leelim; @leelimchoi; @lora; @harko_core; @victor]). In addition, many numerical simulations have been performed to study the gravitational collapse of SFDM/BEC model [@argelia2; @argelia3; @colpi; @fcoluis; @colapsob; @apjpaco; @gleiser]. @chavanisI [@chavanisII] found approximate analytical expression and numerical solutions of the mass-radius relation of SFDM/BEC haloes. Recently, @rindler give constraints on the boson mass to form and mantain more than one vortex in SFDM/BEC haloes. These constraints are in agreement with the ultra light mass founded in previous works (see also [@kaina; @zinner]). Lately, @lora performed N-body simulations to study the dynamics observed in Ursa Minor dwarf galaxy. They model the dark matter halo of Ursa Minor as a SFDM/BEC halo to establish constraints for the boson mass. Moreover, they introduce a dynamical friction analysis with the SFDM/BEC model to study the wide distribution of globular clusters in Fornax. An overall good agreement is found for the ultra light mass $\sim 10^{-22}$eV of bosonic dark matter.
The main objective of this review is to introduce the framework of the SFDM/BEC model that assumes that DM is a scalar field that involves a self-interacting potential of the form $V(\Phi)=m^2\Phi^2/2+\lambda\Phi^4/4$, where $m\sim 10^{-22}$eV is the mass of the scalar field, [@lee_koh], [@further] and [@hu].
We solve the Friedmann equations for the SFDM/BEC model to show that it behaves just like the CDM model. Also, we show that the SFDM/BEC predicts galaxy formation earlier than the CDM model, because they form BEC at a critical temperature $T_c >>$ TeV. So, if SFDM is right, this would imply that we have to see big galaxies at high redshifts. In order to do this, we study the density fluctuations of the scalar field from a hydrodynamical point of view, this will give us some information about the energy density of dark matter halos necessary to obtain the observational results of large-scale structure. Here we will give some tools that might be necessary for the study of the early formation of structure.
In section \[cosmology\] we study the cosmological evolution of the SF with a field and a fluid approach. We obtain a numerical solution for the scalar density as well as analytical expressions for the kinetic and potential energies of the SF. Also, we study the evolution for the density contrast of scalar fluctuations in the linear regime. In section \[galacticSFDM\] we study the rotation curves and the cusp/core discrepancy with SFDM haloes to compare with the predictions of CDM. Finally we give our conclusions in section \[sec:conclusions\].
The Scalar Field Dark Matter Cosmology {#cosmology}
======================================
The framework of the standard cosmological model is of a homogeneous and isotropic Universe whose evolution is best described by Friedmann’s equations that come from general relativity and whose main ingredients can be described by fluids whose characteristics are very similar to those we see in our Universe. Of course, the Universe is not exactly homogeneous and isotropic but this standard model does give us a framework within which we can study the evolution of the expansion rate of the Universe as well as the evolution of small fluctuations that give rise to gravitational structures in the Universe as galaxies and clusters of galaxies.
In this section, we study the cosmological dynamics of the SFDM/BEC model and the evolution of their scalar perturbations with two different frameworks: the field and fluid approach.
Background Universe: The Field Approach {#sec:background}
---------------------------------------
We use the Friedmann-Lemaître-Robertson-Walker (FLRW) metric with scale factor $a(t)$. Our background Universe is composed by SFDM ($\Phi_{0}(t)$) endowed with a scalar potential $V\equiv V(\Phi_0)$, radiation ($z$), neutrinos ($\nu$), baryons ($b$), and a cosmological constant ($\Lambda$) as dark energy. We begin by recalling the basic background equations. From the energy-momentum tensor $\mathbf{T}$ for a scalar field, the scalar energy density $T_0^0$ and the scalar pressure $T_j^i$ are given by $$T_0^0=-\rho_{\Phi_0}=-\left(\frac{1}{2}\dot{\Phi}_0^2+ V\right),
\label{rhophi0}$$ $$T_j^i=P_{\Phi_0}=\left(\frac{1}{2}\dot{\Phi}_0^2-V\right)\delta_j^i,
\label{pphi0}$$ where the dots stand for the derivative with respect to the cosmological time and $\delta^i_j$ is the Kronecker delta. Thus, the Equation of State (EoS) for the scalar field is $P_{\Phi_{0}}=\omega_{\Phi_{0}}\,\rho_{\Phi_0}$ with $$\omega_{\Phi_{0}}= \frac{\frac{1}{2}\dot{\Phi}_{0}^{2}\,-\,V}{\frac{1}{2}\dot{\Phi}_{0}^{2}\,+\,V}.
\label{ec:w}$$
The radiation fields, the baryonic component and the cosmological constant are represented by perfect fluids with baryotropic equation of state $P_{\gamma}=(\gamma-1)\rho_{\gamma}$, where $\gamma$ is a constant, $0\le \gamma \le 2$. For example, $\gamma_{z}=\gamma_{\nu}=4/3$ for radiation and neutrinos, $\gamma_{b}=1$ for baryons, and for a cosmological constant $\gamma_{\Lambda}=0$.
The Einstein-Klein-Gordon equations that describe this Universe are (in units $c=\hbar=1$).
$$\begin{aligned}
\dot H&=&-\frac{\kappa^2}{2}\left(\dot{\Phi}_{0}^2+\frac{4}{3}\rho_z+\frac{4}{3}\rho_{\nu}+
\rho_{b}\right),\\
\ddot{\Phi}_{0} &+& 3\,H \dot{\Phi}_{0}+ V,_{\Phi_{0}}=0,\label{eq:nopertKG}\\
{\dot\rho_{z}}&+&4\,H \rho_{z}=0,\\
{\dot\rho_{\nu}}&+&4\,H \rho_{\nu}=0,\\
{\dot\rho_{b}}&+& 3\,H \rho_{b}=0, \end{aligned}$$
\[eq:back\]
with the Friedmann constraint $$H^2=\frac{\kappa^2}{3}\left(\rho_{\Phi_{0}}+\rho_{z}+
\rho_{\nu}+\rho_{b} + \rho_{\Lambda} \right),
\label{eq:FC}$$ being $\kappa^{2} \equiv 8\pi G$, $H \equiv \dot{a}/a$ the Hubble parameter and the commas stand for the derivative with respect to scalar field. Notice that background scalar quantities at zero order have the subscript $0$. In order to solve the system of equations (\[eq:back\]), we define the following dimensionless variables $$\begin{aligned}
x&\equiv& \frac{\kappa}{\sqrt{6}}\frac{\dot{\Phi}_{0}}{H},\quad
u \equiv \frac{\kappa}{\sqrt{3}}\frac{\sqrt{V}}{H},\nonumber\\
z&\equiv& \frac{\kappa}{\sqrt{3}}\frac{\sqrt{\rho_{z}}}{H},\,\,\,
\nu \equiv \frac{\kappa}{\sqrt{3}}\frac{\sqrt{\rho_{\nu}}}{H},\nonumber\\
b&\equiv& \frac{\kappa}{\sqrt{3}}\frac{\sqrt{\rho_{b}}}{H},\,\,\,
l \equiv \frac{\kappa}{\sqrt{3}}\frac{\sqrt{\rho_{\Lambda}}}{H}.
\label{eq:varb}\end{aligned}$$
Here we take the quadratic scalar potential $V=m^{2}\Phi_{0}^{2}/2$ with $m \sim1 \times10^{-22}$eV. Using these variables, the equations (\[eq:back\]) for the evolution of the background Universe are transformed into
$$\begin{aligned}
x'&=& -3\,x - s u+\frac{3}{2}\Pi\,x, \label{eq:dsb_x}\\
u'&=& s x +\frac{3}{2}\Pi\,u, \label{eq:dsb_u}\\
z'&=&\frac{3}{2}\left(\Pi-\frac{4}{3} \right)\,z, \label{eq:dsb_z}\\
\nu'&=&\frac{3}{2}\left(\Pi-\frac{4}{3} \right)\,\nu, \label{eq:dsb_nu}\\
b'&=&\frac{3}{2}\left(\Pi -1\right)\,b, \label{eq:dsb_b}\\
l'&=&\frac{3}{2}\Pi\,l, \label{eq:dsb_l}\\
s'&=&s_0 \,s^{-k}, \label{eq:dsb_s}\end{aligned}$$
\[eq:dsb\]
where the prime denotes a derivative with respect to the e-folding number $N=\ln a$, and $\Pi$ is defined as $$-\frac{\dot H}{H^2}=\frac{3}{2}(2x^2+b^2+\frac{4}{3} z^2+\frac{4}{3} \nu^2) \equiv \frac{3}{2}\Pi.$$
We have introduced the variable $s\equiv C_{0}/H$, where $C_{0}$ is a constant, as a control parameter for the dynamics of $H$. In Eq. (\[eq:dsb\_s\]), $s_0$ is a constant and the exponent $k$ is $\le0$ (see [@mayra] for more details of the control parameter). With these variables, the density parameters $\Omega_{i}$ for each component $i$ can be written as $$\begin{aligned}
\Omega_{\Phi_{0}}&=&x^2+u^2, \nonumber\\
\Omega_{z}&=&z^2,\nonumber\\
\Omega_{\nu}&=&{\nu}^2,\nonumber\\
\Omega_{b}&=&b^2,\nonumber\\
\Omega_{\Lambda}&=&l^2,\label{eq:dens}\end{aligned}$$ subject to the Friedmann constraint $$\label{eq:fri}
x^2+u^2+z^2+\nu^{2}+b^2+l^2=1.$$ In addition, we may write the EoS of the scalar field as $$\omega_{\Phi_{0}}=\frac{x^{2}-u^{2}}{\Omega_{\Phi_{0}}}.
\label{eq:dlw}$$ Since $\omega_{\Phi_{0}}$ is a function of time, if its temporal average tends to zero, this would imply that $\Phi^2$-dark matter can be able to mimic the EoS for CDM.
We solve the system of equations (\[eq:dsb\]) for the background Universe numerically with a four order Adams-Bashforth-Moulton (ABM) method. We take as initial conditions the best estimates from 5-years [@WMAP5] and 7-years WMAP [@WMAP7] values to $\Omega^{(0)}_{\Lambda}=0.73$, $\Omega^{(0)}_{DM}=0.22994$, $\Omega^{(0)}_b=0.04$, $\Omega^{(0)}_z=0.00004$, $\Omega^{(0)}_{\nu}=0.00002$ and the exponent $k=0$. In Fig.\[fig:phi\] we see the evolution of the scalar field $\Phi_{0}$. This figure shows how the scalar field oscillates very stark about the minimum of the scalar potential $V$, on the bottom of this figure we show a zoom of these oscilations. In Fig.\[fig:Ox\] we show the evolutions of the kinetic ($\dot\Phi^2/2$) and potential energy ($m^2\Phi^2/2$) of the scalar field, as expected, the SF oscilations are translated into very stark oscilations for the kinetic and the potential energies of the SF. However, observe the evolution of the dark matter density of the scalar field, that means $\rho_\Phi=(\dot\Phi^2+m^2\Phi^2)/2$ in Fig.\[fig:back\]. Note the following crucial point, although the kinetic and potential energies show very stark oscillations, the sum of both energies does not oscillates at all, this sum is the density parameter $\Omega_{\Phi_{0}}$, which does not display any oscillation. The important point we have to take from now on into account is that the oscillations are not physical observables at all, they are a feature of the SF, what we observe in fact is the density of the SF which does not oscillates.
Fig.\[fig:back\] shows the numerical evolution of the density parameters in our model[^2]. At early times, radiation dominates the evolution of the Universe. Later on, the Universe has an epoch where the energy density radiation is equal to the dark matter density, at $z_{eq}$, then dark matter begins to dominate the evolution. The recombination era in SFDM model occurs at $z\sim1000$. At later times, the cosmological constant dominates the dynamics of the Universe at $z_{\Lambda} \sim 0.5$. The behavior is exactly the same as in the $\Lambda$CDM model. Furthermore, in the recombination era of the SFDM model, the neutrinos made up $\sim12$ per cent of the Universe completely in agreement with the measurements of WMAP. Fig.\[fig:w0\] shows the evolution of the EoS for the SF. Although the EoS varies with time (oscillations), the temporal average, $\left<\omega_{\Phi_0}\right>$, drops to zero. Therefore, SFDM is like a pressureless fluid and behaves as CDM at cosmological scales ([@turner], [@further], [@phi2], [@matos_open]).
Background Universe: The Fluid Approach {#fondo}
---------------------------------------
In this section we perform a transformation in order to solve the Friedmann equations analytically with the approximation $H<<m$.
Here we take the scalar potential as $V=m^{2}\Phi^2/2\hbar^2+\lambda\Phi^4/4$, (we remain that we use units where $c=\hbar=1$ unless it is stipulated). Then for the ultra-light boson particle we have that $m\sim 10^{-22}$ eV.
Now we express the SF, $\Phi_0$, in terms of the new variables $S$ and $\hat\rho_0$, where $S$ is constant in the background and $\hat{\rho}_0$ will be the energy density of the fluid also in the background. So, our background field is proposed as $$\Phi_0=(\psi_0\, e^{-\, imt/\hbar}+\psi^*_0\, e^{\, imt/\hbar})$$ where, $$\psi_0(t)=\sqrt{\hat{\rho}_0(t)}\, e^{\, iS/\hbar}$$ and with this our SF in the background can be finally expressed as, $$\Phi_0=2\sqrt{\hat{\rho}_0}\cos(S-mt/\hbar),
\label{tri}$$ with this we obtain $$\begin{aligned}
\dot{\Phi}_0^2&=&\hat{\rho}_0\left[\frac{\dot{\hat{\rho}}_0}{\hat{\rho}_0}\cos(S-mt/\hbar)\right.\nonumber\\
&-&\left.2(\dot{S}-m/\hbar)\,\sin(S-mt/\hbar)\right]^2.
\label{backSF}
\end{aligned}$$
To simplify, observe that the uncertanty relation implies that $m\Delta t\sim\hbar$, and for the background in the non-relativistic case the relation $\dot S/m\sim0$ is satisfied. Notice also that for the background we have that the density goes as $(\ln\hat{\rho}_0)\dot{}=-3H,$ but we also have that $H\sim 10^{-33}$ eV $<<m\sim 10^{-22}$ eV, so with these considerations at hand for the background, in (\[backSF\]) we have $$\dot{\Phi}_0^2=4\frac{m^2}{\hbar^2}\hat{\rho}_0\sin^2(S-mt/\hbar)$$
Finally, substituting this last equation and equation (\[tri\]) into (\[rhophi0\]) when taking $\lambda=0$, we obtain $$\rho_{\Phi_0}=2\frac{m^2}{\hbar^2}\hat{\rho}_0[\sin^2(S-mt/\hbar)+\cos^2(S-mt/\hbar)]=2\frac{m^2}{\hbar^2}\hat{\rho}_0.
\label{trigo}$$
Comparing this result with (\[eq:dens\]) we have that the identity $\Omega_{\Phi_0}=2m^2\hat{\rho}_0/\hbar^2$ holds for the background, so comparing with (\[trigo\]), $$x=\sqrt{2\hat{\rho}_0}\frac{m}{\hbar}\sin(S-mt/\hbar)
\label{eq:cinet}$$ $$u=\sqrt{2\hat{\rho}_0}\frac{m}{\hbar}\cos(S-mt/\hbar).
\label{eq:potencial}$$
In terms of the two analytic results, we show the evolution of the kinetic and potential energies (\[eq:cinet\]) and (\[eq:potencial\]) in Fig. \[fig1\], where for the evolution we used the e-folding number $N$ and the fact that $a\sim t^n\rightarrow t\sim\,{e}^{N/n}$.
Observe the excelent accordance with the numerical results in [@phi2] for the kinetic and potential energies of the background respectively. Therefore, we expect that if we sum the analytical kinetic and the potential energies we will obtain the same behavior for the SFDM density shown in Fig. \[fig:back\]. In what follows we will use the analytic expressions (\[eq:cinet\]) and (\[eq:potencial\]) instead of the numerical ones.
Scalar Field Dark Matter Fluctuations
=====================================
Nowadays it is known that our Universe is not exactly isotropic and spatially homogeneous like the FLRW metric describes. There exist small deviations of this model that were the cosmological seeds to lead the large-scale structure formation in the Universe. In this section we compute the growth of the SFDM overdensities $\delta\rho_{\Phi}$ at the stage when the density contrast $\delta\equiv\delta\rho_{\Phi}/\rho_{\Phi_{0}}$ is much smaller than unity. It is believed that the Universe was almost uniform after inflation, with very small density contrast. As the Universe expanded, the small over-densities grew until they began to collapse, leading to the formation of structure in the Universe. Here we consider small enough deviations in the FLRW model, so that they can be treated by linear perturbation theory.
The Linear Regime of SFDM perturbations: The Field approach {#sec:linear}
-----------------------------------------------------------
In the SFDM paradigm, the dark haloes of the galaxies are BECs composed of ultra light scalar particles with $m\sim10^{-22}$eV then it would be about $10^{68}m$ GeV$^{-1}$ scalar particles to follow in a single galaxy and the occupation numbers in galactic haloes are so big that the DM behaves as a coherent classical scalar field $\Phi$ that obeys the Klein-Gordon equation $(\square^2+m^2/\hbar^2)\Phi=0$, where $\Box$ is the D’Alambertian. Therefore, to study the structure formation in the Universe we describe a non-interacting SFDM model such that: i) it can describe it more as a field than as particles and ii) we find a function that only depends on the three spatial coordinates and time.
First, we introduce the perturbed metric tensor in the FLRW background, we only consider scalar perturbations. We then give the equation of energy-momentum conservation and the Einstein field equations for the perturbed metric.
By definition, a perturbation done in any quantity, is the difference between its value in some event in real space-time, and its corresponding value in the background. So, for example for the SF we have $$\Phi=\Phi_0(t)+\delta\Phi(\mbox{\boldmath$x$},t),
\label{pert}$$ where the background is only time dependent, while the perturbations also depend on the space coordinates. Similar cases apply for the metric; $$\begin{aligned}
g_{00}&=&-a^2(1+2\phi),\nonumber\\
g_{0i}&=&a^2B,_i,\nonumber\\
g_{ij}&=&a^2[(1-2\psi)\delta_{ij}+2E,_{ij}].
\end{aligned}$$ Here $\psi$ is a perturbation associated to the curvature and E is asociated to the expansion. We will work under the Newtonian gauge, which is defined when $B=E=0$. An advantage of using this gauge is that here the metric tensor $g_{\mu \nu}$ is diagonal, and so the calculations become much easier. We will only work with scalar perturbations, vector and tensor perturbations are eliminated from the beginning, so that only scalar perturbations are taken into account. Another advantage in using this gauge is that $\phi$ will play the role of the gravitational potential, this will help us to have a simpler physical interpretation, i.e., both potentials $\phi$ and $\psi$ are then related. This metric has already been used in other works (see for example [@bardeen], [@ma] and [@malik]).
The perturbed energy-momentum tensor for the SF can be written as the background value $\mathbf{T}_0\equiv\mathbf{T}_0(t)$ plus a perturbation $\delta\mathbf{T}\equiv\delta\mathbf{T}(x^\mu)$ where $x^{\mu}=[t,x^i]$, i. e. $$\mathbf{T}=\mathbf{T}_0+\delta\mathbf{T}.$$
We now derive the perturbed evolution equations for the different quantities mentioned above; the scalar perturbation $\delta\Phi$ and the scalar potential $\psi$. For the perturbed energy-momentum tensor, we get
$$\begin{aligned}
\delta T^0_0&=&-\delta\rho_{\Phi}=-(\dot{\Phi}_0\dot{\delta\Phi}-\dot{\Phi}_0^2\psi+V,_{\Phi_{0}}\delta\Phi),\label{rhopert}\\
\delta T^0_i&=&-\frac{1}{a}(\dot{\Phi}_0\delta\Phi,_i),\\
\delta T^i_j&=&\delta p_{\Phi}=(\dot{\Phi}_0\dot{\delta\Phi}-\dot{\Phi}_0^2\psi-V,_{\Phi_{0}}\delta\Phi)\delta^i_j.\end{aligned}$$
\[eq:Tperturbado\]
Where in equations (\[eq:Tperturbado\]) the dot denotes differentiating with respect to cosmological time $t$, which is related to conformal time by the simple relation $$\frac{d}{d\eta}=a\frac{d}{dt}.$$
In the Newtonian gauge, the metric tensor $g_{\mu\nu}$ becomes diagonal and from this, in the trace of the Einstein’s equations, the scalar potentials $\psi$ and $\phi$ are identical and, therefore, $\psi$ plays the role of the gravitational potential $$\psi-\phi=0.
\label{cond}$$ Usually this equation contains a term of anisotropic stress, which vanishes in the case of a SF. Altogether, the perturbed Einstein’s equations $\delta G^i_j= \kappa^{2}\delta T^i_j$ to first order for a SF in the Newtonian gauge are $$\begin{aligned}
-8\pi G\delta\rho_{\Phi}&=&6H(\dot{\phi}+H\phi)-\frac{2}{a^2}\nabla^2\phi, \nonumber\\
8\pi G\dot{\Phi}_0\delta\Phi,_i&=&2(\dot{\phi}+H\phi),_i, \nonumber\\
8\pi G\delta p_{\Phi}&=&2[\ddot{\phi}+4H\dot{\phi}+(2\dot{H}+3H^2)\phi].
\label{eq:sfmet}\end{aligned}$$ These equations describe the evolution of the scalar perturbations. For the evolution of the perturbations in the SF we use the perturbed Klein-Gordon equation $$\ddot{\delta\Phi}+3H\dot{\delta\Phi}-\frac{1}{a^2}\nabla^2\delta\Phi+V,_{\Phi\Phi}\delta
\Phi+2V,_{\Phi}\phi-4\dot{\Phi}_0\dot{\phi}=0.
\label{eq:kgl}$$
In order to solve the equations (\[eq:sfmet\]) and (\[eq:kgl\]), we will turn to Fourier’s space. The beauty of this expansion relaying on the fact that each Fourier mode will propagate independently. To first order, the derivation of Fourier’s components is straightforward. The perturbation $\delta\Phi$ relates to its Fourier component $\delta\Phi_k$ by $$\begin{aligned}
\delta\Phi(t,x^i)&=&\int d^3k\delta\Phi(t,k^i)\,\textrm{exp}(ik_ix^i),\nonumber\\&=&\int d^3k\delta\Phi_k\,\textrm{exp}(ik_ix^i),\end{aligned}$$ where $k$ is the wave number. Here the wave number is defined as $k=2\pi /L$, being $L$ the length scale of the perturbation.
The perturbed equations (\[eq:sfmet\]) altogether with the SF transformation read
$$\begin{aligned}
8\pi G(3H\dot{\Phi}_0\delta\Phi_k)+\frac{2k^2}{a^2}\phi&=&-8\pi G(\dot{\Phi}_0\dot{\delta\Phi_k}\nonumber\\&&-\phi\dot{\Phi}_0^2+V,_{\Phi}\delta\Phi_k),\label{eq:fsa}\nonumber\\ \\
2(H\phi+\dot{\phi})&=&8\pi G\dot{\Phi}_0\delta\Phi_k, \label{eq:fsb}\\
2[\ddot{\phi}+4H\dot{\phi}+(2\dot{H}+3H^2)\phi]&=&
8\pi G(\dot{\Phi}_0\dot{\delta\Phi_k}
\nonumber\\&&-\phi\dot{\Phi}_0^2-V_{\Phi}\delta\Phi_k)\label{eq:fsc},\end{aligned}$$
\[eq:fsp\]
and the Klein-Gordon equation (\[eq:kgl\]) transforms into $$\ddot{\delta\Phi}_k+3H\dot{\delta\Phi_k}+\frac{k^2}{a^2}\delta\Phi_k+V,_{\Phi\Phi}\delta\Phi_k-4\dot{\phi}\dot{\Phi}_0+2\phi V,_{\Phi}=0.
\label{eq:kgfou}$$ These set of equations describe the evolution of the perturbations. Eq. (\[eq:fsa\]) makes reference to the evolution of the energy density, Eq. (\[eq:fsb\]) to the evolution of the gravitational potential and finally, Eq. (\[eq:kgfou\]) refers to the perturbations over the SF.
Now, from the time derivative of (\[rhopert\]), we have $$\begin{aligned}
\dot{\delta\rho_{\Phi}}&=&(\ddot{\Phi}_0+V_{\Phi})\dot{\delta\Phi_k}
\nonumber\\&&+(\ddot{\delta\Phi_k}+V,_{\Phi\Phi}\delta\Phi_k-\dot{\Phi}_0\dot{\phi})
\dot{\Phi}_0-2\phi\dot{\Phi}_0\ddot{\Phi}_0.\end{aligned}$$ Performing a Fourier transformation of the above equation and combining it with the unperturbed and perturbed Klein-Gordon equations (\[eq:nopertKG\]) and (\[eq:kgfou\]) respectively, and with the use of equation (\[eq:fsp\]) we arrive at $$\dot{\delta\rho_{\Phi}}=-6H\dot{\Phi}_0\dot{\delta\Phi_k}+6\phi\dot{\Phi}_0^2H-\frac{2k^2}{a^2\kappa^2}(H\phi+\dot{\phi})+3\dot{\phi}
\dot{\Phi}_0^2.$$ And on the other hand we have, $$\delta p_{\Phi}+\delta\rho_{\Phi}=2\dot{\Phi}_0\dot{\delta\Phi_k}-2\dot{\Phi}_0^2\phi,$$ Thus, $$\dot{\delta\rho_{\Phi}}=-3H(\delta p_{\Phi}+\delta\rho_{\Phi})-\frac{2k^2}{a^2\kappa^2}(H\phi+\dot{\phi})+3\dot{\phi}\dot{\Phi}_0^2.$$ This last equation can be expressed in terms of the density contrast $\delta_{\Phi}$ making use of the equations from the background, we have $$\dot\delta_{\Phi}+3H(\frac{\delta p_{\Phi}}{\delta\rho_{\Phi}}-\omega_{\Phi})\delta_{\Phi}=3\dot{\phi}(1+\omega_{\Phi})-G_{\phi},
\label{pro1}$$ where we have defined the function $G_{\phi}$ as $$G_{\phi}=\frac{2k^2}{a^2\kappa^2}\frac{\dot\phi+H\phi}{\rho_{\Phi_{0}}}.$$ It is convenient to define the function $F_{\phi}$ as well, $$F_{\phi}=1+\omega_{\Phi}.$$ Taking the time average of equation (\[pro1\]) we obtain $$\dot\delta_{\Phi}+3H(\frac{<\delta p_{\Phi}>}{<\delta\rho_{\Phi}>}-<\omega_{\Phi}>)\delta_{\Phi}
=3\dot{\phi}<F_{\Phi}>-<G_{\phi}>.
\label{eq:deltlini}$$ In equation (\[eq:deltlini\]) for the radiation and matter dominated eras the first term in the parenthesis goes as $<\delta p_{\Phi}>/< \delta\rho_{\Phi}>\approx 0$, see for example [@further], also we can see from Fig.\[fig:w0\] that $\left<\omega_{\Phi_{0}}\right>\rightarrow 0$. It is easy to see the temporal average of the terms $F_{\Phi}$ and $G_{\phi}$ of the equation (\[eq:deltlini\]). From Fig.\[fig:FPhi0\] we observe that that $\left< F_{\Phi} \right>$ tends to one. On the other hand, since we are using post-newtonian approximation $\left<G_{\phi}\right>$ drops to zero, meaning that the second term on the right-hand side of Eq. (\[eq:deltlini\]) dissapears. So, due to the scalar oscillations around the minimum of the scalar potential, the unperturbed SF behaves very similar to the CDM model, but equation (\[eq:deltlini\]) tell us that their perturbations too, all the growing behavior for the $k$ modes are recovered and preserved so far. Finally, we conclude that Eq. (\[eq:deltlini\]) resembles the equation for the density contrast as in the standard CDM model [@chung]. This means that the SFDM perturbations in the linear regime grow exactly as CDM perturbations.
The rescaled equations for SFDM perturbations
---------------------------------------------
We now study the cosmological evolution of our model through a dynamical system. In order to solve the system of equations for the perturbations given in Eqs. (\[eq:fsp\]) and (\[eq:kgfou\]), we define the following dimensionless variables given by $$\begin{aligned}
U&\equiv&-\frac{\kappa}{\sqrt{6}}\frac{V,_{\Phi_{0}}}{H^2},\,\,\,\,\,\,\,\,\,
l_1\equiv\phi,\nonumber\\
x_2&\equiv&\frac{\dot{\phi}}{H},\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,
z_1\equiv\frac{\kappa}{\sqrt{6}}\delta\Phi_k,\nonumber\\
z_2&\equiv&\frac{\kappa}{\sqrt{6}}\frac{\delta\dot{\Phi}_k}{H}.
\label{eq:dvlr}\end{aligned}$$
Using the variables defined in Eq. (\[eq:dvlr\]), the equations can be transformed into an autonomus system with respect to the e-folding number $N$, then the density contrast can be written in terms of these variables
$$\delta=\frac{2[x(z_2-xl_1)-Uz_1]}{\Omega_{\Phi_{0}}}.
\label{eq:deltala}$$
Therefore the transition between the linear and non-linear regime of the SFDM perturbations can be studied through a numerical code.
The Natural Cut of the mass power spectrum
------------------------------------------
We rewrite equation (\[eq:kgfou\]) in the following form
$$\ddot{\delta\Phi}_k+(\frac{k^2}{a^2}+V,_{\Phi\Phi})\delta\Phi_k=-3H\dot{\delta\Phi_k}+4\dot{\phi}\dot{\Phi}_0-2\phi V,_{\Phi}.
\label{eq:kgfou_2}$$
Equation (\[eq:kgfou\_2\]) is the one of a harmonic oscillator with a damphing term $3H\dot{\delta\Phi}$ and an extra force. This equation contains oscillating, growing and decaying solutions, depending on the factor $\frac{k^2}{a^2}+V,_{\Phi\Phi}$. If this factor is positive, we expect that the fluctuations oscillate, only if the factor is negative we will have growing up fluctuations. This factor contains the scale $k^2/a^2$, which is always positive. Thus, if we expect to have growing up fluctuation of the scalar field, this happen only when the potential $V,_{\Phi\Phi}$ is negative. Then equation (\[eq:kgfou\_2\]) will have growing up solutions for certain values of the scale $k/a$, this is, only for enough small values of the scale, which means for enough big fluctuations. This is the way we can fix the mass of the scalar field, the potential depends on the value of the scalar field mass. Observing the smallest structure we see in the Universe, we are able to fix the mass of the scalar field. Using this observations one can avoid the problem of substructure of the $\Lambda$CDM. In order to do this, in [@further] the mass of the scalar field was fixed in $10^{-22}$eV.
The Linear Regime of SFDM perturbations: The Fluid Approach {#fluct}
-----------------------------------------------------------
The SF $\Phi$ has very stark oscillations from the beginning, these oscillations are transmitted to the fluctuations which apparently seems to grow very fast and are too big. Nevertheless, this behavior is not physical, because we only see the oscillations of the fields, but we can not see clearly the evolution of its density, [@matos_open]. In order to drop out these oscillations, in what follows we perform two transformations. The first one changes the perturbed Klein-Gordon equation into a kind of ’Schrödinger’ equation and the second transforms this last equation into a hydrodynamical system, where we can interpret the physical quantities easier and the observable quantities become much clear.
First, we consider the perturbed Klein-Gordon \[eq:kgl\] where we have set $\dot{\phi}=0$ $$\delta\ddot{\Phi}+3H\delta\dot{\Phi}-\frac{1}{a^2}\hat\nabla^2\delta\Phi+V,_{\Phi\Phi}\delta\Phi+2V,_{\Phi}\phi=0.
\label{K-G}$$ Now we express the perturbed SF $\delta\Phi$ in terms of the field $\Psi$, $$\delta\Phi=\Psi \,{e}^{-\,{i}mt/\hbar}+\Psi^*\,{e}^{\,{i}mt/\hbar},$$ term which oscillates with a frequency proportional to $m$ and $\Psi=\Psi(\mbox{\boldmath$x$},t)$ which would be proportional to a wave function of an ensamble of particles in the condensate. With this equation and the expresion for the potential of the scalar field, (\[K-G\]) transforms into $$-\,{i}\hbar(\dot{\Psi}+\frac{3}{2}H\Psi)+\frac{\hbar^2}{2m}(\square\Psi+9\lambda|\Psi|^2\Psi)+m\phi\Psi=0,
\label{schrodinger}$$ where we have defined $$\square=\frac{\,{d}^2}{\,{d}t^2}+3H\frac{\,{d}}{\,{d}t}-\frac{1}{a^2}\hat\nabla^2.$$ Notice that this last equation could represent a kind of ’Gross-Pitaevskii’ equation in an expanding Universe. The only modification of equation (\[schrodinger\]) in comparison to the Schrödinger or the Gross-Pitaevskii equation is the scale factor $a^{-1}$ associated to the co-moving spatial gradient and that the Laplacian $\hat\nabla^2=\partial^2_{\mbox{\boldmath{x}}}$ transforms into the D’ Alambertian $\Box$.
To explore the hydrodynamical nature of bosonic dark matter, we will use a modified fluid approach. Then, to make the connection between the theory of the field and the condensates waves function, the field is proposed as (Madelung transformation [@pitaevskii]) $$\Psi=\sqrt{\hat{\rho}}\,\,{e}^{\,{i}S},
\label{psi}$$ where $\Psi$ will be the condensates wave function with $\hat{\rho}=\hat{\rho}(\mbox{\boldmath$x$},t)$ and $S=S(\mbox{\boldmath$x$},t)$, [@ginzburg]. Here we have separated $\Psi$ into a real phase $S$ and a real amplitude $\sqrt{\hat{\rho}}$ and the condition $\mid\Psi\mid^2=\Psi\Psi^*=\hat{\rho}$ is satisfied. From (\[psi\]) we have $$\begin{aligned}
\dot{\hat{\rho}}&+&3H\hat{\rho}-\frac{\hbar}{m}\hat{\rho}\Box S+\frac{\hbar}{a^2m}\hat\nabla S\hat\nabla\hat{\rho}
-\frac{\hbar}{m}\dot{\hat{\rho}}\dot{S}=0,\nonumber\\
\hbar\dot S/m&+&\omega\hat{\rho}+\phi+\frac{\hbar^2}{2m^2}\left(\frac{\Box\sqrt{\hat{\rho}}}{\sqrt{\hat{\rho}}}\right)
+\frac{\hbar^2}{2a^2}[\hat\nabla(S/m)]^2\nonumber\\
&-&\frac{\hbar^2}{2}(\dot S/m)^2=0.\label{hidro1}
\label{hidro}
\end{aligned}$$ Now, taking the gradient of (\[hidro1\]) then dividing by $a$ and using the definition $$\mbox{\boldmath$v$}\equiv\frac{\hbar}{ma}\hat\nabla S
\label{vel}$$ we have, $$\begin{aligned}
\dot{\hat{\rho}}&+&3H\hat{\rho}-\frac{\hbar}{m}\hat{\rho}\Box S+\frac{1}{a}\mbox{\boldmath$v$}\nabla\hat{\rho}
-\frac{\hbar}{m}\dot{\hat{\rho}}\dot{S}=0,\nonumber\\
\dot{\mbox{\boldmath$v$}}&+&H\mbox{\boldmath$v$}+\frac{1}{2a\hat{\rho}}\nabla p+\frac{1}{a}\nabla\phi
+\frac{\hbar^2}{2m^2a}\nabla\left(\frac{\Box\sqrt{\hat{\rho}}}{\sqrt{\hat{\rho}}}\right)\nonumber\\
&+&\frac{1}{a}(\mbox{\boldmath$v\cdot$}\nabla)\mbox{\boldmath$v$}-\hbar(\dot{\mbox{\boldmath$v$}}+H\mbox{\boldmath$v$})
(\dot S/m)=0\label{eq:1}.
\end{aligned}$$ where in (\[hidro1\]) $\omega=9\hbar^2\lambda/2m^2$ and in (\[eq:1\]) we have defined $p=\omega\hat{\rho}^2$.
It is worth noting that, to this moment this last set of equations do not involve any approximations with respect to equation (\[schrodinger\]) and can be used in linear and non-linear regimes.
Now, neglecting squared terms, second order time derivatives and products of time derivatives in this last set of equations we get, $$\begin{aligned}
\frac{\partial\hat{\rho}}{\partial t}&+&\nabla\mbox{\boldmath$\cdot$} (\hat{\rho}\mbox{\boldmath$v$})+3H\hat{\rho}=0\label{navier},\\
\frac{\partial\mbox{\boldmath$v$}}{\partial t}&+&H\mbox{\boldmath$v$}+(\mbox{\boldmath$v\cdot$}\nabla)\mbox{\boldmath$v$}-
\frac{\hbar^2}{2m^2}\nabla(\frac{1}{2\hat{\rho}}\nabla^2\hat{\rho})+\omega\nabla\hat{\rho}\nonumber\\
&+&\nabla\phi=0\label{navier1},\\
\nabla^2\phi&=&4\pi G\hat{\rho},\label{poisson}
\end{aligned}$$ where the equation for the gravitational field is given by the Poisson’s equation (\[poisson\]). In these equations we have introduced $\mbox{\boldmath$r$}=a(t)\mbox{\boldmath$x$}$, such that $1/a\hat\nabla=\nabla=\partial_{\mbox{\boldmath{r}}}$.
Equation (\[vel\]) shows the proportionality between the gradient of the phase and the velocity of the fluid. Note that [$v$]{} can represent the velocity field for the fluid and $\hat{\rho}$ will be the particles density number within the fluid. Also there exists an extra term of third order for the partial derivatives in the waves amplitude which goes as the gradient of $\frac{\hbar^2}{2m^2}\frac{\Box\sqrt{\hat{\rho}}}{\sqrt{\hat{\rho}}}$, this term would result in a sort of ’quantum pressure’ that would act against gravity. We remain that $\phi$ represents the gravitational field. These two sets of equations (\[navier\]) and (\[navier1\]) would be analogous to the Euler’s equations of classical ’fluids’, with the main difference that there exists a ’quantum term’, which we will call $Q$ and will be given by $Q=\frac{\hbar^2}{2m^2}\frac{\square\sqrt{\hat{\rho}}}{\sqrt{\hat{\rho}}}$ which can describe a force or a sort of negative quantum pressure.
For equation (\[navier\]) we have that $\hat{\rho}$ will represent the mass density or the particles density number of the fluid, where all the particles would have the same mass. Finally, these equations describe the dynamics of a great number of non-interacting identical particles that manifest themselves in the form of a fluid, also equation (\[schrodinger\]) can describe a great number of non-interacting but self-interacting identical particles in the way of a Bose gas, when the probability density is interpreted as the density number.
Now, these hydrodynamical equations are a set of complicated non-linear differential equations. To solve them we will restrict ourselves to a vecinity of total equilibrium.
For this let $\hat{\rho}_0$ be the mass density of the fluid in equilibrium, the average velocity $\mbox{\boldmath$v$}_0$ will be taken as zero in equilibrium, so we will only have $\mbox{\boldmath$v$}(\mbox{\boldmath$x$},t)$ out of equilibrium. Then, the matter in the Universe will be considered as a hydrodynamical fluid inside an Universe in expansion. This system will then evolve in this Universe and later on they will collapse because of their gravitational attraction.
Then from (\[navier\]) for the mass density of the fluid in equilibrium we have, $$\frac{\partial\hat{\rho}_0}{\partial t}+3H\hat{\rho}_0=0,$$ with solutions of the form $$\hat{\rho}_0=\frac{\rho_{0i}}{a^3},
\label{back}$$ where as we know, in general if we have an equation of state of the form $\hat p=\omega\hat\rho$ and consider CDM or dust as dark matter such that $\hat p=0$ it holds that $\hat\rho\propto a^{-3}.$ Then, when the scale factor was small, the densities were necessarily bigger. Now, the particles density number are inversely proportional to the volume, and must be proportional to $a^{-3}$, therefore the matter energy density will also be proportional to $a^{-3}$, result that is consistent with our expression (\[back\]). In addition, the cosmological behavior of $\hat{\rho}_0$ is in agreement with the numerical results show in Fig. \[fig:back\].
Now for the system out of equilibrium we have $$\begin{aligned}
\frac{\partial\delta\hat{\rho}}{\partial t}&+&3H\delta\hat{\rho}
+\hat{\rho}_0\nabla\mbox{\boldmath$\cdot$}\delta\mbox{\boldmath$v$}=0,\nonumber\\
\frac{\partial\delta\mbox{\boldmath$v$}}{\partial t}&+&H\delta\mbox{\boldmath$v$}
-\frac{\hbar^2}{2m^2}\nabla(\frac{1}{2}\nabla^2\frac{\delta\hat{\rho}}{\hat{\rho}_0})+\omega\nabla\delta\hat{\rho}
+\nabla\delta\phi=0,\nonumber\\
\nabla^2\delta\phi&=&4\pi G\delta\hat{\rho},
\label{outeq}
\end{aligned}$$ equations that are valid in a Universe in expansion. In order to solve system (\[outeq\]) we look for solutions in the form of plane waves, for this the convenient ansatz goes as $$\begin{aligned}
\delta{\hat\rho}&=&\hat\rho_1(t)\exp(\,{i}\mbox{\boldmath$k\cdot x$}/a),\nonumber\\
\delta\mbox{\boldmath$v$}&=&\mbox{\boldmath$v$}_1(t)\exp(\,{i}\mbox{\boldmath$k\cdot x$}/a),\nonumber\\
\delta\phi&=& \phi_1(t)\exp(\,{i}\mbox{\boldmath$k\cdot x$}/a).\nonumber
\end{aligned}$$ where [$x$]{} is the position vector and [$k$]{} is a real wavevector which corresponds to a wavelength $L$. If we substitute these ansatz in the set of equations (\[outeq\]), we then have $$\begin{aligned}
\frac{\,{d}\hat\rho_1}{\,{d} t}&+&3H\hat\rho_1+\,{i}\frac{{\hat\rho}_0}{a}\mbox{\boldmath$k\cdot v$}_1=0,\label{eq1}\\
\frac{\,{d}\mbox{\boldmath$v$}_1}{\,{d} t}&+&H\mbox{\boldmath$v$}_1+\,{i}\frac{\hat\rho_1}{a}\left(\frac{v_{q}^2}{\hat{\rho}_0}-
4\pi G\frac{a^2}{k^2}+\omega\right)\mbox{\boldmath$k$}=0,\\
\phi_1&+&4\pi G\frac{a^2}{k^2}\hat\rho_1=0,
\label{modos}
\end{aligned}$$ where we have defined the velocity $$v^2_{q}=\frac{\hbar^2k^2}{4a^2m^2}.
\label{eq:vq}$$ To solve the system is convenient to rotate the coordinate system so that the propagation of the waves will be along the direction of one of the axes. For this we know that the velocity vector can be divided into longitudinal (parallel to [$k$]{}) and transverse (perpendicular to [$k$]{}) parts, such that $\mbox{\boldmath$v$}_1=\nu\mbox{\boldmath$k$}+\mbox{\boldmath$v$}_2$, where $\mbox{\boldmath$v$}_2$ is the vector perpendicular to the wave propagation vector $\mbox{\boldmath$k\cdot v$}_2=0$. In terms of $\mbox{\boldmath$v$}_2$ for equations (\[eq1\])-(\[modos\]) we have $$\begin{aligned}
\frac{\,{d}\hat\rho_1}{\,{d} t}&+&3H\hat\rho_1+\,{i}\frac{{\hat\rho}_0}{a}k^2\nu=0,\label{eq2}\\
\frac{\,{d}\nu}{\,{d} t}&+&H\nu+\frac{\,{i}}{a}(\frac{v_{q}^2}{\hat{\rho}_0}-4\pi G\frac{a^2}{k^2}+\omega)\hat\rho_1=0,
\label{lambda}
\end{aligned}$$ in addition to an equation for $\mbox{\boldmath$v$}_2$, $\,{d}\mbox{\boldmath$v$}_2/\,{d} t+H\mbox{\boldmath$v$}_2=0$, with solutions $\mbox{\boldmath$v$}_2=C/a$ with $C$ a constant of integration, i.e., perpendicular modes to the wave vector are eliminated with the expansion of the Universe. Now, if we use the result (\[back\]), then equation (\[eq2\]) can be written as $$\frac{\,{d}}{\,{d}t}\left(\frac{\hat\rho_1}{\hat{\rho}_0}\right)=-\frac{\,{i}k^2\nu}{a}. \label{lambda1}$$ System (\[eq2\])-(\[lambda\]) can be treated as in the case of a Universe with no expansion, so combining the two equations and with the aid of (\[lambda1\]), we get $$\frac{\,{d}^2\delta}{\,{d}t^2}+2H\frac{\,{d}\delta}{\,{d}t}+\left[(v_{q}^2+\omega\hat{\rho}_0)\frac{k^2}{a^2}
-4\pi G{\rho}_0\right]\delta=0,
\label{delta}$$ where $\delta=\hat\rho_1/\hat{\rho}_0=\rho_1/{\rho}_0$ is defined as the density contrast. This will be a fundamental equation in the understanding of the evolution of the primordial fluctuations.
Results
=======
First we will give a brief summary of the results for the $\Lambda$CDM model, this will enable us to make a direct comparison with our results.
For CDM the equation for the evolution of the density contrast is given by, $$\frac{\,{d}^2\delta}{\,{d}t^2}+2\,H\frac{\,{d}\delta}{\,{d}t}+\left(c_{s}^2\frac{k^2}{a^2}-4\pi G\hat{\rho}_0\right)\delta=0,
\label{deltaCDM}$$ where $c_{s}$ is defined as the sound velocity (which in our case it is not). Now lets analyze equation (\[deltaCDM\]) at the beginning of the matter dominated era a time just after the epoch of equality, and just before recombination when the radiation has cooled down and the photons do not interact with the electrons anymore, for a relativistic treatment see [@gorini]. In this era, $a\geqslant a_{eq}$, practically all the interesting fluctuation modes are well within the horizon, and the evolution of the perturbations can be well described within the newtonian analysis. At this time, matter behaves like dust with zero pressure. So we have $a\sim t^{2/3}$, $c_{s}^2k^2/a^2\approx 0$ and $\hat{\rho}_0\sim t^{-2}$ therefore $H=(2/3)1/t$. For equation (\[deltaCDM\]) we have $$\frac{\,{d}^2\delta}{\,{d}t^2}+\frac{4}{3}\frac{1}{t}\frac{\,{d}\delta}{\,{d}t}-\frac{2}{3}\frac{1}{t^2}\delta=0.$$
The solutions to this equation are of the form $$\delta(t)\to t^{2/3}C_1+\frac{C_2}{t},
\label{deltat}$$ where $C_1$ and $C_2$ are integration constants, from this solution we can see that we have modes that will disappear as time goes by, and modes that grow proportionally to the expansion of the Universe. This is an important result, because then the density contrast will grow proportionally to the expansion of the Universe when this is dominated by matter. Then, these fluctuations can maybe grow and give life to the galaxies, clusters of galaxies and all the large-scale structure we see now a days.
Now lets see what happens to the SFDM at this epoch ($a\geqslant a_{eq}$). The evolution of the perturbations in this case will be given by equation (\[delta\]).
In general we have that in equation (\[delta\]) the term $v_q$ is very small throughout the evolution of the pertubations ($v_q\leq
10^{-3}ms^{-1}$ for small $k$), so it really does not have a significant contribution on its evolution.
When the condition $\lambda=0$ is taken we can have a BEC that might be or might not be stable, if there exists stability the results of SFDM are consistent with those obtained from CDM (in this case both equations (\[delta\]) and (\[deltaCDM\]) are almost equal), the condition of stability for the BEC in the SFDM case will come from the study of $\lambda$ together with $Q$. $$\frac{\,{d}^2\delta}{\,{d}t^2}+2H\frac{\,{d}\delta}{\,{d}t}+\left(v_{q}^2\frac{k^2}{a^2}
-4\pi G{\rho}_0\right)\delta=0,
\label{deltal}$$
As we can see in Fig. \[fig2\] the perturbations used for the $\Lambda$CDM model grow in a similar way for the SFDM model, when $\lambda=0$, in this case both perturbations can give birth to structures quite similar in size, and this will happen with all the fluctuations as long as $k$ is kept small.
When $\lambda\neq 0$ the results are quite different, so when discussing the evolution of the density perturbations, there are two different cases: i) In the case of $\lambda >0$ the amplitude of the density contrast tends to decrease as $\lambda$ grows bigger and bigger away from zero until the amplitude of the density takes negative values (around $\lambda\sim 10^{8}$), telling us that this kind of fluctuations can not grow in time, and hence do not form a BEC. ii) On the other hand if $\lambda <0$ the fluctuations for the density contrast alway grow despite their size, this results means either than the fluctuations grow and form a stable BEC or than the density grows because it is collapsing into a single point and our BEC might be unstable, the study of the stability of this fluctuations needs then to be studied with non-linear perturbation theory. These results are shown in Fig. \[fig3\], in both figures \[fig2\] and \[fig3\] the initial condition for $\delta$ goes as $\delta\sim 1*10^{-5}$ in accordance with the data obtained from WMAP.
If these fluctuations result stable and because they are big in size, this means that they can only give birth to large structures. These fluctuations can then help for the formation of large clusters or other large-scale structure in the Universe at its early stages (around $a\geqslant a_{eq}$). Then, as these kind of SFDM can only interact with radiation in a gravitational form it is not limited by its interaction with radiation, and the dark matter halos can then create potential wells that will collapse early in time giving enough time for the structures to form. Then if DM is some kind of SFDM, the luminous matter will follow the DM potentials giving birth to large-scale structure.
Finally, a further analysis is required to obtain better results of the growth of scalar perturbations with self-interacting SFDM as well as to study the transition between linear and no linear regime of scalar fluctuations. Nevertheless, we conclude that the SFDM model at cosmological scale is a plausible alternative to DM nature because it preserves all the success of the $\Lambda$CDM model.
Scalar Field Dark Matter in galactic dynamics {#galacticSFDM}
=============================================
In this section we study the SFDM/BEC model at galactic scales, we will focus on the cusp/core discrepancy between the numerical predictions of the standard model and the astronomical observations.
As already said before, recent observations in far and nearby galaxies have shown that the $\Lambda$CDM model faces serious conflicts when trying to explain the galaxy formation at small scales [@victor; @friedmann]. For instance, in the $\Lambda$CDM simulations the halos present rising densities towards the central region behaving as $\rho \sim r^{-1}$ well within 1 kpc [@navarro2010]. On the other hand, several observations suggest that the dynamics (rotation curves) of dwarf and LSB galaxies are more consistent with a constant central density [@kuzio; @deblok1; @sanchez1], this is most commonly known as the cusp/core problem [@deblok1].
Studying a wide range of galaxies of different morphologies and with magnitudes in the interval $-22 \leq M_{B} \leq -8$ @donato fit their rotation curves (RC) using a Burkert profile for the DM [@burkert] and found that $$\log (\mu_{0}/M_{\odot}pc^{-2})= 2.15 \pm 0.2 \label{eq:logmu}$$ remains approximately constant, where $$\mu_{0} = \rho_{0}r_{0} \label{eq:mu}$$ with $\rho_{0}$ the central DM density and $r_{0}$ the core radius. Similar result where found in @kormendy [@spano]. Exploring further the constant value of $\mu_{0}$ for the DM, @gentile found that within $r_{0}$ the DM central surface density in terms of the mass inside it ($M_{< r_{0}}$) is $<\Sigma>_{0,DM} = M_{< r_{0}} / \pi r^{2}_{0}
\approx 72^{+42}_{-27} $M$_{\odot}$ pc$^{-2}$,the gravitational acceleration due to DM felt by a test particle at the radius $r_{0}$ was found to be $$g_{DM}(r_{0})=G \pi <\Sigma>_{0,DM} = 3.2^{+1.8}_{-1.2}\times 10^{-9} \textrm{cm s}^{-2}, \label{eq:gDM}$$ additionally they reported the acceleration due to the luminous matter at $r_{0}$ to be $g_{bar}(r_{0})= 5.7^{+3.8}_{-2.8}\times 10^{-10} \textrm{cm s}^{-2}$.
In the $\Lambda$CDM model the galaxies have evolve through numerous mergers and grew in different environments, the star formation and basic properties of the galaxies are not expected to be a common factor among them. Therefore giving both the constancy of $\mu_{0}$ and the core in the central regions of galaxies seems very unlikely in this model.
Any model trying to become a serious alternative to $\Lambda$CDM has to succeed in reproducing observations in which the standard model fails but also has to keep the solid description at large scale. It is always necesary to test the SFDM/BEC model with the two observations mentioned above, the cusp/core problem and the constant DM central surface density. In order to do this, in [@victor] it was used the Thomas-Fermi approximation and a static BEC DM halo to fit rotation curves of a set of galaxies. However, so far there was no comparison between the density profile and the data, in that work this was done by fitting rotation curves of 13 high resolution low surface brightness (LSB) galaxies and additionally compare the fits to two characteristic density profiles 1) the cuspy Navarro-Frenk-White (NFW) profile that results from N-body simulations using $\Lambda$CDM and 2) the Pseudo Isothermal (PI) core profile. The comparison allow to show that the model is in general agreement with the data and with a core in the central region. For the second goal the meaning of a core is somewhat ambiguous.
Large scale $N$-body simulations using collissionless CDM the internal region of DM halos show a density distribution described by a power law $ \rho \sim r^{\alpha}$ with $\alpha \approx -1$, such behaviour is what is now called a cusp. On the other hand, observations mainly in dwarf and LSB galaxies seem to prefer a central density going as $ \rho \sim r^{0}$. This discrepancy between observation and the CDM model receives the name of cusp/core problem. Among the empirical profiles most frequently used to describe the constant density behavior in this galaxies are the PI [@begeman], the isothermal [@athana] and the Burkert profile [@burkert]. Even though their behavior is similar in the central region and is specified by the central density fitting parameter, their second parameter called the core radius does not represent the same idea. For instance, in the PI profile (eq.(\[eq:PI\])) we see that the core radius will be the distance in which the density is half the central density. For the Burkert profile the core radius $R^{burk}_{c}$ will be when $\rho^{burk}(R^{burk}_{c}) = \rho^{burk}_{0}/4$ and for an isothermal profile (I) [@spano] $\rho^{I}(R^{I}_{c}) = \rho^{I}_{0}/2^{3/2}$. Hence, we see an ambiguity in the meaning of the core radius, they get the same name but the interpretation depends on the profile. If we want to compare the central density of LSB galaxies with that of NFW, it usually suffice to have a qualitative comparison, so far this is what we have been doing by fitting empirical profiles. However, high resolution rotation curves demand a more quantitative comparison. Indeed, if we want to test models by fitting RCs we have to know the specific meaning and size of the core, then we will be able to tell if a model is consistent with a cusp or not by making a direct comparison with the data.
In [@victor] it was proposed a new definition for the core and core radius that allow us to decide when a density profile is cusp or core. Using this definition in the SFDM/BEC model discussed above it is possible to find that the SFDM/BEC model can reproduced the constant value of $\mu_{0}$ and as a crosscheck with the PI profile, in [@victor] was found that the results are in very good agreement with observations. This argues in favor of the model and the core definition.
DM density profiles
-------------------
Following [@victor] we provide the dark matter density profiles that will be used for the analysis. The case in which the dark matter is in the form of a static BEC and the number of DM particles in the ground state is very large was considered in @boharko. Following this paper and assuming the Thomas-Fermi approximation [@dalfovo; @pitaevskii] which neglects the anisotropic pressure terms that are relevant only in the boundary of the condensate, the system of equations describing the static BEC in a gravitational potential $V$ is given by $$\nabla p \biggl( \dfrac{ \rho}{m} \biggr) = - \rho \nabla V, \label{eq:BEC1}$$
$$\nabla^{2} V = 4 \pi G \rho, \label{eq:BEC2}$$
with the following EoS $$p(\rho) = U_{0} \rho^{2},$$ where $ U_{0} = \dfrac{2 \pi \hbar^{2} a}{m^{3}}$, $\rho$ is the mass density of the static BEC configuration and $p$ is the pressure, as we are considering zero temperature $p$ is not a thermal pressure but instead it is produced by the strong repulsive interaction between the ground state bosons. Assuming spherical symmetry and denoting $R$ as the radius at which the pressure and density are zero, the density profile takes the form [@boharko] $$\rho_{B} (r) = \rho^{B}_{0} \dfrac{\sin( k r)}{k r} \label{eq:BECrho}$$ where $k = \sqrt{G m^{3} / \hbar^{2} a}=\pi/R$ and $ \rho^{B}_{0}=\rho_{B}(0)$ is the BEC central density, $m$ is the mass of the DM particle and $a$ is the scattering lenght. The mass at the radius $r$ is given by $$m(r)= \frac{4 \pi \rho^{B}_{0}}{k^{2}} r \biggl( \dfrac{\sin (kr)}{kr } - \cos(kr) \biggr ), \label{eq:BECm}$$ from here the tangential velocity $V$ of a test particle at a distance $r$, is $$V^{2}(r) = \dfrac{4 \pi G \rho^{B}_{0} }{k^{2}} \biggl( \dfrac{\sin (kr)}{kr } - \cos(kr) \biggr ). \label{eq:BECV}$$ The logaritmic slope of a density profile is defined as $$\alpha = \dfrac{d (\log \ \rho)}{d (\log \ r)}
\label{eq:alpha}$$ using (\[eq:BECrho\]) in (\[eq:alpha\]) we obtain $$\alpha(r)= - \biggl [1- \dfrac{\pi r}{R} \cot \biggl(\dfrac{\pi r}{R} \biggr) \biggr].$$
We study now the Pseudo Isothermal (PI) profile. The empirical core profiles that exist in the literature fit two parameters, a scale radius and a scale density. A characteristic profile of this type is $$\rho_{PI} = \dfrac{\rho^{PI}_{0}}{1 \ + \ (r/R_{c})^{2}},\label{eq:PI}$$ this is the PI profile [@begeman]. Here $R_{c}$ is the scale radius and $\rho^{PI}_{0}$ is the central density. The rotation curve is $$V(r)_{PI} = \sqrt{4 \pi G \rho^{PI}_{0} R^{2}_{c} \biggl ( 1-\frac{R_{c}}{r} \arctan \biggl ( \dfrac{r}{R_{c}} \biggr) \biggr ) }.$$
The NFW profile emerges from numerical simulations that use only CDM and are based on the $\Lambda$CDM model [@dubinski; @navarro96; @navarro97]. In addition to this, we have chosen this profile because it is representative of what is called the cuspy behavior ($\alpha \approx -1$) in the center of galaxies due to DM. The NFW density profile[@navarro97] and the rotation curve are given respectively by $$\rho_{NFW}(r) = \dfrac{\rho_{i}}{(r/Rs)(1 \ + \ r/R_{s})^{2}}$$ $$V_{NFW}(r)= \sqrt{4 \pi G \rho_{i} R^{3}_{s}} \sqrt{ \frac{1}{r} \biggl [ ln \biggl (1 \ + \ \dfrac{r}{R_{s}} \biggr ) -
\dfrac{ r/R_{s}}{1\ + \ r/R_{s} } \biggr ]},$$ $\rho_{i}$ is related with the density of the universe at the moment the halo collapsed and $R_{s}^{2}$ is a characteristic radius.
To solve the core radius ambiguity and to unify the concept for future comparison, in [@victor] it was found that a good definition for the $core$ is a region where the density profile presents logarithmic slopes $\alpha \geq -1$ and the core radius will be the radius at which the core begins, that is to say, for radius smaller than the core radius we will have $\alpha \geq -1$, this means that its value $r'$ is determined by the equation $$\alpha(R) = -1.$$ The advantages of this definition are that the interpretation is independent of the profile chosen (also notice that it applies to the total density profile and is not restricted to that of DM) and in virtue of the same definition we can directly tell if a DM model profile is cored or cuspy. With this new definition the specific distance at which the core radius occurs still depends on the profile chosen but now the physical interpretation is only one. In the following when we refer to both the core and core radius we adopt the previous interpretation.
Applying the definition to (\[eq:BECrho\]) we get the core radius for the SFDM/BEC profile $R_{B}$ and for comparison we use (\[eq:PI\]) because in turns out that the parameter $R_{c}$ corresponds to the core radius as defined above. Finally, fitting the NFW profile provides a direct comparison between a cusp and core and hence to the cusp/core problem.
SFDM vs. PI and NFW
-------------------
We see from (\[eq:BECrho\]) that the SFDM/BEC model satisfies $\rho \sim r^{0}$ near the origin, but a priori this does not imply consistency with observed RCs. Therefore we reproduce the fit of the profiles to thirteen high resolution observed RCs of a sample of LSB galaxies given in [@victor]. The RCs were taken from a subsample of @deblok2, we chose galaxies that have at least 3 values within $\sim 1$ kpc, not presenting bulbs and the quality in the RC in H$\alpha$ is good as defined in @mcgaugh. The RCs in this work omit galaxies presenting high asymmetries and included in the error bars are experimental errors in the velocity measurement, inclination and small asymmetries. Because the DM is the dominant mass component for these galaxies we adopt the minimum disk hypothesis (neglects baryon contribution to the observed RC).
As the difference between a core and a cusp is not overlapped only for data values inside $1$kpc and given that in the interval $\sim 1$ to $10$kpc the slopes of core and cusp profiles are very similar, which can lead to the wrong conclusion that cuspy halos are consistent with observations, we determined the logarithmic slope and the uncertainty following @deblok2 with the difference that we fit only the data within $1$ kpc and that there is no need of an uncertain “break radius”.
In Fig. \[fig:rotcur\], we show the fits to the RC data and the density profiles, also shown are the core radius in the BEC (magenta) and in PI (blue) profiles. The gray arrow is the fit that determines $\alpha$, the length denotes the fitted region and is bounded by $R_{1}$ that denotes the nearest radius to $1$kpc where a data point is given. We show $\alpha$ by fitting values inside $R_{1}$ and we also show the core radius for the SFDM/BEC profile $R_{B}$ in order to compare it with $R_{c}$ (see [@victor] for more details of the fitting parameters).
-- -- --
-- -- --
-- -- --
-- -- --
Discussion
----------
The fits of the RCs in Fig \[fig:rotcur\] prove that the solid-body like behavior characterized by a linear increase in the velocity in the central region is more consistent with the core PI and SFDM/BEC profiles than the cuspy NFW. The fits within $R_{1}$ give an average value of $\alpha$=$-$0.27 $\pm$ 0.18 consistent with those obtained in @deblok2 $\alpha$=$-$0.2 $\pm$ 0.2 and with $\alpha$=$-$0.29 $\pm$ 0.07 reported by @ohse analyzing 7 THINGS dwarf galaxies. The case of ESO1870510 might be considered to be consistent with NFW profile, however it is the innermost value that considerably decreases $\alpha$, being an irregular galaxy more central data near the innermost region is required to discard the possibility of any violent event that might have caused such a slope value.
The density profiles corresponding to the RCs fits are also shown in Fig \[fig:rotcur\] for each galaxy. We see that the SFDM/BEC fit slightly deviates from the farthest data points as a result of having a finite radius $R$ that is fixed by the same data. This discrepancy is due to the fact that the halo might be more extended than the value $R$. As a matter of a fact, the more the extended the “flat” outer region in the RCs the more conspicuous the discrepancy. @harkomada recently found finite temperature corrections to (\[eq:PI\]), this suffice to alleviate the latter problem in LSB galaxies and dwarfs but not for bigger galaxies. Some solutions have been proposed including vortex lattices [@rindler; @zinner] and adding more nodes [@sin; @ji_sin] in the solution of system (\[eq:BEC1\]) and (\[eq:BEC2\])), but so far no final conclusion has been reached.
When comparing the SFDM/BEC and PI core radius we find a general difference of $\sim2$kpc, the core size in the PI is approximately $2$ kpc smaller than the BEC core size, but the PI central density is larger. In U4115, U11557 and U11583 both profiles are very similar which results in a similar core and central density values, this can also be taken as a consistency check for our core definition.
Comparing the values of $R_{B}$ we did not find a tendency to a common value. Assuming that the core radius determines the transition where the DM distribution changes from the outer region to the inner constant central density, the lack of the unique value means that there is not a common radius at which this transition takes place.
For the second test [@victor] uses $R_{B}$ to calculate (\[eq:mu\]). We have already seen that $R_{c}$ and $R_{B}$ are generally different and $R_{B}$ is not a fit parameter. Hence *a priori* $R_{B}$ is not expected to correlate with $\rho^{B}_{0}$. With the fitting parameters it was obtained $$\log (\mu^{B}_{0}/ M_{\odot} pc^{-2}) = \log \rho^{B}_{0} R_{B} = 2.05 \pm 0.56, \label{eq:muB}$$ $$\log (\mu^{PI}_{0}/ M_{\odot} pc^{-2}) = \log \rho^{PI}_{0} R_{c} = 2.08 \pm 0.46, \label{eq:muPI}$$ for the average values in the SFDM/BEC and PI profiles respectively. We see the excellent agreement of (\[eq:muB\]) with (\[eq:muPI\]) that was used as a crosscheck and with (\[eq:logmu\]) in which a much bigger sample was used. The agreement has shown that the SFDM/BEC model is capable of reproducing the constancy of the value $\mu_{0}$, something that because of the cuspy nature is not possible in the NFW profile.
In Fig. \[fig:logs\] we plot the above values for each galaxy. If we define the DM central surface density mentioned in the introduction, we have for the BEC profile $$<\Sigma>^{B}_{0,DM} = M_{< R_{B}} / \pi r^{2}_{B},$$ where $M_{< R_{B}}$ is obtained from (\[eq:BECm\]) evaluated at $R_{B}$. We show that for U11748 the value $ log \mu^{B}_{0}$ is considerably above the rest and with the largest uncertainty. For this reason, in the following analysis we omit both, this value and the smallest one that corresponds to ESO1200211. Doing this we get $<\Sigma>^{B}_{0,DM} \approx 191.35$ M$_{\odot}/$pc$^{2}$, and for the acceleration felt by a test particle located in $R_{B}$ due to DM only we have $g_{DM}(R_{B}) \approx $ 5.2 $\times 10^{-9}$ cms$^{-2}$ broadly consistent with (\[eq:gDM\]).
The fact that all galaxies present approximately the same order of magnitude in $g_{DM}(R_{B})$ might suggest that $R_{B}$ represents more than a transition towards a constant density, it can give us information about the close relation between DM and the baryons. Moreover, in view of the lack of a unique core radius, we can interpret the transition in DM distribution as an effect of crossing a certain acceleration scale instead of a radial length scale. Such interpretation reminds us that given in MOND but with the big difference that the acceleration scale found is for DM and is not a postulate of the model.
To determine which interpretation causes the transition whether an acceleration scale or a length scale we will need to study the properties of larger samples of galaxies with the new telescopes.
Conclusions {#sec:conclusions}
===========
In this work, we revisit an alternative paradigm to dark matter nature known as scalar field dark matter or Bose-Einstein condensate dark matter model. In this model a fundamental scalar field plays the role of dark matter in the Universe. The hypothesis is that this scalar field undergoes a phase transition very early in the Universe leading to the formation of Bose-Einstein condensate drops. Therefore, the dark haloes of the galaxies in the Universe are huge drops of scalar field. To explore if this hypothesis is a viable alternative to the standard model we developed both cosmological and galactic studies.
In the cosmological regime, we have shown, with a field approach, that the SFDM/BEC model with an ultralight mass of $10^{-22}$ eV mimics the behavior of the cosmological expansion rate predicted with the $\Lambda$CDM model. Moreover, although in general a scalar field is not a fluid, it can be treated as if it behaved like one. Thus, we performed a fluid approach for the cosmological evolution of the scalar field and we found that the analytical expressions for the kinetic and potential energies of the scalar field is in excellent agreement with our previous numerical results.
The interesting cosmological behavior of the scalar field indicates that their scalar fluctuations can be the appropriate for the purpose of structure formation, because overdense regions of SFDM/BEC can support the formation of galactic structure. Thus, we have revisited the growth of SFDM/BEC fluctuations in the linear regime. Within the linear theory of scalar perturbations we obtain an equation for the evolution of the density contrast. This equation differs from the density contrast equation for CDM, however the extra terms tend to the values of the standard equation. Therefore, as a goal of this model, we obtained that the scalar perturbations grow up exactly as the CDM paradigm. In addition, we study the liner growth of the scalar fluctuations with a fluid approach. In the case of the SFDM/BEC with $\lambda=0$ we have shown that for the matter dominated era and for big structure this model simulates the behavior of CDM because in general in a matter dominated Universe for low-$k$, $v_{q}$ tends to be a very small quantity tending to zero, so from (\[delta\]) we can see that on this era we will have the CDM profile given by (\[deltaCDM\]), i.e., the SFDM density contrast profile is very similar to that of the $\Lambda$CDM model (see Fig. \[fig2\]). On the contrary for $\lambda\neq 0$ both models have different behavior as we can see from Fig. \[fig3\]. The numerical results suggest that linear fluctuations on the SFDM can grow faster than those for CDM around $a\sim 10^{-2}$. Here an important point is that although CDM can grow it does so in a hierarchical way, while from Fig. \[fig3\] we can see that SFDM can have bigger fluctuations just before the $\Lambda$CDM model does, i.e., it might be that no hierarchical model of structure formation is needed for SFDM and is expected that for the non-linear fluctuations the behaviour will be quite the same as soon as the scalar field condensates, in a very early epoch when the energy of the Universe was about $\sim$ TeV. Thus, the standard and the SFDM/BEC model can be contrasted in their predictions concerning the formation of the first galaxies. If in the future we see more and more well formed and massive galaxies at high redshifts, this could be also a new indication in favour of the SFDM/BEC paradigm.
On the other hand, we also studied the implications of a SFDM/BEC model at galactic scales. We find that the SFDM/BEC model gives a constant density profile that is consistent with RCs of dark matter dominated galaxies. The profile is as good as one of the most frequently used empirical core profiles but with the advantage of coming from a solid theoretical frame. We fit data within $1$kpc and found a logarithmic slope $\alpha = -$ 0.27$\pm$ 0.18 in perfect agreement with a core. It is important to notice that the cusp in the central regions is not a prediction that comes from first principles in the CDM model, it is a property that is derived by fitting simulations that use only DM. In addition, we review a new definition for the core and core radius that allows a definite distinction when a density profile is core or cusp. Using this definition we find the core radius in the SFDM/BEC profile to be in most cases over $2$kpc bigger than the core radius in the PI profile. As a second result of the core definition, we show the constant value of $\mu_{0}$ which is proportional to the central surface density. This result is one of several conflicts that jeopardize the current standard cosmological model.
Finally, with all these intriguing results SFDM/BEC model could be a serious alternative to the dark matter problem in the Universe. Although the observational evidence seems to be in favor of some kind of cold dark matter, if we continue to observe even more galaxies at higher redshifts and if higher resolution observations of nearby galaxies exhibit a core density profile, this model can be a good alternative to $\Lambda$CDM. We expect that the close future observations in galaxies surveys can decide the nature of the dark matter.
Acknowledgments {#acknowledgments .unnumbered}
===============
This work was partially supported by CONACyT México under grants CB-2009-01, no. 132400, no. 166212, and I0101/131/07 C-234/07 of the Instituto Avanzado de Cosmologia (IAC) collaboration (http://www.iac.edu.mx/).
[99]{}
Alcubierre M., Guzmán F. S., Matos T., Núñez D., Ureña L. A., 2002 Wiederhold P., Class Quant Grav, 19, 5017
Arbey, A., Lesgourgues, J. and Salati, P., 2001, Phys. Rev. D [**64**]{} 123528 Arbey, A., Lesgourgues, J. and Salati, P., 2002, Phys. Rev. D [**65**]{} 083514 Arbey, A., Lesgourgues, J. and Salati, P., 2003, Phys. Rev. D [**68**]{} 023511 Arbey, A., 2006 EAS Publ. Ser. [**20**]{} 257 Arbey, A., 2006b Phys. Rev. D [**74**]{} 043516 Arbey, A., 2009 EAS Publ. Ser. [**36**]{} 161 Athanassoula E., Bosma A., Papaioannou S., 1987, A&A, 179, 2340
Bardeen J.M., 1980, Phys Rev D, 22, 1882
Begeman K. G., Broeils A. H., Sanders R. H., 1991, MNRAS, 249, 523
Bernal, A., & Guzmán, F. S.,2006, Phys. Rev. D 74, 063504 Bernal, A., & Guzmán, F. S., 2006, Phys. Rev. D 74, 103002 Bernal A., Matos T., Núñez D., 2008, Rev. Mex. A.A., 44, 149
Böehmer C. G., Harko T., 2007, JCAP, 0706, 025
Briscese F., 2011, Phys Lett B, 696, 315
Burkert A., 1995, ApJ, 447, L25
Ma C.P., Bertschinger E., 1995, ApJ, 455, 7.
Clowe D. et al., 2006, ApJ, 648, 2
Colín, P.; Avila-Reese, V.; Valenzuela, O., 2000, ApJ, 542, 622-630
Colpi M., Shapiro S. L. and Wasserman I., 1986, Phys. Rev. Lett., 57, 2485
Copeland E. S., Liddle A. R., Wands D., 1998, Phys. Rev. D, 57, 4686.
Chavanis P. H., 2011, Phys. Rev. D, 84, 043531
Chavanis P. H., 2011, Phys. Rev. D 84, 043532
Chavanis P. H., 2012, A & A, 537, A127 Dalfovo F., Giorgino S., Pitaevskii L. P., Stringari S., 1999, Rev. Mod. Phys. 71, 463
de Blok W.J.G., 2010, Adv. Astron., 2010, Article ID 789293, 14 pages.
de Blok W. J. G. et al., 2001, ApJ, 552, L23
Dehnen H., Rose B., 1993, Astrophys. Sp Sci., 207, 133-144
Donato et al., 2009, MNRAS, 397, 1169
Dubinski J., Carlberg R. G., 1991, ApJ, 378, 496
Friedmann D. E., 2011, arXiv:0912.1668v4
García-Aspeitia, M.A., Magaña, J., Matos, T., 2011, GeRG Online, arXiv:1102.0825
Gentile G. et al., 2009, Naure, 461, 627
Ginzburg V. L., Landau L. D., 1950, Zh Eksp Teor Fiz, 20, 1064
Gleiser M., 1988, Phys. Rev. D, 38, 2376
Gorini V., Kamenshchik A. Y., Moschella U., Piattella O. F., Starobinsky A. A., 2008, JCAP, 0802, 016
Guzmán, F. Siddhartha and Matos, Tonatiuh. Astron. Nachr. 320, 97. Guzman F. S., Matos T., 2000, Class. Quantum Grav., 17, L9-L16. Guzmán F.S., Ureña-López L. A., 2003, Phys. Rev. [**D 68**]{}, 024023. Guzmán F. S., Ureña-López L.A., 2004, Phys. Rev. [**D 69**]{}, 124033. Guzmán F. S., Ureña-López L. A., 2006, ApJ, 645, 814.
Harko T., 2011, MNRAS, 413, 3095 Harko T., 2011, JCAP, 05, 022 Harko T., 2011, Phys. Rev. D, 83, 123515 Harko T., Madarassy E. J. M., 2012, JCAP, 01, 020. Hu W., Barkana R., Gruzinov A., 2000, Phys. Rev. Lett., 85, 1158
Ji S. U., Sin S. J., 1994, Phys. Rev. D, 50, 3650
Kain B., & Ling, H.Y., Phys. Rev. D, 82, 064042 Kain B., & Ling, H.Y., Phys. Rev. D, 85, 023527 Klypin A., Kravtsov A. V., Valenzuela O., Prada F., 1999, ApJ, 522, 82
E. Komatsu et al. 2009, ApJS **180**, 330-376. E. Komatsu et al. 2011, ApJS **192**, 18.
Kormendy J., Freeman K. C., 2004, in Ryder S. D., Pisano D. J., Walker M. A., Freeman K. C., eds, IAU Symp. . 220, Dark Matter in Galaxies. Cambridge Univ. Press, Cambridge, p. 377
Kuzio de Naray R., McGaugh S.S., & de Blok W.J.G., 2008, ApJ, 676, 920
Lee J., Koh I., 1996, Phys Rev D, 53, 2236
Lee J.W., Lim S., Choi D., 2008, arXiv:astro-ph/0805.3827.
Lee, Jae-Weon, 2009, J. Korean Phys.Soc., 54, 2622 Lee, J.W., & Lim, S., 2010, JCAP, 1001, 007. Lee, Komatzu, 2010, ApJ, 718, 60.
Lesgourgues, J., Arbey, A. and Salati, P., 2002, New Astron. Rev. [**46**]{} 791. Loeb, A.; Weiner, N.; 2011, Phys. Rev. Let., 106, 171302
Lora V., Magaña J., Bernal, A., Sánchez-Salcedo F.J., Grebel, E.K., 2011, arXiv:1110.2684, to be published in JCAP
Lundgren A. P., Bondarescu M., Bondarescu R., Balakrishna J., 2010, ApJ, 715, L35
Ma C.-P., E. Bertschinger, 1995, ApJ, 455, 7
Malik K. A., 2009, Phys Rept, 475, 1
Marsh D. & Ferreira P., 2010, Phys. Rev. D, 82, 103528 Matos, T., Guzmán, F. S., 2000, Class. Quant. Grav. 17, L9. arXiv:gr-qc /9810028.
Matos, T., Ureña L. A., 2000, Class Quant Grav, 17, L75
Matos, T., Guzmán F.S., 2001, Class. Quantum Grav., 18, 5055
Matos, T., Ureña L. A., 2001, Phys Rev D, 63, 063506
Matos T., Vázquez-González A.,Magaña J., 2009, MNRAS, 389, 13957
Matos T., Magaña J., Suárez A., 2010, The Open Astron Journal, 3, 94
McGaugh S. S. et al., 2001, ApJ, 122, 2381
Milgrom M., 2010, AIP Conf.Proc. 1241:139 arXiv:0912.2678v2
Moore B., Ghigna S., Governato F., Lake G., Quinn T., Stadel J., Tozzi P., 1999, ApJ, 524, L19
Navarro J. F., Frenk C. S., White S. D. M., 1996, ApJ., 462, 563
Navarro J. F., Frenk C. S., White S. D. M., 1997, ApJ, 490, 493
Navarro J.F. et al., 2010, MNRAS, 402, 21
Oh Se-Heon, Brook Chris et al., 2011, arXiv:1011.2777v2.
Peebles P. J. E., Nusser A., 2010, Nat, 465, 565
Penny S. J., Conselice C. J., De Rijcke S., Held E. V., 2009, MNRAS, 393, 1054
Pitaevskii L., Stringari S., 2003, Bose-Einstein Condensation, Oxford University Press
Ratra B., Peebles P. J. E., 1988, Phys. Rev. D, 37, 3406.
Rindler-Daller T., Shapiro P. R., 2011, arXiv:1106.1256v2
Robles V.H., Matos T., 2012, MNRAS, 392, (2012) in press. arXiv:1201.3032
Rodríguez-Montoya, I., Magaña, J., Matos, T., & Pérez-Lorenzana, A., 2010, ApJ, 721, 1509
Rodríguez-Montoya, I.; Pérez-Lorenzana A.; De La Cruz-Burelo, E.; Giraud-Héraud, Y.; Matos, T.; 2011, arXiv:1110.2751
Sahni V., Wang L., 2000, Phys Rev D, 62, 103527
Sánchez-Salcedo, F. J., Reyes-Iturbide, J. & Hernandez, X., 2006, MNRAS, 370, 1829.
Sanders R. H., 2009, Adv. Astron., 2009, Article ID 752439, 9 pages.
Sin S. J., 1994, Phys. Rev. D, 50, 3650
Spano M. et al., 2008, MNRAS, 383, 297
Spergel D., Steinhardt P., 2000, Phys. Rev. Let., 84, 3760-3763
Suárez A., Matos T., 2011, MNRAS, 416, 87
Thomas, S.A., Abdalla, F., Lahav, O., 2011, Phys. Rev. Let., 106, 241301
Turner M.S., 1983, Phys. Rev. [**D 28**]{}, 1243.
Ureña-López L.A., Reyes-Ibarra M. J., 2009, IJMPD, 18, 621.
Ureña-López L.A., 2009, JCAP [**01**]{}, 014.
Ureña L. A., 2010, AIP Conf Proc, 1318, 82
Velter, H. & Wamba, E.; 2011, arXiv:1111.2032.
Villaescusa-Navarro, F.; Dalal, N.; 2011, JCAP, 03, 024
Wetterich C., 1988, Nucl. Phys. B, 302, 668.
Woo T. P., Chiueh T., 2009, ApJ, 697, 850.
Zinner N. T., 2011, arXiv:1108.4290v1
[^1]: Part of the Instituto Avanzado de Cosmología (IAC) collaboration http://www.iac.edu.mx/
[^2]: The scale factor is such that $a=1$ today, so that it relates to the redshift $z$ by $a=(1+z)^{-1}$.
| {
"pile_set_name": "ArXiv"
} |
---
author:
- 'Tomáš Ježo,'
- 'Jonas M. Lindert,'
- 'Niccolo Moretti,'
- and Stefano Pozzorini
bibliography:
- 'paper.bib'
title: 'New NLOPS predictions for $\boldsymbol{\ttbar+b}$-jet production at the LHC'
---
Introduction
============
At the Large Hadron Collider, searches for $\ttbar H$ production in the $H\to \bbbar$ channel are plagued by a large QCD background, which is dominated by $\ttbb$ production, and the availability of precise theoretical predictions for this multi-particle background process is of crucial importance for the sensitivity of $\ttbar H(\bbbar)$ analyses. The process $pp\to \ttbb$ is also very interesting on its own, as it provides a unique laboratory to explore the QCD dynamics of heavy-quark production and to test state-of-the-art Monte Carlo predictions in a nontrivial multi-scale environment.
As a result of its $\as^4$ dependence, the leading-order (LO) $\ttbb$ cross section is highly sensitive to variations of the renormalisation scale. The uncertainty corresponding to standard factor-two scale variations amounts to 70-80% at LO, and the inclusion of next-to-leading order (NLO) QCD corrections [@Bredenstein:2009aj; @Bevilacqua:2009zn; @Bredenstein:2010rs] is mandatory. At NLO, the scale dependence goes down to 20–30%, and in order to avoid excessively large $K$-factors and potentially large corrections beyond NLO, the renormalisation scale should be chosen in a way that accounts for the fact that the typical energies of the $b$-jet system are far below the hardness of the underlying $pp\to\ttbar$ process [@Bredenstein:2010rs].
The first NLOPS simulation of $pp\to \ttbb$ was carried out in [@TROCSANYI:2014lha; @Garzelli:2014aba] by combining NLO matrix elements in the five-flavour (5F) scheme with parton showers by means of the method [@Nason:2004rx; @Frixione:2007vw]. Shortly after, an NLOPS generator based on four-flavour (4F) $pp\to \ttbb$ matrix elements became available in the framework [@Cascioli:2013era], which implements an improved version [@Hoeche:2011fd] of the MC@NLO matching method [@Frixione:2002ik]. Thanks to the inclusion of $b$-mass effects, $\ttbb$ matrix elements in the 4F scheme are applicable to the full $b$-quark phase space, including regions where one $b$-quark remains unresolved. Thus the 4F scheme guarantees a consistent NLOPS description of inclusive $\ttbar+b$-jet production with one or more $b$-jets. On the contrary, NLOPS $\ttbb$ generators based on 5F matrix elements with massless $b$-quarks suffer from collinear $g\to \bbbar$ singularities that require ad-hoc restrictions of the physical phase space through generation cuts.
In it was pointed out that matching and shower effects play an unexpectedly important role in $\ttbar+b$-jet production. This is due to the fact that two hard $b$-jets can arise from two hard jets involving each a collinear $g\to \bbbar$ splitting. In NLOPS simulations of $pp\to \ttbb$, such configurations result from the combination of a $g\to \bbbar$ splitting that is described at NLO accuracy through $\ttbb$ matrix elements together with a second $g\to \bbbar$ splitting generated by the parton shower. The impact of this so-called double-splitting mechanism can have similar magnitude to the $\ttbar H(\bbbar)$ signal, and the thorough understanding of the related matching and shower uncertainties is very important for $\ttbar H$ analyses.
A first assessment of NLOPS uncertainties was presented in through a tuned comparison of NLOPS $\ttbb$ simulations in [@TROCSANYI:2014lha; @Garzelli:2014aba], [@Cascioli:2013era] and [@Alwall:2014hca]. On the one hand, this study has revealed significant differences between the two generators based on the MC@NLO matching method[^1], i.e. and . Such differences were found to be related to a pronounced dependence on the shower starting scale in . On the other hand, in spite of the fact that and implement different matching methods and different parton showers, the predictions of these two generators turned out to be quite consistent. However, due to the limitations related to the use of the 5F scheme in —which have been overcome only very recently with the 4F upgrade of [@Bevilacqua:2017cru]—the agreement between and did not allow to draw any firm conclusion in the study of .
To date the assessment of theoretical uncertainties in $\ttbb$ production remains an important open problem. In this context one should address the question of which of the various NLOPS methods and tools on the market are more or less appropriate to describe the process at hand. Moreover, in order to address such issues in a systematic way, it is desirable to develop a better picture of the QCD dynamics that drive $\ttbar+b$-jet production. In this spirit, this paper starts with a discussion of the various possible frameworks for theoretical simulations of $\ttbar+b$-jet production at NLOPS accuracy. In particular, we present detailed studies on the role of $g\to \bbbar$ splittings and discuss the advantages and disadvantages of the 4F and 5F schemes. To this end we quantify the relative importance of $g\to \bbbar$ splittings of initial-state (IS) and final-state (FS) type by using approximations based on collinear QCD factorisation, as well as by decomposing $pp\to \ttbb$ matrix elements into diagrams involving IS and FS $g\to \bbbar$ splittings. These studies demonstrate that $\ttbar+b$-jet production is widely dominated by $pp \to \ttbar g$ followed by FS $g\to \bbbar$ splittings. This holds also for observables where initial-state splittings are expected to be enhanced, such as in regions with a single resolved $b$-jet. These findings support the use of NLOPS generators based on $pp\to \ttbb$ matrix elements in the 4F scheme, where $b$-mass effects guarantee a consistent treatment of FS $g\to \bbbar$ splittings. We also consider more inclusive simulations of $\ttbar+$jets production based on multi-jet merging [@Catani:2001cc; @Mangano:2001xp; @Hoeche:2009rj; @Hoeche:2012yf; @Lonnblad:2012ix; @Frederix:2012ps]. In this case we find that $\ttbar+b$-jet observables suffer from an unexpectedly strong dependence on the parton-shower modeling of $g\to \bbbar$ splittings.
Motivated by the above findings we present a new generator for $pp\to \ttbb$ in the 4F scheme.[^2] At variance with the generator of , this new generator is implemented in the framework [@Jezo:2015aia] using , which guarantees a very fast evaluation of the required $2\to 4$ and $2\to 5$ matrix elements. The new generator supports also top-quark decays including spin-correlation effects. Moreover, in order to guarantee a more consistent resummation of QCD radiation, the separation of the so-called singular and finite parts in the is not restricted to initial-state radiation as in but is applied also to final-state radiation (see ).
For what concerns the methodology we pay particular attention to issues related to the multi-scale nature of the process at hand. In particular we point out that the treatment of the recoil associated with NLO radiation can induce sizeable distortions of the underlying $\ttbb$ cross section. This technical inconvenience restricts the domain of applicability of QCD factorisation in a way that can jeopardise the efficiency of event generation and can also lead to unphysical resummation effects. Fortunately, such issues can be avoided by means of a mechanism that restricts the resummation of real radiation to kinematic regions where QCD factorisation is fulfilled within reasonably good accuracy.
Predictions for $pp\to \ttbar+b$-jets at the 13TeV LHC are presented for various cross sections and distributions with emphasis on the discussion of theoretical uncertainties. Besides QCD scale variations, also uncertainties related to the matching method and intrinsic shower uncertainties are analysed in detail. In particular, we consider different approximations for the modelling of $g\to \bbbar$ splitting as well as $\as$ scale uncertainties in . Moreover we compare to . Finally, to gain further insights into the size and the nature of matching and shower uncertainties we present a consistent comparison of generators of $\ttbb$ and inclusive $\ttbar$ production, against corresponding generators based on .
The new $\ttbb$ generator will be soon publicly available on the web site [@PWGBpage].
The paper is organised as follows. In we study the role of $g\to \bbbar$ splittings in $\ttbar+b$-jet production, and we point out the advantages of Monte Carlo generators based on $pp\to \ttbb$ matrix elements in the 4F scheme as compared to more inclusive generators of $\ttbar$ production in the 5F scheme. Technical aspects of the new generator and the setup for numerical simulations are discussed in . In particular, in we review aspects of the method that can play a critical role for multi-scale process like $pp\to \ttbb$. Detailed predictions and uncertainty estimates for cross sections and distributions for $pp\to \ttbb$ with stable and unstable top quarks can be found in Sections \[se:stabletops\] and \[se:decayedtops\], respectively. Our main findings are summarised in .
Anatomy of $\boldsymbol{\ttbar+b}$-jet production and $\boldsymbol{g\to \bbbar}$ splittings {#se:gbbstudies}
===========================================================================================
Events with $\ttbar+b$-jets final states arise from an underlying $pp\to \ttbar$ process that takes place at scales of the order of 500GeV and is accompanied by the production of $b$-jets at typical transverse momenta of a few tens of GeV. The production of $b$-jets is governed by IS or FS $g\to \bbbar$ splittings and is enhanced in kinematic regions where the $p_\rT$ of individual $b$-quarks becomes small or the $\bbbar$ pair becomes collinear and, possibly, also soft.
The understanding of the QCD dynamics that governs $b$-jet production is a crucial prerequisite for a reliable theoretical description of $\ttbar+b$-jet production and related uncertainties. In this spirit, this section compares various theoretical frameworks for the description of $\ttbar+b$-jet production at NLO QCD accuracy with a special focus on the role of $g\to \bbbar$ splittings. Specifically, we compare inclusive or merged simulations of $\ttbar+$multi-jet production in the 5F scheme against a description based on $\ttbb$ matrix elements in the 4F scheme, pointing out the advantages of the latter.
Numerical studies presented in this section are based on the setup specified in Sections \[se:input\] and \[se:cuts\] and have been performed with .
NLOPS $\boldsymbol{\ttbar}$ simulations in the five-flavour scheme {#se:ttnlops}
------------------------------------------------------------------
![Sample $\ttbb$ diagrams with IS (left) and FS (right) $g\to \bbbar$ splittings. In NLOPS simulations of inclusive $\ttbar$ production in the 5F scheme the black subtopologies are described in terms of tree matrix elements, while the orange lines correspond to parton shower emissions.[]{data-label="fig:ttinctopologies"}](graphs/ggttbbMEPSd "fig:"){width="21.00000%"} ![Sample $\ttbb$ diagrams with IS (left) and FS (right) $g\to \bbbar$ splittings. In NLOPS simulations of inclusive $\ttbar$ production in the 5F scheme the black subtopologies are described in terms of tree matrix elements, while the orange lines correspond to parton shower emissions.[]{data-label="fig:ttinctopologies"}](graphs/ggttbbMEPSb "fig:"){width="21.00000%"}
Inclusive NLOPS generators of $\ttbar$ production [@Frixione:2003ei; @Frixione:2007nw] are based on $pp\to
\ttbar$ NLO matrix elements matched to partons showers in the 5F scheme. In this framework, as illustrated in , $\ttbar+b$-jet events are generated starting from $2\to 3$ tree matrix elements of type $gb\to \ttbar b$ or $gg/q\bar q\to \ttbar g$. In the latter case, $\ttbb$ events arise via FS $g\to \bbbar$ shower splittings. Instead, in the case of $gb\to \ttbar b$ the final-state $b$-quark emerges from the matrix element, while $g\to \bbbar$ splittings generate the initial-state $b$-quark through the evolution of the 5F PDFs. The unresolved spectator $b$-quark associated with such IS $g\to \bbbar$ splittings is emitted by the parton shower via backward evolution. The main advantage of the 5F scheme lies in the resummation of potentially large $\as\ln(m_t/m_b)$ terms associated with the evolution of the $b$-quark density. However, such logarithmic effects are typically rather mild at the LHC [@Maltoni:2012pa]. Moreover, as we will show in , $\ttbar+b$-jet production is largely dominated by topologies with FS $g\to \bbbar$ splittings. For this reason, $\ttbar+b$-jet predictions based on NLOPS $\ttbar$ generators suffer from the twofold disadvantage given by the direct dependence on the parton-shower modelling of FS $g\to \bbbar$ splittings plus the LO nature of the underlying $\ttbar g$ matrix element.
$\boldsymbol{\ttbar+}$multi-jet merging in the five-flavour scheme {#se:mepslo}
------------------------------------------------------------------
As a possible strategy to reduce the sensitivity to the parton shower and increase the accuracy of theoretical predictions we consider $\ttbar+$multi-jet merging in the 5F scheme. In this approach, a tower of NLOPS simulations for $\ttbar+0,1,\dots,N$jet production is merged into a single inclusive sample [@Hoeche:2012yf; @Lonnblad:2012ix; @Frederix:2012ps]. This is achieved by clustering QCD partons into jets with a certain $k_\rT$-resolution, $\qcut$, which is known as merging scale. At LO, the phase-space regions with $\Nres=0,1,\dots, \Nmax$ resolved jets ($k_T> \qcut$) are described in terms of -jet LOPS simulations. The LOPS simulation with $\Nmax$ jets fills also the phase space with $\Nres>\Nmax$ resolved jets by means of the parton shower. At NLO, the resolution criterion used to separate regions of different jet multiplicity exactly the same as for LO, while the basic difference with respect to LO merging lies in the fact that jet LOPS simulations are replaced by corresponding NLOPS simulations. Thus in NLO (LO) merging the effective number of resolved jets that is described at NLOPS (LOPS) accuracy is $\Neff=\mathrm{min}\{\Nres,N_\mathrm{\max}\}$, while the $(\Neff+1)^{\mathrm{th}}$ resolved or unresolved jet is described at LOPS (pure PS) accuracy, and all remaining resolved or unresolved jets are described at pure PS accuracy.
Multi-jet merging for $\ttbar+$jet at NLO can be performed in a fully automated way within the [@Gleisberg:2008ta] and [@Alwall:2014hca] frameworks. However, the fact that $\ttbar+b$-jet events constitute only a small fraction of a $\ttbar+$jets sample poses very high requirements in terms of Monte Carlo statistics. Moreover, in order to minimise the dependence on parton-shower modelling, such simulations should be performed using a small merging scale and including a sufficiently high number of NLO jets, $\Nmax$. The required CPU resources grow very fast at large $\Nmax$ and small $\qcut$, and state-of-the-art merged simulations can handle up to two jets at NLO [@Hoeche:2014qda] at present.
The multi-jet merging description of $\ttbb$ events with FS $g\to
\bbbar$ splittings is sketched in for the case of $\ttbar+0,1,2$-jet merging at LO. In regions where the $\bbbar$ pair and/or the parent gluon are emitted at small scales, $g\to \bbbar$ splittings are expected to be generated by the parton shower, while hard $b$-jet pairs are expected to arise from $\ttbb$ matrix elements. However, as we will see, typical $\ttbar+b$-jet events involve additional light jets that are emitted at harder scales with respect to the $g\to
\bbbar$ branching. In that case, $2\to 4$ matrix elements account only for light jets, and $g\to \bbbar$ splittings are left to the parton shower. In general, the relative importance of matrix elements and parton shower depends on the resolution scale $\qcut$, and using a finite resolution is mandatory in the 5F scheme, since collinear $g\to \bbbar$ splittings are divergent, $\ttbb$ matrix elements cannot be used in the full phase space.
In we analyse the matrix-element content of a $\ttbar+0,1,2$-jet merged simulation at LO. These studies are based on the method [@Hoeche:2009rj] implemented in , but the main findings are expected to hold also for other merging methods. Besides the $b$-jet multiplicity distribution, in we plot differential observables in the presence of ttbb cuts, i.e. requiring $\Nb\ge 2$ $b$-jets as defined in . For jets we apply the acceptance cuts and, in order to maximise the possibility to resolve jets at matrix-element level, we choose a merging scale lower than the jet-$p_\rT$ threshold, $\qcut=20$GeV.
As expected, in we find that $\ttbar+b$-jet observables with $\Nb\ge 2$ resolved $b$-jets are largely dominated by $\ttbar+$2-parton matrix elements. This holds also for $\Nb\ge 1$. However, the breakdown of the merged sample into contributions from matrix elements with different $b$-quark multiplicity in reveals that, in spite of the low merging scale, the cross section for producing one or more $b$-jets is dominated by matrix elements with zero $b$-quarks. The contribution of $\ttbb$ matrix elements hardly exceeds 50% even in the region of large $b$-jet $\pt$ or large invariant mass of the $b$-jet pair. This counterintuitive feature can be attributed to the fact that, in $\ttbar+$jet events that involve $g\to \bbbar$ splittings, the two hardest QCD branchings are typically associated with the emission of the parent gluon of the $\bbbar$ pairs and/or with the production of other light jets. As a consequence, in LOPS merged samples with $\Nmax=2$, $g\to \bbbar$ splittings are left to the parton shower. We have verified that the contribution of $\ttbb$ matrix elements remains relatively low even when the merging procedure is extended up to 3 or 4 jets with $\qcut=20$GeV. Moreover, we have checked that increasing the merging scale leads to a further suppression of the contribution of $\ttbb$ matrix elements.
d demonstrates that $\ttbar+2\,b$-jet events are indeed accompanied by abundant emission of extra light jets, and the importance of $\ttbb$ matrix elements decreases with increasing light-jets multiplicity. Moreover, even in the bin with zero additional light jets it turns out that the contribution of $\ttbb$ matrix elements remains below 50%. This is probably due to the fact that $g\to \bbbar$ splittings tend to take place at branching scales below the jet-$p_\rT$-threshold of 25GeV.
In summary, contrary to naive expectations, $\ttbar+$jets samples based on LOPS merging do not guarantee a matrix-element description of $b$-jet production but largely rely on the parton shower modelling of $g\to
\bbbar$ splittings. In the case of multi-jet merging at NLO, to a certain extent $g\to \bbbar$ shower splittings should be matched to $\ttbb$ and $\ttbb g$ tree matrix elements. Nevertheless, based on the above observations, the theoretical accuracy in the description of $b$-jet production is expected to remain between the LOPS and the pure PS level.[^3] Finally, in the light of the presence of abundant light-jet radiation with a typical hardness beyond the one of $b$-jets, the role of hard radiation on top of the $\ttbb$ system should be studied with great care also in the context of NLOPS simulations of $\ttbb$ production (see ).
$\boldsymbol{\ttbb}$ production in the four-flavour scheme {#se:ttbbISvsFS}
----------------------------------------------------------
In order to minimise the dependence on parton-shower modelling and to maximise the use of higher-order matrix elements, in the following we will adopt a description of $\ttbar+b$-jet production based on $ttbb$ matrix elements in the 4F scheme. In this scheme, $b$-quarks are treated as massive partons, and $g\to
\bbbar$ splittings are free from collinear singularities. Thus $\ttbb$ matrix elements can be used in the entire phase space. Generic $\ttbb$ topologies where $b$-quarks emerge from IS and FS splitting processes are illustrated in . In the case of FS $g\to \bbbar$ splittings, $\ttbb$ matrix elements with $m_b>0$ can be extended to the collinear regime, where the $\bbbar$ pair becomes unresolved within a single $b$-jet. Similarly, 4F $\ttbb$ matrix elements describe also collinear IS $g\to
\bbbar$ splittings, where the spectator $b$-quark is emitted in the beam direction and remains unresolved, while the $bg\to \ttbar b$ sub-process with a single $b$-jet corresponds to the description of $\ttbar+b$-jet production at LO in the 5F scheme. Thus, $\ttbb$ matrix elements provide a fully inclusive description of $\ttbar+b$-jet production, and NLO predictions in the 4F scheme yield NLO accuracy both for observables with two $b$-jets and for more inclusive observables with a single resolved $b$-jet.
\
The inclusion of $m_b$ effects in $g\to \bbbar$ splittings represents a clear advantage of the 4F scheme with respect to the 5F scheme. However, the 4F scheme has the disadvantage that potentially large $\as \ln(m_b/Q)$ terms that arise from IS $g\to \bbbar$ splittings are not resummed through the PDF evolution. In the following, in order to assess the relevance of this limitation, we decompose the $pp\to \ttbb$ LO cross section into contributions from IS and FS $g\to \bbbar$ splittings. Since the $q\bar q$ channel involves only FS $g\to \bbbar$ splitting, we focus on the $gg$ channel and we first consider a naive diagrammatic splitting of the $gg\to\ttbb$ matrix element, $$\label{eq:ISFSMEsplitting}
\mathcal{M}_{\ttbb}=
\mathcal{M}_{\mathrm{IS},\ttbb}
+ \mathcal{M}_{\mathrm{FS},\ttbb}+
\mathcal{M}_{\mathrm{rem},\ttbb}.$$ The terms $\mathcal{M}_{\mathrm{IS},\ttbb}$ and $\mathcal{M}_{\mathrm{FS},\ttbb}$ correspond, respectively, to 18 diagrams with IS $g\to \bbbar$ splittings and 16 diagrams with FS $g\to \bbbar$ splittings. Generic diagrams with IS and FS splittings are depicted in a–b. The term $\mathcal{M}_{\mathrm{rem},\ttbb}$ corresponds to two remaining diagrams where an $s$-channel gluon splits into an on-shell $b$-quark and an off-shell $b$-line coupled to the $\ttbar$ system. Its numerical impact turns out to be negligible. Based on we split the $\ttbb$ cross section into three terms, $$\label{eq:ISFSXSsplitting}
\rd\sigma_{\ttbb}= \rd\sigma_{\IS,\ttbb}
+\rd\sigma_{\FS,\ttbb}
+\rd\sigma_{\mathrm{int},\ttbb},$$ where the IS and FS parts are defined as $$\begin{aligned}
\label{eq:ISFSXSdef}
\rd\sigma_{\IS,\ttbb}=\frac{|\mathcal{M}_{\IS,\ttbb}|^2}{|\mathcal{M}_{\ttbb}|^2}\,
\rd\sigma_{\ttbb},
\qquad
\rd\sigma_{\FS,\ttbb}=\frac{|\mathcal{M}_{\FS,\ttbb}|^2}{|\mathcal{M}_{\ttbb}|^2}\,
\rd\sigma_{\ttbb},\end{aligned}$$ while $\rd\sigma_{\mathrm{int},\ttbb}$ consists of the interference between $\mathcal{M}_{\mathrm{IS},\ttbb}$ and $\mathcal{M}_{\mathrm{FS},\ttbb}$ plus a minor contribution from $\mathcal{M}_{\mathrm{rem},\ttbb}$.
In order to check the soundness of the above gauge-dependent separation we compare it to an alternative definition of IS and FS $g\to \bbbar$ contributions based on the collinear limits of the $gg\to \ttbb$ matrix element. In this case we define $$\begin{aligned}
\label{eq:ISFScoll}
\rd\sigma_{\IS\otimes\ttbar}=
\frac{|\mathcal{M}_{\IS\otimes\ttbar}|^2}
{|\mathcal{M}_{\ttbb}|^2}\,\rd\sigma_{\ttbb},
\qquad
\rd\sigma_{\FS\otimes\ttbar}=
\frac{|\mathcal{M}_{\FS\otimes\ttbar}|^2}
{|\mathcal{M}_{\ttbb}|^2}\,\rd\sigma_{\ttbb}.\end{aligned}$$ Here $|\mathcal{M}_{\IS\otimes\ttbar}|^2$ and $|\mathcal{M}_{\FS\otimes\ttbar}|^2$ describe the collinear limits of the topologies depicted in c and \[fig:ttbbtopologies\]d, respectively. Note that, for simplicity, we consider only the leading collinear enhancements where the $\bbbar$ system originates either through the combination of $g\to \bbbar$ and $b\to gb$ IS splittings ($\mathcal{M}_{\IS\otimes\ttbar}$) or via IS $g\to gg$ plus FS $g\to \bbbar$ splittings ($\mathcal{M}_{\FS\otimes\ttbar}$). For events with external momenta $$\label{eq:ggttbbkin}
g(p_1)\,g(p_2) \to t (p_3)\, \bar t(p_4)\, b(p_5)\, \bar b(p_6),$$ the collinear limits take the general form $$\begin{aligned}
\label{eq:ISFSapprox}
\big|\mathcal{M}_{\mathrm{IS}\otimes\ttbar/\mathrm{FS}\otimes\ttbar}\big|^2
&=&(8\pi\as)^2\,
\max_{\alpha=1,2 \atop ij=56,65}\left\{
\frac{K_{\IS/\FS}(p_\alpha,p_i,p_j)}{(p_\alpha-p_i-p_j)^2}\,
\big|\mathcal{M}_{gg\to\ttbar}\big|^2_{p_\alpha\to z p_\alpha}\right\},\end{aligned}$$ where $\alpha\in\{1,2\}$ and $ij\in\{56,65\}$ specify, respectively, the IS gluon emitter and the ordering of the $\bbbar$ pair as depicted in c and \[fig:ttbbtopologies\]d. $K_{\IS/\FS}(p_\alpha,p_i,p_j)$ are the corresponding splitting kernels. The choice of $\alpha$ and $ij$ specifies a particular topology, and the maximum in defines $\big|\mathcal{M}_{\mathrm{IS}\otimes\ttbar}\big|^2$ and $\big|\mathcal{M}_{\mathrm{FS}\otimes\ttbar}\big|^2$ as the collinear limit of the most likely topology of IS and FS type. The splitting kernels read $$\begin{aligned}
\label{eq:ISapproxa}
K_{\mathrm{IS}}(p_\alpha,p_i,p_j)
&=&\frac{1}{(p_\alpha-p_i)^2-m_b^2}
\frac{P_{gq}(\xis)}{\xis}
\frac{P_{qg}(\yis)}{\yis},\nonumber\\
K_{\mathrm{FS}}(p_\alpha,p_i,p_j)
&=&\frac{-1}{(p_i+p_j)^2}
\frac{P_{gg}(z)}{z}
P_{gq}(\xfs)
,\end{aligned}$$ where $$\begin{aligned}
\label{eq:ISapproxc}
P_{gq}(x)&=T_R\left[1-2x(1-x)\right], & P_{qg}(x)& =C_F\left[\frac{1+(1-x)^2}{x}\right],
\nonumber\\
P_{gg}(x)&= 2C_A\left[\frac{x}{1-x}+\frac{1-x}{x}+x(1-x)\right], &\end{aligned}$$ with $T_R=1/2$, $C_F=4/3$ and $C_A=3$. The various momentum fractions are set to $$\begin{aligned}
\label{eq:ISapproxb}
\xis & =\frac{E_\alpha-E_i}{E_\alpha} & \yis& =\frac{E_\alpha-(E_i+E_j)}{E_\alpha-E_i},
\nonumber\\
\xfs & =\frac{E_j}{E_i+E_j}, & z& = \xis\,\yis\,. \end{aligned}$$ Finally, the underlying $gg\to \ttbar$ squared matrix element in reads $$\begin{aligned}
\label{eq:ttME}
%\hspace{-4mm}
\big|\mathcal{M}_{gg\to\ttbar}\big|^2_{p_\alpha\to z p_\alpha}
&=&
\left(4\pi\as\right)^2
\theta\left(z -\frac{m_{\ttbar}}{p_1\cdot p_2}\right)
\Bigg\{
\left[\frac{\pp{1}{2}^2}{6\pp{1}{3}\pp{2}{3}}-\frac{3}{8}\right]
\nonumber\\
&\times&
\left[\frac{\pp{1}{3}^2}{\pp{1}{2}^2}
+\frac{\pp{2}{3}^2}{\pp{1}{2}^2}
-\frac{2m_t^2}{\pp{1}{2}}
+\frac{m_t^4}{\pp{1}{3}\pp{2}{3}}
\right]
\bigg\}_{p_\alpha\to z p_\alpha}\hspace{-4mm},\hspace{4mm}\end{aligned}$$ where helicity/colour sums and average factors are included, and the momentum of the IS emitter has to be rescaled by $z$.
Numerical results for the diagrammatic decomposition and the collinear decomposition of $pp\to\ttbb$ at $\sqrt{s}=13$TeV are shown in . The first two plots display the leading $b$-jet $p_\rT$ distribution in the presence of ttb and ttbb cuts as defined in , requiring $\Nb\ge 1$ and $\Nb\ge 2$ $b$-jets, respectively. In the ttbb phase space, topologies with $g\to \bbbar$ FS splittings turn out to be surprisingly close to the full matrix element, with deviations that do not exceed the 10% level in the entire spectrum. This agreement remains remarkably good also in the inclusive ttb phase space, where IS splitting processes with one unresolved $b$-quark are expected to be more pronounced. Actually, with ttb and ttbb cuts the pure IS contribution ranges between 20–40% and 15–75%, respectively, but is almost entirely cancelled by the IS–FS interference.
The fact that the collinear approximations agree rather well with the corresponding squared Feynman diagrams up to relatively high $p_\rT$ confirms that $\ttbb$ production is dominated by the topologies in c and \[fig:ttbbtopologies\]d. On the other hand, the importance of interference effects provides strong motivation for using exact $\ttbb$ matrix elements, while collinear approximations such as those in or in the parton-shower modelling of $g\to \bbbar$ splittings should be used with due caution.
The above considerations apply also for the $m_{\bbbar}$ and $\Delta R_{\bbbar}$ distribution in c and d. In particular, we observe that topologies with FS $g\to \bbbar$ splittings are very close to the full matrix element in the whole $m_{\bbbar}$ spectrum as well as for $\Delta R_{\bbbar}<2$ . At the same time, for $50\,\GeV< m_{\bbbar}<
200\,\GeV$ and $1<\Delta R_{\bbbar}<2.5$, i.e. in the range of interest for $\ttbar H(\bbbar)$ analyses, we observe that IS splitting contributions and negative interference effects grow fast and tend to become very sizable. Thus a naive separation into contributions from IS and FS splittings is not applicable at large $m_{\bbbar}$ and $\Delta R_{\bbbar}$. On the other hand, in the region of moderate invariant mass and $\Delta R$ separation, which contains the bulk of the $\ttbb$ cross section, interference effects are rather small, and $\bbbar$ pairs turn out to originate almost entirely from FS $g\to \bbbar$ splittings.
In summary, given that the 5F scheme is based on the LO process $gb\to
\ttbar b$, where FS $g\to \bbbar$ splittings and interference effects are entirely neglected, the above observations provide strong motivation for a description of $\ttbar+b$-jet production based on $\ttbb$ matrix elements in the 4F scheme.
Technical aspects and setup of NLOPS $\boldsymbol{\ttbb}$ simulations {#se:technical}
=====================================================================
In this section we introduce a new generator based on $\ttbb$ matrix elements in the 4F scheme. Special emphasis is devoted to some technical aspects of the method that turn out to play an important role for multi-scale processes like $pp\to \ttbb$. In addition we describe the setup used for the $\ttbar+b$-jet simulations presented in , all relevant input parameters, scale choices and parton shower settings, as well as the treatment of theoretical uncertainties and the definitions of physics objects and selection cuts. Finally we provide details on the treatment of top-quark decays.
The new $\ttbb$ generator is implemented in the framework [@Jezo:2015aia], and the relevant LO and NLO matrix elements are computed by [@Cascioli:2011va; @OLhepforge; @Buccioni:2017yxi] through its interface [@Jezo:2016ujg]. For the evaluation of one-loop integrals employs the library [@Denner:2014gla; @Denner:2002ii; @Denner:2005nn; @Denner:2010tr] and, alternatively, [@Ossola:2007ax; @Ossola:2006us] together with the library [@vanHameren:2010cp]. While we do not apply the resonance-aware method [@Jezo:2015aia], the framework allows us to make use of new technical features, such as the automated implementation of scale variations , a [@Buckley:2010ar] interface and the option to unweight events partially. This $\ttbb$ generator will be soon publicly available on the webpage [@PWGBpage].
Powheg methodology {#se:methods}
------------------
In the following, we briefly review the method [@Nason:2004rx; @Frixione:2007vw] with emphasis on the separation of radiation into singular and finite parts. In this context we discuss technical subtleties that arise in the case of multi-scale processes like $pp\to \ttbb$.
The master formula for the description of NLO radiation in the approach consists of two contributions, $$\begin{aligned}
\mathd \sigma & = &
\mathd \sigma_\sing +
\mathd \sigma_\fin,
\label{eq:powhegsplit1}\end{aligned}$$ which arise from the splitting of real emission into singular (s) and finite (f) parts, $$R(\Phi_{\mathrm{R}}) = R_\sing(\Phi_{\mathrm{R}})+R_\fin(\Phi_{\mathrm{R}}).
\label{eq:powhegsplit2}$$ Here $R(\Phi_{\mathrm{R}})$ should be understood as squared real-emission matrix element, and $\Phi_{\mathrm{R}}$ as the corresponding phase space. Similarly, Born and virtual contributions in the Born phase space are denoted as $B(\Phi_{\mathrm{B}})$ and $V(\Phi_{\mathrm{B}})$. The splitting is implemented as $$\begin{aligned}
R_{\sing}(\Phi_{\mathrm{R}})
=
F(\Phi_{\mathrm{R}})\,
R(\Phi_{\mathrm{R}}),
\qquad
R_{\fin}(\Phi_{\mathrm{R}})
=
\left[1-F(\Phi_{\mathrm{R}})\right]\,
R(\Phi_{\mathrm{R}}),
\qquad
\label{eq:powhegsplit3}\end{aligned}$$ where $F(\Phi_{\mathrm{R}})\in [0,1]$ is a damping function that fulfils $F\to 1$ and $F\to 0$, respectively, in the infrared and hard regions of phase space (see below).
The singular part of real radiation is resummed according to the formula $$\begin{aligned}
\mathd \sigma_\sing & = & \bar{B} (\Phi_{\mathrm{B}}) \,\mathd \Phi_{\mathrm{B}} \left[
\Delta (q_{\tmop{cut}}) + \sum_{\alpha} \Delta (k_{\rT,\alpha})
\frac{R_{\sing,\alpha} (\Phi_{\alpha}
(\Phi_{\mathrm{B}}, \Phi_{\tmop{rad}})
)}{B
(\Phi_{\mathrm{B}})} \,\mathd \Phi_{\tmop{rad}} \right],
\label{eq:powheg1}\end{aligned}$$ where real emission is further split into FKS sectors [@Frixione:1995ms], $$R_\sing = \sum_{\alpha} R_{\sing,\alpha},\qquad
R_\fin = \sum_{\alpha} R_{\fin,\alpha},
\label{eq:fks1}$$ which isolate collinear singularities arising from individual emitters. In each sector, the emission phase space $\Phi_{\mathrm{R}}$ is factorised into the Born phase space $\Phi_{\mathrm{B}}$ and a one-particle radiation phase space $\Phi_{\tmop{rad}}$ through an appropriate FKS mapping, $$(\Phi_{\mathrm{B}}, \Phi_{\tmop{rad}})
\;\longrightarrow\;
\Phi_{\mathrm{R}} = \Phi_{\alpha} (\Phi_{\mathrm{B}}, \Phi_{\tmop{rad}}).
\label{eq:fks2}$$ The term within squared brackets in generates the hardest radiation according to an emission probability $R/B$. The parameter $k_{\rT,\alpha}=k_{\rT,\alpha}(\Phi_\rad)$ stands for the hardness of the radiated parton, and radiation harder than $k_{\rT,\alpha}$ is excluded by means of corresponding Sudakov form factors, $$\begin{aligned}
\Delta (q) & = &
\exp \left[ - \sum_\alpha\,\int_{k_{\rT,\alpha}\, >\, q}
\frac{R_{\sing,\alpha} (\Phi_{\alpha}
(\Phi_{\mathrm{B}}, \Phi_{\tmop{rad}})
)}{B
(\Phi_{\mathrm{B}})}\, \mathd \Phi_{\tmop{rad}} \right].
\label{eq:Sudakov}\end{aligned}$$ The term $\Delta(q_{\tmop{cut}})$ in represents the no-emission probability above the infrared cutoff $q_{\tmop{cut}}$, and Sudakov form factors account for unresolved multiple emissions in a way that cancels infrared singularities while preserving the differential NLO cross section $\bar{B} (\Phi_{\mathrm{B}})$ in the Born phase space. The latter is defined by integrating out the singular part of real radiation, $$\begin{aligned}
\bar{B} (\Phi_{\mathrm{B}}) & = & B (\Phi_{\mathrm{B}}) + V (\Phi_{\mathrm{B}}) +
\sum_{\alpha} \int R_{\sing,\alpha} (\Phi_{\alpha}
(\Phi_{\mathrm{B}}\,,\Phi_{\tmop{rad}})
) \,\mathd \Phi_{\tmop{rad}}\,.
\label{eq:powheg2}\end{aligned}$$ Here infrared cancellations between $V$ and $R_\sing$ are controlled via FKS subtraction. The remaining finite part of NLO radiation is treated as in fixed-order calculations, $$\begin{aligned}
\mathd \sigma_\fin
&=&
\sum_\alpha
R_{\fin,\alpha} (\Phi_{\alpha} (\Phi_{\mathrm{B}}, \Phi_{\tmop{rad}}))\,
\mathd \Phi_{\mathrm{B}},
\mathd \Phi_{\tmop{rad}}\,.
\label{eq:powheg3}\end{aligned}$$ Note that flux and symmetry factors as well as the convolution with PDFs are implicitly understood in and .
Let us now come back to the details of the separation of the singular and finite parts of real emission in . Technically, the damping function $F$ is implemented based on the kinematics of the actual FKS sector, $$\begin{aligned}
R_{\sing,\alpha}(\Phi_{\alpha})
= R_{\alpha}(\Phi_{\alpha}) -
R_{\fin,\alpha} (\Phi_{\alpha})
=F_{\alpha} (\Phi_{\alpha})
R_{\alpha}(\Phi_{\alpha}).\end{aligned}$$ The default functional form of $F$ in [@Alioli:2010xd; @Alioli:2008tz] is $$\begin{aligned}
F_{\alpha} (\Phi_{\alpha})
=
F_{\damp,\alpha} (\Phi_{\alpha})\,
F_{\bzd, \alpha} (\Phi_{\alpha}),
\label{eq:dampingdef}\end{aligned}$$ where $$\begin{aligned}
F_{\damp,\alpha} (\Phi_{\alpha})
=
\frac{\hdamp^2}{\hdamp^2+k_{\rT,\alpha}^2}
\label{eq:hdampdef}\end{aligned}$$ is the usual factor that smoothly shifts the weight of real radiation from $R_\sing$ to $R_\fin$ when the hardness of the emission, $k_{\rT,\alpha}$, becomes of the order of $\hdamp$ or higher. Note that the freely adjustable $\hdamp$ parameter in plays an analogous role as the resummation scale $\mu_Q$ in the MC@NLO method. This is because both parameters act as a $k_{\rT}$-threshold that separates the radiative phase space into a hard region, which is described by fixed-order matrix elements, from a singular region, where large logarithms of soft and collinear origin are resummed to all orders by means of Sudakov form factors. More precisely, in the MC@NLO approach the factor $R_{\sing,\alpha}/B$ and the terms $\Delta$ in correspond, respectively, to the parton-shower emission probability and the associated Sudakov form factors. Thus, in the MC@NLO framework, equation corresponds to the weight of so-called soft events supplemented by the probability of the first parton-shower emission and its no-emission counterpart. In analogy with , also the first MC@NLO shower emission is modulated by a certain damping function. The related reference scale $\mu_Q$, i.e. the MC@NLO counterpart of the scale $\hdamp$ in , corresponds to the upper bound of the first shower emission in MC@NLO. Thus MC@NLO predictions are sensitive to the choice of the shower starting scale.[^4] On the contrary, the first emission in – is entirely determined by the matrix element, which also dictates the scale at which the shower starts emitting further partons. Thus the method has the advantage of being essentially independent on the shower starting scale. More generally, thanks to the fact that the first emission is completely independent of the parton shower, predictions are characterised by a rather mild sensitivity to systematic uncertainties associated with the parton shower.
In addition to the well-known $\hdamp$-dependent damping mechanism , the also implements a theta function[^5] of the form [@Alioli:2008gx; @Alioli:2010xd] $$\begin{aligned}
F_{\bzd, \alpha} (\Phi_{\alpha})
=
\theta\left(\hbzd -\frac{R_{\alpha}(\Phi_{\alpha})}{\calR_{\alpha}(\Phi_{\alpha})}\right),
\label{eq:bzddef}\end{aligned}$$ where $\calR_{\alpha}$ corresponds to the infrared (soft and collinear) approximation of the full matrix element. Schematically it has the factorised form $$\begin{aligned}
\calR_{\alpha}(\Phi_\alpha) = \calK_{\alpha}(\Phi_{\mathrm{rad}}) B(\Phi_{\mathrm{B}}),
\label{eq:fks3}\end{aligned}$$ with an FKS kernel $\calK_{\alpha}(\Phi_{\mathrm{rad}})$ and an underlying Born contribution $B(\Phi_{\mathrm{B}})$, whose kinematics is determined by the inverse of the mapping in the actual sector $\alpha$. By default, the cut-off parameter $\hbzd$ in is set equal to 5. In this way, in the vicinity of IR singularities, where $R_\alpha/\calR_{\alpha}\to 1$, radiative contributions are attributed to $R_\sing$ and resummed according to . On the contrary, when the real emission matrix element largely exceeds the IR approximation , the resummation of the full $R/B$ kernel according to is not well justified, and corresponding events are attributed to the finite remnant through the theta function . In the standard , and in , the damping function is applied only to initial-state radiation. However, in the present $\ttbb$ generator we have extended it to all (massless or massive) final-state emitters, that have a FKS sector associated with it, in order to ensure a consistent resummation of QCD radiation off $b$-quarks.
The requirement $R_{\alpha}(\Phi_{\alpha}) <
\hbzd\,\calK_{\alpha}(\Phi_{\mathrm{rad}})\,B(\Phi_{\mathrm{B}})$ was originally introduced in order to avoid possible divergences of $R(\Phi_\alpha)/B(\Phi_\rB)$ due to so-called Born zeros, i.e. phase space regions where $B(\Phi_{\mathrm{B}})\to 0$. Such divergences cancel in the $\bar B/B$ ratio, i.e. they are not physical, and are not related to IR radiation. Corresponding $\Phi_\rB$ regions should thus be attributed to the finite remnant. Otherwise they could lead to dramatic inefficiencies in the event generation. More generally, the damping factor can play an important role also in case of multi-scale processes where the Born cross section involves enhancement mechanisms at scales well below the hard energy of the full process. Such enhancements can compete with the ones due to soft and collinear QCD radiation in a way that is somewhat analogous to Born zeros.
In the case of $pp\to \ttbb$, such effects can arise from the interplay of soft and collinear enhancements due to NLO light-jet radiation and to the generation of the $\bbbar$ system in regions with $m_{\bbbar}\ll m_{\ttbb}$ and/or $p_{\rT,\bbbar}\ll m_{\ttbb}$. For example, let us consider a $gg\to \ttbb g$ event with a gluon emission of ISR type. Its kinematics is generated starting from a $gg\to \ttbb$ Born event through a mapping of type , which creates the required gluon recoil by boosting the final state of the $gg\to\ttbb$ Born event in the transverse direction. The relevant boost factor, $\gamma=1/(1-\beta^2)^{1/2}$, is determined by $p_{\rT,j}=p'_{\rT,\ttbb}=\gamma\beta E_{\ttbb}$, where $E_{\ttbb}$ is the $\ttbb$ energy of the $\bbbar$ system in the Born event. If we assume, for simplicity, that he gluon is emitted in the same azimuthal direction as the $\bbbar$ system in the Born event, then the $\bbbar$ transverse momentum of the radiative event becomes $p'_{\rT,\bbbar}=\gamma(p_{\rT,\bbbar}-\beta E_{\bbbar})$, where $p_{\rT,\bbbar}$ and $E_{\bbbar}$ are the $\bbbar$ transverse momentum and energy in the Born event. Thus, the FKS mapping can lead to a very significant reduction of $p_{\rT,\bbbar}$. More precisely, for radiative events with $$\frac{p_{\rT,j}}{p_{\rT,\bbbar}}=(1+\epsilon)\frac{E_{\ttbb}}{E_{\bbbar}},
\label{eq:pTbbboosta}$$ the effect of the FKS boost on the $\bbbar$ system amounts to $$\frac{p'_{\rT,\bbbar}}{p_{\rT,\bbbar}}=\tilde \epsilon=\gamma-1-\epsilon.
\label{eq:pTbbboostb}$$ Thus, since the bulk of the $\ttbb$ cross section is characterised by $E_{\ttbb}
\gg E_{\bbbar}$ and $\gamma\sim 1$, in the case of hard QCD radiation with $\tilde \epsilon\ll 1$ the FKS mapping can lead to a drastic reduction of the $p_{\rT}$ of the $\bbbar$ system. As a result, in the region of small $m_{\bbbar}$, the ISR boost can enhance the $R_\alpha(\Phi_\alpha)/\calR_\alpha(\Phi_\alpha)$ ratio by up to a factor[^6] $(p_{\rT,\bbbar}/p'_{\rT,\bbbar})^2\sim 1/{\tilde\epsilon}^2$. This violates the main assumption that justifies the formula , namely $R_\alpha(\Phi_\alpha)/B(\Phi_B)\sim
\calR_\alpha(\Phi_\alpha)/B(\Phi_B)={\cal K}_\alpha(\Phi_\rad)$, which requires a sufficiently hard $\ttbb$ process as compared to the $k_\rT$ of NLO radiation. In particular, due to the sensitivity of the Born amplitude to scales of the order ${p_{\rT,\bbbar}}\sim (E_{\bbbar}/E_{\ttbb})\, p_{\rT,j}\ll p_{\rT,j}$, the factorisation formula is not fulfilled.
Fortunately, this problematic behaviour emerges only in relatively hard regions of the $\Phi_\rad$ phase space.[^7] Thus, as a remedy it is natural to shift such events into the finite remnant by means of the damping factor . In fact, in the case of $\ttbb$ production we have found that the $\hbzd$-dependent cut plays an important role for the efficiency generation of Les Houches events (LHEs) as well as for a consistent scale dependence. Moreover, applying a large $\hbzd$ cut we have observed a significant enhancement of the QCD scale dependence. This can be attributed to the fact that scale variations in the soft term are restricted to the $\bar B$ factor , where the unphysical distortions of the $\bbbar$ kinematics induced by the FKS mappings can jeopardise the natural cancellation of virtual and real contributions associated with a given Born configuration.
As discussed in , predictions for ttbb observables are rather stable with respect to variations of $\hbzd$. Thus, in order to avoid an unphysical enhancement of the scale dependence, we have reduced $\hbzd$ from its default value of 5 to $\hbzd=2$. This guarantees a more reasonable consistency with the fixed-order scale dependence without shifting an excessive fraction of the cross section from $\rd\sigma_\sing$ to $\rd\sigma_\fin$.
Input parameters, PDFs and scale choices {#se:input}
----------------------------------------
The predictions in are based on the following input parameters, scale choices and PDFs. Heavy-quark mass effects are included throughout using m\_[t]{} = 172.5 ,m\_[b]{} = 4.75 . All other quarks are treated as massless in the perturbative part of the calculations. Since we use massive $b$-quarks, for the PDF evolution and the running of $\as$ we adopt the 4F scheme. Thus, for consistency, we renormalise $\as$ in the decoupling scheme, where top- and bottom-quark loops are subtracted at zero momentum transfer. In this way, heavy-quark loop contributions to the evolution of the strong coupling are effectively described at first order in $\as$ through the virtual corrections.
For the calculation of hard cross sections at LO and NLO, as well as for the generation of the first emission, we use the [NNPDF30\_nlo\_as\_0118\_nf\_4]{} parton distributions [@Ball:2014uwa] as implemented in the LHAPDFs [@Buckley:2014ana] and the corresponding $\asfour$.[^8] To assess PDF uncertainties we re-evaluate the weights of LHEs with 100 different PDF replicas, while using the nominal PDF set for parton showering.
Since it scales with $\as^4$, the $\ttbb$ cross section is highly sensitive to the choice of the renormalisation scale $\mur$, and this choice plays a critical role for the stability of perturbative predictions. Following [@Cascioli:2013era; @deFlorian:2016spz], we adopt a scale choice of the form $$\label{eq:mur}
\mur=\xir \sqrt{\mu_{\ttbar}\,\mu_{\bbbar}},$$ with the scale-variation factor $\xir\in[0.5,2]$. This dynamic scale choice accounts for the fact that $\ttbb$ production is characterised by two widely separated scales, which are related to the $\ttbar$ and $\bbbar$ systems and are chosen as the geometric average of the respective transverse energies, $$\label{eq:mubbtt}
\mu_{\bbbar}=\sqrt{E_{\rT,b}E_{\rT,\bar b}},
\qquad
\mu_{\ttbar}=\sqrt{E_{\rT,t}E_{\rT,\bar t}}\;.$$ The transverse energies $E_{\rT,i}=\sqrt{m_i^2+p^2_{\rT,i}}$ are defined in terms of the rest masses $m_i$ and the transverse momenta $p_{\rT,i}$ of the bare heavy quarks. The scales are computed according to physical kinematics, i.e. without projecting real emission events to the underlying Born phase space. The choice (\[eq:mur\]) is applied to all (N)LO matrix elements apart from the $\as$ factor that results from the $R/B$ ratio in . In that case $\as$ is evaluated at the transverse momentum of the hardest emission, and that $\as(k_{\rT,\alpha})$ factor is not subject to scale variations.
For the factorisation scale $\muf$ we use[^9] $$\label{eq:muf}
\muf=\xif\,\frac{H_\rT}{2}
= \frac{\xif}{2} \sum_{i=t,\bar t, b, \bar b,j} E_{\rT,i}\;,$$ where $\xif\in [0.5,2]$, and the total transverse energy of the $\ttbb$ system, $H_\rT$, is computed in terms of bare-quark transverse momenta including also QCD radiation at NLO. Our nominal predictions correspond to $\xir=\xif=1$, and to quantify scale uncertainties we take the envelope of the seven-point variation $(\xir,\xif)=(0.5,0.5)$, $(0.5,1)$, $(1,0.5)$, $(1,1)$, $(1,2)$, $(2,1)$, $(2,2)$.
For the parameters $\hbzd$ and $\hdamp$, which control the resummation of NLO radiation according to – as discussed in , we set $$\label{eq:bzd}
\hbzd=2$$ and $$\hdamp=\frac{H_\rT}{2}
= \frac{1}{2} \sum_{i=t,\bar t, b, \bar b} E_{\rT,i}\;.
\label{eq:hdampchoice}$$ Here the various $E_{\rT,i}$ are defined in the underlying Born phase space. To account for the uncertainties associated with these choice we apply the independent variations $\hbzd=2$, $5$, $10$ and $\hdamp={H_\rT}/{4},{H_\rT}/{2},H_\rT,1.5\,m_t$, varying both parameters one at a time.[^10]
The above choices for $\mur,\muf$ and $\hdamp$, as well as the employed PDFs correspond to the setup recommended in [@deFlorian:2016spz].
Parton shower settings and variations {#se:shower}
-------------------------------------
By default, LHEs are showered with 8.2 using the A14 tune[^11], where ISR and FSR parameters as well as the MPI activity have been tuned in a single step using most of the available $\ttbar$ ATLAS data from Run1 [@ATL-PHYS-PUB-2014-021]. In the A14 tune, $m_b=4.75$GeV and $\asfive(M_Z)=0.127$, both for ISR and FSR, while in the default Monash tune $\asfive(M_Z)=0.13650$. Since the shower evolution is implemented in the 5F scheme we shower events using 5F PDFs. Specifically, we choose the [NNPDF30\_nlo\_as\_0118]{} 5F PDFs.[^12]
The interplay between and is controlled by the [scalup]{} parameter, which describes the hardness of radiation in LHEs and may be taken as starting scale for . However, in order to avoid inconsistencies due to the fact that the evolution variable does not coincide with the definition of hardness in , we apply the following two-step procedure based on the class. Instead of starting below [scalup]{}, we instruct to generate radiation up to the kinematic limit by setting
pythia.readString("SpaceShower:pTmaxMatch = 2");
pythia.readString("TimeShower:pTmaxMatch = 2");
Then, to guarantee the correct ordering of emissions in and , we apply a veto on each emission that is harder than [scalup]{} according to the definition of hardness. This is achieved by setting
pythia.readString("POWHEG:nFinal = 4");
pythia.readString("POWHEG:veto = 1");
The remaining settings are left to their default values.
At LOPS level we set the shower starting scale equal to $H_{\rT}/2$ and vary it up and down by a factor two in order to assess the related uncertainty. At NLOPS, the shower starting scale is dictated by the kinematics of real emission matrix elements in the method. Thus, at variance with NLOPS predictions based on the MC@NLO method, predictions are free from uncertainties related to the choice of the shower starting scale.
In order to assess uncertainties due to the parton-shower modelling of $g\to \bbbar$ splittings we vary the parameter [TimeShower:weightGluonToQuark]{}, which permits to select various optional forms of the $g\to Q\bar Q$ splitting kernel in 8. The default is option 4, which corresponds to the splitting probability [@pythianote] $$\label{eq:gbbmesplitting}
\rd P_{g\to \bbbar} = \frac{\as(p_T^2)}{2\pi}\frac{\rd m_{\bbbar}^2}{m_{\bbbar}^2}
\frac{\beta_b}{2}\left[z^2+(1-z)^2+8r_b z(1-z)\right](1-\delta)^3,$$ where $r_b=m_b^2/m_{\bbbar}^2$, $\beta_b=\sqrt{1-4r_b}$ and $\delta=m^2_{\bbbar}/m^2_{\mathrm{dipole}}$. The factor $(1-\delta)^3$, which suppresses the production of high-mass $\bbbar$ pairs, is derived from the $H\to g\bbbar$ matrix element by interpreting $m_H$ as the mass of a gluon dipole, $m_{\mathrm{dipole}}$. Omitting the factor $(1-\delta)^3$ in corresponds to option 2 and results in a DGLAP splitting probability of type $\gamma^*\to
\bbbar$ with mass effects. More precisely, in option 2, $g\to \bbbar$ splittings are generated based on massless kinematics, and the $r_b$ mass correction is implemented through reweighting, while massive kinematics is restored through momentum reshuffling. Option 3, which implements massive DGLAP splittings in a more realistic way, involves an additional $(1+\delta)/(1-\delta)$ factor that leads to a significant enhancement of the $g\to \bbbar$ rate. This option is excluded by LEP/SLC data and also by direct measurements of $\ttbar+b$-jet production [@Aad:2015yja]. Finally, option 1 corresponds to option 2 with $r_b=0$ and yields very similar results. Thus, for the assessment of $g\to \bbbar$ shower uncertainties we will compare options 4 and 2.
In addition to the functional form of the heavy-quark splitting kernel we also vary the scale of $\as$ in the parton shower. To this end, we set [TimeShower:weightGluonToQuark]{} to 6 and 8, which corresponds to options 2 and 4 with $\as(p_T^2)$ replaced by $\as(m_{\bbbar}^2)$ in the heavy-quark splitting kernel . Moreover, using [TimeShower:renormMultFac]{}, we vary $\as(p_\rT^2)\to \as(\xi p_\rT^2)$ with prefactors $\xi=0.1,1,10$ both for options 2 and 4. This latter variation is applied to all final-state QCD splittings, i.e. also splittings of type $g\to gg$, $q\to q g$, etc.
Comparisons against alternative generators {#se:gencompsettings}
------------------------------------------
In order to assess systematic uncertainties related to the parton shower and the matching scheme, in we compare predictions of $\ttbb$ production against corresponding predictions generated with and with [@Cascioli:2013era]. The and generators of $\ttbb$ production are also compared against corresponding generators of inclusive $\ttbar$ production in the 5F scheme[^13].
In the case of [@Bellm:2015jjp] we apply the angular ordered shower using version 7.1, setting $m_b = 4.75$GeV, and leaving the strong coupling to its default value, $\alpha_S(m_Z) = 0.126234$. To restrict the hardness of emissions according to the value of [scalup]{} in the LHEs we set
set /Herwig/Shower/ShowerHandler:RestrictPhasespace Yes
set /Herwig/Shower/ShowerHandler:MaxPtIsMuF Yes
In the case of we use version 2.2.4 with its default tune[^14] for ’s dipole shower [@Schumann:2007mg]. The relevant one-loop matrix elements are computed with , and matching to the parton shower is based on the implementation [@Hoeche:2011fd] of the MC@NLO method [@Frixione:2002ik], dubbed SMC@NLO. As for the hard cross section we use the same input parameters, PDFs and scale settings as specified in for the case of . Moreover, as motivated in we identify the resummation scale $\mu_Q$ in with the $\hdamp$ parameter in , we set $\mu_Q=H_\rT/2$. In the simulation the [NNPDF30\_nlo\_as\_0118\_nf\_4]{} PDF set is used throughout, also for paton showering.
For and simulations of inclusive $\ttbar$ production we use the same setup as for the corresponding $\ttbb$ generators, with the only exceptions being the QCD scales, $\mur=\muf=0.5\sqrt{E_{\rT,t}E_{\rT,\bar t}}$, and the choice of the [NNPDF30\_nlo\_as\_0118\_nf\_5]{} PDF set. In this setup the inclusive NLO cross section amounts to $\sigma_{\ttbar}=815$pb, which is only 2% below the NNLO prediction of $832^{+45}_{-50}$pb [@Czakon:2013goa].
Simulations with stable or decayed top quarks {#se:topdecaytreatment}
---------------------------------------------
In predictions for $\ttbb$ production are presented both for the case of stable top quarks and with spin-correlated top decays. Simulations with stable top quarks permit to avoid the combinatorial complexity that results from the presence of four $b$-quarks in decayed $\ttbb$ events. In this way one can focus on the production of the $\bbbar$ pair that is governed by QCD dynamics and which represents the main source of theoretical uncertainty in $pp\to \ttbb$. Moreover, results with stable top quarks can be compared to the benchmarks of . Since top quarks do not hadronise, when we switch off top decays we disable hadronisation and, following , we also deactivate multi-parton interactions (MPIs) and QED radiation in the parton shower. This is achieved by setting
pythia.readString("6:mayDecay = off");
pythia.readString("-6:mayDecay = off");
pythia.readString("SpaceShower:QEDshowerByQ = off");
pythia.readString("SpaceShower:QEDshowerByL = off");
pythia.readString("TimeShower:QEDshowerByQ = off");
pythia.readString("TimeShower:QEDshowerByL = off");
pythia.readString("PartonLevel:MPI = off")
pythia.readString("HadronLevel:All = off");
For the case of decaying top quarks we show results both with hadronisation and MPI switched off or on, while the QED shower is always activated and hadrons are kept stable throughout. For the implementation of spin-correlated decays in the framework we follow the approach of , which has already been employed in the framework in . More precisely, we use resonant tree matrix elements for the full $2\to 8$ Born processes $q\bar q/gg\to t (\to b ij)\bar{t}(\to \bar b kl)\bbbar$, where $ij$ and $kl$ stand for the leptons or quarks from $W$ decays, and corresponding $2\to 9$ processes with an additional external gluon at the level of the $pp\to\ttbb$ sub-process. In the $2\to 8(9)$ matrix elements we include only topologies with two intermediate top resonances. This accounts for spin correlations as well as for off-shell effects associated with the top and the $W$ propagators. Technically, top decays are generated starting from on-shell $\ttbb$ events with a veto algorithm based on the ratio between $2\to 8(9)$ matrix elements and corresponding $2\to 4(5)$ matrix elements for the underlying $pp\to\ttbb$(+jet) process.
As additional input parameters for top decays we use [@Patrignani:2016xqp] $$\label{eq:ewinputs}
M_{W} = 80.385 \;\GeV,\qquad G_{\rF} = 1.1663787\cdot 10^{-5}\; \GeV^{-2},$$ the total widths $$\label{eq:inputwidths}
\Gamma_t=1.329\;\GeV,\qquad
\Gamma_W=2.089\;\GeV\,,$$ and the branching ratios $$\begin{aligned}
\label{eq:GammaWa}
\mathrm{BR}_{t\to b\ell_i\nu_j}&=&
\mathrm{BR}_{W\to \ell_i\nu_j}=
\frac{\delta_{ij}}{3}\mathrm{BR}_{W\to \mathrm{lept}},
\\
\mathrm{BR}_{t\to b u_i d_j}&=&
\mathrm{BR}_{W\to u_i d_j}=
\frac{|V_{ij}|^2}{2}\mathrm{BR}_{W\to \mathrm{had}},\end{aligned}$$ where we assume a 100% branching ratio for $t\to bW$ decays. For the total $W$-boson branching ratios into leptons and hadrons we use the values [@Patrignani:2016xqp] $$\begin{aligned}
\label{eq:GammaWb}
\mathrm{BR}_{W\to \mathrm{had}}=0.675,\qquad
\mathrm{BR}_{W\to \mathrm{lept}}=0.325\,,\end{aligned}$$ which include state-of-the-art higher-order corrections.
Jet observables and acceptance cuts {#se:cuts}
-----------------------------------
For the reconstruction of jets we use the anti-$k_\rT$ [@Cacciari:2008gp] algorithm with $R=0.4$. We select jets that fulfil $$\label{eq:jetcuts}
p_{\rT}>25\,\GeV,\qquad
|\eta|<2.5,$$ both for the case of light jets and $b$-jets. At parton level, we define as $b$-jet a jet that contains at least a $b$-quark, jets that contain a $\bbbar$ pair arising from a collinear $g\to \bbbar$ splitting are also tagged as $b$-jets. At particle level, i.e. when hadronisation is switched on, we tag as $b$-jets those jets that are matched to a $B$-hadron using the ghost method as implemented in [@Cacciari:2011ma].
When studying $\ttbb$ production with stable top quarks, in Sections \[se:gbbstudies\] and \[se:stabletops\], we categorise events according to the number $\Nb$ of $b$-jets that do not arise from top decays and fulfil the acceptance cuts . For the analysis of cross sections and distributions we consider an inclusive selection with $N_b\ge 1$ and a more exclusive one with $\Nb\ge 2$. We refer to them as ttb and ttbb selections, respectively.
In we present predictions for $\ttbb$ production with top-quark decays in the dilepton channel. In this case we require two oppositely charged leptons, $\ell=e$ or $\mu$, with $$p_{\rT,\ell}>20\,\GeV,
\qquad
|\eta_{\ell}|< 2.5.
\label{eq:leptoncuts}$$ Charged leptons are dressed with collinear photon radiation within a cone of radius $0.1$. We do not apply any cut on missing transverse energy. Jets are defined as for the case of stable top quarks, and we select events with at least four $b$-jets that fulfill the acceptance cuts .
Predictions for $\boldsymbol{\ttbb}$ production with stable top quarks {#se:stabletops}
======================================================================
In this section we present numerical predictions for $pp\to \ttbb$ at $\sqrt{s}=13$TeV in the 4F scheme. The presented results have been obtained with using the setup of . Top quarks are kept stable throughout as specified in , and we study cross sections and distributions in the inclusive ttb phase space with $\Nb\ge 1$ $b$-jets, as well as in the ttbb phase space with $\Nb\ge 2$.
NLOPS predictions with perturbative uncertainties {#se:pertplots}
-------------------------------------------------
In this section we compare (N)LO and (N)LOPS predictions focusing on NLO and matching effects as well as perturbative and PDF uncertainties. presents cross sections in the ttb and ttbb phase space, as well as in the presence of an additional cut, $m_{b_1b_2}>100\,$GeV, on the invariant mass of the two hardest $b$-jets. At fixed order we find perfect agreement with the NLO results of . The various phase space regions feature similar NLO uncertainties, around 25–30%, while corresponding LO scale variations are roughly a factor two larger. Both at LO and NLO, scale uncertainties are strongly dominated by $\mur$ variations. The large $\sigma_{\mathrm{ttb}}/\sigma_{\mathrm{ttbb}}$ ratio, which exceeds a factor 5, reflects the appearance of large logarithms of $m_b$ when a $b$-quark becomes unresolved. As shown in , such logarithms are mainly due to FS $g\to \bbbar$ splittings. Thus the use of 4F PDFs, where $\ln(m_b/Q)$ effects of IS origin are not resummed in the PDF evolution is well justified. Note also that $\ln(m_b/Q)$ effects in $\sigma_{\mathrm{ttb}}$ are present already at LO. Thus they do not jeopardise the convergence of the perturbative expansion. In fact, $\sigma_{\mathrm{ttb}}/\sigma_{\mathrm{ttbb}}$ turns out to be very stable with respect to NLO corrections. The same hold for $\sigma_{\mathrm{ttbb}}/\sigma_{\mathrm{ttbb}_{100}}$.
At variance with [@Cascioli:2013era], where LO calculations were performed using LO PDFS and the corresponding value of $\as$, here, in order to obtain a more realistic picture of the convergence of the $\as$-expansion, we use NLO inputs throughout.[^15] This approach increases the NLO $K$-factors from 1.15–1.25 [@Cascioli:2013era] to 1.80–1.95. This observation raises some concerns regarding the possible presence of significant higher-order corrections beyond NLO and calls for a better understanding of the origin of the large $K$-factor at NLO. This question as well as the search for possible improvements is deferred to future studies.
Comparing fixed-order (N)LO cross sections against (N)LOPS ones we find that matching and showering effects are almost negligible in $\sigma_{\mathrm{ttb}}$, while in the case of $\sigma_{\mathrm{ttbb}}$ they slightly exceed 10%, and in the Higgs-signal region, $m_{b_1b_2}>100\,\GeV$, they approach 30% . As pointed out in , such effects can be understood in terms of $\ttbar+2b$-jet production via double $g\to \bbbar$ splittings. In practice, one of the $b$-jets results from a $g\to \bbbar$ splitting in the $\ttbb$ matrix element, while the second one is created by the parton shower via a further $g\to \bbbar$ collinear splitting. This interpretation is confirmed by the fact that the enhancement at hand is not present in the LHE-level cross sections presented in . In fact, double splittings are generated only at NLOPS level through parton showering. Double-splitting enhancements in behave in a qualitatively similar way as in , but their size turns out to depend on the employed NLOPS generator. As compared to , we observe that the NLOPS/NLO correction to ${\sigma_\mathrm{ttbb}}$ in (+12%) is twice as large as in (+6%), very close to [^16] (+13%) and well below the prediction of (+41%).
For what concerns scale variations, in we see that their impact at NLOPS tends to be 5–10% higher as compared to fixed-order NLO. This is consistent with the behaviour of and in , while features a significantly lower scale uncertainty. Such differences may be an artefact of the incomplete implementation of scale variations in the various NLOPS tools. In the case of , as anticipated in we have found that increasing $\hbzd$ can lead to unphysical enhancements of the scale uncertainty. This effect is mostly visible in the ttbb phase space, where the maximum scale variation amounts to $+40\%$ for $\hbzd=2$ and grows up to $+45\%$ and $+54\%$ when setting $\hbzd=5$ and 50, respectively. Based on these observations, as default for our $\ttbb$ simulations we have set $\hbzd=2$. This choice guarantees a decent consistency with fixed-order scale variations without altering the matching procedure in a drastic way. In particular, when $\hbzd$ is reduced from its standard value of $5$ down to $2$, we have checked that the fraction of the $\ttbar+b$-jet cross section that is shifted from the singular part to the finite remnant amounts to only 10–20%. This holds for all considered distributions in the ttb and ttbb phase space.
Differential observables with ttb and ttbb cuts are presented in . The inclusive $b$-jet multiplicity distribution in a extends the results of , which correspond to $\Nb\ge 1,2$, to the bins with $\Nb\ge 3,4$. The latter are populated by events that result from the interplay of real-emission matrix elements and $g\to \bbbar$ parton-shower splittings. Thus they feature an enhanced scale dependence.
For kinematic distributions that are inclusive with respect to NLO QCD radiation, NLOPS scale variations have a minor impact on shapes and amount essentially to a normalisation shift, similar to what observed at the level of the ttb and ttbb cross sections. In contrast, in the case of the light-jet $p_\rT$ spectra, scale variations increase from about 30% in the soft region up to 100% in the hard tails. This is consistent with the fact that such observables are only LOPS accurate and depend on $\as^5(\mur)$. The effect of PDF variations is clearly subleading as compared to scale uncertainties and has little impact on shapes.
Comparing (N)LOPS predictions to the respective fixed-order (N)LO results, we observe that matching and shower effects remain almost negligible also at the level of distributions in the ttb phase space. As for the ttbb region, the NLOPS effects of order 10% observed in $\sigma_{\mathrm{ttbb}}$ turn out to be quite sensitive to the kinematics of $b$-jets. In particular, as expected from the QCD dynamics of double $g\to \bbbar$ splittings [@Cascioli:2013era], the most pronounced effects are observed in the tails of the $m_{b_1b_2}$ and $\Delta R_{b_1b_2}$ distributions, where the NLOPS/NLO ratio approaches a factor two. In the Higgs signal region, $m_{b_1b_2}\sim 125$GeV, the NLOPS enhancement is around 1.25 and well consistent with .
Comparing fixed-order NLO predictions to LO ones we find that, in spite of the fairly large $K$-factors observed in , the shapes of distributions turn out to be quite stable with respect to higher-order QCD corrections. In the case of (N)LOPS predictions, the situation is different, especially for the shape of the light-jet $p_\rT$ spectra, which receives significant NLO distortions. This is not surprising, since at LOPS the light-jet $p_\rT$ is entirely generated by the parton shower. Thus the NLOPS/LOPS ratio should be regarded as a LO matrix-element correction to the parton-shower approximation, rather than a NLO correction in the perturbative sense.
Significant differences between NLOPS and LOPS shapes are observed also in the $m_{b_1b_2}$ and $\Delta R_{b_1b_2}$ distributions. Since the respective NLO and LO shapes are very similar, this behaviour can be attributed to the parton shower. More precisely, it can be understood as a side effect of the above-mentioned NLOPS/LOPS correction to the light-jet $p_\rT$ spectra, which is converted into a double-splitting effect by $g\to \bbbar$ splittings inside the light jet.
Shower uncertainties {#se:PSplots}
--------------------
In we study the sensitivity of (N)LOPS predictions to parton-shower and matching uncertainties for the same observables considered in .
The ratios displayed in the upper frames illustrate the net effect of parton showering by comparing full NLOPS predictions against results at LHE level. In addition, to assess parton-shower uncertainties, NLOPS predictions based on are compared to the corresponding results obtained with .
In the ttb phase space, apart from a mild distortion of the light-jet spectrum, the net effect of parton showering is essentially negligible. In contrast, in the ttbb phase space it increases the cross section by about 5% and tends to grow in the tails of distributions. The most sizable shower effects are observed in the $m_{b_1b_2}$ and $\Delta
R_{b_1b_2}$ distributions, where they reach up to 50–100%. This behaviour is well consistent with the enhancement of the NLOPS/NLO ratio observed in , and the fact that it is driven by the parton shower provides further support to its interpretation in terms of double $g\to
\bbbar$ splittings.
In spite of the important role of parton showering, it is reassuring to observe that the sensitivity of NLOPS predictions to the choice of parton shower is very small. In fact, the typical agreement between results based on and is at the level of a few percent both in the ttb and ttbb selections. Sizeable deviations at the level of 20% are observed only when requiring more than three $b$-jets. As discussed in , the very mild sensitivity of predictions to the choice of parton shower is due to the fact that the first emission is completely independent of the parton shower in the approach.
The ratios shown in the central frames of illustrate (N)LOPS uncertainties related to the modelling of $g\to \bbbar$ splittings and variations of $\as$ in (see ). At LOPS also variations of the shower starting scale ([scalup]{}) are shown.
The fact that $\ttbb$ 4F matrix elements populate the whole $\bbbar$ phase space restricts the effect of $g\to \bbbar$ shower splittings to events with four or more $b$-quarks. Thus, only the cross sections with $\Nb\ge 3,4$ $b$-jets suffer from sizable shower uncertainties. Vice versa, all considered observables with ttb or ttbb cuts turn out to be very stable, with typical shower uncertainties of a few percent at NLOPS. This holds also for the observables that are most sensitive to double splittings, i.e. $m_{b_1b_2}$ and $\Delta R_{b_1b_2}$, the only exception being the tail of the $\Delta R_{\bbbar}$ distribution, where double-splitting effects can reach 50% of the NLOPS cross section, while $g\to \bbbar$ shower uncertainties can reach 15%.
Predictions at LOPS depend also on the choice of the shower starting scale. This uncertainty is especially sizeable in the case of the light-jet spectrum, where [scalup]{} acts as a cutoff. A sizeable [scalup]{} dependence is visible also in the LOPS predictions for the $p_\rT$-distributions of $b$-jets, which indicates that such observables are rather sensitive to QCD radiation. Let us recall that the [scalup]{} dependence disappears completely in NLOPS simulations based on the approach.
Ratios plotted in the lower frames of show the dependence of NLOPS predictions with respect to the choice of the $\hdamp$ and $\hbzd$ parameters, which control the separation of the first emission into events of soft and hard type in the framework (see ). The $\hdamp$ band is obtained by varying $\hdamp=H_\rT/4$, $H_\rT/2$, $H_\rT$, $1.5m_t$ with the value of $\hbzd$ fixed to 2, while the $\hbzd$ band is obtained by varying $\hbzd=2$, $5$, $10$ with fixed $\hdamp=H_\rT/2$. Observables that are inclusive with respect to light-jet radiation reveal a remarkably small dependence, typically of the order of a few percent, on the choice of $\hdamp$ and $\hbzd$. Non-negligible but moderate uncertainties are found only in the light-jet spectra, which are enhanced by up to 20% when $\hbzd$ is increased from 2 to 10. Investigating simultaneous variations of $\hdamp$ and $\hbzd$ (not plotted) we have found that the size of the $\hdamp$ variation band is fairly stable with respect to the value of $\hbzd$ within the considered range.
Comparisons against other $\boldsymbol{\ttbb}$ and $\boldsymbol{\ttbar}$ generators {#se:gencomp}
-----------------------------------------------------------------------------------
In we compare $\ttbar+b$-jet predictions based on and . This comparison is done both for (N)LOPS $pp\to \ttbb$ generators in the 4F scheme and for corresponding generators of inclusive $\ttbar$ production in the 5F scheme. Specifically, in the case of we use [hvq]{} [@Frixione:2007nw]. As detailed in , input parameters, QCD scales and matching parameters are chosen as coherently as possible across all generators. In this spirit, the parameter $\hdamp=H_\rT/2$ in is identified with the resummation scale $\mu_Q$ in the SMC@NLO framework of . Instead, for what concerns the parton showers we simply use standard settings, we do not try to improve the agreement between generators by tuning the and showers.
The ratios in the upper frames of show $pp\to \ttbb$ predictions normalised to corresponding predictions at LOPS and NLOPS accuracy. The bands describe the combination in quadrature of all matching and shower uncertainties[^17] in (referred to shower uncertainties in the following), while only nominal predictions are considered in the ratios. Comparing LOPS predictions gives direct insights into the different modelling of radiation in and . For observables that are inclusive with respect to jet radiation we find deviations between 10–40% and comparably large shower uncertainties. In contrast, in the jet-$p_\rT$ distributions the LOPS predictions of are far above the ones by , with differences that can reach a factor 2.5 in the tails. These differences are perfectly consistent with LOPS shower uncertainties, which are dominated by variations of the starting scale.
Moving to NLOPS reduces the direct dependence on the parton shower. At the same time, differences between the and SMC@NLO matching methods come into play. In practice, at NLOPS we observe a drastic reduction of shower uncertainties, especially in the light-jet and $b$-jet $p_\rT$-distributions. Also the differences between and become very small at NLOPS. The ttb and ttbb cross sections agree at the percent level, and differential $b$-jet observables deviate by more than 5% only in the tails of the $m_{b_1b_2}$ and $\Delta R_{b_1b_2}$ distributions. Even the light-jet spectra in the ttb and ttbb phase space deviate by less than 10–20% up to high $p_\rT$, in spite of the limited formal accuracy (LOPS) of such observables. In the light of these results, NLOPS theoretical uncertainties related to the matching scheme and the parton shower seem to be well under control in $pp\to \ttbb$. In particular, their impact appears to be clearly subleading as compared to QCD scale uncertainties.
In the central frames of we compare (N)LOPS generators of inclusive $\ttbar$ production based on and . In this case, the $g\to \bbbar$ final-state splittings that give rise to $\ttbar+b$-jet signatures are entirely controlled by the parton shower. At LOPS, also the parent gluon that splits into $\bbbar$ is generated by the parton shower. Nevertheless, the ttb and ttbb LOPS cross sections predicted by and deviate by less than 30%–40%. Instead, as expected, the shapes of $\ttbar+b$-jet observables vary very strongly, and in all considered light-jet and $b$-jet distributions results exceed ones by a factor of two and even more. This excess is well consistent with the estimated LOPS shower uncertainties. At NLOPS, only $g\to \bbbar$ splittings are controlled by the parton shower, while the emission of their parent gluon is dictated by LO matrix elements. Consequently, we observe a drastic reduction of shower uncertainties as compared to LOPS. Also the differences between and are largely reduced at NLO, nevertheless they remain quite significant in various distributions.
To provide a more complete picture of the uncertainties of inclusive $\ttbar$ simulations, in the lower frames of we compare generators of inclusive $\ttbar$ production and $\ttbb$ production. Shower uncertainties are shown only for the $\ttbar$ generator. At LOPS, the $\ttbar$ generator is strongly sensitive to the modelling of $pp\to \ttbar g$ through initial-state gluon radiation in . As a result, the $\ttbar$ generator overestimates the ttb and ttbb cross sections by about 90% and 50%, respectively. This excess is strongly sensitive to [scalup]{}, and in the $p_\rT$-distributions it is confined to the regions below 100–200GeV, while the tails are strongly suppressed. Also the $m_{b_1b_2}$ and $\Delta R_{b_1b_2}$ distributions feature strong shape differences as compared to LOPS $\ttbb$ predictions.
Such differences go down significantly at NLOPS. The ttb and ttbb cross sections predicted by the $\ttbar$ generator overshoot $\ttbb$ results by only 15–20%, and also $b$-jet observables feature an improved agreement with $\ttbb$ predictions. Nevertheless, in $b$-jet observables we find quite significant shape differences, especially for the $m_{b_1b_2}$ and $\Delta R_{b_1b_2}$ distributions, and shower uncertainties remain far above the ones of the $\ttbb$ generator (see upper frame). As for the light-jet spectra, $\ttbar$ predictions turn out to lie above $\ttbb$ ones by about a factor of two in the tails. In principle, with the help of parton shower tuning NLOPS $\ttbar$ generators may be amenable to a reasonable description of inclusive $\ttbar+b$-jet observables. However, in the light of the above results it should be clear that NLOPS $\ttbb$ generators are mandatory in order to achieve an acceptable level of shower systematics.
$\boldsymbol{\ttbb}$ production with top-quark decays {#se:decayedtops}
=====================================================
In this section we present NLOPS results of the $\ttbb$ generator with leptonic top-quark decays. More precisely we consider final states with oppositely charged leptons and/or muons. By default hadronisation and MPI are deactivated in , and their effect is shown separately. As detailed in , our implementation of top decays is based on resonant $pp\to \ttbb\to 2\ell2\nu \bbbar$(+j) matrix elements, where spin correlations are consistently taken into account.
Top-quark decays are due to weak interactions and, up to small corrections of $\mathcal{O}(\Gamma_t/m_t)$, their effect factorises with respect to $\ttbb$ production. Thus, while they strongly increase the complexity of $\ttbar+b$-jet events, top decays are not expected to interfere with the QCD dynamics of $pp\to \ttbb$ in a significant way. In order to verify this hypothesis, in we compare NLOPS $pp\to
\ttbb$ simulations with stable and decayed top quarks. To this end, based on Monte Carlo truth, all reconstructed jets are split into two subsets associated with $\ttbb$ production and top decays. Specifically, jets that contain a parton originating from showered top-decay products are attributed to top decays, otherwise to $\ttbb$ production.[^18] At the level of top decays we require two $b$-jets and two charged leptons within the acceptance cuts –, while for the “reconstructed” $\ttbb$ system we consider the same cuts and observables as for the case of stable top quarks. In order to mimic the leptonic branching ratio and the efficiency of acceptance cuts on top-decay products, the normalisation of the $\ttbb$ simulation with stable top quarks is adapted to the predictions with decayed top quarks. This is done at the level of the ttbb cross section through a constant normalisation factor.
As shown in , $\ttbar+b$-jet observables with stable top quarks and reconstructed top decays turn out to agree quite well: $b$-jet cross sections and distributions deviate by only 5–10%, and also in the light-jet $\pt$-distribution decay effects hardly exceed 10%. These differences can be understood as indirect effect of the acceptance cuts on top-decay products, which result from the correlation between the kinematics of the $\ttbar$ system and the additional jets.
Keeping in mind that realistic $b$-jet observables consist of a combinatorial superposition of $b$-jets from $\ttbb$ production and from top decays, the fact that Monte Carlo truth acceptance cuts on top decays have only a minor effect on the production of $b$-jets suggests that the essential features observed in $pp\to \ttbb$ production, such as double-splitting effects, are expected to show up also in the presence of top decays.
In order to assess the importance of spin correlations, in we also compare spin-correlated top decays to isotropic decays generated by . At the level of reconstructed $\ttbb$ observables this comparison does not reveal any significant effect of spin correlations.
A more realistic analysis of $\ttbb$ production and decay is presented in , where $b$-jet and leptonic observables are defined at the level of the full final state, and two charged leptons and four $b$-jets within the acceptance cuts – are required, without any distinction between $\ttbb$ production and decay.
Comparing spin-correlated and isotropic top decays, in $b$-jet observables we find no significant deviation, and significant spin-correlation effects show up only in the azimuthal correlation of the two charged leptons.
In we also assess the relative impact of hadronisation and multi-parton interactions (MPI). It turns out that $b$-jet observables are very stable with respect to hadronisation, with differences between parton and hadron level that do not exceed the few percent level. The same holds for MPI effects.
The above results indicate that insights on the QCD dynamics of $\ttbb$ production gained through studies with stable top quarks at parton level should hold true also in the presence of top decays and hadronisation.
Summary and conclusions {#se:conclusions}
=======================
Searches for $\ttbar H$ production in the $H\to \bbbar$ channel call for a precise theoretical description of the irreducible $\ttbar+b$-jet background. To shed light on the QCD dynamics that governs this nontrivial multi-scale process, in the first part of this paper we have analysed the relative importance of the various mechanisms that lead to the radiation of $b$-quarks off $pp\to \ttbar$ events. To this end we have compared the role of $pp\to \ttbb$ topologies involving initial-state and final-state $g\to \bbbar$ splittings. Using a naive diagrammatic splitting, as well as gauge-invariant collinear approximations, we have demonstrated that the $\ttbar+b$-jet cross section is strongly dominated by $b$-jet production via final-state $g\to \bbbar$ splittings. This holds both for phase space regions with two or only one resolved $b$-jets. These findings support the usage of NLOPS generators based on $pp\to \ttbb$ matrix elements in the four-flavour scheme, while we have pointed out that $\ttbar+b$-jet predictions based on $\ttbar+$multi-jet merging rely very strongly on the parton-shower modelling of $g\to \bbbar$ splittings.
Motivated by these observations we have introduced a new $pp\to \ttbb$ generator in the 4F scheme. This tool is based on the framework, and all relevant matrix elements are computed with . When applied to a multi-scale process like $pp\to\ttbb$, the method can lead to subtle technical issues. In particular, we have pointed out that the FKS mappings which generate the recoil associated with the first emission can enhance the amplitude of the underlying $\ttbb$ Born process in a way that leads to anomalously large weights as compared to the behaviour expected from the factorisation of soft and collinear radiation. Fortunately, such anomalies arise only from events with finite transverse momenta and not in the soft and collinear limits. Moreover, the framework disposes of a mechanism that automatically attributes such events to the so-called finite remnant, where QCD radiation is handled as in fixed-order NLO calculations. This mechanism, which is controlled by the $\hbzd$ parameter in , plays an important role for the efficiency of event generation. Moreover, it permits to avoid artefacts that can result from the application of QCD factorisation and resummation far away from their validity domain.
We have discussed predictions of the new generator and theoretical uncertainties for various $\ttbar+b$-jet cross sections and distributions at the 13TeV LHC. At variance with previous studies, in order to provide a better picture of the perturbative convergence, we have evaluated QCD corrections using the same $\as$ value and the same PDFs at LO and NLO. The resulting NLO $K$-factors turn out to be close to two, even if the renormalisation scale is chosen in a way that is expected to absorb large logarithms associated with the running of $\as$. The question of the origin of such large higher-order effects and the search for possible remedies, such as improved scale choices, deserve to be addressed in future studies.
Scale uncertainties at fixed-order NLO amount to 25–30% and are dominated by renormalisation-scale variations. At NLOPS they tend to increase in a similar way as in , while in they tend to decrease [@deFlorian:2016spz]. However this behaviour may be an artefact of the incomplete implementation of scale variations in NLOPS generators.
Comparing predictions at NLO, LHE and LOPS level reveals significant shower effects at the level of 10% in the ttbb cross section and up to 30% or more for the invariant-mass and $\Delta R$ distributions of $b$-jet pairs. These effects can be attributed to double $g\to \bbbar$ splittings [@Cascioli:2013era] and are qualitatively and quantitatively consistent with the findings of . For the $p_\rT$-distribution of light-jet radiation, the predictions of the new generator are quite close to fixed-order NLO and also quite stable with respect to variations of the parameters $\hdamp$ and $\hbzd$, which separate real radiation into singular and finite parts. This good stability is guaranteed by the $\hbzd$-dependent mechanism mentioned above.
To assess pure shower uncertainties we have compared samples generated with 8 and 7. In addition, we have considered systematic uncertainties due to the modelling of $g\to \bbbar$ splittings and the choice of $\as$ in . At NLOPS, all shower uncertainties turn out to be rather small and clearly subleading with respect to QCD scale variations. As a further independent estimate of matching and shower uncertainties we have compared NLOPS $\ttbb$ generators based on and finding remarkable agreement both for $\ttbar+b$-jet cross sections and distributions. We have also shown that matching and shower uncertainties increase considerably if NLO corrections are not taken into account. The same holds for NLOPS generators of inclusive $\ttbar$ production as compared to $\ttbb$ generators.
Finally, we have presented predictions for $pp\to\ttbb$ with spin-correlated top decays. In this context we have show that hadronisation and MPI effects are almost negligible. Thus, the key features of the QCD dynamics of $\ttbb$ production at parton level are expected to hold true also at particle level after top decays.
The new $\ttbb$ generator will be made publicly available in the near future, and its application to experimental analyses may lead to significant steps forward in the understanding of the QCD dynamics of $\ttbar+b$-jet production and in the control of the theoretical uncertainties that plague $\ttbar H(\bbbar)$ searches.
TJ would like to thank Emanuele Bagnaschi, Silvia Ferrario Ravasio, Alessandro Vicini and especially Paolo Nason for enlightening discussions. This research was supported in part by the Swiss National Science Foundation (SNF) under contracts BSCGI0-157722, PP00P2-153027, and CRSII2-160814, as well as from the Research Executive Agency of the European Union under the Grant Agreement PITN–GA–2012–316704 ([*HiggsTools*]{}).
[^1]: Here one should keep in mind that the and implementations are not identical. For instance, only the latter guarantees an exact $\mathcal{O}(\alpha_S)$ description of soft radiation, which is achieved by upgrading the first shower emission and the related Catani-Seymour counterterm to full-colour accuracy.
[^2]: Preliminary results of this project have been presented in 2017 at the 25th International Workshop on Deep Inelastic Scattering and Related Topics [@talkDIS17] and QCD@LHC 2017 [@talkQCDatLHC17] conferences.
[^3]: This is due to the fact that the jet-resolution criterion employed for LO and NLO merging is exactly the same, while regions that are described at pure PS accuracy in LO merging, i.e. regions with $N>\Nmax$, can be either promoted to LOPS accuracy or remain at pure PS accuracy in the case of NLOPS merging (see above).
[^4]: Note that in the MC@NLO implementation of [@Alwall:2014hca] the scale $\mu_Q$ is also taken as starting scale for the showering of so-called hard events, which represent the MC@NLO counterpart of . Instead, in such events are showered starting from the actual $k_\rT$ of the first emission.
[^5]: The damping functions – are implemented in the [bornzerodamp]{} routine of and can be controlled through the flags named [withdamp]{} and [bornzerodamp]{}. The former activates the overall damping factor, $F_\damp\to 1$ if [withdamp]{} is set to 0. The same happens if [hdamp]{} is not explicitly set by the user. The remaining [bornzerodamp]{} flag controls the $\hbzd$-dependent theta function . By default [bornzerodamp]{}=[withdamp]{}, and $F_\bzd\to 1$ if [bornzerodamp]{} is set to 0.
[^6]: The $1/p_{\rT,\bbbar}$ dependence arises from the collinear singularity associated with the emission of the parent gluon in the topologies of type (d) in .
[^7]: This holds also for similar issues due to final-state radiation.
[^8]: More precisely, $\asfour$ is taken from the PDFs everywhere except for the evaluation of the Sudakov form factors in , where the corresponding implementation of $\asfour$ is used. In both implementations $\asfour(M_Z)=0.112$, which corresponds to $\asfive(M_Z)=0.118$.
[^9]: This choice does not coincide with the scale $\muf=\frac{1}{2}\sum_{i=t, \bar t} E_{T,i}$ adopted in [@Cascioli:2013era]. However, this difference has a rather minor impact on our predictions.
[^10]: The choice $\hdamp=1.5\,m_t$ corresponds to the default setting used for inclusive $\ttbar$ production in ATLAS.
[^11]: More precisely we use the 8.219 version and the specific A14 tune for NNPDFs, which is based on the NNPDF2.3 LO PDFs.
[^12]: We have checked that showering LHEs with the [NNPDF30\_nlo\_as\_0118\_nf\_4]{} (used for the NLO hard cross section) or the [NNPDF30\_lo\_as\_0130\_nf\_4]{} instead of the 5F NNPDFs does not induce any significant change in our results.
[^13]: In the case of , the well known [hvq]{} generator [@Frixione:2007nw] is used.
[^14]: More precisely, in order to be consistent with the 2.1 benchmarks presented in we have used the shower recoil scheme proposed in , which was the default in 2.1. This corresponds to setting [CSS\_KIN\_SCHEME=0]{}, while [CSS\_KIN\_SCHEME=1]{}, which became the new default in 2.2, leads to slightly more significant differences with respect to the $\ttbar+b$-jet predictions of . More precisely, comparing against 2.2 with the new recoil scheme we observe differences at the level of 10% in the ttbb cross section and up to about 40% in the light-jet $p_\rT$ spectrum. Such differences are well consistent with QCD scale variations. For comparison they are three times smaller with respect to the differences between and in . Note that [CSS\_KIN\_SCHEME]{} acts only on the second and subsequent shower emissions, it does not affect the SMC@NLO matching procedure.
[^15]: For processes like $\ttbb$ production, whose LO cross section scales with $\as^4$, evaluating the $K$-factor with LO inputs for the LO cross section results is a very strong dependence on the LO value of $\as$. The latter can depend very strongly on the employed PDF set. In particular, the two existing NNPDF 4F LO sets, which correspond to $\as(M_Z)=0.118$ and $\as(M_Z)=0.130$, can result in factor 1.5 ambiguity in the $\ttbb$ $K$-factor. We also note that the local $K$-factor in the matching formula, i.e. the $\bar B/B$ ratio in , is computed using NLO input throughout.
[^16]: We note in passing that for other observables, especially in the ttb phase space, results in this paper can deviate more significantly from the predictions of . This can be attributed to differences in the settings and, most importantly, to the fact that the generator used in was based on 5F $\ttbb$ matrix elements with $m_b=0$ and made use of technical generation cuts in order to avoid collinear singularities from $g\to \bbbar$ splittings. This limitation has now been overcome by upgrading the $\ttbb$ generator to the 4F scheme [@Bevilacqua:2017cru].
[^17]: Note that QCD scale uncertainties are not shown here.
[^18]: Since top quarks carry colour charge, a separation of production and decay is only possible at parton level, while at hadron level QCD radiation from production and decay is merged via colour reconnection.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'An $n$-simplex is called circumscriptible (or edge-incentric) if there is a sphere tangent to all its $n(n+1)/2$ edges. We obtain a closed formula for the radius of the circumscribed sphere of the circumscriptible $n$-simplex, and also prove a double inequality involving the circumradius and the edge-inradius of such simplices. Among this inequality settles affirmatively a part of a problem posed by the authors.'
address:
- 'Department of Mathematics, Zhejiang Xinchang High School, Shaoxing, Zhejiang 312500, People’s Republic of China'
- 'Department of Mathematics, Shili Senior High School in Zixing, Chenzhou, Hunan 423400, People’s Republic of China'
author:
- 'Yu-Dong Wu'
- 'Zhi-Hua Zhang'
title: 'On the circumradius of a special class of $n$-simplices'
---
Introduction
============
A (non-degenerate) $n$-simplex $\Omega = [A_0,A_1, . . .,A_{n}], n
\geq 1$, is defined as the convex hull of $n + 1$ affinely independent points (or position vectors) $A_0,A_1, . . .,A_n$ in Euclidean $n$-space. The points $A_0,A_1, . . .,A_n$ are the vertices of $\Omega$, and the line segments $a_{ij}$ joining two different vertices $A_i$ and $A_j$ are its edges.
Every $n$-simplex has a circumscribed sphere passing through its $n+1$ vertices and an inscribed sphere tangent to each of its $n+1$ facets. For the circumradius $R$ and the inradius $r$, we have the celebrated Euler’s inequality as follows $$\label{eq001}
R\geq nr.$$
An $n$-simplex is circumscriptible (or edge-incentric) if there is a sphere tangent to all its $n(n+1)/2$ edges. Considering such a simplex, we call this the edge-tangent sphere of the $n$-simplex, and note $\rho$ as the edge-inradius of this sphere. Of course, not every $n$-simplex $(n\geq 3)$ has an edge-tangent sphere. However, we have the following sufficient and necessary condition given by Lin and Zhu [@lzc01] (see also Hajja [@hajja01 p. 242, Theorem 4.1]).
\[t01\] The $n$-simplex $\Omega$ has an edge-tangent sphere if and only if there exist [(]{}i.e., so-called the balloon radii[)]{} $x_{i}>0$ with $0\leq i\leq n$ satisfying $a_{ij}=x_{i}+x_{j}$ for $0\leq i<
j\leq n$ or $$x_i=\frac{1}{n(n-1)}\left(n\sum_{\substack{i=0\\(i\neq j)}}^{n}a_{ij}-\sum_{0\leq i<j\leq n}a_{ij}\right).$$
In 2006, Hajja [@hajja01] derived many geometrical properties of the circumscriptible $n$-simplex. He also proved a closed formula involving the edge-inradius (i.e., the radius of the edge-tangent sphere):
\[t02\][([@hajja01 p. 249, Theorm 7.2 (d)])]{} The edge-inradius $\rho$ of $\Omega$ is given by $$\label{02}
{\rho}^2=\frac{2(n-1)}{\left(\sum_{i=0}^{n}\limits
{\frac{1}{x_{i}}}\right)^2-(n-1)\sum_{i=0}^{n}\limits{\frac{1}{x_{i}^2}}}.$$
The original for Theorem \[t02\] is based on the following generalized formula of the edge-inradius of a circumscriptible $n$-simplex in terms of its edge-lengths given by Ivanoff [@ivan01], Lin and Zhu [@lzc01].
\[t03\] Given a circumscriptible $n$-simplex $\Omega$, we have $${\rho}^2=-\frac{|A|}{2|A_1|},$$ where $$A=\left(\begin{array}{cccc}
-2x_{0}^2 & 2x_{0}x_{1} &\cdots & 2x_{0}x_{n}\\
2x_{0}x_{1} & -2x_{1}^2 &\cdots & 2x_{1}x_{n}\\
\cdots &\cdots &\cdots &\cdots \\
2x_{0}x_{n} & 2x_{1}x_{n} &\cdots & -2x_{n}^2\\
\end{array}\right),$$ and $$A_{1}=\left(\begin{array}{cccc}
0 & 1 & \cdots & 1\\
1 & \cdot & \cdots & \cdot\\
\vdots & \vdots & A & \vdots\\
1 & \cdot & \cdots & \cdot\\
\end{array}\right).$$
For the circumradius $R$ of a circumscriptible $n$-simplex $\Omega$, in 2007, an interesting problem stems naturally from the above investigations and the following Theorem \[t04\] by Hajja [[@hajja01 p. 261]]{}: Finding a closed formula for the circumradius of a circumscriptible $n$-simplex in terms of its balloon radii $x_i$ with $0\leq i\leq n$ as similarly .
[(see [@ivan01] and also [@wx01])]{}\[t04\] For a circumscriptible $n$-simplex $\Omega$, then we have $${R}^2=-\frac{|D|}{2|D_1|},$$ where $$D=\begin{pmatrix}
0 & (x_{1}+x_{0})^2& (x_{2}+x_{0})^2 &\cdots & (x_{n}+x_{0})^2\\
(x_{0}+x_{1})^2 & 0 & (x_{2}+x_{1})^2 &\cdots & (x_{n}+x_{1})^2\\
(x_{0}+x_{2})^2 & (x_{1}+x_{2})^2& 0 & \cdots & (x_{n}+x_{2})^2\\
\vdots &\vdots &\vdots & &\vdots\\
(x_{0}+x_{n})^2 & (x_{1}+x_{n})^2 &(x_{2}+x_{n})^2& \cdots & 0\\
\end{pmatrix},$$ and $$D_1=\left(\begin{array}{cccc}
0 & 1 & \cdots & 1\\
1 & \cdot & \cdots & \cdot\\
\vdots & \vdots & D & \vdots\\
1 & \cdot & \cdots & \cdot\\
\end{array}\right).$$
A double inequality for the radius of the circumscriptible tetrahedron $\Omega=A_{0}A_{1}A_{2}A_{3}A_{4}$ and sharpening Euler’s inequality is proved in [@lz01] and [@wz01] $$\label{01}
R\geq \sqrt{3}\rho\geq 3r.$$
As a generalization of inequality , Wu and Zhang [@wz01] posed an analogous problem for the circumscriptible $n$-simplex.
\[c02\] In a circumscriptible $n$-simplex $\Omega$, prove or disprove that $$\label{03}
R\geq \sqrt{\frac{2n}{n-1}}\rho \geq nr.$$
Recently, Wu et al. [@wzw01] proved the right hand of double inequality .
In this paper, we will give a closed formula for the circumradius of a circumscriptible $n$-simplex in terms of its balloon radii $x_i$ with $0\leq i\leq n$ as similarly , and settle the left hand of double inequality affirmatively.
Main Results
============
\[t05\] The radius $R$ of the circumscribed sphere of a circumscriptible $n$-simplex $\Omega$ is given by $$\label{341}
\left(\frac{R}{\rho}\right)^2=\frac{[MP-(n-1)(n-3)]^2-[M^2-(n-1)N][P^2-(n-1)Q]}{16(n-1)^2},$$ and $$\label{34}
R^2=\frac{[MP-(n-1)(n-3)]^2-[M^2-(n-1)N][P^2-(n-1)Q]}{8(n-1)[P^2-(n-1)Q]},$$ where $$\label{zzh001}
M=\sum_{i=0}^{n}{x_{i}},~~ N=\sum_{i=0}^{n}{{x_{i}}
^{2}},~~P=\sum_{i=0}^{n}{\frac{1}{x_{i}}},~~\mbox{and}~~
Q=\sum_{i=0}^{n}{\frac{1}{x_{i}^{2}}}.$$
When $n=2$, then, in the triangle, we have the well known formula $$\begin{aligned}
R&=\frac{(x_0+x_1)(x_1+x_2)(x_2+x_0)}{4\sqrt{x_0x_1x_2(x_0+x_1+x_2)}}\\
&=\frac{a_{01}a_{12}a_{02}}{\sqrt{(a_{01}+a_{12}+a_{02})
(a_{01}+a_{12}-a_{02})(a_{12}+a_{02}-a_{01})(a_{01}+a_{02}-a_{12})}}.\end{aligned}$$
In 2007, Hajja [[@hajja01 p. 261]]{} said:“the questions regarding non-regular edge-incentric d-simplices in which the circumcenter and the incenter coincide were not considered. We expect these questions to be rather difficult, since we were unable to find a closed formula for the circumradius of an edge-incentric d-simplex in terms of its balloon radii. Such a formula in the form of a quotient of two determinants is given in [[@ivan01]]{}".
For the given formula , in our private communication, Hajja also said:“I am really very impressed that you succeeded in finding a closed formula for $R$ — I have tried to do so last year but never was able to. Actually, I asked a colleague in Germany for help but he failed too".
\[t06\] For circumscriptible $n$-simplex $\Omega$, we have $$\label{04}
0\leq R^2-\frac{2n}{n-1}\rho^2\leq (n+1)^2|OG|^2,$$ where $O$ and $G$ are the circumcenter and the centroid of the circumscriptible $n$-simplex $\Omega$, respectively.
The left hand of double inequality is just the left hand of double inequality .
Preliminary Results
===================
Throughout this section, let $A,~A_1,~D,~D_1$ and $x_{i}$ for $0\leq
i\leq n$ be defined by the above section, $V$ be the volume of $\Omega$, and $$B_{1}=
\begin{pmatrix}
x_{0}^2 &1\\
x_{1}^2 &1\\
x_{2}^2 &1\\
\vdots &\vdots\\
x_{n}^2 &1\\
\end{pmatrix},
B_{2}=
\begin{pmatrix}
1 & 1 & 1 &\cdots & 1 \\
x_{0}^2 & x_{1}^2 & x_{2}^2 &\cdots & x_{n}^2\\
\end{pmatrix}.$$
[([@wzw01 Lemma 1])]{}\[l02\]We have $$\label{27}
|A|=(-1)^n(n-1)2^{2n+1}\left(\prod_{i=0}^{n}x_{i}\right)^2,$$ and $$\label{281}
A^{-1}=
\begin{pmatrix}
\frac{2-n}{4n-4}\cdot\frac{1}{x_{0}^2} & \frac{1}{4n-4}\cdot\frac{1}{x_{1}x_{0}} & \frac{1}{4n-4}\cdot\frac{1}{x_{2}x_{0}} &\cdots & \frac{1}{4n-4}\cdot\frac{1}{x_{n}x_{0}}\\
\frac{1}{4n-4}\cdot\frac{1}{x_{0}x_{1}} & \frac{2-n}{4n-4}\cdot\frac{1}{x_{1}^2} & \frac{1}{4n-4}\cdot\frac{1}{x_{2}x_{1}} &\cdots & \frac{1}{4n-4}\cdot\frac{1}{x_{n}x_{1}}\\
\frac{1}{4n-4}\cdot\frac{1}{x_{0}x_{2}} & \frac{1}{4n-4}\cdot\frac{1}{x_{1}x_{2}} & \frac{2-n}{4n-4}\cdot\frac{1}{x_{2}^2} &\cdots & \frac{1}{4n-4}\cdot\frac{1}{x_{n}x_{2}}\\
\vdots &\vdots &\vdots & &\vdots\\
\frac{1}{4n-4}\cdot\frac{1}{x_{0}x_{n}} & \frac{1}{4n-4}\cdot\frac{1}{x_{1}x_{n}} & \frac{1}{4n-4}\cdot\frac{1}{x_{2}x_{n}} &\cdots & \frac{2-n}{4n-4}\cdot\frac{1}{x_{n}^2}\\
\end{pmatrix}.$$
It is clear that $$\begin{aligned}
|A|&=2^{n+1}\left(\prod_{i=0}^{n}x_{i}\right)^2
\begin{vmatrix}
-1 & 1 & 1 &\cdots & 1\\
1 & -1 & 1 &\cdots & 1\\
1 & 1 & -1 &\cdots & 1\\
\vdots &\vdots &\vdots & &\vdots\\
1 & 1 & 1 &\cdots & -1\\
\end{vmatrix}\end{aligned}$$$$\begin{aligned}
\\&=2^{n+1}\left(\prod_{i=0}^{n}x_{i}\right)^2
\begin{vmatrix}
-1 & 1 & 1 &\cdots & 1\\
2 & -2 & 0 &\cdots & 0\\
2 & 0 & -2 &\cdots & 0\\
\vdots &\vdots &\vdots & &\vdots\\
2 & 0 & 0 &\cdots & -2\\
\end{vmatrix}
\\
&=2^{n+1}\left(\prod_{i=0}^{n}x_{i}\right)^2
\begin{vmatrix}
n-1 & 1 & 1 &\cdots & 1\\
0 & -2 & 0 &\cdots & 0\\
0 & 0 & -2 &\cdots & 0\\
\vdots &\vdots &\vdots & &\vdots\\
0 & 0 & 0 &\cdots & -2\\
\end{vmatrix}
\\&=2^{n+1}\left(\prod_{i=0}^{n}x_{i}\right)^2\cdot(-1)^n(n-1)2^n
\\&=(-1)^n(n-1)2^{2n+1}\left(\prod_{i=0}^{n}x_{i}\right)^2.\end{aligned}$$
Let us to compute $A^{-1}$. We know that the adjoint of $A$ is $$A^{*}=
\begin{pmatrix}
A_{11} & A_{12} & \cdots & A_{1,n+1}\\
A_{21} & A_{22} & \cdots & A_{2,n+1}\\
\vdots & \vdots & \vdots & \vdots \\
A_{n+1,1} & A_{n+1,2} & \cdots & A_{n+1,n+1}\\
\end{pmatrix},$$ where $A_{ij}$ is the cofactor of the element $2x_{i-1}x_{j-1}(i\neq
j)$ or $-2x_{i-1}^2(i=j)$ of $A$.
From the process of computing $|A|$ above, for $0\leq j\leq n$, it is easily to obtain that $$A_{jj}=(-1)^{n-1}(n-2)2^{2n-1}\left(\prod_{\begin{subarray}{c}{i=0,}\\i\neq
j-1\end{subarray}}^{n}x_{i}\right)^2=\frac{2-n}{4(n-1)}\cdot\frac{1}{x_j^2}\cdot|A|.$$ Now, we compute $A_{ij}$ for $0\leq i<j\leq n$ because of $A_{ij}=A_{ji}$ with $A=A'$. That is $$\begin{aligned}
A_{ij}&=(-1)^{i+j}
\begin{vmatrix}
-2x_{0}^2 & 2x_{1}x_{0}& \cdots & 2x_{j-2}x_{0} &2x_{j}x_{0} &\cdots & 2x_{n}x_{0}\\
2x_{0}x_{1} & -2x_{1}^2 & \cdots & 2x_{j-2}x_{1} &2x_{j}x_{1} &\cdots & 2x_{n}x_{1}\\
\vdots &\vdots& & \vdots & \vdots & &\vdots\\
2x_{0}x_{i-2} & 2x_{1}x_{i-2}& \cdots &2x_{j-2}x_{i-2} &2x_{j}x_{i-2} &\cdots & 2x_{n}x_{i-2}\\
2x_{0}x_{i} & 2x_{1}x_{i}& \cdots & 2x_{j-2}x_{i} &2x_{j}x_{i} &\cdots & 2x_{n}x_{i}\\
\vdots &\vdots & & \vdots & \vdots & &\vdots\\
2x_{0}x_{n} & 2x_{1}x_{n}& \cdots & 2x_{j-2}x_{n} & 2x_{j}x_{n} &\cdots & -2x_{n}^2\\
\end{vmatrix}\\
&=(-1)^{i+j}2^{n}\frac{\left(\prod_{i=0}^{n}\limits
x_{i}\right)^2}{x_{i-1}x_{j-1}}
\begin{vmatrix}
-1 & 1 & 1& 1& \cdots&1\\
1 & -1 & 1& 1& \cdots&1\\
\vdots & \vdots & \ddots & & &\vdots\\
1& 1& & C_{1} & &1\\
\vdots&\vdots& & & \ddots &\vdots\\
1& 1& 1& 1& \cdots & -1\\
\end{vmatrix}
\\&
=(-1)^{i+j}2^{n}\frac{\left(\prod_{i=0}^{n}\limits
x_{i}\right)^2}{x_{i-1}x_{j-1}}
\begin{vmatrix}
-2 & & & & &\\
& -2 & & & &\\
& & \ddots & & &\\
& & & C_{2} & &\\
& & & & \ddots &\\
& & & & & -2\\
\end{vmatrix}
\\
&=(-1)^{i+j}2^{n}(-2)^{n-(j-i)}|C_{2}|\frac{\left(\prod_{i=0}^{n}\limits
x_{i}\right)^2}{x_{i-1}x_{j-1}}\\&
=(-1)^{i+j}2^{n}(-2)^{n-(j-i)}(-1)^{j-i+1}(-2)^{j-i-1}\frac{\left(\prod_{i=0}^{n}\limits x_{i}\right)^2}{x_{i-1}x_{j-1}}\\\end{aligned}$$ $$\begin{aligned}
&=(-1)^{2j+1}2^{n}(-2)^{n-1}\frac{\left(\prod_{i=0}^{n}\limits
x_{i}\right)^2}{x_{i-1}x_{j-1}}\\&
=(-1)^{n}2^{2n-1}\frac{\left(\prod_{i=0}^{n}\limits
x_{i}\right)^2}{x_{i-1}x_{j-1}}
=\frac{1}{4(n-1)}\cdot\frac{1}{x_{i-1}x_{j-1}}\cdot|A|,\end{aligned}$$ where $$C_{1}=
\begin{pmatrix}
1&-2&1&\cdots&1\\
1&1&-2&\cdots&1\\
\vdots&\vdots&\vdots&\ddots&\vdots\\
1&1&1&\cdots&-2\\
1&1&1&\cdots&1\\
\end{pmatrix}_{(j-i)\times(j-i)},$$ and $$C_{2}=
\begin{pmatrix}
0&-2&0&\cdots&0\\
0&0&-2&\cdots&0\\
\vdots&\vdots&\vdots&\ddots&\vdots\\
0&0&0&\cdots&-2\\
1&0&0&\cdots&0\\
\end{pmatrix}_{(j-i)\times(j-i)}.$$ It follows that is true from the above $A^*,A_{ij}$ and $$\label{28}
A^{-1}=\frac{1}{|A|}A^{*}.$$
Wu et al. [[@wzw01]]{} directly gave $A^{-1}$ without computing process, and we here give the complete proof of this lemma. And the proof of is more simple than Hajja [[@hajja01 p. 250]]{}.
\[l03\]For $x_i>0$ with $0\leq i\leq n$, $$\begin{aligned}
\begin{split}
|D|=(-1)^n&\frac{2^{2n-3}}{n-1}\left(\prod_{i=0}^{n}x_{i}\right)^2\\&\cdot\{[MP-(n-1)(n-3)]^2-[M^2-(n-1)N][P^2-(n-1)Q]\},
\end{split}\end{aligned}$$ where$M, N, P,$ and $Q$ are given by .
We shall give two proofs of Lemma \[l03\] as follows.
Denoting the $j-$th column by $P_j$ and the $j-$th row by $Q_j$, we perform the following operations on $|D|$:
1. We accession a new row $Q_1=\begin{pmatrix} 1& x_0^2 & x_1^2&
x_2^2 &\cdots & x_n^2
\end{pmatrix}.$
2. We subtract $Q_1$ from $Q_{j+1}$ for $j = 1,\cdots, n+1$.
3. We accession a new column $P_1=\begin{pmatrix} 1&0& x_0^2 & x_1^2&
x_2^2 &\cdots & x_n^2
\end{pmatrix}^{'}.$
4. We subtract $P_1$ from $P_{j+2}$ for $j = 1,\cdots, n+1$.
5. We divide $Q_{j+2}$ and $P_{j+2}$ by taking appropriate common factor $x_{j-1}$ for $j = 1,\cdots, n+1$.
6. We add $P_{j+3}$ from $P_3$ for $j = 1,\cdots, n$ and divide $P_3$ by taking a common factor $n-1$.
7. We subtract $P_3$ from $P_{j+3}$ for $j = 1,\cdots, n$.
8. We add $\frac{1}{4}\left(\frac{1}{n-1}P-\frac{1}{x_j}\right)Q_{j+3}$ from $Q_1$ and $\frac{1}{4}\left({x_j}-\frac{1}{n-1}M\right)Q_{j+3}$ from $Q_2$ for $j =1, 2,\cdots, n$, and also add $\frac{1}{2}\left(\frac{1}{n-1}P
-\frac{1}{x_0}\right)Q_{3}$ from $Q_1$ and $\frac{1}{2}\left({x_0}-\frac{1}{n-1}M\right)Q_{3}$ from $Q_2$.
It is clear to see that $$\begin{aligned}
|D|&=
\begin{vmatrix}
0 & (x_{1}+x_{0})^2& (x_{2}+x_{0})^2 &\cdots & (x_{n}+x_{0})^2\\
(x_{0}+x_{1})^2 & 0 & (x_{2}+x_{1})^2 &\cdots & (x_{n}+x_{1})^2\\
(x_{0}+x_{2})^2 & (x_{1}+x_{2})^2& 0 & \cdots & (x_{n}+x_{2})^2\\
\vdots &\vdots &\vdots & &\vdots\\
(x_{0}+x_{n})^2 & (x_{1}+x_{n})^2 &(x_{2}+x_{n})^2& \cdots & 0\\
\end{vmatrix}\\
&=
\begin{vmatrix}1& x_0^2 & x_1^2& x_2^2 &\cdots & x_n^2\\
0& 0 & (x_{1}+x_{0})^2& (x_{2}+x_{0})^2 &\cdots & (x_{n}+x_{0})^2\\
0& (x_{0}+x_{1})^2 & 0 & (x_{2}+x_{1})^2 &\cdots & (x_{n}+x_{1})^2\\
0& (x_{0}+x_{2})^2 & (x_{1}+x_{2})^2& 0 & \cdots & (x_{n}+x_{2})^2\\
\vdots & \vdots &\vdots &\vdots & &\vdots\\
0& (x_{0}+x_{n})^2 & (x_{1}+x_{n})^2 &(x_{2}+x_{n})^2& \cdots & 0\\
\end{vmatrix}\\
&=
\begin{vmatrix}1& x_0^2 & x_1^2& x_2^2 &\cdots & x_n^2\\
-1& -x_0^2 & x_{0}^2+2x_{0}x_{1}& x_{0}^2+2x_{0}x_{2} &\cdots & x_{0}^2+2x_{0}x_{n}\\
-1& x_{1}^2+2x_{0}x_{1} & -x_1^2 & x_{1}^2+2x_{1}x_{2} &\cdots & x_{1}^2+2x_{1}x_{n}\\
-1& x_{2}^2+2x_{0}x_{2}& x_{2}^2+2x_{1}x_{2}& -x_2^2 & \cdots & x_{2}^2+2x_{2}x_{n}\\
\vdots & \vdots &\vdots &\vdots & &\vdots\\
-1& x_{n}^2+2x_{0}x_{n}& x_{n}^2+2x_{1}x_{n}&x_{n}^2+2x_{1}x_{n} & \cdots & -x_n^2\\
\end{vmatrix}\\
&=
\begin{vmatrix}1&0& 0 & 0& 0 &\cdots & 0\\
0&1& x_0^2 & x_1^2& x_2^2 &\cdots & x_n^2\\
x_0^2&-1& -x_0^2 & x_{0}^2+2x_{0}x_{1}& x_{0}^2+2x_{0}x_{2} &\cdots & x_{0}^2+2x_{0}x_{n}\\
x_1^2&-1& x_{1}^2+2x_{0}x_{1} & -x_1^2 & x_{1}^2+2x_{1}x_{2} &\cdots & x_{1}^2+2x_{1}x_{n}\\
\vdots &\vdots & \vdots &\vdots &\vdots & &\vdots\\
x_n^2&-1& x_{n}^2+2x_{0}x_{n}& x_{n}^2+2x_{1}x_{n}&x_{n}^2+2x_{1}x_{n} & \cdots & -x_n^2\\
\end{vmatrix}
\\&=
\begin{vmatrix}1&0& -1 & -1& -1 &\cdots & -1\\
0&1& x_0^2 & x_1^2& x_2^2 &\cdots & x_n^2\\
x_0^2&-1& -2x_0^2 & 2x_{0}x_{1}& 2x_{0}x_{2} &\cdots & 2x_{0}x_{n}\\
x_1^2&-1& 2x_{0}x_{1} & -2x_1^2 & 2x_{1}x_{2} &\cdots & 2x_{1}x_{n}\\
\vdots &\vdots & \vdots &\vdots &\vdots & &\vdots\\
x_n^2&-1& 2x_{0}x_{n}& 2x_{1}x_{n}&2x_{1}x_{n} & \cdots & -2x_n^2\\
\end{vmatrix}
\\
&=\prod_{i=0}^n x_i^2
\begin{vmatrix}1&0& -\frac{1}{x_0} & -\frac{1}{x_1}& -\frac{1}{x_2}&\cdots & -\frac{1}{x_n}\\
0&1& x_0 & x_1& x_2 &\cdots & x_n\\
x_0&-\frac{1}{x_0}& -2 & 2& 2 &\cdots & 2\\
x_1&-\frac{1}{x_1}& 2 & -2 & 2 &\cdots & 2\\
\vdots &\vdots & \vdots &\vdots &\vdots & &\vdots\\
x_n&-\frac{1}{x_n}& 2& 2&2 & \cdots & -2\\
\end{vmatrix}\\\end{aligned}$$ $$\begin{aligned}
&=(n-1)\prod_{i=0}^nx_i^2
\begin{vmatrix}1&0& -\frac{1}{n-1}P & -\frac{1}{x_1}& -\frac{1}{x_2}
&\cdots & -\frac{1}{x_n}\\
0&1&\frac{1}{n-1}M & x_1& x_2 &\cdots & x_n\\
x_0&-\frac{1}{x_0}& 2 & 2& 2 &\cdots & 2\\
x_1&-\frac{1}{x_1}& 2 & -2 & 2 &\cdots & 2\\
x_2&-\frac{1}{x_2}& 2& 2& -2 & \cdots & 2\\
\vdots &\vdots & \vdots &\vdots &\vdots & &\vdots\\
x_n&-\frac{1}{x_n}& 2& 2&2 & \cdots & -2\\
\end{vmatrix}\\
&=(n-1)\prod_{i=0}^nx_i^2
\begin{vmatrix}1&0& -\frac{1}{n-1}P &
\frac{1}{n-1}P-\frac{1}{x_1}&\cdots
& \frac{1}{n-1}P-\frac{1}{x_n}\\
0&1&\frac{1}{n-1}M & x_1-\frac{1}{n-1}M& \cdots
& x_n-\frac{1}{n-1}M\\
x_0&-\frac{1}{x_0}& 2 & 0 &\cdots & 0\\
x_1&-\frac{1}{x_1}& 2 & -4 &\cdots & 0\\
\vdots &\vdots & \vdots &\vdots & &\vdots\\
x_n&-\frac{1}{x_n}& 2& 0 & \cdots & -4\\
\end{vmatrix}\\
&=(n-1)\prod_{i=0}^nx_i^2
\begin{vmatrix}{\frac{1}{4(n-1)}}X_1
& {\frac{1}{4(n-1)}}X_2&0 & 0& 0&\cdots &0\\
{\frac{1}{4(n-1)}}X_3& {\frac{1}{4(n-1)}}X_1&0& 0& 0&\cdots
& 0\\ x_0&-\frac{1}{x_0}& 2 & 0& 0 &\cdots & 0\\
x_1&-\frac{1}{x_1}& 2 & -4 & 0 &\cdots & 0\\
x_2&-\frac{1}{x_2}& 2& 0& -4 & \cdots & 0\\
\vdots &\vdots & \vdots &\vdots &\vdots & &\vdots\\
x_n&-\frac{1}{x_n}& 2& 0&0 & \cdots & -4\\
\end{vmatrix}\\&
=(-1)^n\frac{2^{2n-3}}{n-1}\left(\prod_{i=0}^{n}x_{i}\right)^2\cdot(X_1^2-X_2X_3).\end{aligned}$$ where $$\begin{aligned}
\begin{split}\label{zzhe3320}
X_1&=MP-(n-1)(n-3),\\
X_2&=P^2-
(n-1)Q,\\
X_3&=M^2-(n-1)N, \end{split}\end{aligned}$$ and $M, N, P, Q$ are given by .
It is easily to find that $$\begin{aligned}
\label{20}\begin{split}
\begin{vmatrix}
A & B_{1}\\
-B_{2} & E_{2}\\
\end{vmatrix}
&=\begin{vmatrix}
A & B_{1}\\
0 & E_{2}+B_{2}A^{-1}B_{1}\\
\end{vmatrix}
=|A|\cdot|E_{2}+B_{2}A^{-1}B_{1}|,\end{split}\end{aligned}$$ and $$\begin{aligned}
\label{21}\begin{split}
\begin{vmatrix}
A & B_{1}\\
-B_{2} & E_{2}\\
\end{vmatrix}
&=\begin{vmatrix}
A+B_{1}B_{2} & B_{1}\\
0 & E_{2}\\
\end{vmatrix}
=|A+B_{1}B_{2}|,\end{split}\end{aligned}$$ where $E_{2}=\begin{pmatrix}1 &0 \\ 0 &1
\end{pmatrix}$.
By means of , we obtain $$\begin{aligned}
\label{zzh33331}\begin{split}
&E_{2}+B_{2}A^{-1}B_{1}=
\begin{pmatrix}
\frac{1}{4(n-1)}X_1 & \frac{1}{4(n-1)}X_2\\
\frac{1}{4(n-1)}X_3& \frac{1}{4(n-1)}X_1
\end{pmatrix}.\end{split}\end{aligned}$$
From and –, it is deduced that $$\begin{aligned}
|D|=&|A+B_{1}B_{2}|=|A|\cdot|E_{2}+B_{2}A^{-1}B_{1}|
\\=&(-1)^n\frac{2^{2n-3}}{n-1}\left(\prod_{i=0}^{n}x_{i}\right)^2(X_1^2-X_2X_3).\end{aligned}$$ where $X_1,X_2$ and $X_3$ are given by .
This evidently completes the proof of Lemma \[l03\].
[([@wx01 Corollary 2, p. 96])]{}\[l04\] For $n$-simplex $\Omega$, we have $$(n!)^2V^2R^2=-\det\left(-\frac{1}{2}a_{ij}^2\right)=\frac{(-1)^n|D|}{2^{n+1}}.$$
[([@lzc01 Corollary 1])]{}\[l05\] Given a circumscriptible $n$-simplex $\Omega$, we have $$\label{11}
(n!)^2V^2\rho^2=2^n(n-1)\left(\prod_{i=0}^{n}x_{i}\right)^2.$$
[([@wx01 (3.5.11), p. 112])]{}\[l06\] Let $O$ and $G$ are the circumcenter and the centroid of the $n$-simplex $\Omega$, respectively. Then we have $$|OG|^2=R^2-\frac{1}{(n+1)^2}\sum_{0\leq i<j\leq n}a_{ij}^2.$$
The Proof of Theorem \[t05\]
============================
This follow straightforwardly from Theorem **II** and Lemmas \[l03\]–\[l05\] by standard arguments.
The Proof of Theorem \[t06\]
============================
We will prove Theorem \[t06\] with two steps.
1. Firstly, we prove the left hand of inequality .
Let $M, N, P,$ and $Q$ are given by . By using the well-known power mean inequality and Cauchy inequality, then we have $N\geq \frac{M^2}{n+1},Q\geq \frac{P^2}{n+1},$ and $MP\geq (n+1)^2$. Further considering $P^2-(n-1)Q>0$ follows that $$\begin{aligned}
\label{zzh30}\begin{split}
[M^2&-(n-1)N][P^2-(n-1)Q]\\&\leq
\left[M^2-\frac{n-1}{n+1}M^2\right]\left[P^2-\frac{n-1}{n+1}P^2\right]=\frac{4}{(n+1)^2}M^2P^2.
\end{split}\end{aligned}$$ From , $1-\frac{4}{(n+1)^2}>0$ and the function $$y=\left[1-\frac{4}{(n+1)^2}\right]\left[x-\frac{(n-3)(n+1)^2}{n+3}\right]^2-\frac{4(n-1)(n-3)^2}{n+3}$$ is increasing on interval $\left(\frac{(n-3)(n+1)^2}{n+3},+\infty\right)$, and $MP\geq
(n+1)^2\geq \frac{(n-3)(n+1)^2}{n+3}$ for $n\geq 2$, we obtain $$\begin{aligned}
\label{zzh31}
&[MP-(n-1)(n-3)]^2-[M^2-(n-1)N][P^2-(n-1)Q]\nonumber\\ \geq
&[MP-(n-1)(n-3)]^2-\frac{4}{(n+1)^2}M^2P^2\nonumber\\
=&\left[1-\frac{4}{(n+1)^2}\right]\left[MP-\frac{(n-3)(n+1)^2}{n+3}\right]^2
-\frac{4(n-1)(n-3)^2}{n+3} \\ \geq
&\left[1-\frac{4}{(n+1)^2}\right]\left[(n+1)^2-\frac{(n-3)(n+1)^2}{n+3}\right]^2
-\frac{4(n-1)(n-3)^2}{n+3}\nonumber\\ =&32n(n-1).\nonumber\end{aligned}$$
According to and , we get $$\left(\frac{R}{\rho}\right)^2\geq\frac{32n(n-1)}{16(n-1)^2}=\frac{2n}{n-1}.$$ It is clear to show that the left of inequality holds.
2. Secondly, with Lemma \[l06\], the right hand of inequality is $$\label{05}
(n+1)^2R^2-\sum_{0\leq i<j\leq n}{a_{ij}^2}\geq
R^2-\frac{2n}{n-1}\rho^2.$$ For $n=3$, the required result is proved in [@lz01].
Now we prove that inequality holds when $n\geq 4$. Obviously, from Theorems \[t02\]–\[t05\], inequality is equivalent to $$\begin{aligned}
n(n+2)R^2&+\frac{2n}{n-1}\rho^2 \geq \sum_{0\leq i<j\leq
n}{a_{ij}^2},\end{aligned}$$ or $$\begin{aligned}
\frac{n(n+2)}{8(n-1)}&\cdot\frac{[MP-(n-1)(n-3)]^2-[M^2-(n-1)N][P^2-(n-1)Q]}{P^2-(n-1)Q}\nonumber\\
&+\frac{4n}{P^2-(n-1)Q}\geq M^2+(n-1)N,\end{aligned}$$ that is $$\begin{aligned}
\label{52}\begin{split}
&n(n+2)[MP-(n-1)(n-3)]^2+32n(n-1)\\
\geq &[(n^2+10n-8)M^2-(n-1)(n-2)(n-4)N][P^2-(n-1)Q],\end{split}\end{aligned}$$ where $M, N, P,$ and $Q$ are given by .
From the proof of Theorem \[t05\], we have $N\geq \frac{M^2}{n+1}$ and $Q\geq \frac{P^2}{n+1}$. Hence, in order to prove inequality , we only need to prove the following inequality $$\begin{aligned}
n(n+2)[MP-(n-1)(n-3)]^2+32n(n-1)\geq \frac{12n(3n-2)}{(n+1)^2}M^2P^2\end{aligned}$$ or $$\begin{aligned}
\begin{split}\label{54}
(n^2+5n-26)&\left[MP-\frac{(n+2)(n-3)(n+1)^2}{n^2+5n-26}\right]^2\\&-\frac{4(3n-10)^2(n+1)^4}{n^2+5n-26}\geq
0.\end{split}\end{aligned}$$ When $n\geq 4(n\in N)$, it’s clear that $(n+1)^2>\frac{(n+2)(n-3)(n+1)^2}{n^2+5n-26}$, and the function $$f(x)=(n^2+5n-26)\left[x-\frac{(n+2)(n-3)(n+1)^2}{n^2+5n-26}\right]^2-\frac{4(3n-10)^2(n+1)^4}{n^2+5n-26}$$ is increasing on interval $[(n+1)^2, +\infty)$. Thus, from $MP\geq
(n+1)^2$, we get $$\begin{aligned}
f(MP)\geq f((n+1)^2)=0.\end{aligned}$$ It is just as inequality . Further, inequality or holds.
The proof of Theorem \[t06\] is thus completed.
Acknowledgements
================
The authors would like to thank Professor M. Hajja for his kindly help in sending his valuable paper to them.
[7]{} M. Hajja, *Coincidences of Centers of Edge-Incentric, or Balloon, Simplices*, Results Math. **49** (2006), 237–263.
V. F. Ivanoff, *The circumradius of a simplex*. Math. Mag. **43** (1970), 71–72.
Z.-Ch. Lin, *The Tangent Sphere of the Edge in $n-$dimension Simplex*, Math. Practice Theory. **4** (1995), 90–93. (in Chinese)
Z.-Ch. Lin and H.-F. Zhu, *Research in the Tangent Sphere of the Edge in the Tetrahedron*. Geometric Inequalities in China. Jiangsu Educational Press, Jiangsu Province, China, 1996, 175–187. (in Chinese)
W.-X. Shen, *Introducition to Simplices*. Hunan Normal University Press, Changsha, Hunan Province, China, 2000. (in Chinese)
Y.-D. Wu and Zh.-H. Zhang, *The Edge-Tangent Sphere of a Circumscriptible Tetrahedron*, Forum Geom. **7** (2007), 19–24.
Y.-D. Wu, Zh.-H. Zhang and Zh.-G. Wang, *On a Special Class of $n$-Simplices*, Results Math. **54** (2009), 389–396.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We classify complete permutation monomials of degree $\frac{q^n-1}{q-1}+1$ over the finite field with $q^n$ elements, for $n+1$ a prime and $(n+1)^4 < q$. As a corollary, a conjecture by Wu, Li, Helleseth, and Zhang is proven. When $n+1$ is a power of the characteristic we provide some new examples. Indecomposable exceptional polynomials of degree $8$ and $9$ are also classified.'
author:
- 'D. Bartoli, M. Giulietti, L. Quoos, and G. Zini'
title: Complete permutation polynomials from exceptional polynomials
---
[**Keywords:**]{} Permutation polynomials, complete permutation polynomials, exceptional polynomials, bent-negabent boolean functions.
Introduction
============
Let $\mathbb{F}_{\ell}$ denote the finite field of order $\ell$ and characteristic $p$. A *permutation polynomial* (or PP) $f(x) \in \mathbb{F}_\ell[x]$ is a bijection of $\mathbb{F}_\ell$ onto itself. If $f(x)\in\mathbb F_{\ell}$ is a permutation polynomial over $\mathbb F_{\ell^m}$ for infinitely many $m$, then $f(x)$ is said to be an *exceptional polynomial* over $\mathbb F_{\ell}$. A polynomial $f(x) \in \mathbb{F}_\ell[x]$ is a *complete permutation polynomial* (or CPP) of $\mathbb F_{\ell}$ if both $f(x)$ and $f(x) + x$ are permutation polynomials of $\mathbb{F}_\ell$.
CPPs are also related to bent and negabent functions which are studied for a number of applications in cryptography, combinatorial designs, and coding theory; see for instance [@SGCG; @MN; @C; @ZQ2015]. The most studied class of CPPs is the monomial one. If there exists a complete permutation monomial of degree $d$ over $\mathbb F_\ell$, then $d$ is called a CPP exponent over $\mathbb F_\ell$. Complete permutation monomials have been investigated in a number of recent papers, especially for $\ell=q^n$ and $d=\frac{\ell-1}{q-1}+1$. Note that for an element $\alpha\in \mathbb{F}_\ell^*$, the monomial $\alpha x^d$ is a CPP of $\mathbb F_\ell$ if and only if $\gcd(d, \ell -1) = 1$ and $\alpha x^d +x$ is a PP of $\mathbb F_\ell$. In [@BZ; @BZ2015; @WLHZ; @WLHZ2015] PPs of type $f_b(x)=x^{\frac{q^n-1}{q-1}+1} +bx$ over $\mathbb{F}_{q^n}$ are thoroughly investigated for $n=2$, $n=3$, and $n=4$. For $n=6$, sufficient conditions for $f_b$ to be a PP of $\mathbb{F}_{q^6}$ are provided in [@WLHZ; @WLHZ2015] in the special cases of characteristic $p\in\{2,3,5\}$, whereas in [@BGZ] all $a$’s for which $ax^{\frac{q^6-1}{q-1}+1}$ is a CPP over $\mathbb{F}_{q^6}$ are explicitly listed. The case $p=n+1$ is dealt with in [@MZFG].
In this paper we discuss monomials of degree $d=\frac{q^n-1}{q-1}+1$ for general $n$. The starting point of our investigation is the observation that $b^{-1}x^d \in \mathbb F_{q^n}[x]$ is a CPP of $\mathbb F_{q^n}$ if and only if $b,b^q,\ldots,b^{q^{n-1}}$ are the roots of $$v_g(x)=\frac{g(-x)-g(0)}{-x} \in \mathbb F_q[x]$$ for some permutation polynomial $g(x)$ of degree $n+1$ over $\mathbb F_q$ such that the first-degree term is not zero. If for a root $b$ of $v_g(x)$ the monomial $b^{-1}x^d$ is a CPP over $\mathbb F_{q^n}$, then $g(x)$ will be called a [*good*]{} PP over $\mathbb F_q$; in this case, all roots of $v_g(x)$ have the same property. Clearly, a PP $g(x)$ over $\mathbb F_q$ is good if and only if the roots of $v_g(x)$ in the algebraic closure of $\mathbb F_q$ form a unique orbit under the action of the Frobenius map $x\mapsto x^q$.
Our aim is to classify good permutation polynomials over $\mathbb F_{q} $. Here we achieve this goal for all $n$, $(n+1)^4 <q$, with the exception of the cases $n+1=p^r$, with $r>1$, and $n+1=p^r(p^r-1)/2$, with $p\in\{2,3\}$. For $n+1=p^{r}$ we provide several examples. Proposition \[dickson\] shows that, if $q=p^k$ and $n+1$ is a prime different from $p$ satisfying $\gcd(n,k)=\gcd(n+1,p^2-1)=1$, then $d=\frac{q^n-1}{q-1}+1$ is a CPP exponent over $\mathbb F_{q^n}$. This solves a conjecture by Wu, Li, Helleseth, and Zhang, see [@WLHZ2015 Conjecture 4.18 and Proposition 4.19]. Our classification implies a result by Bhattacharya and Sarkar (see [@BS Theorem 1.1]) which determines the PPs of type $f_b$ when $p=2$, $n$ is a power of $2$, and $b\in\mathbb{F}_{q^2}$.
Note that since every permutation polynomial with degree less than $q^{1/4}$ is exceptional (see [@MP2013 Theorem 8.4.19]), condition $(n+1)^4<q$ allows us to consider only exceptional polynomials. A key tool in our investigation is the classification of indecomposable exceptional polynomials of degree different from $p^r$ for some $r>1$; see [@MP2013 Section 8.4].
If $g(x)$ is a good PP over $\mathbb F_q$ then it is easily seen that $c\cdot g(c'x)+e$ is a good PP over $\mathbb F_q$ for each $c,c',e\in \mathbb F_q$ with $cc'\neq 0$. In this paper two PPs $g(x)$ and $h(x)$ over $\mathbb F_q$ will be called *CPP-equivalent* if there exist $c,c',e\in \mathbb F_q$ with $cc'\neq 0$ such that $h(x)=c\cdot g(c'x)+e$. Note that for $g(x)$ a PP over $\mathbb F_q$ and $k\in\mathbb F_q$, the permutation polynomials $g(x+k)$ and $g(x)$ are equivalent in the usual sense but not CPP-equivalent; in fact, it’s possible that one of them is good but the other is not. Note that, when $g'(x)$ ranges over the CPP-equivalence class of $g(x)$, the roots of $v_{g'}(x)$ range over the roots of $v_g(x)$ and their multiples by non-zero elements in $\mathbb F_q$. We will consider only one polynomial in a CPP-equivalence class. In particular, we assume that $g(x)$ is monic and that $g(0)=0$. Since exceptional polynomials only exist for degrees coprime with $q-1$, when $n$ is odd we assume that $p=2$.
Our first result is that if $g$ is decomposable, that is $g$ is a composition of two exceptional polynomials with degree grater than one, then $g$ is not good; see Proposition \[Indecomponibili\].
If $g(x) \in \mathbb{F}_{q}[x]$ is a monic indecomposable exceptional polynomial of degree $n+1$ with $g(0)=0$, then, up to CPP-equivalence, one of the following holds [@MP2013 Section 8.4].
- $n+1$ is a prime different from $p$ not dividing $q-1$, and
1. $g(x)=(x+e)^{n+1}-e^{n+1}$, with $e\in \mathbb F_q$, or
2. $g(x)=D_{n+1}(x+e,a)-D_{n+1}(e,a)$, where $a,e\in \mathbb F_q$, $a\ne0$, $n+1\nmid q^2-1$, and $D_{n+1}(x,a)$ denotes a Dickson polynomial of degree $n+1$.
- $n+1=p$ and $g(x)=(x+e) ((x+e)^{\frac{p-1}{r}}-a)^r-e(e^{\frac{p-1}{r}}-a)^r$, with $r\mid p-1$, $a,e\in \mathbb F_q$, and $a^{r(q-1)/(p-1)}\neq 1$.
- $n+1=s(s-1)/2$, where $p\in \{2,3\}$, $q=p^m$, $s=p^r>3$, and $(r,2m)=1$.
- $n+1=p^r$ with $r>1$.
For the case $n+1=p^r$, $r>1$, Guralnick and Zieve conjectured in [@GZ] that there are no examples of indecomposable exceptional polynomials other than those described in [@MP2013 Propositions 8.4.15, 8.4.16, 8.4.17]. The paper is organized as follows. We classify good exceptional polynomials of type A) and B) in Sections \[section:A\] and \[section:B\]; see Theorems \[typeA\] and \[Caso\_primo\]. We show in Section \[section:C\] that certain exceptional polynomials of type C) are not good; see Proposition \[typeC\]. We describe in Section \[section:D\] some good exceptional polynomials of type D); see Propositions \[linearizzati\], \[lin2\], and \[lin3\]. Finally, we determine all the exceptional polynomials of degree $8$ and $9$ (see Propositions \[$n=7$\] and \[exceptional\]); in this way we provide a proof of the above mentioned Guralnick-Zieve conjecture for the special cases $n=7,8$. As a byproduct, we obtain all the CPPs with $n+1=8$ and $n+1=9$; see Corollaries \[coro1\] and \[coro2\] in Section \[section:89\].
Preliminaries
=============
Throughout the paper, $q$ will be a power $p^m$ of a prime $p$ and $\zeta_s$ will denote a $s$-th primitive root of unity, for $s \geq 1$. We begin this section by rephrasing a result by Wu, Li, Helleseth, and Zhang [@WLHZ2013]. For $b\in \mathbb F_{q^n}$, let $
A_i(b)\in \mathbb F_q
$ denote the evaluation of the $i$-th elementary symmetrical polynomial in $b,b^q,\ldots,b^{q^{n-1}}$, that is, $$A_i(b)=\sum_{0\le j_1<j_2<\ldots<j_i\le n-1}b^{q^{j_1}+q^{j_2}+\ldots+q^{j_i}}.$$ As a matter of notation let $A_0(b)=1$. Recall that $b,b^q,\ldots,b^{q^{n-1}}$ are the roots of the polynomial $$(-1)^nA_n(b)+(-1)^{n-1}A_{n-1}(b)T+\ldots+(-1)^{n-i}A_{n-i}(b)T^{i}+\ldots+T^n.$$ By [@WLHZ2013 Lemma 5] we have the following result.
\[CPPexceptional\] Assume that $(n+1)^4<q$. The monomial $b^{-1}x^{\frac{q^n-1}{q-1}+1}$ is a CPP over $\mathbb F_{q^n}$ if and only if $\gcd(n+1,q-1)=1$ and $\sum_{i=0}^n A_{n-i}(b)x^{i+1}$ is an exceptional polynomial over $\mathbb F_{q}$.
Let $g(x)=\sum_{i=0}^{n+1}\lambda_{n+1-i}x^i$ be an exceptional polynomial over $\mathbb F_q$, and assume that $\lambda_n\neq 0$ and $\lambda_0=1$. Consider the polynomial $$h_g(x)=\frac{g(x)-g(0)}{x}=\frac{\sum_{i=1}^{n+1} \lambda_{n+1-i}x^i}{x}=\sum_{i=0}^{n}\lambda_{n-i}x^i.$$ Then $v_g(x):=h_g(-x)=\sum_{i=0}^{n}(-1)^i\lambda_{n-i}x^i$. Note that if $n$ is even, then $h_g(-x)$ can be written as $\sum_{i=0}^{n}(-1)^{n-i}\lambda_{n-i}x^i$. If $n$ is odd, then $p=2$ and the same relation holds.
This means that, for any root $b$ of $v_g(x)$, the monomial $b^{-1}x^{\frac{q^n-1}{q-1}+1}$ is a CPP over $\mathbb F_{q^n}$ if and only if the roots of $v_g(x)$, or equivalently $h_g(-x)$, form a unique orbit under the Frobenius map $x\mapsto x^q$. This motivates the following definition.
An exceptional polynomial $g(x) \in \mathbb{F}_q[x]$ with $g(0)=0$ and $g'(0) \ne 0$ is said to be good if the roots of $\frac{g(-x)}{-x}$ form a unique orbit under the Frobenius map $x \mapsto x^q$.
Therefore, the following has been proved.
Assume that $(n+1)^4<q$. Then the elements $b\in\mathbb F_{q^n}\setminus \mathbb F_q$ such that $b^{-1}x^{\frac{q^n-1}{q-1}+1}$ is a CPP over $\mathbb F_{q^n}$ are the roots of polynomials $\frac{g(-x)}{-x}$, for $g$ ranging over good exceptional polynomials of degree $n+1$ over $\mathbb F_q$, with $g(0)=0$ and $g'(0) \ne 0$.
Note that $h_g(x)$ can be viewed as the bivariate polynomial $\frac{g(x)-g(y)}{x-y}$ evaluated at $y=0$. So, assume that we know the factorization of $\frac{g(x)-g(y)}{x-y}$ into absolutely irreducible factors defined over the algebraic closure of $\mathbb F_q$, say $\frac{g(x)-g(y)}{x-y}=\prod_{k=1}^s \ell_k(x,y)$. Then $$h_g(x)=\prod_{k=1}^s \ell_k(x,0).$$ Obviously, this can be extremely useful to establish whether an exceptional polynomial $g$ is good or not. Recall that an exceptional polynomial $g(t)$ is decomposable if there exist exceptional polynomials $g_1$, $g_2$ with degree greater than $1$ such that $g(x)=g_1(g_2(x))$.
\[Indecomponibili\] If $g(x)$ is a good exceptional polynomial, then $g(x)$ is not decomposable.
Suppose that $g(x)$ is decomposable and write $g(x)=g_1 (g_2(x))$, with polynomials $g_1,g_2$ such that $\deg(g_1),\deg(g_2)>1$. Then $$v_g(x)=\frac{g_1(g_2(-x))-g_1(g_2(0))}{-x}=\frac{g_2(-x)-g_2(0)}{-x}\lambda(g_2(-x)),$$ with $$\lambda(g_2(-x)) = \prod_{i=1}^{\deg(g_1)-1}(g_2(-x)-\beta_i)$$ for some $\beta_i\in\overline{\mathbb F}_q$. Since $\frac{g_2(-x)-g_2(0)}{-x}$ is a factor of positive degree defined over $\mathbb{F}_q$, the only possibility for the roots of $v_g(x)$ to form a unique orbit under the Frobenius map is that $v_g(x)$ is a power of $\frac{g_2(-x)-g_2(0)}{-x}$. Note that $0$ cannot be a root of $v_g(x)$, since for $b=0$ the monomial $bx^{\frac{q^n-1}{q-1}+1}$ is not a CPP. On the other hand, any root of a factor $g_2(-x)-\beta_i$ must be a root of $g_2(-x)-g_2(0)$, that is $\beta_i=g_2(0)$. Therefore, $$v_g(x)= \left(\frac{g_2(-x)-g_2(0)}{-x}\right)^{\deg(g_1)} (-x)^{\deg(g_1)-1},$$ which is impossible since $\deg(g_1)>1$.
CPPs from exceptional polynomials of type A) {#section:A}
============================================
Throughout this section we assume that $n+1\geq3$ is a prime different from $p$. We denote by $T_{q^{n/2}}$ the absolute trace map $\mathbb{F}_{q^{n/2}}\to\mathbb{F}_2$, $x\mapsto x+x^2+x^4+\ldots+x^{(q^{n/2})/2}$. We are going to prove the following result.
\[typeA\] Assume that $(n+1)^4<q$. For $i\in\{1,\ldots,n/2\}$ let $$\alpha_i=\zeta_{n+1}^i+\zeta_{n+1}^{-i} \text{ and } \beta_{i}=\zeta_{n+1}^i-\zeta_{n+1}^{-i}.$$ Then the monomial $b^{-1}x^{\frac{q^n-1}{q-1}+1}$ is a CPP of $\mathbb F_{q^n}$ precisely in the following cases.
- If $p\ne2$:
i) the order of $q$ modulo $n+1$ is $n$ and, up to multiplication by a non-zero element in $\mathbb F_q$, $b$ is as follows:
1. $b=\zeta_{n+1}^i-1$, for some $i \in \{1, \dots , n\}$;
2. for $n/2$ even, $b=e(\alpha_i-2)\pm\sqrt{\beta_i^2(e^2-4a)}$ for some $i \in \{1,\ldots n/2\}$, $a\in \mathbb F_q^*$, and $e \in \mathbb F_q$;
3. for $n/2$ odd, $b=e(\alpha_i-2)\pm\sqrt{\beta_i^2(e^2-4a)}$ for some $i \in \{1,\ldots n/2\}$, $a\in \mathbb F_q^*$, and $e\in\mathbb F_q$ such that $e^2-4a$ is a square in $\mathbb F_q$.
ii) the order of $q$ modulo $n+1$ is $n/2$, $n$ is not divisible by $4$, and, up to multiplication by a non-zero element in $\mathbb F_q$, $b=e(\alpha_i-2)\pm\sqrt{\beta_i^2(e^2-4a)}$ for some $i \in \{1,\ldots n/2\}$, $a\in \mathbb F_q^*$, and $e \in \mathbb F_q$ such that $e^2-4a$ is $0$ or a non-square in $\mathbb F_q$.
- If $p=2$:
i) the order of $q$ modulo $n+1$ is $n$ and, up to multiplication by a non-zero element in $\mathbb F_q$, $b=\zeta_{n+1}^i-1$ for some $i\in\{1,\ldots,n\}$;
ii) the order of $q$ modulo $n+1$ is $n$ or $n/2$, and $$b= z_i:= \varepsilon\delta_i^{2} + (\varepsilon+\varepsilon^2)\delta_i^{4} + \ldots + (\varepsilon+\varepsilon^2+\ldots+\varepsilon^{q^n/4})\delta_i^{q^n/2} \quad \textrm{or} \quad b=z_i+1,$$ where $\varepsilon\in\mathbb{F}_{q^n}$ satisfies $T_{q^{n/2}}(\varepsilon)=1$ and, for some $i\in\{1,\ldots,n\}$, $\delta_i=\frac{1}{\alpha_i}+\frac{a}{e^2}$ and $T_{q^{n/2}}(\delta_i)=1$.
By Propositions \[CPPexceptional\] and \[Indecomponibili\], the determination of the CPPs of type $b^{-1}x^{\frac{q^n-1}{q-1}+1}$ over $\mathbb F_{q^n}$ relies on the classification of indecomposable exceptional polynomials, which is given in [@MP2013 Section 8.4]. In particular, by [@MP2013 Theorem 8.4.11], Theorem \[typeA\] is implied by the results of Sections \[typeA1\] and \[typeA2\].
CPPs from exceptional polynomials of type A1) {#typeA1}
---------------------------------------------
Throughout this subsection we also assume that $n+1$ does not divide $q-1$. Note that for each $e\neq 0$ the polynomial $g(x)=(x+e)^{n+1}-e^{n+1}$ has a non-zero term of degree one. Also, the $n$ distinct roots of $
h_g(-x)=\frac{(-x+e)^{n+1}-e^{n+1}}{-x}
$ are $$-e(\zeta_{n+1}^i-1),\quad i=1,\ldots,n.$$
Assume that $e \in \mathbb F_q^*$. The polynomial $(x+e)^{n+1}-e^{n+1}$ is a good exceptional polynomial over $\mathbb F_q$ if and only if the order of $q$ modulo $n+1$ is equal to $n$.
The roots of $h_g(-x)$ form a unique orbit under the Frobenius map if and only if $\zeta_{n+1}$ does not belong to any proper subfield of $\mathbb F_{q^n}$. This is equivalent to the order of $q$ modulo $n+1$ being equal to $n$.
Assume that the order of $q$ modulo $n+1$ is equal to $n$. Then for $b=e(\zeta_{n+1}^i-1)$ the monomial $b^{-1}x^{\frac{q^n-1}{q-1}+1}$ is a CPP of $\mathbb F_{q^n}$, for each $e\in \mathbb F_q^*$ and $i\in\{1,\ldots,n\}$.
CPPs from exceptional polynomials of type A2) {#typeA2}
---------------------------------------------
Throughout this subsection we further assume that $n+1$ does not divide $q^2-1$. We begin by considering Dickson polynomials $D_{n+1}(x,a)\in\mathbb F_q[x]$. Recall that $$D_{n+1}(x,a)=\sum_{k=0}^{n/2}\frac{n+1}{n+1-k}\binom{n+1-k}{k}(-a)^kx^{n+1-2k}\,.$$
Note that $D_{n+1}(x,a)$ has a non-zero term of degree $1$, for each $a\neq 0$. In [@BZ1999 Theorems 7 and 8] Bhargava and Zieve provide the factorization of $\frac{D_{n+1}(x+e,a)-D_{n+1}(y+e,a)}{x-y}$, $e\in \mathbb{F}_q$.
\[dickson\] The polynomial $g(x)=D_{n+1}(x+e,a)-D_{n+1}(e,a)$, with $a,e\in\mathbb F_q$, $a\ne0$, and $D'_{n+1}(e,a)\ne 0$, is a good exceptional polynomial over $\mathbb{F}_{q}$ if and only if one of the following cases occurs:
i) $p\ne2$, $n/2$ is even and the order of $q$ modulo $n+1$ is $n$;
ii) $p\ne2$, $n/2$ is odd and either $e^2-4a$ is $0$ or a non-square in $\mathbb{F}_q$ and the order of $q$ modulo $n+1$ is $n/2$, or $e^2-4a$ is a square in $\mathbb{F}_q$ and the order of $q$ modulo $n+1$ is $n$;
iii) $p=2$, the order of $q$ modulo $n+1$ is $n$ or $n/2$, and $T_{q^{n/2}}(\delta_1)=1$, where $\delta_i=\frac{1}{\alpha_i}+\frac{a}{e^2}$.
In Cases *[i)]{} and *[ii)]{}, the roots of $h_{g}(-x)$ are $$b=-\frac{1}{2}\cdot\left(e(\alpha_i-2)\pm\sqrt{\beta_i^2(e^2-4a)}\right).$$ In Case *[iii)]{}, let $\varepsilon\in\mathbb{F}_{q^n}$ with $T_{q^{n/2}}(\varepsilon)=1$. Then the roots of $h_g(-x)$ are $$b=\varepsilon\delta_i^{2} + (\varepsilon+\varepsilon^2)\delta_i^{4} + \ldots + (\varepsilon+\varepsilon^2+\ldots+\varepsilon^{q^n/4})\delta_i^{q^n/2} \quad \textrm{and} \quad b+1.$$***
By [@BZ1999 Theorem 7] we have $$D_{n+1}(x+e,a)-D_{n+1}(y+e,a)=(x-y)\prod_{i=1}^{n/2}\left((x+e)^2-\alpha_i(x+e)(y+e)+(y+e)^2+\beta_i^2a\right),$$ where $\alpha_i=\zeta_{n+1}^i+\zeta_{n+1}^{-i}$ and $\beta_i=\zeta_{n+1}^i-\zeta_{n+1}^{-i}$. Then $$h_{g}(-x)=\frac{D_{n+1}(-x+e,a)-D_{n+1}(e,a)}{-x}=\prod_{i=1}^{n/2}\left((-x+e)^2-\alpha_ie(-x+e)+e^2+\beta_i^2a\right),$$ that is, since $\alpha_i^2=\beta_i^2+4$, $$h_{g}(-x)=\prod_{i=1}^{n/2}\left(x^2+xe(\alpha_i-2) +(\alpha_i-2)((\alpha_i+2)a-e^2) \right).$$
Note that the values $e(\alpha_i-2)$ are pairwise distinct for $i=1,\ldots,n$; hence, the sets of roots of two distinct quadratic factors of $h_{g}(-x)$ are disjoint.
Assume $p\ne2$. Since $\alpha_i^2=\beta_i^2+4$, the roots of $h_{g}(-x)$ are $$-\frac{1}{2}\cdot\left(e(\alpha_i-2)\pm\sqrt{\beta_i^2(e^2-4a)}\right).$$
Since $(\beta_i^2(e^2-4a))^{q^j}=\left(\beta_{iq^j \pmod {n+1}}\right)^2(e^2-4a)$, if the roots of $h_{g}(-x)$ form a unique orbit under the Frobenius map then the order $ord_{n+1}(q)$ of $q$ in $\mathbb{Z}_{n+1}^*$ must be either $n$ or $n/2$. Thus, we check when $\beta_i^2(e^2-4a)$ is a non-square in $\mathbb F_{q^{n/2}}$, so that the $(n/2)$-th power of the Frobenius map permutes the roots of $h_{g}(-x)$. Note that if $ord_{n+1}(q)=n/2$, then $\beta_i^{q^{n/2}}=\beta_i$ and therefore $\beta_i^2$ is a square in $\mathbb{F}_{q^{n/2}}$; if on the contrary $ord_{n+1}(q)=n$, then $\beta_i^{q^{n/2}}=-\beta_i$ and $\beta_i^2$ is a non-square in $\mathbb{F}_{q^{n/2}}$. Also, $n/2$ even implies that $(e^2-4a)$ is always a square in $\mathbb{F}_{q^{n/2}}$, whereas if $n/2$ is odd then $(e^2-4a)$ is a square in $\mathbb{F}_{q^{n/2}}$ if and only if it is a square in $\mathbb F_q$.
If $e^2-4a=0$, then $h_g(-x)$ is a square and its roots form a unique orbit under Frobenius. This completes the proof for $p\ne2$.
For $p=2$, similar computations using the solutions of quadratic equations in characteristic $2$ provide the claim.
CPPs from exceptional polynomials of type B) {#section:B}
============================================
Throughout this section we assume that $n+1=p$. For $p=2$, it is straightforward that there exist no exceptional polynomials of type B); hence, we assume that $p\ne2$. We denote by $\mathbb{N}_{\mathbb{F}_q/\mathbb{F}_p}$ the norm map $\mathbb{F}_q\rightarrow\mathbb{F}_p$, $x\mapsto x^{1+p+p^2+\cdots+q/p}$.
\[Caso\_primo\] Assume that $(n+1)^4<q$. The monomial $b^{-1}x^{\frac{q^n-1}{q-1}+1}$ is a CPP of $\mathbb F_{q^n}$ if and only if, for some divisor $r$ of $n$, one of the following cases occurs:
i) $b$ is an element of $\left\{-\zeta_r^i\alpha \ | \ i\in\{0,\ldots,r-1\}, \ \alpha^r=\zeta_{q-1}^j, \ \gcd(r,j)=1\right\}$, or
ii) $b$ is an element of $$\Big\{(v_0-\lambda u_0)^\frac{p-1}{r}-e \,\Big|\, \lambda \in \mathbb F_p^*, \ e,u_0^{p-1} \in \mathbb{F}_q^*, \ u_0^{\frac{(p-1)(q-1)}{r}}\neq 1,$$ $$\hspace{1.8 cm} \ v_0^\frac{p-1}{r}=e,\ ord\left(\mathbb{N}_{\mathbb{F}_q/\mathbb{F}_p}\left(\frac{u_0^{p-1}}{e^{r}}\right)\right)=p-1\Big\}.$$
Up to CPP-equivalence, the only indecomposable exceptional polynomials of degree $p$ over $\mathbb F_q$ are the polynomials $$g(x)=(x+e)\left((x+e)^r-a\right)^{k},$$ where $r$ is a divisor of $n$ and $k=n/r$, with $a,e \in \mathbb{F}_q$, $a^{\frac{q-1}{r}}\neq 1$; see [@MP2013 Theorem 8.4.14]. Hence, $$h_g(-x)=\frac{1}{-x}\Big((-x+e)\left((-x+e)^r-a\right)^{k}-e\left(e^r-a\right)^k\Big).$$ We distinguish a number of cases.
- $a=0$. In this case the polynomial $g(x)=(x+e)^p$ is not good.
- $e=0$ and $a\ne0$. We have that $h_g(-x)=((-x)^r-a)^k$ has $r$ distinct roots with multiplicity $k$, namely $-\zeta_r^i\alpha$, where $\alpha^r=a$ and $i=0, \dots, k-1$. They form a single orbit under the Frobenius map if and only if $x^r-a$ is irreducible over $\mathbb F_q$. By [@LN Theorem 3.75], this is equivalent to require that $a=\zeta_{q-1}^j$ with $\gcd(r,j)=1$.
- $e\neq0$ and $a\ne0$. Fix $u_0$, $v_0$ such that $u_0^{p-1}=a$ and $v_0^k=e$. It is straightforward to check that the set of roots of $h_g(-x)$ contains $$R=\left\{(v_0-\lambda u_0)^k-e\mid \lambda \in \mathbb F_p^*\right\}.$$ Note that $e^r\ne a$, since $a^{\frac{q-1}{r}}\neq 1= \left(e^r\right)^{\frac{q-1}{r}}$. We show that $R$ actually consists of the $p-1$ distinct roots of $h_g(-x)$. Assume on the contrary that $(v_0-\lambda u_0)^k-e=(v_0-\lambda' u_0)^k-e$ for some $\lambda\neq \lambda'$. Then $v_0-\lambda u_0=\mu(v_0-\lambda' u_0)$ for some $\mu$ with $\mu^k=1$, and hence $
v_0(1-\mu)=u_0(\lambda-\mu \lambda').
$ Since $k$ divides $p-1$, both $\mu$ and $\mu-1$ lies in $\mathbb F_p$. As $\lambda \neq \lambda'$ we have $\mu\neq 1$ and hence $
1=(v_0/u_0)^{p-1}=e^{\frac{p-1}{k}}/a=e^r/a,
$ a contradiction.
In the following we prove that the elements of $R$ are in the same orbit under the Frobenius map if and only if $$ord\left(\mathbb{N}_{\mathbb{F}_q/\mathbb{F}_p}\left(\frac{a}{e^{r}}\right)\right) =p-1.$$
Let $i\in\{1,\ldots,p-1\}$ be the smallest positive integer such that $((v_0-\lambda u_0)^k-e)^{q^i}=(v_0-\lambda u_0)^k-e$, so that the elements of $R$ are in the same orbit under the Frobenius map if and only if $i=p-1$.
Since $u_0^{q^i}=u_0a^{(q^i-1)/(p-1)}$ and $v_0^{q^i}=v_0e^{(q^i-1)/k}$, the condition $(v_0-\lambda u_0)^{kq^i}=(v_0-\lambda u_0)^k$ holds if and only if $$(v_0e^{(q^i-1)/k}-\lambda u_0a^{(q^i-1)/(p-1)})^{k}=(v_0-\lambda u_0)^k\,,$$ which is equivalent to $$\left(v_0-\lambda u_0\frac{a^{(q^i-1)/(p-1)}}{e^{(q^i-1)/k}}\right)^{k}=(v_0-\lambda u_0)^k\,,$$ that is, $$\left(v_0-\lambda u_0\frac{a^{(q^i-1)/(p-1)}}{e^{(q^i-1)/k}}\right)=\xi(v_0-\lambda u_0)\,,$$ where $\xi^k=1$. Suppose $\xi \neq 1$, then $$v_0/u_0=\lambda\frac{\frac{a^{(q^i-1)/(p-1)}}{e^{(q^i-1)/k}}-\xi}{1-\xi}\in \mathbb{F}_p^*\,,$$ and hence $(v_0/u_0)^{p-1}=1$; this implies $a=e^{(p-1)/k}=e^r$, impossible. This means $\xi=1$, that is $$\label{condizione}
\frac{a^{(q^i-1)/(p-1)}}{e^{(q^i-1)/k}}=1\,.$$ Since $$\frac{q^i-1}{p-1}\equiv \frac{i(q-1)}{p-1}\pmod{q-1} \quad\textrm{ and }\quad \frac{q^i-1}{k}\equiv \frac{i(q-1)}{k}\pmod{q-1}\,,$$ Equation is equivalent to $$\frac{a^{i(q-1)/(p-1)}}{e^{i(q-1)/k}}=1\,,$$ that is, $$\left(\mathbb{N}_{\mathbb{F}_q/\mathbb{F}_p}\left(\frac{a}{e^{(p-1)/k}}\right)\right)^i=1\,.$$ Therefore, $i=p-1$ if and only if $ord\left(\mathbb{N}_{\mathbb{F}_q/\mathbb{F}_p}\left(\frac{a}{e^{(p-1)/k}}\right)\right)=p-1$. The thesis follows.
CPPs from exceptional polynomials of type C) {#section:C}
============================================
In this section we deal with one of the three classes of exceptional polynomials of type C), namely the third class in [@MP2013 Theorem 8.4.12 with $e=1$].
\[typeC\] Let $p=3$, $s=p^r>3$, $\gcd(r,2m)=1$. The exceptional polynomial $$f_e(x)=(x+e)((x+e)^{2}-a)^{(s+1)/4}\left( \frac{((x+e)^{2}-a)^{(s-1)/2}+a^{(s-1)/2}}{(x+e)^{2}}\right)^{(s+1)/2},$$ where $a$ is a non-square in $\mathbb{F}_q^*$, is not good over $\mathbb{F}_q$.
Following [@Zieve1998 Prop. 2], consider $\tau(y)= (Ey+F)/(\overline{F}y +\overline{E})$, with $E, F,\overline{E},\overline{F} \in \mathbb{F}_{q^2}$ and $E\overline{E}-F\overline{F}=1$. The points $(x,y)$ of the curve with equation $\frac{f_0(x)-f_0(y)}{x-y}$ are exactly the points such that $x=\tau(y)$, where the choice of $(E, F)$ is unique up to replacing $(E,F)$ by $(-E,-F)$ and one of the following cases occurs:
- $F\overline{F}=-1/2$ ;
- $\left(EF\overline{E}\overline{F}\right)^{(q-1)/2}=-1$ and $F\overline{F}\neq-1/2$ ;
- $\left(EF\overline{E}\overline{F}\right)^{(q-1)/2}=1$ .
For $e\in \mathbb{F}_q$, a zero of $h_{f_{e}}(-x)=\frac{f_{e}(-x)-f_e(0)}{-x}$ corresponds to the point $(-x+e,e)$ of the curve defined by $\frac{f_0(x)-f_0(y)}{x-y}=0$, that is $x=y-\tau(y)$ for some $\tau$ as described above. Since $E, F, \overline{E},\overline{F}\in\mathbb{F}_{q^2}$, we have $\left(\tau(y)\right)^{q^2}=\tau(y)$ and $x^{q^2}=x$. Therefore, the roots of $h_{f_{e}}(-x)$ are not in a unique orbit under the Frobenius map.
CPPs from exceptional polynomials of type D) {#section:D}
============================================
Throughout this section we assume that $n+1=p^r$ with $r>1$. No complete classification of indecomposable exceptional polynomials of type D) is known. The following propositions deal with the cases related to linearized polynomials.
\[linearizzati\] Let $j,k\geq1$ and $H(x)\in\mathbb F_q[x]$ such that $L(x)=x^jH(x^k)$ is a linearized polynomial of degree $n+1$. For $e\in\mathbb F_q$ we have that $S_e(x)=(x+e)^jH^k(x+e)-e^jH^k(e)$ is a good exceptional polynomial over $\mathbb F_q$ if and only if the elements $e-(e_0-\ell)^k$ belong to a unique orbit under the Frobenius map, where $e_0$ is a fixed $k$-th root of $e$ and $\ell$ ranges over the roots of $L(x)\setminus\{0\}$.
Following [@Cohen1990 Theorem 2.1] we give the factorization of the curve defined by $S_0(x^k)-S_0(y^k)=0$. Let $N:=\deg(H)=\frac{(n+1)-j}{k}$ and write $$H(t)=\prod_{h=1}^{N}(t-\gamma_h),$$ where $\gamma_h \in \overline{\mathbb{F}}_q$. Then the roots of $H(t)$ and $L(x)=x^jH(x^k)$ are $\mathcal{H}=\{\gamma_h : h=1,\ldots,N\}$ and $\mathcal{L}=\{\zeta_k^i \gamma_h : i=0,\ldots,k-1, \ h=1,\ldots,N\}\cup \{0\},$ respectively.
Since $S_0(x^k)=(L(x))^k$, we have $$S_0(x^k)-S_0(y^k)=\left(L(x)\right)^k-\left(L(y)\right)^k=\prod _{i=0}^{k-1}\left(L(x)-\zeta_k^i L(y)\right)=\prod _{i=0}^{k-1}L(x-\zeta_k^i y)$$ $$=\left(x^k- y^k\right)^j\prod _{\alpha=0}^{d-1}\prod _{\beta=0}^{d-1}\prod _{h=1}^{N}\left(y-\zeta_k^\alpha x -\zeta_k^{\beta} \gamma_h\right).$$
Consider the curve $\mathcal{C}_S$ defined by $S_0(x)-S_0(y)=0$. Clearly, the points $(x,y)$ of $\mathcal{C}_S$ satisfy $\overline{y}=\zeta_k^{\alpha}\overline{x}+\zeta_k^{\beta}\gamma_h$, where $h\in\{1,\ldots,N\}$, $\alpha,\beta\in\{0,\ldots,k-1\}$, $\overline{x}^k=x$, $\overline{y}^k=y$.
Now consider the polynomial $h_{S_e}(-x)=\frac{S_e(-x)-S_e(0)}{-x}$. The zeros of $h_{S_e}(-x)$ correspond to the points $(-x+e,e)$ of $\mathcal{C}_S$, $x\neq 0$. Fix $e_0$ such that $e_0^k=e$; then the zeros of $h(-x)$ are $\left \{e-(e_0 -\ell)^k\ | \ \ell \in \mathcal{L}\setminus \{0\} \right\}$.
In general, it is not easy to establish when the elements $e-(e_0-\ell)^k$ belong to the same orbit under the Frobenius map. The following propositions provide two families of good exceptional polynomials arising from linearized polynomials.
\[lin2\] Let $q=p^m$ and $L(x)=x^{p^r}-\zeta_{q-1} x \in\mathbb F_q[x]$. If $r$ divides $m$, then $L(x)$ is good exceptional over $\mathbb F_q$.
Let $N=p^r-1$, and let $\eta\in\mathbb F_{q^{N}}$ be a root of $h_L (-x)=x^{N}-\zeta_{q-1}$. Then the roots of $h_L (-x)$ are $\{\lambda\eta \mid \lambda\in\mathbb F_{p^r}^*\}$. The hypothesis $r \mid m$ is equivalent to require that $N$ divides $(q-1)$, and this implies that $N(q-1)\mid q^N-1$. Hence we can choose $\eta=\omega^{\frac{q^N-1}{N(q-1)}}$, where $\omega$ is a primitive element of $\mathbb F_{q^N}$. The thesis is proved by showing that $\eta$ is not an element of any proper subfield of $\mathbb F_{q^{N}}$. Suppose that $\eta\in\mathbb F_{q^{k}}$ with $k\mid N$. Then $\omega^{\frac{q^N-1}{N(q-1)}(q^k-1)}=1$, that is $N\mid\frac{q^k-1}{q-1}$; since $q\equiv1\pmod N$, this is equivalent to $N\mid k$, and hence to $N=k$.
\[lin3\] If $\gcd(m,p^r-1)$ is a divisor of $r$, then there exists a linearized polynomial $L(x)\in\mathbb F_q[x]$ of degree $p^r$ which is good exceptional over $\mathbb F_q$.
Let $\epsilon=\gcd(m,p^r-1)$ and $\ell(x)\in\mathbb F_q[x]$ be a primitive polynomial of degree $r/\epsilon$ over $\mathbb F_{p^\epsilon}$, so that $\ell(x)$ is irreducible over $\mathbb F_{p^\epsilon}$ and has order $p^r-1$. Let $L(x)\in\mathbb F_q[x]$ be the linearized $p^\epsilon$-associate of $\ell(x)$. Then, by [@LN Theorem 3.63], the polynomial $L(x)/x$ is irreducible over $\mathbb F_{p^\epsilon}$. Let $\alpha$ be a non-zero root of $L(x)$. Then the field extension $\mathbb F_{p^\epsilon}(\alpha) : \mathbb F_{p^\epsilon}$ has degree $p^r-1$, while the extension $\mathbb F_q : \mathbb F_{p^\epsilon}$ has degree $m/\epsilon$. The field $\mathbb F_q(\alpha)$ is the compositum of $\mathbb F_q$ and $\mathbb F_{p^\epsilon}(\alpha)$; since $\gcd(m/\epsilon,p^r-1)=1$, we have that $[\mathbb F_q(\alpha):\mathbb F_q]=p^r-1$. Then $L(x)/x=h_L(-x)$ is irreducible over $\mathbb F_q$, and the thesis follows.
The cases $n+1=8$ and $n+1=9$ {#section:89}
=============================
The aim of this section is to study the cases $n+1=8$ (with $p=2$) and $n+1=9$ (with $p=3$). Since no complete classification of exceptional polynomials is known, we study the existence of good exceptional polynomials via algebraic curves associated to a PP. In fact, a polynomial $f(x)$ is a PP of $\mathbb F_q$ if and only if the algebraic curve $\mathcal C_{f}$ of degree $n$ with equation $$\frac{f(x)-f(y)}{x-y}=0$$ has no $\mathbb F_q$-rational points off the ideal line and the line $x=y$. For $q$ large enough with respect to the degree of $f(x)$, by the Hasse-Weil bound this is only possible when $\mathcal C_f$ splits into components not defined over $\mathbb F_q$; see for instance [@BGZ]. On the other hand, if $\mathcal C_f$ has no absolutely irreducible component defined over $\mathbb F_q$, with the only possible exception of the line $x=y$, then $f(x)$ is a permutation polynomial over an infinite number of extensions of $\mathbb F_q$, that is $f(x)$ is exceptional over $\mathbb F_q$; see [@C1970] and [@MP2013 Chapter 8.4].
$n+1=8$, $p=2$
--------------
\[$n=7$\] Let $q=2^{m}$, $n+1=8$. The polynomial $f(x)=x^8+\sum_{i=1}^{7} A_{i}x^{8-i}\in\mathbb F_q[x]$ is exceptional over $\mathbb F_q$ if and only if $A_1=A_2=A_3=A_5=0$ and the polynomial $g(x)=x^7+A_4x^3+A_6x+A_7$ has no roots in $\mathbb{F}_q^*$. Also $f(x)$ is good exceptional if and only if $g(x)$ is irreducible over $\mathbb{F}_q$.
The equation of the curve $\mathcal{C}_f$ reads $$\begin{aligned}
\label{Curva8}
& (x+y)^7+A_1(x^6+x^5y+x^4y^2+x^3y^3+x^2y^4+xy^5+y^6)+A_2(x^5+x^4y+x^3y^2+x^2y^3+xy^4+y^5)\\&+A_3(x^4+x^3y+x^2y^2+xy^3+y^4)
+A_4(x+y)^3+A_5(x^2+xy+y^2)+A_6(x+y)+A_7=0.\end{aligned}$$
Applying the Frobenius automorphism $t \mapsto t^q $ to the factors of $\mathcal{C}_f$ it is easy to conclude that if the curve $\mathcal{C}_f$ does not have absolutely irreducible components defined over $\mathbb{F}_q$, then the curve contains either two conics and three lines or seven lines. The unique ideal point of $\mathcal{C}_f$ is $(1:1:0)$. A line $\ell$ that is a component of the curve $\mathcal{C}_f$ has equation $\ell : y=x+\alpha$ and $$\left\{\begin{array}{l}
A_1=0\\
A_2\alpha+A_3=0\\
A_2\alpha^3+A_5=0\\
A_3\alpha^2+A_5=0\\
\alpha^7+A_2\alpha^5+A_3\alpha^4+A_4\alpha^3+A_5\alpha^2+A_6\alpha+A_7=0.
\end{array}\right.$$ If the line $\ell$ is not defined over $\mathbb{F}_q$ then $\alpha\in \overline{\mathbb{F}}_q\setminus \mathbb{F}_q$; this yields $A_2=A_3=A_5=0$, and the last equality becomes $\alpha^7+A_4\alpha^3+A_6\alpha+A_7=0$. It is easily seen that if $A_2=A_3=A_5=0$ then the curve $\mathcal{C}_f$ contains the seven lines $y+x+\alpha_i=0$, $i=1,\ldots,7$, where $\alpha_i^7+A_4\alpha_i^3+A_6\alpha_i+A_7=0$, and therefore $\mathcal C_f$ cannot split in two conics and three lines.
Thus, the only open case occurs when $\mathcal{C}_f$ splits in seven lines either not defined over $\mathbb{F}_q$ or equal to $x-y=0$.
Therefore $f(x)$ is exceptional if and only if $T^7+A_4T^3+A_6T+A_7$ has no roots in $\mathbb{F}_q$ and it is good exceptional if and only if $T^7+A_4T^3+A_6T+A_7$ is irreducible over $\mathbb{F}_q$ since all the roots must be in the same orbit.
Let $q=2^{m}$, $n+1=8$ and suppose $3$ divides $m$. The polynomial $x^8+A_7x$ is the only good exceptional polynomial over $\mathbb F_q$.
Since $3$ divides $m$ we have that $\zeta_7 \in \mathbb{F}_q$. From Cyclic extensions theory, $T^7+A_4T^3+A_6T+A_7$ is irreducible over $\mathbb{F}_q$ if and only if $A_4=A_6=0$. The thesis follows from Proposition \[$n=7$\].
The exceptional polynomials of Proposition [\[$n=7$\]]{} are linearized, and hence described in [[@MP2013 Prop. 8.4.15]]{}. Also, Proposition [\[$n=7$\]]{} confirms the conjecture [[@MP2013 Remark 8.4.18]]{} for the special case $n+1=8$.
\[coro1\] Assume that $q=2^r>8^4$. The monomial $b^{-1}x^{\frac{q^7-1}{q-1}+1}$ is a CPP of $\mathbb F_{q^8}$ if and only if $b$ is, up to a scalar multiple in $\mathbb{F}_q^*$, a root of some $F(x)=x^7+\alpha x^3+\beta x+\gamma \in \mathbb{F}_q[x]$, irreducible over $\mathbb{F}_q$.
$n+1=9$, $p=3$
--------------
\[exceptional\] Let $q=3^h$. The polynomial $$F(x)=x^9+A_1x^8+A_2x^7+A_3x^6+A_4x^5+A_5x^4+A_6x^3+A_7x^2+A_8x$$ is exceptional over $\mathbb{F}_q$ if and only if one of the following cases occurs.
i) $$\label{pol1} F(x)=x^9+A_3x^6+A_6x^3$$
and $T^6+A_3T^3+A_6\in\mathbb F_q[T]$ has no roots in $\mathbb{F}_q^*$;
ii) $$\label{pol2} F(x)=x^9+A_6x^3+A_8x$$
and $T^8+A_6T^2+A_8\in\mathbb F_q[T]$ has no roots in $\mathbb{F}_q^*$;
iii) $$F(x)=x^9+A_3x^6+A_4x^5+A_5x^4+\left(A_3^2 + A_3 \frac{A_5^3}{A_4^3} + \frac{A_5^2}{A_4}\right)x^3$$ $$\label{pol3}\qquad\quad+\left(2A_3 A_4 +2 \frac{A_5^3}{A_4^2}\right)x^2+\left(2A_3A_5 + A_4^2 +2 \frac{A_5^4}{A_4^3}\right)x,$$ where
1. $A_4\neq 0$,
2. the polynomial $T^4+2A_3T+2A_4\in\mathbb F_q[T]$ has no roots in $\mathbb F_{q}$;
iv) $$F(x)=x^9+A_2x^7+A_3x^6+A_5x^4+\left(A_2^3 + \frac{A_3 A_5}{A_2}\right)x^3+$$ $$\label{pol4}
\left(2A_2 A_5 +2\frac{A_3^3}{A_2}\right)x^2+\left(A_2^4+A_3A_5+\frac{A_5^2}{A_2}+\frac{A_3^4}{A_2^2}\right)x,$$ where $2A_2$ is not a square in $\mathbb{F}_q$.
The curve $\mathcal{C}_f$ associated to the polynomial $F(x)=x^9+\sum_{i=1}^8 A_{8-i}x^i$ is $$\sum_{i=0}^{8} A_{8-i}\frac{x^{i+1}-y^{i+1}}{x-y}=0,$$ where $A_i\in \mathbb{F}_q$, $i=1,\ldots,8$, $A_0=1$, and $A_8\neq 0$.
- Suppose that $\mathcal{C}_f$ contains a line $\ell$ with equation $\ell: y=x+\alpha$, where $\alpha=0$ or $\alpha\notin \mathbb{F}_q$. Then $\alpha $ satisfies: $$\left\{\begin{array}{l}
2A_1=0\\
2\alpha A_1+A_2=0\\
2 \alpha^2 A_1=0\\
\alpha^3A_1+2 \alpha^2 A_2 + A_4=0\\
2 \alpha^4 A_1+2\alpha^3 A_2 + 2 \alpha^2 A_3+\alpha A_4+A_5=0\\
\alpha^5 A_1 + \alpha^2 A_4=0\\
2\alpha^6 A_1 + \alpha^8 A_2 + \alpha^3 A_4 + \alpha^2 A_5 + 2A_7=0\\
\alpha^8 A_0+\alpha^7 A_1+ \alpha^6 A_2 + \alpha^5 A_3 + \alpha^4 A_4+\alpha^3 A_5 +\alpha^2 A_6 +\alpha A_7+ A_8=0
\end{array}\right..$$ This implies $A_1=A_2=0$. We distinguish two cases.
- $\alpha=0$. Then $A_4=A_5=A_7=A_8=0$. The curve becomes $$(x-y)^2((x-y)^6+A_3(x-y)^3+A_6)=0.$$ We have to require that the polynomial $T^6+A_3T^3+A_6$ has no roots in $\mathbb{F}_q^*$.
- $\alpha\neq 0$. Then $A_4=0$. The conditions above read $$A_5\alpha^2 + 2A_7=0, \quad \alpha^8 + A_3\alpha^5 + A_5 \alpha^3 + A_6 \alpha^2 + A_7\alpha + A_8=0,\quad 2A_3\alpha^2 + A_5=0.$$ If $A_5=0$ then $A_3=A_7=0$ and $\alpha^8 + A_6\alpha^2 + A_8 =0$ and the curve splits in $8$ lines. They are not defined over $\mathbb{F}_q$ or equal to $x-y=0$ if and only if the polynomial $T^8 + A_6T^2 + A_8 =0$ has no roots in $\mathbb{F}_q^*$.
If $A_3=0$ then $A_5=A_7=0$ and $\alpha^8 + A_6\alpha^2 + A_8 =0$, as above.
Suppose now $A_3,A_5\neq 0$. Then $A_5=A_3\alpha^2$, $A_7 =A_5^2/A_3$, and $A_8 = 2A_5A_6/A_3 + 2A_5^4/A_3^4$. Since $\alpha^2=A_5/A_3$, we have that $A_5/A_3$ is not a square in $\mathbb{F}_q$, otherwise the lines $y=x+\xi_1$ and $y=x+\xi_2$, where $\xi_i^2=A_5/A_3$, are $\mathbb{F}_q$-rational lines and the polynomial $F(x)$ is not exceptional. Let $a_3,a_5\in\mathbb F_{q^2}$ be such that $a_3^2=A_3$ and $a_5^2=A_5$. In this case the curve splits in $$(a_3x -a_3y + a_5)(a_3x -a_3y - a_5)\cdot$$ $$\cdot(a_3^6 (x-y)^6 + a_3^4 a_5^2 (x-y)^4 + a_3^8 (x+y)^3 + a_3^2 a_5^4 (x-y)^2 -a_3^6 a_5^2 (x+y) + a_3^6 A_6 + a_5^6)=0.$$
Since the sextic is defined over $\mathbb{F}_q$, it must split either in three conics or in two cubics. In the first case it is easily seen that all of them must be fixed by $\psi$.
If a conic of equation $(x-y)^2+\alpha(x+y)+\beta=0$ is contained in the sextic then in particular $A_3^2 =\alpha^3$ from which we get $a_3^{32} a_5^{12}=0$, impossible.
Suppose now that the sextic splits in two cubics.
If they are fixed by $\psi$ then they have equations $$(x-y)^3+\alpha_1 x^2+\alpha_2 x y+\alpha_1 y^2+\beta_1 x+\beta_1 y+\gamma_1=0$$ and $$(x-y)^3+\alpha_3 x^2+\alpha_4 x y+\alpha_3 y^2+\beta_2 x+\beta_2 y+\gamma_2=0.$$ Then in particular $\alpha_4=\alpha_3=-\alpha_1$, $\alpha_2=\alpha_1$. If $\alpha_1=0$ then $a_3^2 = \gamma_1 + \gamma_2$ and $a_3^2 = -\gamma_1 - \gamma_2$, which imply $a_3=0$, impossible. If $\alpha_1\neq 0$ then $\beta_1=\beta_2$ and again $a_3^2 = \gamma_1 + \gamma_2$ and $a_3^2 = -\gamma_1 - \gamma_2$, which imply $a_3=0$, impossible.
If they are switched by $\psi$ then they have equations $$(x-y)^3+\alpha_1 x^2+\alpha_2 x y+\alpha_3 y^2+\beta_1 x+\beta_2 y+\gamma_1=0$$ and $$\lambda((y-x)^3+\alpha_3 x^2+\alpha_2 x y+\alpha_1 y^2+\beta_2 x+\beta_1 y+\gamma_1)=0.$$ Then in particular $\lambda= -a_3^6$, $\alpha_3=\alpha_1$. If $\alpha_2=-\alpha_1$, then $a_3^2 \alpha_1^2 + a_3^2 \beta_1 -a_3^2\beta_2 -a_5^2=0$ and $\alpha_1=0$, which implies $a_3=0$, impossible. If $\alpha_2=\alpha_1$, then $\alpha_1(\beta_1+\beta_2)=0$. In both the cases $a_3=0$, impossible.
- Suppose that $\mathcal{C}_f$ splits in four absolutely irreducible conics. There are three distinct possibilities, depending on the number of components fixed by $\psi$.
1. All the conics are fixed by $\psi$. In this case the four conics are defined by $$\label{theconics} \mathcal{C}_i : (x-y)^2+\alpha_i(x+y)+\beta_i=0,$$ for $i=1,2,3,4$. This gives immediately $A_1=A_2=0$.
The condition $A_4= 0$ implies $A_5=A_7=0$ and $A_3A_8=0$, that is the polynomial is either of type or .
Suppose $A_4\neq 0$. Then, by direct computation, $A_6 = A_3^2 + A_3 A_5^3/A_4^3 + A_5^2/A_4$, $A_7 = 2A_3 A_4 +2 A_5^3/A_4^2$, $A_8 = 2A_3A_5 + A_4^2 +2 A_5^4/A_4^3$; also, the $\alpha_i$’s are roots of $\ell_1(x)=x^4+2A_3x +2 A_4$, and $\beta_i= \alpha_i^2+A_5/A_4\alpha_i.$ On the other hand if all these conditions are satisfied then the curve splits in the four conics defined in . Finally, the four conics are not defined over $\mathbb{F}_q$ if and only if the polynomial $T^4+2A_3T +2 A_4$ has no roots in $\mathbb{F}_q$.
2. Two conics are fixed by $\psi$ and two are switched. We can assume $$\mathcal{C}_1 : (x-y)^2+\alpha_1 (x+y)+\beta_1=0, \quad \mathcal{C}_2 : (x-y)^2+\alpha_2 (x+y)+\beta_2=0,$$ $$\mathcal{C}_3 : (x-y)^2+\alpha_3 x+\alpha_4 y+\beta_3=0, \quad \mathcal{C}_4 : (x-y)^2+\alpha_4 x+\alpha_3 y+\beta_3=0.$$ By direct compuation, we get immediately $A_1=A_2=A_4=A_5=A_7=0$ and $A_3A_8=0$, and hence $F(x)$ is of type or .
3. No conic is fixed by $\psi$. We can assume $$\mathcal{C}_1 : (x-y)^2+\alpha_1 x+\alpha_2 y+\beta_1=0, \quad \mathcal{C}_2 : (x-y)^2+\alpha_2 x+\alpha_1y+\beta_1=0,$$ $$\mathcal{C}_3 : (x-y)^2+\alpha_3 x+\alpha_4y+\beta_2=0, \quad \mathcal{C}_4 : (x-y)^2+\alpha_4 x+\alpha_3y+\beta_2=0.$$ Also in this case we get $A_1=A_2=A_4=A_5=A_7=0$ and $A_3A_8=0$, and hence $F(x)$ is of type or .
- Suppose that $\mathcal{C}_f$ splits in two absolutely irreducible quartics $\mathcal Q_1$ and $\mathcal Q_2$. The automorphism $\psi$ can either switch or fix the two components.
In the former case, $\mathcal Q_1$ and $\mathcal Q_2$ have the form $$\mathcal{Q}_1 : (x-y)^4+\alpha_1x^3+\alpha_2x^2y+\alpha_3xy^2+\alpha_4y^3+\beta_1x^2+\beta_2xy+\beta_3y^2+\gamma_1x+\gamma_2y+\delta=0,$$ $$\mathcal{Q}_2 : (x-y)^4+\alpha_4x^3+\alpha_3x^2y+\alpha_2xy^2+\alpha_1y^3+\beta_3x^2+\beta_2xy+\beta_1y^2+\gamma_2x+\gamma_1y+\delta=0.$$ We get $A_1=A_2=A_3=A_4=A_5=A_7=0$; hence, we have case .
In the latter case, $\mathcal Q_1$ and $\mathcal Q_2$ have the form $$\mathcal{Q}_1 : (x-y)^4+\alpha_1x^3+\alpha_2x^2y+\alpha_2xy^2+\alpha_1y^3+\beta_1x^2+\beta_2xy+\beta_1y^2+\gamma_1(x+y)+\delta_1=0,$$ $$\mathcal{Q}_2 : (x-y)^4+\alpha_3x^3+\alpha_4x^2y+\alpha_4xy^2+\alpha_3y^3+\beta_3x^2+\beta_4xy+\beta_3y^2+\gamma_2(x+y)+\delta_2=0.$$ Since $A_1=0$, we obtain $A_2A_4=0$.
- Suppose $A_2=0$ and $A_4\neq 0$. Then $A_6 = A_3^2 + A_3A_5^3/A_4^3 + A_5^2/A_4$, $A_7 = 2 A_3A_4 + 2A_5^3/A_4^2$, $A_8= 2A_3A_5 + A_4^2 +A_5^4/A_4^3$ and we have case .
- Suppose now $A_2\neq 0$ and $A_4= 0$. Then $A_6 = A_2^3+ A_3A_5/A_2$, $A_7 = 2A_2A_5 + 2A_3^3/A_2$, $A_8= A_2^4 + A_3A_5 + A_5^2/A_2 + A_3^4/A_2^2$. Also, $\alpha_1^2=2A_2$, $\alpha_2=\alpha_3=-\alpha_4=-\alpha_1$, $\beta_1=-\beta_3=2A_3/\alpha_1$, $\beta_2=-A_3/\alpha_1 -\alpha_1^2$, $\beta_4=A_3/\alpha_1 -\alpha_1^2$, $\gamma_1=A_3+\alpha_1^3$, $\gamma_2=A_3-\alpha_1^3$, $\delta_1 = A_3\alpha_1 + A_3^2/A_2 + 2A_5\alpha_1/A_2 + 2\alpha_1^6/A_2$, $\delta_2 = -A_3\alpha_1 + A_3^2/A_2 + A_5\alpha_1/A_2 + 2\alpha_1^6/A_2$. Note that $\alpha_i,\beta_i,\gamma_i,\delta_i$ are not defined over $\mathbb{F}_q$ if and only if $2A_2$ is not a square in $\mathbb{F}_q$. The quartics $\mathcal Q_1$ and $\mathcal Q_2$ read $$(x-y)^4+\alpha_1x^3+2\alpha_1x^2y+2\alpha_1xy^2+\alpha_1y^3+2A_3/\alpha_1 x^2+2(A_3/\alpha_1 +\alpha_1^2)xy$$ $$+2A_3/\alpha_1y^2+(A_3+\alpha_1^3)(x+y)+A_3\alpha_1 + A_3^2/A_2 + 2A_5\alpha_1/A_2 + 2\alpha_1^6/A_2=0,$$ $$(x-y)^4+2\alpha_1x^3+\alpha_1x^2y+\alpha_1xy^2+2\alpha_1y^3+A_3/\alpha_1x^2+(A_3/\alpha_1 +2\alpha_1^2)xy$$ $$+A_3/\alpha_1y^2+(A_3+2\alpha_1^3)(x+y)+2A_3\alpha_1 + A_3^2/A_2 + A_5\alpha_1/A_2 + 2\alpha_1^6/A_2=0,$$ and $\mathcal Q_1$ and $\mathcal Q_2$ are switched by the Frobenius map.
- Finally, $A_2=A_4= 0$ implies $A_5=A_7=0$ and $A_3A_8=0$. As above, this gives types or .
By direct computation, the exceptional polynomials of Proposition [\[exceptional\]]{} are equivalent to exceptional polynomials described in [[@MP2013 Prop. 8.4.15]]{}.
In fact, if $F(x)$ satisfies Case $i)$ or $ii)$, then $F(x)$ is a linearized polynomial.
If $F(x)$ satisfies Case $iii)$, then $F(x)=L_1\circ S\circ L_2(x)$, where $L_1(x)$ and $L_2(x)$ are linear, and $S(x)\in\mathbb F_q[x]$ has the form $x(a_2x^4+a_1x+a_0)^2$.
If $F(x)$ satisfies Case $iv)$, then $F(x)=L_1\circ S\circ L_2(x)$, where $L_1(x)$ and $L_2(x)$ are linear, and $S(x)\in\mathbb F_q[x]$ has the form $S(x)=x(a_2x^4+a_1x+a_0)^2$ when $A_2^2 A_5 + A_3^3\ne0$, or $S(x)=x(a_2x^2+a_0)^4$ when $A_2^2 A_5 + A_3^3=0$.
This confirms the conjecture [[@MP2013 Remark 8.4.18]]{} for the special case $n+1=9$.
\[goodexceptional\] Let $q=3^h$. The polynomial $$F(x)=x^9+A_1x^8+A_2x^7+A_3x^6+A_4x^5+A_5x^4+A_6x^3+A_7x^2+A_8x$$ is good exceptional over $\mathbb{F}_q$ if and only if one of the following cases occurs.
- $$F(x)=x^9+A_6x^3+A_8x$$ and $x^8+A_6x^2+A_8$ is irreducible over $\mathbb{F}_q$;
- $$F(x)=x^9+A_3x^6+A_4x^5+A_5x^4+\left(A_3^2 + A_3 \frac{A_5^3}{A_4^3} + \frac{A_5^2}{A_4}\right)x^3$$ $$\qquad\quad+\left(2A_3 A_4 +2 \frac{A_5^3}{A_4^2}\right)x^2+\left(2A_3A_5 + A_4^2 +2 \frac{A_5^4}{A_4^3}\right)x,$$ where
- $A_4\neq 0$,
- the polynomial $x^8+2A_3x^2+2A_4\in\mathbb F_q[x]$ has no roots in $\mathbb F_{q^4}$;
- $$F(x)=x^9+A_2x^7+A_3x^6+A_5x^4+\left(A_2^3 + \frac{A_3 A_5}{A_2}\right)x^3+$$ $$\left(2A_2 A_5 +2\frac{A_3^3}{A_2}\right)x^2+\left(A_2^4+A_3A_5+\frac{A_5^2}{A_2}+\frac{A_3^4}{A_2^2}\right)x,$$ where
- $2A_2$ is not a square in $\mathbb{F}_q$,
- $h(-x)=(x^4+2\alpha x^3+2A_3/\alpha x^2+2(A_3+2\alpha A_2)x+A_3\alpha + A_3^2/A_2 + 2A_5\alpha/A_2 + A_2^2)(x^4+\alpha x^3+A_3/\alpha x^2+2(A_3+\alpha A_2)x+2A_3\alpha + A_3^2/A_2 + A_5\alpha/A_2 + A_2^2)$, where $\alpha^2=2A_2$, is irreducible over $\mathbb{F}_q$.
We use the classification of exceptional polynomials of degree $9$ given in Proposition \[exceptional\].
- Let $F(x)$ be as in Case $i)$ of Proposition \[exceptional\]. Then $A_8=0$; hence, $F(x)$ is not good.
- Let $F(x)$ be as in Case $ii)$ of Proposition \[exceptional\]. Then $h_F(-x)=x^8+A_6x^2+A_8$; since $h_F(-x)$ cannot be a square in $\mathbb F_q[x]$, we have that $F(x)$ is good if and only if $h_F(-x)$ is irreducible over $\mathbb F_q$.
- Let $F(x)$ be as in Case $iii)$ of Proposition \[exceptional\]. The factors of $h_F(-x)$ are $x^2-\alpha_i x+\beta_i$, $i=1,\ldots,4$, where the $\alpha_i$’s are roots of $\ell_1(x)=x^4+2A_3x+2A_4$ and $\beta_i=\alpha_i^2+A_5/A_4\alpha_i$; hence, $\ell_1(x)$ must be irreducible over $\mathbb F_q$ in order for $F(x)$ to be good. Also, the roots of $h(-x)$ are $-\alpha_i\pm \sqrt{-A_5/A_4\alpha_i}$. Since $-A_5/A_4$ is an element of $\mathbb{F}_q$ and hence a square in $\mathbb{F}_{q^4}$, the roots of $h(-x)$ are in the same orbit under the Frobenius map if and only if $\alpha_i$ is not a square in $\mathbb{F}_{q^4}$, that is the polynomial $x^8+2A_3x^2 +2 A_4\in\mathbb F_q[x]$ has no roots in $\mathbb{F}_{q^4}$.
- Let $F(x)$ be as in Case $iv)$ of Proposition \[exceptional\]. The polynomial $h_F(-x)$ reads $$(x^4+2\alpha x^3+2A_3/\alpha x^2+2(A_3+2\alpha A_2)x+A_3\alpha + A_3^2/A_2 + 2A_5\alpha /A_2 + A_2^2)\cdot$$ $$\cdot(x^4+\alpha x^3+A_3/\alpha x^2+2(A_3+\alpha A_2)x+2A_3\alpha + A_3^2/A_2 + A_5\alpha /A_2 + A_2^2),$$ where $\alpha^2=2A_2$. Hence, the roots of $h_F(-x)$ are in a unique orbit under the Frobenius map if and only if $h_F(-x)$ is irreducible over $\mathbb F_q$.
We give two families of good exceptional polynomials arising from Proposition [\[goodexceptional\]]{}. Let $q=3^h$ with $h$ even, and $\eta$ be an odd number; by [[@LN Theorem 3.75]]{}, the polynomial $x^8+2\zeta_{q-1}^\eta\in\mathbb F_q[x]$ is irreducible over $\mathbb F_q$. Therefore, by Case $ \mathit{i)}$ in Proposition [\[goodexceptional\]]{}, the polynomial $F(x)=x^9+2\zeta_{q-1}^\eta x$ is good exceptional over $\mathbb F_q$. Also, by Case $\mathit{ii)}$ in Proposition [\[goodexceptional\]]{}, the polynomial $$F(x)=x^9+\zeta_{q-1}^\eta x^5 + ax^4 + \frac{a^2}{\zeta_{q-1}^\eta}x^3 + 2
\frac{a^3}{\zeta_{q-1}^{2\eta}}x^2 + \left(\zeta_{q-1}^{2\eta}+2\frac{a^4}{\zeta_{q-1}^{3\eta}}\right)x$$ is good exceptional over $\mathbb F_q$, for any $a\in\mathbb F_q$.
\[coro2\] Assume that $q=3^r>8^4$. The monomial $b^{-1}x^{\frac{q^8-1}{q-1}+1}$ is a CPP of $\mathbb F_{q^8}$ if and only if $b$ is, up to a scalar multiple in $\mathbb{F}_q^*$, a root of some $(F(-x+e)-F(e))/(-x)\in\mathbb F_q[x]$, where $e\in \mathbb{F}_q$ and $F(x)\in\mathbb F_q[x]$ satisfies Case $i)$, $ii)$, or $iii)$ in Proposition [\[goodexceptional\]]{}.
Acknowledgements
================
The research of D. Bartoli, M. Giulietti, and G. Zini was partially supported by Ministry for Education, University and Research of Italy (MIUR) (Project PRIN 2012 “Geometrie di Galois e strutture di incidenza” - Prot. N. 2012XZE22K$_-$005) and by the Italian National Group for Algebraic and Geometric Structures and their Applications (GNSAGA - INdAM).
The research of L. Quoos was partially supported by CNPq – Proc. 200434/2015-2. This work was done while L. Quoos was enjoying a sabbatical at the UniversitÃà degli Studi di Perugia.
[99]{}
D. Bartoli, M. Giulietti, G. Zini, *On monomial complete permutation polynomials*, Finite Fields Appl. [**41**]{} (3) (2016) 132–158.
L.A. Bassalygo, V.A. Zinoviev, *On one class of permutation polynomials over finite fields of characteristic two*, Mosc. Math. J. [**15**]{} (4) (2015) 703–713.
L.A. Bassalygo, V.A. Zinoviev, *Permutation and complete permutation polynomials*, Finite Fields Appl. [**33**]{} (2015) 198–211.
M. Bhargava, M. E. Zieve, *Factoring Dickson Polynomials over Finite Fields*, Finite Fields Appl. [**5**]{} (1999) 103–111.
S. Bhattacharya, S. Sarkar, *On some permutation binomials and trinomials over $\mathbb{F}_{2^n}$*, Des. Codes Cryptogr. [**82**]{} (2017) 149–160.
Y. Laigle-Chapuy, *Permutation polynomials and applications to coding theory*, Finite Fields Appl. [**13**]{} (1) (2007) 58–70.
S.D. Cohen, *The distribution of polynomials over finite fields*, Acta Arith. [**17**]{} (1970) 255–271.
S.D. Cohen, *Exceptional polynomials and the reducibility of substitution polynomials*, L’Enseignement Mathématique [**36**]{} (1990) 53–65.
S.D. Cohen, R.W. Matthews, *Exceptional polynomials over finite fields*, Finite Fields Appl. [**1**]{} (1995) 261–277.
R.M. Guralnick, M.E. Zieve, *Polynomials with ${\rm PSL}(2)$ monodromy*, Annals of Math. [**172**]{} (2010) 1321–1365.
G.L. Mullen, H. Niederreiter, *Dickson polynomials over finite fields and complete mappings*, Can. Math. Bull. [**30**]{} (1) (1987) 19–27.
R. Lidl, H. Niederreiter, *Introduction to Finite Fields and their Applications*, Cambridge University Press, 1986.
J. Ma, T. Zhang, T. Feng, G. Ge, *New results on permutation polynomials over finite fields*, arXiv:1506.05525.
G.L. Mullen, D. Panario, *Handbook of Finite Fields*, Chapman and Hall, 2013.
P. Stanica, S. Gangopadhyay, A. Chaturvedi, A.K. Gangopadhyay, S. Maitra, *Investigations on bent and negabent functions via the nega-Hadamard transform*, IEEE Trans. Inf. Theory [**58**]{} (6) (2012) 4064–4072.
G. Wu, N. Li, T. Helleseth, Y. Zhang, *Some classes of monomial complete permutation polynomials over finite fields of characteristic two*, Finite Fields Appl. [**28**]{} (2014) 148–165.
G. Wu, N. Li, T. Helleseth, Y. Zhang, *Some classes of complete permutation polynomials over $\mathbb F_q$*, Sci. China Math. [**58**]{} (10) (2015) 2081–2094.
G. Wu, N. Li, T. Helleseth, Y. Zhang, *More Classes of Complete Permutation Polynomials over $\mathbb F_q$*, arXiv:1312.4716.
M. Zieve, *Bivariate factorizations via Galois theory, with application to exceptional polynomials*, J. Algebra [**210**]{} (1998) 670–689.
Y. Zhou, L. Qu, *Constructions of negabent functions over finite fields*, Cryptography and Communications (2015) 1–16.
Daniele Bartoli and Massimo Giulietti are with\
Dipartimento di Matematica e Informatica\
Università degli Studi di Perugia\
Via Vanvitelli, 1 - 06123 Perugia - Italy\
[*emails:*]{} [daniele.bartoli,[email protected]]{}\
Giovanni Zini is with\
Dipartimento di Matematica e Informatica “Ulisse Dini”\
Università degli Studi di Firenze\
Viale Morgagni, 67/A - 50134 Firenze - Italy\
[*email:* ]{}[[email protected]]{}\
Luciane Quoos is with\
Instituto de Matemática\
Universidade Federal do Rio de Janeiro\
Rio de Janeiro 21941-909 - Brazil\
[*email:* ]{} [[email protected]]{}.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'A brief history of the study of fluctuations in high energy nuclear collisions at the CERN SPS performed by NA49 is presented. The ideas and the corresponding experimental data on fluctuations are discussed from the point of view of their sensitivity to the onset of deconfinement.'
address: |
University of Frankfurt, Germany and\
Swietokrzyska Academy, Kielce, Poland
author:
- Marek Gazdzicki
title: Fluctuations and NA49
---
.
Introduction
============
What are the phases of strongly interacting matter? This question and in particular, the hypothesis that at high energy densities the matter is in the form of quark-gluon plasma (QGP) [@qgp] rather than a gas of hadrons motivated the first stage of the broad experimental program of study of ultra-relativistic nucleus–nucleus collisions [@exp]. The results on the energy dependence of the mean hadron multiplicities and the shape of the transverse mass spectra obtained by the NA49 experiment at the CERN SPS indicate that the properties of the created matter change rapidly in the region of low SPS energies [@onset] ($\sqrt{s}_{NN} \approx$ 6–12 GeV). The observed effects confirm predictions of the transition from hadron gas to QGP [@gago] and thus indicate that in fact a new form of strongly interacting matter exists in nature at sufficiently high energy densities.
Do results on energy dependence of event-by-event fluctuations confirm the conjecture of the onset of deconfinement at low SPS energies? This report summarizes the current status of the NA49 data on this subject.
The paper starts with a brief historical review of the event-by-event physics in NA49 (Section 2). The experimental status of the basic ideas relating the onset of deconfinement and the fluctuations in nuclear collisions are presented and discussed in Section 3. A summary given in Section 4 closes this report.
A brief history of fluctuations in NA49
=======================================
The NA49 large acceptance spectrometer [@nim] was constructed with the aim to study event-by-event fluctuations. In a letter of intend [@loi] it was argued: “the focus on an analysis of macroscopic and microscopic observables, referring to the [*individual events*]{}, stems from our expectation that the phase transition, due to the statistical fluctuations in the collision dynamics, may not uniformly occur in the [*average*]{} central Pb+Pb collision”. The first data on central Pb+Pb collisions at 158$A$ GeV were taken during the NA49 running period 1994-1996 and soon after the pioneering analysis of event-by-event fluctuations started. It appeared to be more difficult than originally expected. In particular the problems were caused by:
- a need to control the experimental biases on event-by-event basis,
- a difficulty to remove the influence of trivial fluctuations in collision geometry,
- the absence of quantitative models predicting event-by-event fluctuations in case of the deconfinement phase transition.
Numerous measures of fluctuations were proposed, see e.g. Refs. [@measures], and their advantages and disadvantages were vividly discussed. Clearly a choice of the measure depends on the physics goals. It should maximize the sensitivity to the effects which we are looking for and minimize the sensitivity to the biasing contributions (experimental and physical).
Due to all these problems the first results on transverse momentum fluctuations [@Appelshauser:1999ft] and the fluctuations of the kaon to pion ratio [@Afanasev:2000fu] in central Pb+Pb collisions were published only in 1999 and 2000, respectively. The measured fluctuations (see Figs. \[fig\_mpt\] and \[fig\_kpi\]) are approximately reproduced by a “mixed event” model in which uncorrelated production of hadrons is assumed. They clearly demonstrate that possible fluctuations of the early stage energy density within the analyzed event sample, if existent, do not lead to significant changes of the event properties. All collisions are “similar”.
![\[fig\_kpi\] Distribution of the event-by-event kaon to pion ratio estimated using a maximum likelihood method (points) for central Pb+Pb collisions at 158$A$ GeV. As a reference, the same procedure was applied to a “mixed event” sample (histogram).](mpt_publi.eps){width="14pc"}
![\[fig\_kpi\] Distribution of the event-by-event kaon to pion ratio estimated using a maximum likelihood method (points) for central Pb+Pb collisions at 158$A$ GeV. As a reference, the same procedure was applied to a “mixed event” sample (histogram).](kpi_publi.eps){width="13pc"}
In 1997 and 1998 the data on C+C and Si+Si collisions at 158$A$ and 40$A$ GeV were registered. Together with the data on all inelastic Pb+Pb interactions at these energies, they allow to study system size dependence of event-by-event fluctuations. In particular, the transverse momentum [@kasia] and multiplicity [@maciek] fluctuations were analyzed. The data on mean $p_T$ fluctuations, see Fig. \[mpt\], indicate an increase of fluctuations for C+C, Si+Si and semi-peripheral Pb+Pb collisions. A similar effect is also observed at RHIC energies [@rhic]. It is still lacking a well established interpretation.
![\[mpt\] The system size dependence of transverse momentum fluctuations at 158$A$ GeV. The points correspond to the NA49 experimental data [@kasia], whereas the solid line to the predictions of the string-hadronic model Hijing [@hijing]. ](mpt.eps){width="20pc"}
The hypothesis that the onset of deconfinement is located at low SPS energies [@gago] motivated the energy scan program of the NA49 Collaboration [@add1]. Within this program data on central Pb+Pb collisions at 20$A$, 30$A$, 40$A$ and 80$A$ GeV were registered in runs in 1999, 2000 and 2002. The predicted [@gago] rapid changes in the energy dependence of hadron properties averaged over central Pb+Pb collisions, like pion and kaon mean multiplicities and transverse mass spectra, were observed [@onset]. The search for the expected anomalous behavior of other quantities is in progress. Among them event-by-event fluctuations are of particular importance. This is because they are expected to be sensitive to the physics properties of the system which are not accessible in the study of the mean quantities. In the following section we review current status of the ideas and the experimental results on the energy dependence of electric charge, multiplicity and strangeness fluctuations in the domain of the onset of deconfinement.
Fluctuations and the onset of deconfinement
===========================================
Net-electric charge fluctuations
--------------------------------
A simple estimate of the net electric charge fluctuations show that they are much smaller in a Quark-Gluon Plasma than in a hadron gas [@jeon]. This difference is caused by smaller charge units in the QGP (fractional charges of quarks) than in the hadron gas. Thus a decrease of the net-charge fluctuations is expected when the collision energy crosses the threshold for the deconfinement phase transition. A schematic sketch of this naive prediction is presented in Fig. \[nc\].
![\[nc\] A naive expectation for the energy dependence of the net-electric charge fluctuations in the region of the onset of deconfinement.](sketch_charge.eps){width="36pc"}
The experimental results on central Pb+Pb collisions at 20$A$, 30$A$, 40$A$, 80$A$ and 158$A$ GeV shown in Fig. \[nc\_data\] suggest only a very weak, if any, dependence on energy. They are close to those expected for a gas of pions correlated only by global charge conservation ($\Delta \Phi_q = 0$) and significantly above the naive prediction for the QGP $\Delta \Phi_Q \approx -0.5$ [@Zaranek:2001di].
![\[nc\_data\] The energy dependence of the net-charge fluctuations measured by $\Delta \Phi_Q$ [@Zaranek:2001di] in central Pb+Pb collisions for two different rapidity intervals $\Delta y = 1.2$ (left) and $\Delta y = 3$ (right). For more details see the original publication [@Alt:2004ir]. ](charge_energy.eps){width="36pc"}
These results do not necessarily exclude reduced fluctuations in the QGP because these can be masked by contributions from resonance decays. In fact a simple model of the QGP hadronization and resonance decay can quantitatively explain the magnitude of the measured fluctuations, see Fig. \[nc\_data1\].
![\[nc\_data1\] The dependence of $\Delta \Phi_q$ on the fraction of accepted particles in central Pb+Pb collisions at 20-158$A$ GeV. The prediction for the ideal QGP is indicated by the dashed curve (QGP), whereas the prediction for the QGP including hadronization and resonance decay is shown by the dotted curve (QGP+hadronization). For more details see the original publication [@Alt:2004ir]. ](charge_acc.eps){width="20pc"}
The influence of resonance decays on charge fluctuations depends on the size of the rapidity interval, $\Delta {\rm y}$, in which fluctuations are calculated. If $\Delta {\rm y}$ is much bigger than the typical distance in rapidity of the daughter particles, $\Delta {\rm y_{SMEAR}}$, the charge within the interval will not be changed by the decay and therefore the charge fluctuations should not be affected. On the other hand, if $\Delta {\rm y}$ is small, a large fraction of daughter particles will leave the interval and the initial net-charge will be significantly changed. The mean rapidity difference of two pions originating from decays of $\rho(770)$ meson is approximately one unit of rapidity. Therefore in order to minimize the decay effect $\Delta {\rm y}$ should be much larger than 1. However, this constraint is difficult to fulfill at SPS and lower energies because the width of the rapidity distribution of all produced particles, $\Delta {\rm y_{TOT}}$ is not much broader than 1. Hence a rapidity interval which is large enough to be unaffected by the resonance decays would contain almost all particles produced in a collision. The net-charge in this interval would then reflect the number of participant protons and the fluctuations would be determined by fluctuations of the collision centrality and not the particle production mechanism. A sketch which illustrates these considerations is shown in Fig. \[sketch\_smear\].
![\[sketch\_smear\] A sketch illustrating different rapidity scales important in the study of the net-charge fluctuations as a possible signal of QGP. $\Delta {\rm y}$ is the acceptance window in rapidity, $\Delta {\rm y_{SMEAR}}$ is a typical distance in rapidity of decay products of resonances and $\Delta {\rm y_{TOT}}$ is the width of the rapidity distribution of all particles. The observation of the predicted suppression of the net-charge fluctuations in QGP requires: $\Delta {\rm y_{SMEAR}} < \Delta {\rm y} < \Delta {\rm y_{TOT}}$. This condition is not fulfilled at the SPS energies where $\Delta {\rm y_{SMEAR}} \approx \Delta {\rm y} \approx \Delta {\rm y_{TOT}}
\approx 1$. ](sketch_smear.eps){width="36pc"}
At the SPS and AGS energies $\Delta y \approx \Delta y_{SMEAR} \approx \Delta y_{TOT} \approx 1 $ and consequently the net-charge fluctuations are not sensitive to the initial QGP fluctuations. Hence their energy dependence may be not affected by the onset of deconfinement.
Multiplicity fluctuations
-------------------------
It was suggested that properly filtered multiplicity fluctuations should be sensitive to the equation of state at the early stage of the collision and thus to its changes in the deconfinement phase transition region [@mfluct]. This idea follows from the expectation that the early stage energy density changes from collision to collision and thus it causes fluctuations of the thermodynamical parameters of the matter like temperature, pressure, entropy and strangeness. Obviously the relation between energy density fluctuations and fluctuations of the other thermodynamical parameters depends on the equation of state. This opens a possibility to study equation of state via properly defined analysis of event-by-event fluctuations. This basic idea is sketched in Fig. \[sketch\_temp\] for the case of the temperature fluctuations in the phase transition domain.
![\[sketch\_temp\] A sketch illustrating how the early stage energy density fluctuations lead to temperature fluctuations which depend on the equation of state. The solid line shows equation of state with the 1st order phase transition. The width of the vertical bands indicates early stage energy density fluctuations for three different collision energies, whereas the width of the horizontal bands gives the corresponding temperature fluctuations. ](sketch_temp.eps){width="36pc"}
As the entropy is closely related to the particle multiplicity, and it is expected to be approximately conserved during the evolution of the matter created at the early stage, the fluctuations of entropy are of primary interest. Based on the second principle of thermodynamics it can be shown [@mfluct] that the relative entropy fluctuations are given by: $$\frac {(\delta S)^2} {S^2} =
(1+ \frac {p} {\epsilon})^{-2} \frac {(\delta E)^2} {E^2},$$ where $S$ and $E$ denote entropy and energy, respectively, $p$ is pressure and $\epsilon$ energy density. For the experimental study the measure of multiplicity fluctuations, $R_e$ was proposed [@mfluct]: $$R_e = \frac {(\delta \bar n)^2/\bar n^2} {(\delta \bar E)^2/ \bar E^2},$$ where $\delta \bar n$ and $\delta \bar E$ are dynamical fluctuations of particle multiplicity and energy, respectively. They are related to the variance of particle multiplicity $V(n)$ as: $$V(n) =
(\delta \bar n)^2 + \langle (\delta n)^2 \rangle ,$$ where $\langle (\delta n)^2 \rangle$ are dynamically averaged statistical fluctuations which can be measured using the sub-event method [@volo]. The measure $R_e$ is directly related to the early stage equation of state [@mfluct]: $$R_e = (1+ \frac {p} {\epsilon})^{-2},$$ and therefore it should be sensitive to the onset of deconfinement. As it is defined as a ratio of relative multiplicity and energy fluctuations it is insensitive to the magnitude of the early stage energy density fluctuations. One should underline that the method assumes that the early stage entropy fluctuations are proportional to the dynamical multiplicity fluctuations, but it does not require knowledge of the value of the proportionality factor. This is a clear advantage in regard to the methods of studing the equation of state based on the analysis of the average quantities.
The energy dependence of $R_e$ calculated with the Statistical Model of the Early Stage (SMES) [@gago] is presented in Fig. \[shark\] [@mfluct].
![\[w\] The scaled variance for positively (+) and negatively (-) charged particles calculated within canonical and grand canonical ensembles as a function of the mean multiplicity [@m1]. The results corresponds to a system with net charge equal to zero. ](shark.eps){width="18pc"}
![\[w\] The scaled variance for positively (+) and negatively (-) charged particles calculated within canonical and grand canonical ensembles as a function of the mean multiplicity [@m1]. The results corresponds to a system with net charge equal to zero. ](w.eps){width="20pc"}
It shows a characteristic ’shark fin’ structure caused by large fluctuations in the mixed phase region.
This simple idea motivated further theoretical and experimental effort directed toward understanding of multiplicity fluctuations in nucleus-nucleus collisions at high energies.
The effect of conservation laws on the multiplicity fluctuations were studied within canonical and micro-canonical ensembles. This led to an unexpected observation [@m1] that the scaled variance of the multiplicity distribution, $w = V(n)/\langle n \rangle$, is different in various statistical ensembles (grand canonical, canonical and micro-canonical) even in the thermodynamical limit, $V \rightarrow \infty$. The simplest example is presented in Fig. \[w\], where the scaled variance for positively and negatively charged particles calculated within grand canonical and canonical ensembles is plotted as a function of mean particle multiplicity [@m1]. A detailed study of multiplicities fluctuations in different statistical ensembles is in progress [@m2; @m3].
An important contribution to multiplicity fluctuations may come from the fluctuations in the collision geometry. Clearly an interaction with a small number of participating nucleons will result in much lower particle multiplicity than a central collision in which almost all nucleons are involved in the reaction process.
![\[maciek\] The scaled variance of the multiplicity distribution for the negatively charged hadrons produced in Pb+Pb collisions (circles) and p+p interactions (triangle) at 158$A$ GeV as a function of the number of participants nucleons from the projectile [@maciek]. The solid line indicates predictions of a string-hadronic model Hijing [@hijing]. ](maciek.eps){width="20pc"}
Thus in order to extract the dynamical effects a precise control of the geometrical fluctuations is needed. In NA49 this is achieved by selecting events with fixed number of projectile participants, $N^{PROJ}_P$, measured by the forward VETO calorimeter [@veto]. The presented results are corrected for the resolution of the VETO calorimeter and the finite bin size in the measured forward energy. In contrary to the naive expectations [@hijing] it was found [@maciek] that the scaled variance of the multiplicity distribution increases with a decreasing centrality of the Pb+Pb collisions at 158$A$ GeV. The preliminary results[^1] are shown in Fig. \[maciek\]. The observed non-trivial centrality dependence focused the efforts on its tests and understanding and consequently the analysis of the energy dependence of multiplicity fluctuations was delayed.
Strangeness fluctuations
------------------------
Similar to entropy (multiplicity) fluctuations, strangeness fluctuations are also sensitive to the early stage energy density fluctuations and via these to the equation of state at the early stage of the collisions [@sfluct].
![\[sdata\] The energy dependence of the dynamical strangeness fluctuations in central Pb+Pb collisions at the SPS energies [@na49_s] and central Au+Au collisions at the top RHIC energy [@star_s]. ](sfluct.eps){width="17pc"}
![\[sdata\] The energy dependence of the dynamical strangeness fluctuations in central Pb+Pb collisions at the SPS energies [@na49_s] and central Au+Au collisions at the top RHIC energy [@star_s]. ](sdata.eps){width="18pc"}
The measure $R_{s/e}$ proposed to study them is defined as: $$R_{e/s} = \frac {(\delta \bar n_K)^2/\bar n_K^2}
{(\delta \bar n)^2/ \bar n^2},$$ where $\delta \bar n_K$ and $\delta \bar n$ are dynamical fluctuations of the multiplicity of kaons and pions, respectively. The energy dependence of the measure $R_{s/e}$ calculated within the SMES model [@gago; @sfluct] is shown in Fig. \[sfluct\]. The observed rapid change of the behavior is caused by the onset of deconfinement. The first results of NA49 [@na49_s] and STAR [@star_s] concerning strangeness fluctuations in central Pb+Pb collisions at the SPS energies are presented in Fig. \[sdata\]. The data seem to also suggest a change of energy dependence in SPS energies. However, a direct comparison between the data and the model predictions is not possible as the used experimental measure of fluctuations is different than $R_{s/e}$.
Summary
=======
The basic ideas concerning fluctuation signals of the onset of deconfinement at the SPS energies were reviewed. The status of their experimental tests based on the NA49 data was presented.
The naively expected significant suppression of the net-charge fluctuations is not observed. This may be due to dilution of the initial charge fluctuations by resonance decays.
The study of energy dependence of multiplicity fluctuations is in progress. It focuses on the search for the “shark fin” structure expected at the onset of deconfinement. In this respect two unexpected observations were made recently. Theory: the scaled variance of multiplicity distribution calculated within various statistical ensembles (micro-canonical, canonical and grand canonical) was found to be different even in the thermodynamical limit. Experiment: the scaled variance of multiplicity distribution for Pb+Pb collisions at 158$A$ GeV increases significantly when going from central to semi-peripheral collisions.
The preliminary results on the dynamical fluctuations of the kaon to pion ratio in central Pb+Pb collisions indicate a change of behavior at the SPS energies. The relation of this interesting effect to the onset of deconfinement is still unclear. This is because a proper comparison between the data and the model predictions is still missing.
I would like to thank the organizers of the Workshop on Correlations and Fluctuations in Relativistic Nuclear Collisions (MIT, April 23-25, 2005) for the interesting and inspiring meeting. This work was supported by the Virtual Institute VI-146 of Helmholtz Gemeinschaft, Germany.
[9]{}
J. C. Collins and M. J. Perry, Phys. Rev. Lett. [**34**]{}, 1353 (1975). for a recent review see: Proceedings of the 17th International Conference on Ultra-Relativistic Nucleus-Nucleus Collisions, Quark Matter 2004, Oakland, California, USA, 11–17 January 2004, editors H. G. Ritter and X.-N. Wang, J. Phys. [**G30**]{}, S633 (2004).
S. V. Afanasiev [*et al.*]{} \[The NA49 Collaboration\], Phys. Rev. C [**66**]{}, 054902 (2002) \[arXiv:nucl-ex/0205002\],\
M. Gazdzicki [*et al.*]{} \[NA49 Collaboration\], J. Phys. G [**30**]{}, S701 (2004) \[arXiv:nucl-ex/0403023\]. M. Gazdzicki and M. I. Gorenstein, Acta Phys. Polon. B [**30**]{}, 2705 (1999) \[arXiv:hep-ph/9803462\],\
M. I. Gorenstein, M. Gazdzicki and K. A. Bugaev, Phys. Lett. B [**567**]{}, 175 (2003) \[arXiv:hep-ph/0303041\]. S. Afanasev [*et al.*]{} \[NA49 Collaboration\], Nucl. Instrum. Meth. A [**430**]{}, 210 (1999). A. Karabarbounis [*et al.*]{}, CERN/SPSC 89-73, SPSC/I-173 (1989)
M. Gazdzicki and S. Mrowczynski, Z. Phys. C [**54**]{}, 127 (1992), M. Gazdzicki, Eur. Phys. J. C [**8**]{}, 131 (1999) \[arXiv:nucl-th/9712050\], L. Stodolsky, Phys. Rev. Lett. [**75**]{}, 1044 (1995), E. V. Shuryak, Phys. Lett. B [**423**]{}, 9 (1998) \[arXiv:hep-ph/9704456\], M. A. Stephanov, K. Rajagopal and E. V. Shuryak, Phys. Rev. D [**60**]{}, 114028 (1999) \[arXiv:hep-ph/9903292\], S. Mrowczynski, Phys. Lett. B [**430**]{}, 9 (1998) \[arXiv:nucl-th/9712030\], G. Baym and H. Heiselberg, Phys. Lett. B [**469**]{}, 7 (1999) \[arXiv:nucl-th/9905022\], S. A. Voloshin, V. Koch and H. G. Ritter, Phys. Rev. C [**60**]{}, 024901 (1999) \[arXiv:nucl-th/9903060\], T. A. Trainor, arXiv:hep-ph/0001148, H. Appelshauser [*et al.*]{} \[NA49 Collaboration\], Phys. Lett. B [**459**]{}, 679 (1999) \[arXiv:hep-ex/9904014\]. S. V. Afanasev [*et al.*]{} \[NA49 Collaboration\], Phys. Rev. Lett. [**86**]{}, 1965 (2001) \[arXiv:hep-ex/0009053\]. T. Anticic [*et al.*]{} \[NA49 Collaboration\], Phys. Rev. C [**70**]{}, 034902 (2004) \[arXiv:hep-ex/0311009\]. M. Rybczynski [*et al.*]{} \[NA49 Collaboration\], J. Phys. Conf. Ser. [**5**]{}, 74 (2005) \[arXiv:nucl-ex/0409009\]. J. Adams [*et al.*]{} \[STAR Collaboration\], arXiv:nucl-ex/0308033, M. J. Tannenbaum \[PHENIX collaboration\], J. Phys. G [**30**]{}, S1367 (2004) \[arXiv:nucl-ex/0403048\]. M. Gazdzicki [*et al.*]{} \[NA49 Collaboration\], J. Phys. G [**30**]{}, S701 (2004) \[arXiv:nucl-ex/0403023\]. J. Bächler et al. (NA49 Collab.), [*Searching for QCD Phase Transition*]{}, Addendum–1 to Proposal SPSLC/P264, CERN/SPSC 97 (1997).
S. Jeon and V. Koch, Phys. Rev. Lett. [**85**]{}, 2076 (2000) \[arXiv:hep-ph/0003168\], M. Asakawa, U. W. Heinz and B. Muller, Phys. Rev. Lett. [**85**]{}, 2072 (2000) \[arXiv:hep-ph/0003169\]. J. Zaranek, Phys. Rev. C [**66**]{}, 024905 (2002) \[arXiv:hep-ph/0111228\]. C. Alt [*et al.*]{} \[NA49 Collaboration\], Phys. Rev. C [**70**]{}, 064903 (2004) \[arXiv:nucl-ex/0406013\]. M. Gazdzicki, M. I. Gorenstein and S. Mrowczynski, Phys. Lett. B [**585**]{}, 115 (2004) \[arXiv:hep-ph/0304052\]. S. A. Voloshin, V. Koch and H. G. Ritter, Phys. Rev. C [**60**]{}, 024901 (1999) \[arXiv:nucl-th/9903060\]. V. V. Begun, M. Gazdzicki, M. I. Gorenstein and O. S. Zozulya, Phys. Rev. C [**70**]{}, 034901 (2004) \[arXiv:nucl-th/0404056\]. V. V. Begun, M. I. Gorenstein, A. P. Kostyuk and O. S. Zozulya, Phys. Rev. C [**71**]{}, 054904 (2005) \[arXiv:nucl-th/0410044\], V. V. Begun, M. I. Gorenstein and O. S. Zozulya, arXiv:nucl-th/0411003, A. Keranen, F. Becattini, V. V. Begun, M. I. Gorenstein and O. S. Zozulya, J. Phys. G [**31**]{}, S1095 (2005) \[arXiv:nucl-th/0411116\], V. V. Begun, M. I. Gorenstein, A. P. Kostyuk and O. S. Zozulya, arXiv:nucl-th/0505069. J. Cleymans, K. Redlich and L. Turko, Phys. Rev. C [**71**]{}, 047902 (2005) \[arXiv:hep-th/0412262\], J. Cleymans, K. Redlich and L. Turko, arXiv:hep-th/0503133. H. Appelshauser [*et al.*]{} \[NA49 Collaboration\], Eur. Phys. J. A [**2**]{}, 383 (1998). M. Gyulassy and X. N. Wang, Comput. Phys. Commun. [**83**]{}, 307 (1994) \[arXiv:nucl-th/9502021\]. M. I. Gorenstein, M. Gazdzicki and O. S. Zozulya, Phys. Lett. B [**585**]{}, 237 (2004) \[arXiv:hep-ph/0309142\]. C. Roland [*et al.*]{} \[NA49 Collaboration\], J. Phys. G [**30**]{}, S1381 (2004) \[arXiv:nucl-ex/0403035\]. S. Das \[STAR Collaboration\], arXiv:nucl-ex/0503023.
[^1]: Note that an error was found in the analysis of multiplicity fluctuations in C+C and Si+Si collisions at 158$A$ GeV, and thus the corresponding preliminary results shown in Refs. [@maciek; @marek] should not be used any more.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: |
Using results of Pitowsky and Gupta, we show in a direct, elementary fashion that, in the case of three spins, Bell’s inequalities indeed provide a representation of the tetrahedron of all spin correlation matrices as intersection of half-spaces.\
\
Mathematics Subject Classification\[2010\]:52B12,62H20,81P40
author:
- 'Paolo Dai Pra, Michele Pavon, Neeraja Sahasrabudhe'
title: 'A note on the geometric interpretation of Bell’s inequalities'
---
[^1]
Introduction
============
The non-local character of quantum correlations is manifested by the violation of Bell’s inequality [@Bell], and, more generally, of the so-called CHSH inequality [@CHSH; @Cir]. The experimental violation of these inequalities [@ASPECT_GRANGIER_ROGER] constitutes by now an important part of modern scientific culture. This property and its relation to entanglement has in fact become one of the cornerstones of quantum information theory, see e.g. [@WW]. The relation between Bell’s inequalities and convex geometry is also well-known: It has been studied somewhat independently in the mathematical physics literature [@Pitowsky1991], [@WW Appendix A], [@Pitowsky2001], [@Pitowsky2008] and in the statistics literature [@Balasubramanian_Gupta_Parthasarathy; @JCG]. In particular, Bell’s inequalities are known to be related to the representation of a polytope as intersection of halfspaces. Nevertheless, the problem of finding a minimal set of Bell-like inequalities is in general still open as it relates to the hard problem of finding all the facets of a polytope [@Pitowsky1991; @Pitowsky2001; @WW].
The purpose of this note is to show that, in the simple but important original case of $n=3$ spin random variables, the fact that Bell’s inequalities provide a facets representation of the polytope of all spin correlation matrices can be proven in a direct, straightforward way using the available description of the extreme points [@Pitowsky1991; @JCG]. This elementary derivation, which is apparently missing in the large body of literature on this topic, only works in the $n=3$ case. In such a case, and for no other $n\ge 3$, the polytope is actually a simplex, so that it constitutes a [*tetrahedron*]{}, see Figure \[TETRAHEDRON\].
![Tetrahedron[]{data-label="TETRAHEDRON"}](Tetrahedron.pdf){width="4cm"}
Background
==========
Let $E$ be a $p$-dimensional real vector space. We denote by $\aff A$ the [*affine hull*]{} of $A\subseteq E$, namely the intersection of all linear manifolds in $E$ containing $A$. It can be shown that $$\aff A=\{\sum_{i=1}^m\lambda_ix^i | x^i\in A, \sum_{i=1}^m\lambda_i=1, m \;{\rm finite}\}.$$ We denote by $\con A$ the [*convex hull*]{} of $A$, namely the intersection of all convex sets containing $A$. It is easy to see that $$\con A=\{\sum_{i=1}^m\lambda_ix^i | x^i\in A, \sum_{i=1}^m\lambda_i=1, \lambda_i\ge 0\}.$$
A convex set $D\subseteq E$ is a $d$-[*simplex*]{} if $D=\con(x_0,x_1,\ldots,x_d)$ and the points $x_0,x_1,\ldots,x_d$ are [*affinely independent*]{}, i.e. $x_i\not\in\aff(x_0,x_1,\ldots,x_{i-1},x_{i+1},\ldots,x_d), \forall i$. Necessarily, $d\le p$.
If $D$ is a $d$-[*simplex*]{}, every point $x\in D$ admits a [*unique*]{} representation as convex combination of the points $x_0,x_1,\ldots,x_d$. Moreover, $\ext D$, the set of [*extreme points*]{} of $D$, is indeed $\ext D=\{x_0,x_1,\ldots,x_d\}$.
A set $P\subseteq E$ which is the convex hull of a finite number of points is called a [*(convex) polytope*]{}.
Consider a family of spin random variables $\{\xi_i,1\le i\le n\}$ with ${\mathbb{P}}(\xi_i=-1)={\mathbb{P}}(\xi_i=1)=1/2$ and form the corresponding column vector $\xi$, where $\xi^T=(\xi_1,\ldots,\xi_n)$, $T$ denoting transposition. The random vector $\xi$ has zero expectation and takes values in $\Omega_n=\{-1,1\}^n$. Let $$\Sigma={\mathbb{E}}\{\xi\xi^T\}=\left(\sigma_{ij}\right)_{i,j=1}^n$$ denote the corresponding covariance/correlation matrix. Given the $\sigma_{ij}$, the question whether the corresponding matrix can be realized as the correlation matrix of spin random variables is a nontrivial one. Indeed, while any symmetric, positive semidefinite matrix $\Sigma$ can be realized as the covariance of a Gaussian random vector, the conditions on $\Sigma$ become much more stringent if we require realization through spin random variables. Since [@Bell], it is known that for three spins variables, the $\sigma_{ij}$ need to satisfy Bell’s inequalities to be realizable. Actually, it was shown in [@Balasubramanian_Gupta_Parthasarathy] that, in the case $n=3,4$, Bell’s inequalities $$\label{Bellsinequalities}
1+\epsilon_i\epsilon_j \sigma_{ij}+ \epsilon_i\epsilon_k \sigma_{ik} + \epsilon_j\epsilon_k \sigma_{jk}\geq 0, \quad\forall \quad1\le i<j<k\le n,\quad
\epsilon_i\in\{-1,\,+1\},$$ are necessary and sufficient for the matrix $\Sigma={\mathbb{E}}\{\xi\xi^T\}$ to be the correlation matrix of $n$ spin random variables. Notice that a correlation matrix is symmetric and has ones on the main diagonal. It is therefore determined by the elements above the main diagonal. Hence, these matrices are in one to one correspondence with a $n(n-1)/2$-dimensional Euclidean space. For $n\ge 2$, let ${\mathcal C}_n$ denote the set of correlation matrices that can be realized through spin random variables. Pitowsky [@Pitowsky1991] and Gupta [@JCG Theorem3.1] showed that ${\mathcal C}_n$ is a [*polytope*]{}. It consists namely of all convex combinations of a finite number of correlation matrices. They also gave an explicit description of the extreme points (matrices) of this set. These are rank-$1$ matrices of the form $\omega^{(n)}(\omega^{(n)})^T$, where the $n$-dimensional vector $\omega^{(n)}$ has entries equal to $1$ or $-1$. Notice that $\omega^{(n)}$ and $-\omega^{(n)}$ generate the same correlation matrix so that there are $2^{n-1}$ different such matrices. Since a polytope simply consists of all convex combinations of its extreme points, this completely characterizes ${\mathcal C}_n$.
Geometric meaning of Bell’s inequalities
========================================
By the Minkowski-Weyl theorem [@Z], [@WW Appendix A], a polytope, besides being the convex hull of its extreme points (V-representation), can also be dually described as the intersection of half-spaces through a system of linear inequalities (H-representation). For instance, a triangle in the plane is the intersection of three half-planes determined by the straight lines containing its sides. It is known [@Pitowsky2001; @WW] that the latter description for ${\mathcal C}_n$ is related to the Bell’s inequalities (\[Bellsinequalities\]). Consider now the $n=3$ case, where (\[Bellsinequalities\]) take the form $$\begin{aligned}
\label {Bell1}1+ \sigma_{12} +\sigma_{13}+\sigma_{23}\ge 0,\\ 1- \sigma_{12} +\sigma_{13}-\sigma_{23}\ge 0,\\ 1+ \sigma_{12} -\sigma_{13}-\sigma_{23}\ge 0,\\ 1- \sigma_{12} -\sigma_{13}+\sigma_{23}\ge 0.\label{Bell4}\end{aligned}$$
The set ${\mathcal C}_3$ is a $3$-simplex, namely a tetrahedron. Bell’s inequalities (\[Bell1\])-(\[Bell4\]) provide the H-representation of the tetrahedron ${\mathcal C}_3$. Thus, any element in ${\mathcal C}_3$ satisfies Bell’s inequalities (\[Bell1\])-(\[Bell4\]). Conversely, any correlation matrix satisfying Bell’s inequalities belongs to ${\mathcal C}_3$.
Consider the four vectors $$v_1=\left(\begin{matrix}+1\\+1\\+1\end{matrix}\right),\quad v_2=\left(\begin{matrix}+1\\-1\\+1\end{matrix}\right),\quad v_3=\left(\begin{matrix}+1\\+1\\-1\end{matrix}\right),\quad v_4=\left(\begin{matrix}-1\\+1\\+1\end{matrix}\right).$$ By [@Pitowsky1991; @JCG], the extreme points of ${\mathcal C}_3$ are given by the four matrices $$\begin{aligned}
&&\Sigma_1=v_1v_1^T=\left[\begin{matrix}+1& +1& +1\\ +1& +1 & +1\\ +1 & +1 & +1\end{matrix}\right],\quad \Sigma_2=v_2v_2^T=\left[\begin{matrix}+1& -1& +1\\ -1& +1 & -1\\ +1 & -1 & +1\end{matrix}\right],\\
&&\Sigma_3=v_3v_3^T=\left[\begin{matrix}+1& +1& -1\\ +1& +1 & -1\\ -1 & -1 & +1\end{matrix}\right],\quad\Sigma_4=v_4v_4^T=
\left[\begin{matrix}+1& -1& -1\\ -1& +1 & +1\\ -1 & +1 & +1\end{matrix}\right].\end{aligned}$$ Consider now a symmetric matrix $\Sigma$ with ones on the main diagonal. As observed before, they can be bijectively mapped onto a $3$-dimensional space. Let $$A:=\left[\begin{matrix}+1& +1& +1&+1\\ +1& -1 & +1&-1\\ +1 & +1 & -1&-1\\+1 & -1 & -1&+1\end{matrix}\right],\quad X=\left[\begin{matrix}1\\ \sigma_{12}\\ \sigma_{13}\\ \sigma_{23}\end{matrix}\right].$$ Knowledge of $X$ is equivalent to knowledge of $\Sigma$. Define $$Y=AX=\left[\begin{matrix}1+ \sigma_{12} +\sigma_{13}+\sigma_{23}\\ 1- \sigma_{12} +\sigma_{13}-\sigma_{23}\\ 1+ \sigma_{12} -\sigma_{13}-\sigma_{23}\\ 1- \sigma_{12} -\sigma_{13}+\sigma_{23}\end{matrix}\right].$$ Then Bell’s inequalities (\[Bell1\])-(\[Bell4\]) simply state that the vector $Y$ belongs to the positive orthant $ Y\in{\mathbb{R}}^4_+$, it has namely all nonnegative components. Consider a symmetric, $3\times 3$ matrix $\Sigma$ with ones on the main diagonal expressed as [*linear*]{} combination of the $\Sigma_i$: $$\label{form}\Sigma=\lambda_1 \Sigma_1 + \lambda_2 \Sigma_2 + \lambda_3\Sigma_3 + \lambda_4\Sigma_4,\quad \lambda_i\in{\mathbb{R}}.$$ Since $\sigma_{ii}=1$ and the $\Sigma_i, i=1,2,3,4$ have the same property, it follows that necessarily $$\label{sum}
\sum_{i=1}^4\lambda_i=1,$$ namely $\Sigma$ is in $\aff\{ \Sigma_1, \Sigma_2,\Sigma_3,\Sigma_4\}$. From (\[form\]), a simple calculation yields $$X=\left[\begin{matrix}1\\ \sigma_{12}\\ \sigma_{13}\\ \sigma_{23}\end{matrix}\right]=A\left[\begin{matrix}\lambda_1\\ \lambda_2\\ \lambda_3\\ \lambda_4\end{matrix}\right].$$ Observe that $\frac{1}{2}A$ is involutory, namely $A^2=4 I_4$. Since $A$ is invertible, it follows that any symmetric matrix with ones on the main diagonal may be expressed as in (\[form\]). In particular, every element in the convex hull of $\{\Sigma_1,\Sigma_2,\Sigma_3,\Sigma_4\}$ admits a unique representation. Hence, the $\Sigma_i, i=1,2,3,4$ are affinely independent and generate a simplex, namely a tetrahedron. Moreover, we have $$Y=AX=A^2\left[\begin{matrix}\lambda_1\\ \lambda_2\\ \lambda_3\\ \lambda_4\end{matrix}\right]=4\left[\begin{matrix}\lambda_1\\ \lambda_2\\ \lambda_3\\ \lambda_4\end{matrix}\right].$$ Thus Bell’s inequalities are [*equivalent*]{} to having $\lambda_i\ge 0, i=1,2,3,4$. Taking (\[form\]) and (\[sum\]) into consideration, we conclude that a convex combination of the $\Sigma_i$ is a correlation matrix satisfying Bell’s inequalities. Namely, any element in ${\mathcal C}_3$ satisfies Bell’s inequalities (\[Bell1\])-(\[Bell4\]). Conversely, any correlation matrix satisfying Bell’s inequalities belongs to the convex hull of the $\Sigma_i$, i.e. it belongs to ${\mathcal C}_3$. [$\Box$ 2ex]{}
Closing comments
================
The above elementary argument does not extend to higher dimensions. In particular, ${\mathcal C}_n$ is not a simplex for $n>3$. For instance, we know from [@Balasubramanian_Gupta_Parthasarathy] that also in the case $n=4$ Bell’s inequalities are necessary and sufficient for $\Sigma$ to be realizable as correlation matrixof spin random variables. There are $16$ such inequalities. The set ${\mathcal C}_4$ has eight extreme points whereas $\Sigma$ is determined by six elements (those above the main diagonal). But, in a $6$-dimensional space, a simplex can have at most $7$ extreme points. Thus the polytope ${\mathcal C}_4$ is not a simplex. More generally, the equation $$2^{n-1}=\frac{n(n-1)}{2}+1$$ relating the number of extreme points of ${\mathcal C}_n$ and the maximum number of affinely independent points in a space of dimension $n(n-1)/2$ has no solution for $n>3$. Moreover, for $n\ge 5$, Bell’s inequalities (\[Bellsinequalities\]) are known not to be sufficient for a correlation matrix to be the correlation matrix of spin random variables [@JCG].
[1]{}
[[[ A. Aspect, A. P. Grangier and G. Roger]{}]{}, [[*Experimental tests of realistic local theories via Bell’s theorem*]{}]{}, Physical Review Letters [[vol. [**47**]{}]{}]{}, [[pp. 460-463]{}]{}, 1981]{}.
[[[ K. Balasubramanian, J.C. Gupta and K.R. Parthasarathy]{}]{}, [[*Remarks on Bell’s Inequalities for Spin Correlations*]{}]{}, Sankhya: The Indian Journal of Statistics [[vol. [**60**]{}]{}]{}, [[pp. 29-35]{}]{}, 1998]{}.
[[[ J. S. Bell]{}]{}, [[*On the Einstein Podolsky Rosen Paradox*]{}]{}, Physics [[vol. [**1**]{}]{}]{}, [[pp. 195-200]{}]{}, 1964]{}.
[[[ B. S. Cirel'son]{}]{}, [[*Quantum generalizations of Bell’s inequality*]{}]{}, Lett. Math. Phys. [[vol. [**4**]{}]{}]{}, [[pp. 93-100]{}]{}, 1980]{}.
.
[[[ J. C. Gupta]{}]{}, [[*Characterisation of Correlation Matrices of Spin Variables*]{}]{}, Sankhya: The Indian Journal of Statistics [[vol. [**61**]{}]{}]{}, [[pp. 282-285]{}]{}, 1999]{}.
[[[ I. Pitowsky]{}]{}, [[*Correlation Polytopes: Their Geometry and Complexity*]{}]{}, Mathematical Programming [[vol. [**50**]{}]{}]{}, [[pp. 395-414]{}]{}, 1991]{}.
[[[ I. Pitowsky and K. Svozil]{}]{}, [[*Optimal tests of quantum nonlocality*]{}]{}, Physical Review A [[vol. [**64**]{}]{}]{}, [[pp. 014102]{}]{}, 2001]{}.
[[[ I. Pitowsky]{}]{}, [[*On the geometry of quantum correlations*]{}]{}, Physical Review A [[vol. [**77**]{}]{}]{}, [[pp. 062109-1-5 ]{}]{}, 2008]{}.
[[[ R. F. Werner and M. M. Wolf]{}]{}, [[*Bell inequalities and entanglement*]{}]{}, Quant. Inform. Comput. [[vol. [**1**]{}]{}]{}, [[pp. 1-25]{}]{}, 2001]{}.
[[[ G. Ziegler]{}]{}, [[*Lectures on polytopes*]{}]{}, Springer-Verlag, New York, 1997]{}.
[^1]: The work of M. Pavon was partially supported by the QuantumFuture research grant of the University of Padova and by an Alexander von Humboldt Foundation fellowship at the Institut für Angewandte Mathematik, Universität Heidelberg, Germany.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Based on an analogy with topologically ordered new state of matter in condensed matter systems, we propose a low energy effective field theory for quark-gluon plasma in quantum chromodynamics (QCD) system. It shows that below a QCD gap which is several times of the critical temperature, the quark-gluon plasma behaves like a topological fluid.'
author:
- 'M.J.Luo'
title: 'Effective Field Theory for Quark-Gluon Plasma by Bosonization'
---
Introduction
============
Quantum Chromodynamics (QCD) based on non-Abelian gauge field theory was well established as the fundamental theory of the strong interaction. However, because of its strong coupling nature at low energy or long distance, from the day the exact theory was written down, its low energy behavior is still being poorly understood by solving the first principle theory. Although we already have a low energy effective theory based on the Goldstone’s theorem under the condition of chiral symmetry breaking (for a review, see: [@scherer2003introduction]), there is no analog foundation to formulate a low energy effective theory with neither symmetry breaking nor order parameter, more specifically, applying to the disordered quark-gluon plasma phase.
A crucial phenomenological clue to the problem comes from the experimental studies of the Relativistic Heavy Ion Collision (RHIC), in which new state of matter of strong interaction is produced. It is discovered, at sufficiently high temperature and density, that the nucleons in the collision are deconfined into constituent quarks and gluons and form a soup of quark-gluon plasma. Important experimental facts of the quark-gluon plasma are the following: (i) The long distance low frequency behavior of the quark-gluon plasma around the critical temperature $T_{c}$ is well described by the theory of fluid mechanics (hydrodynamics) [@kolb2001elliptic; @kolb2003transverse]. (ii) The quark-gluon plasma around $T_{c}$ looks like a strong correlated and near-perfect liquid [@shuryak2005rhic], the pressure, numbers of transport properties (e.g. conductivity, diffusion and viscosity) are small, strongly deviating from the behavior of a weakly coupled ideal gas predicted from the perturbative QCD. (iii) The charge of the deconfined quark or parton excitation in the quark-gluon plasma is fractionalized [@jeon2000charged]. The purpose of the paper is to propose a topological dominant effective field theory for the quark-gluon plasma by bosonizing the gauge invariant QCD that I feel are consistent with the known experimental facts.
The connection between a pure gluon plasma in hard thermal loop limit and the topological Chern-Simons theory has been discussed in literature [@efraty1992action; @efraty1993chern]. The remarkable relation between the quark-gluon plasma and topological field theories can be generalized to a more generic level. The quark-gluon plasma phase is a deconfined phase where the bound state hadrons are decomposed into quarks and gluons at high temperature and density. The situation is analogous with the case in condensed matter physics such as the gauge symmetric Heisenberg spin system that at high enough temperature and chemical potential where the system is well into a deconfined phase, the Heisenberg spins are decomposed into constituent holons and spinons and form a plasma phase which enjoy a new type of order called topological order/phase (see, for example, [@wen1991mean; @wen2002quantum] at low dimensions, [@diamantini2011charge; @vishwanath2013physics; @chan2013effective] in 3+1 dimensions). The ground state of the topologically ordered matter is not associated with spontaneous symmetry breaking, protected by a gap, and topological fluid (see, for example, [@frohlich1991large; @zee1992long; @wen1992classification] in 2+1 dimensions) as the low energy degree of freedom in many aspects is analogous with the fluid-like plasma of quark-gluon around the critical temperature and chemical potential. The topological fluid is in general incompressible, with excitation carrying fractional charges [@laughlin1983anomalous; @haldane1983fractional; @jain1989incompressible], characterized by having non-trivial ground state degeneracies [@wen1990ground; @wen1991mean] on topological non-trivial compact manifold.
Bosonization is a proper description to such bosonic fluids system, especially at finite chemical potential because of a more controllable sign structure than fermion [@wu2008sign; @luo2014gauge]. The validity of bosonization approach in the real 3+1 dimensional QCD is based on the facts: (i) lattice data show that quark (baryon) masses seem to be large in the strongly coupled quark-gluon plasma near $T_{c}$ [@liao2006lattice], (ii) a massive fermionic theory and the bosonized descriptions share the same macroscopic limit which is dominated by topological degrees of freedom [@chan2013effective].
The structure of the paper is as follows. In section II, we construct a basic formulation of a bosonized low energy effective field action for the quark-gluon plasma. In subsection A of the section III, we derive the classical equations of motion and argue the dyonic nature of the fluids. In subsection B, a hard-thermal-loops effective action related to Chern-Simons theory is reproduced. In subsection C, we explain why the hydrodynamics description is so successful. In subsection D, linear responses of the theory are studied, more specifically, the ratio of shear viscosity over entropy density of an homogeneous and isotropic quark-gluon plasma is derived at Bogomolnyi bound. In subsection E, the fractionalization of charge for quark/parton in the plasma is explained. In section IV, we summarize the results.
Bosonic Effective Action
========================
Both for mathematical and physical reasons, in a strongly coupled many-body system, what we observed are gauge invariant responses to external (in practice electromagnetic) probe, and hence the notion of response is more proper than the unobserved microscopic fundamental particles. The bosonization approach starts from replacing the fermionic fundamental quarks by the responses to an external weak U(1) gauge field being a source minimally coupling to them,$$\left\langle j_{\mu_{1}}(x_{1})j_{\mu_{2}}(x_{2})...j_{\mu_{n}}(x_{n})\right\rangle =\frac{\delta}{\delta A_{\mu_{1}}^{ex}(x_{1})}\frac{\delta}{\delta A_{\mu_{2}}^{ex}(x_{2})}...\frac{\delta}{\delta A_{\mu_{n}}^{ex}(x_{n})}\ln Z[A^{ex}],\label{eq:current-correlation}$$ where$$Z[A^{ex}]=\int\mathcal{D}\bar{\psi}\mathcal{D}\psi\mathcal{D}G\exp\left(-\int d^{4}x\bar{\psi}\left[i\gamma_{\mu}\left(\partial_{\mu}+iG_{\mu}+\mu\delta_{\mu0}+iA_{\mu}^{ex}\right)\right]\psi-S_{YM}[G]\right).$$ The $\psi$ are quarks, $G_{\mu}$ are gluon gauge fields, $\mu$ is the chemical potential, $A_{\mu}^{ex}$ are external U(1) source, and $S_{YM}[G]$ is the Yang-Mills action for gluons. The form of $Z[A^{ex}]$ can be determined as follows. By using the U(1) gauge invariance $Z[A^{ex}]=Z[A^{ex}+a]$, where $a$ is a pure gauge, i.e. $f_{\alpha\beta}(a)=\partial_{\alpha}a_{\beta}-\partial_{\beta}a_{\alpha}=0$. The pure gauge condition can be imposed by inserting a delta function into the functional integral,$$\begin{aligned}
Z[A^{ex}] & =\int\mathcal{D}aZ[A^{ex}+a]\left(\prod_{x}\prod_{\mu<\nu}\epsilon_{\mu\nu\alpha\beta}\delta[f_{\alpha\beta}(a(x))]\right).\end{aligned}$$ The delta function can be exponentiated by introducing an auxiliary anti-symmetric 2-form gauge fields $b_{\mu\nu}$ as Lagrangian multipliers,
$$Z[A^{ex}]=\int\mathcal{D}a\mathcal{D}bZ[A^{ex}+a]\exp\left(-\frac{i}{4\pi}\int d^{4}x\epsilon_{\mu\nu\alpha\beta}b_{\mu\nu}f_{\alpha\beta}(a)\right),$$
in which the term in the exponent is purely topological known as the BF action (for a review see: [@birmingham1991topological]), and the prefactor $\frac{1}{4\pi}$ is for convention. The prefactor can be in general defined as $\frac{\mathsf{m}}{4\pi}$, $\mathsf{m}\in\mathbb{Z}$ plays a similar role of a Chern-number in quantum Hall systems characterizing its topological order, the integer-valued $\mathsf{m}$ so defined characterizes the topological order of the ground state of the system, a value $\mathsf{m}\neq1$ represents a fractionalized topological order.
The auxiliary gauge field $b$ is of ($D-2$)-form where $D$ is spacetime dimensions. In the paper we will focus on the real QCD system $D=4$, so $b$ field is of a 2-form. By doing a variable replacement $a\rightarrow a-A^{ex}$, we obtain$$Z[A^{ex}]=\int\mathcal{D}a\mathcal{D}bZ[a]\exp\left(-\frac{i}{4\pi}\int d^{4}x\epsilon_{\mu\nu\alpha\beta}b_{\mu\nu}\left[f_{\alpha\beta}(a)-f_{\alpha\beta}(A^{ex})\right]\right).\label{eq:identity}$$ By using Eq.(\[eq:current-correlation\]), the above partition function indicates a bosonization correspondence for the flow$$j_{\mu}(x)=\frac{1}{2\pi}\epsilon_{\mu\nu\alpha\beta}\partial_{\nu}b_{\alpha\beta}(x),\label{eq:charge-current}$$ which is gauge invariant under the transformations$$b_{\mu\nu}\rightarrow b_{\mu\nu}+\partial_{\mu}\eta_{\nu}-\partial_{\nu}\eta_{\mu}.$$ Note that the flow is automatically conserved $$\partial_{\mu}j_{\mu}\equiv0.\label{eq:charge-current-conservation}$$ The non-dynamical component $a_{0}$ coupled to the density $j_{0}$ plays the role of a Lagrangian multiplier similar with a chemical potential.
Beside $a_{0}$, there is another non-dynamical Lagrangian multiplier $b_{0i}=-b_{i0}$, ($i=1,2,3$), which implies the existence of a spin/angular momentum density tensor $$s_{\mu\nu}(x)=\frac{1}{2\pi}\epsilon_{\mu\nu\alpha\beta}\partial_{\alpha}a_{\beta}(x).\label{eq:spin-density}$$ It can be interpreted as a spin or angular momentum density because it is a response to an external magnetic field which will be shown as following. If we add a kinetic term of $b_{\mu\nu}$ with infinite mass into the parenthesis of the Eq.(\[eq:identity\]) as a regulator,$$S\rightarrow S+S_{reg},\quad S_{reg}=\lim_{\lambda\rightarrow\infty}\frac{1}{12\lambda^{2}}\int d^{4}xh_{\mu\nu\rho}h_{\mu\nu\rho},\label{eq:kinetic-b}$$ where $h_{\mu\nu\rho}=\partial_{\mu}b_{\nu\rho}+\partial_{\nu}b_{\rho\mu}+\partial_{\rho}b_{\mu\nu}$ is the field strength of $b_{\mu\nu}$, and $\lambda$ is a mass parameter. By integrating out the $b_{\mu\nu}$ field, we obtain an effective coupling between $s_{\mu\nu}$ and external source field, $$S_{int}=\frac{1}{4\pi}\int d^{4}xs_{\mu\nu}\tilde{F}_{\mu\nu}^{ex},$$ in which $\tilde{F}_{\mu\nu}^{ex}=\epsilon_{\mu\nu\alpha\beta}\partial_{\alpha}A_{\beta}^{ex}$ is an external magnetic field. So we prove that the tensor $s_{\mu\nu}$ is a response to the external magnetic field,$$\left\langle s_{\mu_{1}\nu_{1}}(x_{1})s_{\mu_{2}\nu_{2}}(x_{2})...s_{\mu_{n}\nu_{n}}(x_{n})\right\rangle =(4\pi)^{n}\frac{\delta}{\delta\tilde{F}_{\mu_{1}\nu_{1}}^{ex}(x_{1})}\frac{\delta}{\delta\tilde{F}_{\mu_{2}\nu_{2}}^{ex}(x_{2})}...\frac{\delta}{\delta\tilde{F}_{\mu_{n}\nu_{n}}^{ex}(x_{n})}\ln Z[A^{ex}].$$ In the rest of the paper we will not distinguish the terminologies spin and angular momentum, but using “spin” for convention.
A spin current can be defined by the spin density tensor$$M_{[\mu\nu]\alpha}=\partial_{\alpha}s_{\mu\nu}.$$ It is easy to verify that the spin current satisfies 6 continuity equations$$\partial_{\alpha}M_{[\mu\nu]\alpha}\equiv0.\label{eq:spin-current-conservation}$$ In addition, there is another continuity equation, note that a Pauli-Lubanski pseudovector constructed by the spin current$$\begin{aligned}
W_{\mu}(x) & =\frac{1}{2}\epsilon_{\mu\nu\alpha\beta}M_{[\alpha\beta]\nu},\end{aligned}$$ is also automatically satisfied $$\partial_{\mu}W_{\mu}\equiv0.\label{eq:Pauli-conservation}$$
The independence of the continuity equations for charge and spin currents manifests an important fact that the carriers of charge and spin in the fluids are no longer identified, in contrast to the conventional quasi-particle picture in which it carries both. In other words, there are two types of currents in the fluids [@diamantini2011charge], which is known as charge-spin separation in condensed matter physics.
So far the bosonization recipe is quite general and is able to apply to a gapped or gapless system, but only in 1+1 dimensions the $Z[a]$ can be written down exactly, in other dimensions $Z[a]$ can only be approximately evaluated. The $Z[a]$ now encodes almost all the physics of QCD, which needs some guess work. Benefited from an intrinsic gap [@gribov1978quantization] $\Lambda$ in QCD, the $Z[a]$ can be obtained at low energy by inverse gap expansion. As is shown in [@zwanziger2005equation], the gap must be as large as several times of $T_{c}$. The gap defines a boundary of the validity of our low energy effective theory. Fortunately, even without knowing the exact value of the gap $\Lambda$, two leading actions can be explicitly written down based on the principle of gauge invariance. The two leading terms have no dimensional parameter, so they are marginal in the RG sense. We have$$Z[a]=\exp\left(-\frac{i\theta}{16\pi^{2}}\int d^{4}x\epsilon_{\mu\nu\alpha\beta}f_{\mu\nu}(a)f_{\alpha\beta}(a)-\frac{1}{4g^{2}}\int d^{4}xf_{\mu\nu}(a)f^{\mu\nu}(a)+\mathcal{O}(\partial^{4}/\Lambda^{2})\right).\label{eq:Z[a]}$$ The first term is the $\theta$-term which is topological and always marginal at low energy, the second term is the dynamical Maxwell term in which the coupling constant becomes large at low energy and hence its contribution is suppressed compared with the topological terms. The parameter $\theta$ is an angle with periodicity $2\pi$, and the $\theta$-term generally breaks the parity and time-reversal symmetries unless it is quantized as $\theta=\nu\pi$, $\nu\in\mathbb{Z}$ is the topological winding number. It is worth stressing that the experimental bounds on the $\theta$ parameter in QCD (from the dipole moment of the neutron) is either exceedingly small $\theta<10^{-9}$ or very close to $\theta=\pi$. The following discussions are based on a nontrivial choice $\theta=\pi$ in the quark-gluon plasma.
There are two remarks to $Z[a]$ taking the form as Eq.(\[eq:Z\[a\]\]): (1) such expansion form works only at low energy below the gap $\Lambda$, and the results in the rest of the paper are achieved below the gap; (2) it is such a form that is remarkably successful in re-interpretating the gauge fields $a_{\mu}$,$b_{\mu\nu}$ and their classical equations of motion as certain hydrodynamic fluids in the ground state, which satisfy a series of hydrodynamic equations phenomenologically proposed. In other words, such form of $Z[a]$ gives a simple hydrodynamic re-interpretation of the QCD ground state, which will be shown in the section-III.
So put everything together, it contains only relevant and marginal terms and should be compatible with the gauge invariance, the bosonized low energy effective action is given by$$Z[A^{ex}]=\int\mathcal{D}a\mathcal{D}b\exp\left(-S_{eff}\right),$$ with$$S_{eff}(a,b)=\frac{i}{4\pi}\int d^{4}x\epsilon_{\mu\nu\alpha\beta}b_{\mu\nu}\left[f_{\alpha\beta}(a)-f_{\alpha\beta}(A^{ex})\right]+\frac{i\theta}{16\pi^{2}}\int d^{4}x\epsilon_{\mu\nu\alpha\beta}f_{\mu\nu}(a)f_{\alpha\beta}(a)+\frac{1}{4g^{2}}\int d^{4}xf_{\mu\nu}(a)f^{\mu\nu}(a),\label{eq:effective-action}$$ in which the fermionic quarks are bosonized by two types of gauge fields: 1-form $a_{\mu}$ and 2-form $b_{\mu\nu}$. The first two terms are topological term, and always take imaginary $i$ in its prefactor no matter in Euclidean or Minkovski metric since the topological term is metric independent. Here the imaginary-defined topological terms after the bosonization play the role of a non-trivial sign structure of fermions, which is insteadly responsible for the notorious sign problem for the original fermionic system at finite chemical potential. While the third term is non-topological and metric dependent, it is real in Euclidean metric or imaginary in Minkovski metric, the paper is using the Euclidean metric for convention.
One may wonder that since the system is gapped, there seems no Fermi surface. But remind that the system is bosonized, the chemical potential $\mu$ can also be well-defined which formally shifts the non-dynamical time component of gauge field $A_{0}^{ex}$, as if the quarks fill the Fermi sphere in the equivalent bosonized QCD.
General Features
================
Plasma with both Electric and Magnetic Charges
----------------------------------------------
At criticality, the fixed point coupling constant $g$ is large, then the first two topological terms in Eq.(\[eq:effective-action\]) dominate the low energy theory at the criticality, while the dynamical Maxwell term can be seen as a perturbation. The classical equations of motion of the topological action are given by,$$\begin{aligned}
\frac{\delta S_{eff}}{\delta a_{\mu}}=0:\quad j_{\mu}+\frac{\theta}{2\pi}j_{\mu}^{m} & =0.\label{eq:EOM-1}\end{aligned}$$ $$\frac{\delta S_{eff}}{\delta b_{\mu\nu}}=0:\quad f_{\alpha\beta}(a)-f_{\alpha\beta}(A^{ex})=0.$$ The first equation of motion shows a relation between the electric current and the magnetic monopole current $j_{\mu}^{m}=\frac{1}{2\pi}\partial_{\nu}\tilde{f}_{\mu\nu}$ in the fluid, i.e. $e=q_{m}\theta/2\pi$ [@witten1979dyons], where $e$ is the fundamental electric charge unit and $q_{m}$ a magnetic charge. In the theory with $\theta=\pi\neq0$, the fluid carries both electric and magnetic charges [@liao2007strongly; @chernodub2007magnetic]. As is well known that in the 2+1 dimensional topological Chern-Simons theory gauge field plays the role of attaching fluxes to electrons, in 3+1 dimensions the gauge field $a_{\mu}$ is attaching a monopole to the electron (known as dyon). We will see in the latter discussion that this dyonic property is crucial for a small value of shear viscosity over entropy density of the quark-gluon plasma. The second equation of motion is a constraint because the $b$ field plays the role of a Lagrangian multiplier, or equivalently, the mass of the $b$ field can be considered infinitely heavy shown as Eq.(\[eq:kinetic-b\]).
Hydrodynamic Equations
----------------------
As discussed in Section II, the topological action describes a hydrodynamic effective field theory for two types of topological fluids, the charge current and spin current. Here we have 1+6 continuity equations, one continuity equation for the charge current Eq.(\[eq:charge-current-conservation\]) and 6 continuity equations for the spin currents Eq.(\[eq:spin-current-conservation\]),$$\partial_{\mu}j_{\mu}=0,\quad\partial_{\alpha}M_{[\mu\nu]\alpha}=0.$$ In general, there is a relation connecting an energy-momentum tensor to the spin current $$M_{[\mu\nu]\alpha}(x)=\delta x_{\mu}T_{\nu\alpha}(x)-\delta x_{\nu}T_{\mu\alpha}(x),$$ in which $\delta x$ is an infinitesimal deviation from point $x$. Then the continuity equations for the spin current are given by$$\partial_{\alpha}M_{[\mu\nu]\alpha}=T_{\nu\mu}+\delta x_{\mu}\partial_{\alpha}T_{\nu\alpha}-T_{\mu\nu}-\delta x_{\nu}\partial_{\alpha}T_{\mu\alpha}=0.$$ Under the condition that the energy-momentum tensor is symmetric $T_{\mu\nu}=T_{\nu\mu}$, one can recognize that the 6 continuity equations for the spin currents naturally contain the 4 continuity equations for energy-momentum, $$\partial_{\mu}T_{\mu\nu}=0.\label{eq:stress-conservation}$$
In addition, the classical equation of motion Eq.(\[eq:EOM-1\]) $j_{\mu}=-\frac{\theta}{2\pi}M_{[\mu\nu]\nu}$ implies a relation between the current and the stress-energy tensor so defined, for $\theta=\pi$, we have $T_{\mu\nu}=\frac{1}{2}\left(\partial_{\mu}j_{\nu}+\partial_{\nu}j_{\mu}\right)+g_{\mu\nu}T_{\alpha\alpha}$. The relation naturally leads to a traceless condition $T_{\alpha\alpha}=0$, so the stress-energy tensor is related to current by $$T_{\mu\nu}=\frac{1}{2}\left(\partial_{\mu}j_{\nu}+\partial_{\nu}j_{\mu}\right).\label{eq:T-j}$$ If we define the flow $v_{\mu}$ as $$j_{i}=\rho v_{i}=-D\partial_{i}\rho,\label{eq:definition-j}$$ where $D$ is a diffusion constant with dimension of length and $\rho=j_{0}$ is the flow density, we obtain $$\partial_{0}\left(\rho v_{j}\right)=D\partial_{i}T_{ij},\quad(i,j=1,2,3).\label{eq:v-DT}$$ By using Eq.(\[eq:T-j\] and \[eq:definition-j\]), we have $$D\partial_{i}T_{ij}=\frac{1}{2}D\partial_{i}\left[\partial_{i}\left(\rho v_{j}\right)+\partial_{j}\left(\rho v_{i}\right)\right]=-\partial_{i}\left(\rho v_{i}v_{j}\right)+D\rho\nabla^{2}v_{j},$$ so the Eq.(\[eq:v-DT\]) (the classical equation of motion Eq.(\[eq:EOM-1\])) are actually the Navier-Stokes equations for incompressible fluid without pressure,$$\partial_{0}\left(\rho v_{j}\right)+\partial_{i}\left(\rho v_{i}v_{j}\right)=\eta\nabla^{2}v_{j},\quad(i,j=1,2,3).\label{eq:NS-equ}$$ in which $\eta=D\rho$ here is a viscosity constant.
In summary, the ground state of the theory is described by topological fluids satisfying hydrodynamical equations: continuity equation of current, conservation of stress-energy tensor Eq.(\[eq:stress-conservation\]), the traceless of stress-energy tensor, and the Navier-Stokes equations Eq.(\[eq:NS-equ\]). These relations explain the fact (i) of quark-gluon plasma.
Nearly Perfect Liquid
---------------------
It is worth stressing that since the low energy effective theory is dominated by the topological terms, so in principle, there is no strict concepts of metric and energy, one may wonder why we can talk about the the energy-momentum tensor defined above. The reason is that the energy-momentum tensor defined above just describes classical flows or distributions of the charge and spin currents in the ground state gapped from the excitations, which are perfect fluids. Like the Hall fluids which have vanishing viscosity, if the low energy behavior of quark-gluon fluids here are solely governed by the topological terms, they have no viscosity either (completely perfect). When the excitation states coming from the non-topological terms are taken into account, internal frictions appear and the fluid becomes dissipation. At temperature well below the gap which protects the topological order, the deviation from viscousless is small, so the fluids still behave nearly perfect. This property of the theory is associated with the fact (ii).
The terminology “incompressible” of the fluids is tantamount to “topological”, more precisely, since the partition function given by the topological fixed point action only depends on the topology of the manifold, so it is exactly responseless to the compressing or expanding the 3-volume of the manifold. As a result, the quantities like pressure and bulk viscosity of quark-gluon plasma are identically zero predicted from the lowest order topological theory.
To consider the transport beyond the lowest order result, one can first integrate out the $a$ and $b$ field, the effective fixed point action becomes$$S_{eff}(A^{ex})=\frac{i\theta}{8\pi^{2}}\int d^{4}x\epsilon_{\mu\nu\alpha\beta}\partial_{\mu}A_{\nu}^{ex}(t,\mathbf{x})\partial_{\alpha}A_{\beta}^{ex}(t,\mathbf{x})+\frac{1}{4g^{2}}\int d^{4}xF_{\mu\nu}(A^{ex})F^{\mu\nu}(A^{ex}),\label{eq:action-by-external-A}$$ which can be viewed as an action for the U(1) fields in a quark-gluon plasma medium. The first term is the $\theta$-term of external U(1) source which is topological, the second term is the dynamical Maxwell term which is also expected being suppressed at low energy since the integration over QCD fields gives a Debye gap. The transport of the effective theory e.g. the conductivity, charge susceptibility and shear viscosity, therefore take small values that are of order of the dynamical Maxwell term. The following calculations are doing at zero temperature, and the finite temperature results can be generalized by standard Matsubara summation procedures.
Since the fluid is incompressible if we only consider the topological term, the topological term only gives a current transport on the boundary of the quark-gluon plasma, there is no net current in its bulk. The current transport or conductivity in the bulk solely gives rise from the Maxwell term,$$\sigma=\lim_{\omega\rightarrow0}\frac{1}{\omega}\frac{\delta^{2}S_{eff}(A^{ex})}{\delta A_{i}^{ex}(\omega,k=0)\delta A_{j}^{ex}(0,0)}P_{ij}^{\bot}=\lim_{\omega\rightarrow0}\frac{1}{2g^{2}}\frac{\omega^{2}}{\omega}=\frac{\omega}{2g^{2}},$$ in which $P_{ij}^{\bot}=\delta_{ij}-k_{i}k_{j}/\left|\mathbf{k}\right|^{2}$ is the transverse projector. Note that the chemical potential $\mu$ can be seen as a shift of the non-dynamical $A_{0}^{ex}$, i.e. $A_{0}^{ex}(\omega=0,\mathbf{k})$, then the charge susceptibility can be calculated by$$\chi=\frac{\delta^{2}S_{eff}}{\delta\mu^{2}}=\lim_{\left|\mathbf{k}\right|\rightarrow0}\frac{\delta^{2}S_{eff}(A^{ex})}{\delta A_{0}^{ex}(\omega=0,\mathbf{k})\delta A_{0}^{ex}(0,0)}=\frac{\left|\mathbf{k}\right|^{2}}{2g^{2}}.$$ One can find that the ratio between the conductivity and susceptibility matches the Einstein relation,$$\frac{\sigma}{\chi}=\frac{\omega}{\left|\mathbf{k}\right|^{2}}=D,\label{eq:einstein-relation}$$ in which $D$ is the charge diffusion constant. For a special interest, the value of $D$ can be determined on self-dual configurations shown as follows.
An important observation to the Maxwell action of the system is that a lower bound can be realized on self-dual configurations $F_{\mu\nu}=\tilde{F}_{\mu\nu}=\frac{1}{2}\epsilon_{\mu\nu\rho\sigma}F_{\rho\sigma}$, it is known as the Bogomol’nyi bound, $$\frac{1}{4g^{2}}\int d^{4}xF_{\mu\nu}F^{\mu\nu}\geqslant S_{S.D.}\overset{\mathrm{Self-Dual}}{\equiv}\frac{8\pi^{2}\left|Q\right|}{g^{2}},$$ where $Q=\frac{1}{16\pi^{2}}\int d^{4}xF_{\mu\nu}\tilde{F}_{\mu\nu}\in\mathbb{Z}$ is the instanton number.
As a first consequence of the self-duality, it leads to a relation $\omega^{2}=\left|\mathbf{k}\right|^{2}$ in Euclidean metric for the self-dual zero mode. Thus the charge diffusion constant $D$ in Eq.(\[eq:einstein-relation\]) equals to $D=1/\omega$ on the self-dual configurations, which at the lowest Matsubara frequency $\omega=2\pi T$ takes value $D=\frac{1}{2\pi T}$, agreeing with [@policastro2002ads]. The smallness of the diffusion constant is because it stems from the zero mode of the fluids, but rather the excited states.
If we set $g=4\pi/N$ and replace $\omega$ by the lowest Matsubara frequency $\omega=2\pi T$, then the conductivity and susceptibility agree with [@policastro2002ads]$$\sigma=\frac{N^{2}T}{16\pi},\quad\chi=\frac{N^{2}\left|\mathbf{k}\right|^{2}T}{16\pi\omega}.$$
The second consequence of the self-duality is about the shear viscosity, which is related to component of correlation function of stress-energy tensor being a response to metric. Differ from the topological terms, the Maxwell term is the only term in the effective action containing metric $g_{\mu\nu}$. Strictly speaking, all physics of QCD are encoded in $Z[a]$, to derive shear viscosity of quark-gluon liquids, we need to vary with respect to the metric perturbations of Maxwell term in $Z[a]$, but when $a$ and $b$ fields are integrated out, it is equivalent to vary with respect to the metric perturbation in the effective action $S_{eff}(A^{ex})=-\ln Z[A^{ex}]$. We will rewrite the action by explicitly including the metric (in order to distinguish between the gauge coupling and metric, we insteadly use $g_{*}$ to represent the gauge coupling),$$S=\frac{1}{4g_{*}^{2}}\int d^{4}x\sqrt{g}F_{\alpha\beta}F^{\alpha\beta},$$ then doing the functional derivative with respect to the metric and finally taking the limit of flat metric back $$\lim_{\sqrt{g}\rightarrow1}\frac{\delta^{2}S_{eff}(A^{ex})}{\delta g_{xy}\delta g_{xy}}=\frac{1}{4g_{*}^{2}}\left(F_{xy}F_{xy}-\frac{1}{2}F_{\alpha\beta}F^{\alpha\beta}\right).$$
By considering the quark-gluon plasma is homogeneous and isotropic ($\left|F_{xy}\right|=\left|F_{xz}\right|=\left|F_{yz}\right|$), so that the component satisfies $F_{xy}F_{xy}=\frac{1}{6}F_{\alpha\beta}F^{\alpha\beta}$. For the field strengths in the action are homogeneous, the two-point correlation function of stress-energy tensor $G_{xy,xy}$, is just given by the average action density,$$G_{xy,xy}(\omega,k)=\lim_{\sqrt{g}\rightarrow1}\frac{\delta^{2}\ln Z[A^{ex}]}{\delta g_{xy}(\omega,k)\delta g_{xy}(0,0)}\geqslant\frac{8\pi^{2}\left|Q\right|}{3g_{*}^{2}V_{4}},$$ in which $V_{4}$ is a Euclidean 4-volume defined as $\ln Z(x)=V_{4}\int\frac{d^{4}k}{(2\pi)^{4}}\ln Z(k)$. Then, by using the Kubo formula, the minimal value of action gives lower bound to a shear viscosity$$\begin{aligned}
\eta & =\lim_{\omega\rightarrow0}\frac{1}{2\omega}G_{xy,xy}(\omega,k)\geqslant\frac{2\pi\left|Q\right|}{3g_{*}^{2}V_{3}}.\end{aligned}$$ in which we have used $\lim_{\omega\rightarrow0}\frac{1}{\omega V_{4}}=\frac{1}{2\pi V_{3}}$. An interesting observation is that the shear viscosity is proportional to the instanton density $\rho_{ins}=\left|Q\right|/V_{3}$, or related gluon condensate.
On the other hand, considering the thermodynamic relation for the entropy density $s=\partial p/\partial T$, moreover, because the the system is isotropic, $p_{x}=p_{y}=p_{z}=p$, then the pressure is related to the energy density by $\epsilon=3p$, which can be given by the minimal value of the action $\epsilon=-\frac{T}{V_{3}}\ln Z=\frac{T}{V_{3}}S_{S.D.}$. Finally we have $$s=\frac{8\pi^{2}\left|Q\right|}{3g_{*}^{2}V_{3}}.$$ The entropy density averaged shear viscosity is a dimensionless quantity, which measures a pure quantum (zero temperature) originated intrinsic viscosity,$$\frac{\eta}{s}=\frac{1}{4\pi},$$ in a quark-gluon plasma with self-dual gauge field configurations.
It is worth emphasizing that the self-duality here is crucial for not only achieving the small result of diffusion constant $D$ but also the $\eta/s$. It reflects the fact that the small value of these quantities are closely related to a mixture of equal importance of the electric and magnetic components of the plasma, which has been discussed in [@liao2007strongly; @liao2008magnetic-1; @liao2008magnetic-2; @liao2008magnetic-3]. This small value of $\eta/s$ is also obtained from another self-dual theory [@policastro2001shear; @policastro2002ads], the AdS/CFT correspondence, in which people conjectures that the constant is a universal lower bound. The assumption of isotropy is also important for the result, since without isotropic the lower bound of the action can not constrain each component of the shear viscosity, and hence certain component may be lower than $1/4\pi$ and break the conjecture.
Fractionalized Charge
---------------------
In this section we show that the fact (iii) is also a natural consequence of the effective theory. Charge fractionalization is an important mathematical feature of the topological action, and hence be a characteristic of the quark-gluon plasma phase being a topological phase. Following the form of the effective action Eq.(\[eq:effective-action\]), we drop the Maxwell term and focus on the topological action. Considering effective field action for each parton field (labelled by flavor index $I=1,2,...,\mathsf{m}$) are in its fractionalized topological order characterized by coefficient $\mathsf{m}$,$$\begin{aligned}
S^{(I)} & =\frac{i\mathsf{m}}{2\pi}\int d^{4}x\epsilon_{\mu\nu\alpha\beta}b_{\mu\nu}^{(I)}\partial_{\alpha}a_{\beta}^{(I)}+\frac{i\theta}{8\pi^{2}}\int d^{4}x\epsilon_{\mu\nu\alpha\beta}\partial_{\mu}a_{\nu}^{(I)}\partial_{\alpha}a_{\beta}^{(I)}+\int d^{4}xj_{\mu}^{(I)}A_{\mu}^{ex}\end{aligned}$$ in which $j_{\mu}^{(I)}=\frac{\mathsf{m}}{2\pi}\epsilon_{\mu\nu\alpha\beta}\partial_{\nu}b_{\alpha\beta}^{(I)}$ is the parton current. Like the Chern-Simons theory of the fractional quantum Hall fluids in which the Chern-number characterizes its topological order, the fractionalized topological order of each parton in the plasma is characterized by the integer number $\mathsf{m}$ in its action, which must be odd for the fermion statistic of each parton.
Since the effective coefficient $\mathsf{m}_{eff}$ in an effective action for total partons follows the composition law $\mathsf{m}_{eff}^{-1}=\sum_{I}\mathsf{m}_{I}^{-1}$ from the coefficient of each parton, hence we can prove that the effective action is given by$$Z[A^{ex}]=\int\prod_{I}\mathcal{D}a^{(I)}\mathcal{D}b^{(I)}\exp\left(-\sum_{I}S^{(I)}\right)=\int\mathcal{D}a\mathcal{D}b\exp\left(-S_{eff}\right)$$ in which$$S_{eff}=\frac{i}{2\pi}\int d^{4}x\epsilon_{\mu\nu\alpha\beta}b_{\mu\nu}\partial_{\alpha}a_{\beta}+\frac{i\theta}{8\pi^{2}}\int d^{4}x\epsilon_{\mu\nu\alpha\beta}\partial_{\mu}a_{\nu}\partial_{\alpha}a_{\beta}+\int d^{4}xj_{\mu}A_{\mu}^{ex},$$ with fractionalized charges and parton fields satisfying $$e=\sum_{I}e_{I}=1,\quad a_{\mu}=\sum_{I}a_{\mu}^{(I)},\quad b_{\mu\nu}=\sum_{I}b_{\mu\nu}^{(I)},\quad j_{\mu}=\sum_{I}j_{\mu}^{(I)}.$$
The observed effective charge $e$ is quantized by fundamental unit, but here the current $j_{\mu}$ seems as a composite current and can be splited into constitutes with fractionalized charges $e_{I}=e/\mathsf{m}$ in its fractionalized topological order. In this sense, the plasma phase is a deconfined phase. In the soup of quark-gluon plasma, the baryon with integer electric charge is splited into three individual quarks with fractionalized electric charges. In this framework, the exotic assumption that quarks have fractional electric charge ($e/3$) is a natural consequence of the deconfined topologically ordered phase characterized by the odd integer $\mathsf{m}=3$.
Reminding that there are two independent fluids in the plasma, the charge current and spin current. It is worth noting that the fractionalization does not apply to the spin current. Opposite to the direct coupling between the charge current and the external gauge field $A^{ex}$, the coupling between the spin density and the external magnetic field is indirect, mediating by the infinitely massive $b$ fields. The $b$ field, as is proved, could be splited into constitutes $b^{(I)}$ and hence the corresponding coupling between $b^{(I)}$ and external magnetic field $\tilde{F}^{ex}$ is fractionalized, however, the $\tilde{F}^{ex}$ can not couple to each individual spin constitute $s_{\mu\nu}^{(I)}=\frac{\mathsf{m}}{2\pi}\epsilon_{\mu\nu\alpha\beta}\partial_{\alpha}a_{\beta}^{(I)}$ since all intermediate constitute $b^{(I)}$ fields must be integrated out. Thus the coupling constant between spin density and external magnetic field is not fractionalized into smaller units, i.e. spin or angular momentum does not split.
Conclusions
===========
In this paper, we show remarkable connection of low energy effective field theory of quark-gluon plasma with topological quantum field theory, and with hydrodynamic theory. In many aspects the deconfined quark-gluon plasma phase of strong interaction system is analogous to the topological phase in condensed matter systems. In fact, the central argument of the paper is that it is indeed a topological phase [@Zhitnitsky:2013hs]. The low energy degrees of freedoms in the topological phase are incompressible quantum fluids which are analogous with the nearly perfect quark-gluon liquid around $T_{c}$ discovered at RHIC. Bosonization approach makes an effective description to the topological phase and low energy topological fluid possible, due to the fact that a gauge theory with massive fermions and the bosonized description share the same topological limit. Below the massive scale, which is expected several times of $T_{c}$, the bosonization approach is valid, and the effective action Eq.(\[eq:effective-action\]) is the central result of the paper, which is suggested as an infrared fixed point action for the QCD system coupling to an U(1) external source. In such a theory, the quark-gluon plasma is effectively described by two types of fluids, more precisely, by hydrodynamics continuity equations of charge current and spin/angular momentum current. The smallness of pressure and bulk viscosity of the quark-gluon plasma around $T_{c}$, as well as the charge fractionalization of the excitations are naturally explained from its topological nature. Both the electric and magnetic components are equally important and half-half mixing in the quark-gluon liquids, which is also topologically originated. Indeed, the electric-magnetic self-duality is closely related to achieving a small value of diffusion constant and $\eta/s$, making the quark-gluon liquids most perfect liquids.
This work was supported in part by the National Science Foundation of China (NSFC) under Grant No.11205149.
[34]{} natexlab\#1[\#1]{}bibnamefont \#1[\#1]{}bibfnamefont \#1[\#1]{}citenamefont \#1[\#1]{}url \#1[`#1`]{}urlprefix\[2\][\#2]{} \[2\]\[\][[\#2](#2)]{}
, in ** (, ), pp. .
, , , , ****, ().
, ****, ().
, ****, ().
, ****, ().
, ****, ().
, ****, ().
, ****, ().
, ****, ().
, ().
, ****, ().
, , , , ****, ().
, ****, ().
, ****, ().
, ****, ().
, ****, ().
, ****, ().
, ****, ().
, ****, ().
, , , ().
, ().
, ****, ().
, , , , ****, ().
, ****, ().
, ****, ().
, ****, ().
, ****, ().
, ****, ().
, , , ****, ().
, ().
, ****, ().
, ****, ().
, , , ****, ().
, ****, (), .
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Within our thick brane approach previously used to obtain the cosmological evolution equations on a thick brane embedded in a five-dimensional Schwarzschild Anti-de Sitter spacetime it is explicitly shown that the consistency of these equations with the energy conservation equation requires that, in general, the thickness of the brane evolves in time. This varying brane thickness entails the possibility that both Newton’s gravitational constant $G$ and the effective cosmological constant $\Lambda_4$ are time dependent.'
author:
- |
Samad Khakshournia$^{1}$ [^1]\
$^{1}$ Nuclear Science and Technology Research Institute (NSTRI),Tehran, Iran
title: '**A Note on the Generalized Friedmann Equations for a Thick Brane**'
---
$$$$ $$$$ $$$$
**Introduction**
================
Cosmological models in which the observed universe is realized as a four-dimensional thin brane embedded in a higher dimensional spacetime (bulk) are of current interest. In these theories the cosmological evolution on the brane is described by the effective Friedmann equations that incorporates nontrivially the effects of the bulk. The most important feature that distinguishes brane cosmology from the standard scenario is the fact that at high energies Friedmann equation modifies by an extra term quadratic in the energy density of matter on the brane [@Bin99]. Several features and variations of this scenario including thick brane configurations, in particular, in the cosmological context have been considered. In the approach used in Ref. [@Moun] to derive generalized Friedmann equations, the four-dimensional effective brane quantities are obtained by integrating the corresponding five-dimensional ones along the extra-dimension over the brane thickness. These cosmological equations describing a brane of finite thickness interpolate between the case of an infinitely thick brane corresponding to the familiar Kaluza-Klein picture and the opposite limit of an infinitely thin brane giving the unconventional Friedmann equation, where the energy density enters quadratically. The latter case is then made compatible with the conventional cosmology at late times by introducing and fine tuning a negative cosmological constant in the bulk and an intrinsic positive tension in the brane [@whole1]. Authors in Ref. [@Nav05] considered a thick codimension 1 brane including a matter pressure component along the extra dimension in the energy-momentum tensor of the brane. By integrating the 5D Einstein equations along the fifth dimension, while neglecting the parallel derivatives of the metric in comparison with the transverse ones, they wrote the equations relating the values of the first derivatives of the metric at the brane boundary with the integrated componentes of the brane energy-momentum tensor. These, so called matching conditions were then used to obtain the cosmological evolution of the brane which is of a non-standard type, leading to an accelerating universe for special values of the model parameters. On the other hand the cosmological implications of brane world scenario in which fundamental physical constants are variable functions of time have been investigated [@Gvar1; @Gvar2]. For instance, in Ref. [@Gvar1] was shown that the introduction of a scalar function varying with time in the brane-world theory yields a number of cosmological models which do not admit constant values for the Newton’s gravitational coupling $G$ and the effective cosmological term $\Lambda_4$.\
We are interested in the effects arising from taking into account the thickness of the brane. In this way, we have presented a general equation to study the dynamics of codimension one brane of finite thickness immersed in an arbitrary bulk spacetime. That was obtained in a general setting by imposing the Darmois junction conditions on the brane boundaries with the two embedding spacetimes and then using an expansion scheme for the extrinsic curvature tensor at the brane boundaries in terms of the proper thickness of the brane. Using this approach we have derived the generalized Friedmann equations written up to the first order of the brane proper thickness governing the cosmological evolution of a thick brane embedded in a five-dimensional Schwarzschild Anti-de Sitter spacetime with a $Z_{2}$ symmetry [@ghkhm].\
The aim of this note is to address the presence of the brane with a time dependent thickness as a general feature of our thick brane cosmological solutions.\
In section 2 we give a brief review of the generalized cosmological equations as obtained in Ref. [@ghkhm]. In section 3 we derive an evolution equation for the brane thickness as required from the self-consistency of our equations. In section 4 conclusion is presented.
**Generalized Cosmological Equations**
======================================
We consider a thick brane embedded in a five-dimensional Schwarzschild Anti-de Sitter spacetime with the metric $$\label{metout}
ds^2=-f(r)dT^2+\frac{dr^2}{f(r)}+r^2d\Omega_{k}^{2},$$ with $$\label{f}
f(r)=k-\frac{\Lambda}{6}r^2-\frac{C}{r^2},$$ where $d\Omega_{k}^{2}$ is the metric of the 3D hypersurfaces $\Sigma$ of constant curvature that is parameterized by $k=0, \pm
1$; $\Lambda$ is the bulk cosmological constant, and constant $C$ is identified with the mass of a black hole located at $r=0$. We then take the following ansatz for the metric of the brane being written in a Gaussian normal coordinate system in the vicinity of the core of the thick brane situated at $y=0$ $$\label{metin}
ds^2=-n^2(t,y)dt^2+dy^2+a^2(t,y)d\Omega_{k}^{2},$$ where $n^2(t,y)$ and $a^2(t,y)$ are some unknown functions, and $y$ is the normal coordinate of the extra dimension. Compatibility with the cosmological symmetries requires that the energy-momentum tensor of the matter content in the brane takes the simple form $$\label{stress}
T^{\mu}_{\nu}=(-\rho,P_L,P_L,P_L,P_T),$$ where the energy density $\rho$, the longitudinal pressure $P_L$, and the transverse pressure $P_{T}$ are functions of $t$ and $y$.\
The generalized first Friedmann equation written up to first order of the brane thickness takes the following form within our formalism [@ghkhm] $$\label{f1}
\sqrt{f_{0}+\dot{a}_{0}^2}=\frac{w}{a_{0}}\left(\frac{-\Lambda}{3}a_{0}^2
+\frac{\kappa^2}{3}\rho_{0}a_{0}^2\right),$$ where the subscript “0” means evaluation at the center of the thick brane and the dot stands for the derivative with respect to the proper time $\tau$ on the brane center. Taking the square of Eq. (\[f1\]), substituting the expression (\[f\]) and rearranging, we arrive at the following equation $$\label{eqs}
H_{0}^2+\frac{k}{a_{0}^2} = \frac{8\pi
G}{3}\varrho+\frac{\kappa^4}{36}\varrho^2
+\frac{\Lambda_4}{3}+\frac{C}{a_{0}^4},$$ where $H_0=\frac{\dot{a}_{0}}{a_{0}}$, and the effective four-dimensional energy density $\varrho$ associated to the five-dimensional energy density $\rho$ has been defined as $$\label{rho}
\varrho=\int_{-w}^{w} \rho dy\simeq 2w\rho_{0}+O(w^2),$$ and the following identifications have been made $$\begin{aligned}
\label{effect}
\frac{\Lambda_4}{3}=\frac{\Lambda}{6}
+\frac{w^2\Lambda^2}{9},\hspace{1cm} 8\pi G =
\frac{\kappa^2w(-\Lambda)}{3}.\end{aligned}$$ We see that there is a linear in addition to a quadratic term in the matter density, due to the non-vanishing of the thickness $w$. There is no need of introducing an ad hoc tension for the brane, and splitting it from the matter density on the brane.\
The generalized second Friedmann equation written up to first order of the brane thickness takes the following form within our formalism [@ghkhm] $$\begin{aligned}
\label{acc}
\frac{\frac{\ddot{a}_{0}}{a_{0}}-\frac{\Lambda}{6}
+\frac{C}{a_{0}^4}}{\sqrt{f_{0}+
\dot{a}_{0}^2}}&=&\frac{-w}{a_{0}}\left(
\frac{2\kappa^2}{3}\left(\rho_{0}
+\frac{3}{2}P^{0}_{L}-P^{0}_{T}+\frac{3}{a_{0}^4}\tilde{P_T}\right)
+\frac{\Lambda}{6}+\frac{3E}{a_{0}^4} +\frac{3C}{a_{0}^4}\right),\end{aligned}$$ where $E>0$ is an integration constant, $P^{0}_{L}=P_{L}(t,y=0)$, and $P^{0}_{T}=P_{T}(t,y=0)$, and we have also defined $\tilde{P_T}\equiv\int_{0}^{\tau} P^{0}_T\:a^4_0\:H_0d\tau$. Note now that the time component of the covariant derivative of the brane energy-momentum tensor (\[stress\]), using the metric (\[metin\]), leads to the familiar energy conservation condition on the core of the thick brane: $$\label{conser}
\dot{\rho_{0}}+3H_{0}(\rho_{0}+P^0_{L})=0.$$ Defining the four-dimensional effective quantities associated to the five-dimensional longitudinal and transverse pressures $P_{L}$ and $P_{T}$ in the form $$\label{pl}
p_{L}=\int_{-w}^{+w}P_{L}\:dy\simeq 2wP^0_{L}+O(w^2),$$ $$\label{pt}
p_{T}=\int_{-w}^{+w}P_{T}\:dy\simeq 2wP^0_{T}+O(w^2),$$ Let us assume the arbitrary effective equations of state of the form $$\label{sta}
p_{L}=\omega_{L}\varrho,\hspace{1cm}p_{T}=\omega_{T}\varrho,$$ with constants $\omega_{L}$ and $\omega_{T}$. The conservation equation (\[conser\]) can then be integrated with the result as usual $$\label{conser2}
\rho_{0}=\rho_{i}a_{0}^{-3(1+\omega_{L})},$$ where $\rho_{i}$ is a constant. Then $\tilde{P_T}$ can be computed as $$\label{tildePevol}
\tilde{P_T}=\frac{\rho_{i}\omega_{T}}{1-3\omega_{L}}a_{0}^{1-3\omega_{L}}.$$
**Evolution of Brane Thickness**
================================
Let us now write down the time derivative of the first Fridmann equation (\[f1\]) while allowing the brane thickness $w$ to evolve in time. We get $$\begin{aligned}
\label{deriv}
\frac{\frac{\ddot{a}_{0}}{a_{0}}-\frac{\Lambda}{6}
+\frac{C}{a_{0}^4}}{\sqrt{f_{0}+
\dot{a}_{0}^2}}&=&\frac{-w}{a_{0}}\left(
\frac{2\kappa^2}{3}\left(\rho_{0} +\frac{3}{2}P^{0}_{L}\right)
+\frac{\Lambda}{3}\right)+\frac{\dot{w}}{\dot{a}_{0}}
\left(-\frac{\Lambda}{3}+\frac{\kappa^2}{3}\rho_{0}\right),\end{aligned}$$ where the energy conservation equation (\[conser\]) has been used. Comparing this with the second Fridmann equation (\[acc\]) we infer $$\begin{aligned}
\label{deriv3}
\frac{\dot{w}}{w}=\frac{-H_{0}}{\frac{-\Lambda
}{3}+\frac{\kappa^2}{3}\rho_{0}}\left(\frac{2\kappa^2\rho_{0}}{3}
\left(\frac{3\omega_{T}}{1-3\omega_{L}}-\omega_{T}\right)+\left(\frac{3E}{a_{0}^4}
+\frac{3C}{a_{0}^4}+\frac{-\Lambda}{6}\right)\right),\end{aligned}$$ where we have used Eqs. (\[rho\]), (\[pl\]), (\[pt\]), (\[sta\]), (\[conser2\]), and (\[tildePevol\]). This is an evolution equation for the brane thickness indicating, in general, the presence of the brane with a time dependent thickness. Depending on the magnitude and sign of the different terms within the bracket in Eq. (\[deriv3\]), the brane thickness may increase or decrease with time. For instance, we observe that for $\omega_{T}$=0, the brane thickness extended along the extra dimension always decreases with time while its ordinary spacial dimensions expand if $k=0$, or $-1$. On the other hand for $\omega_{T}<0$ leading to an accelerating brane cosmology at late time [@ghkhm], from Eq. (\[deriv3\]) one can see that it is possible to have an increasing brane thickness with time. Furthermore, as the equation (\[deriv3\]) poses, the parameters within the bracket must be finely tuned to yield a time constant brane thickness.
**Conclusion**
==============
In this note we showed that the requirement of compatibility our thick brane cosmological equations written up to the first order of the brane thickness with the familiar energy conservation equation leads to an evolution equation for the brane thickness. In the absence of the pressure along the extra dimension in the brane energy-momentum tensor the thickness of the brane decreases with time while at late time a negative transverse pressure can lead to an increase in the brane thickness. We have realized that our two generalized Friedmann equations (\[f1\]) and (\[acc\]) together with the energy conservation equation (\[conser\]) form the three independent equations to determine the evolution of the three unknown functions $a_{0}$, $\rho_{0}$, and $w$ once one gets the equations of state (\[sta\]) for brane matter. According to the identifications made in (\[effect\]) a time dependent brane thickness induces time variations in the Newton’s gravitational coupling $G$ and the effective cosmological term $\Lambda_4$. This suggests our thick brane cosmological model as a general framework in which one enables to study the simultaneous variation of $G$ and $\Lambda_4$.
Acknowledgments
===============
The author would like to thank Professor R. Mansouri and S. Ghassemi for fruitful comments that led to the development of this work.
[99]{} P. Binétruy, C. Deffayet and D. Langlois, Nucl. Phys. B [**565**]{}, 269 (2000) \[arXiv:hep-th/9905012\]. P. Mounaix and D. Langlois, Phys. Rev. D [**65**]{}, 103523 (2002). P. Binétruy, C. Deffayet, U. Ellwanger and D. Langlois, Phys. Lett. B [**477**]{}, 285 (2000) \[arXiv:hep-th/9910219\]; D. Ida, JHEP [**0009**]{}, 014 (2000) \[arXiv:gr-qc/9912002\]. I. Navarro, J. Santiago, JCAP [**0603**]{}, 015 (2006) \[arXiv:hep-th/0505156\]. J. Ponce de Leon, Class. Quant. Grav. [**20**]{}, 5321 (2003) \[arXiv:gr-qc/0305041\]. Soma Nath, Subenoy Chakraborty, and Ujjal Debnath \[arXiv:gr-qc/0411111\]. Sima Ghassemi, Samad Khakshournia, and Reza Mansouri, JHEP [**08**]{}, 019 (2006) \[arXiv:gr-qc0605094\].
[^1]: Email Address: [email protected]
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Experimental evidence from RHIC indicates that matter having an energy density far in excess of the value required for the creation of a deconfined phase is produced in ultrarelativistic Au+Au collisions at $\sqrt{s_{_{\rm NN}}}$=200 GeV. This matter thermalizes rapidly, is strongly interacting, and displays hydrodynamic properties akin to a fluid with very low viscosity. Studies of the interaction of hard scattered partons with this matter provide an important probe of its properties. In fact, studies of the suppression of high-p$_T$ hadrons have been one of the key arguments used by the RHIC experiments that the matter created is something fundamentally new and exciting. I review the current experimental picture that has emerged from the data collected at RHIC, with an emphasis on the effort to extract quantitative measurements from the suppression of high-$p_T$ hadrons.'
author:
- John Lajoie$^1$
title: 'Jets and Parton Energy Loss at RHIC: Experimental Status'
---
Jets As Partonic Probes
=======================
The hard scattering of partons in the initial state of an ultrarelativistic heavy-ion collision holds great promise as a probe of the hot, dense matter created in these collisions. The initial hard scatterings occur very early in the collision process ($\tau <$ 1 fm/c) and are sensitive to the evolution of the surrounding, color-charged medium primarily through energy loss mechanisms (either collisional or radiative)[@white_papers]. These hard-scattered partons are not observed by themselves in the final state, however, as they fragment into collimated “jets” of color-singlet hadrons, clustered around the initial parton direction. Experimentally, jets can be identified and studied through the correlation of hadrons in the final state, or through the single inclusive spectra of high-$p_T$ particles. In this contribution I will limit the discussion to what we have learned about the properties of the medium created in heavy ion collisions through the use of single inclusive spectra. The properties of the medium that can be inferred through particle correlations are discussed in another contribution to this conference [@Olga].
Hard Probes and pQCD
--------------------
In order to use hard scattering as a probe of nuclear matter we need to have a solid understanding of three main components: initial state effects, the parton-level cross section, and final state effects (exclusive of the effect of the matter itself). In the collinear factorization ansatz, the cross section for the production of high-$p_T$ particles can be written as:
$$\begin{aligned}
\lefteqn{\frac{d\sigma^{h}_{n_{1}n_{2}}}{dyd^{2}p_{T}} = } \nonumber \\
& & K\sum_{abcd}\int dx_{a}dx_{b}f_{a/n_{1}}(x_{a},Q^{2})f_{b/n_{2}}(x_{b},Q^{2}) \times \nonumber \\
& & \frac{D_{h/c}(z_{c},Q^{2})}{\pi z} \frac{d\sigma}{d\hat{t}}(ab\rightarrow cd).\end{aligned}$$
Initial state effects include the distribution of partons in the initail nuclei, as represented by the function $f_{a/n}(x_{a},Q^{2})$. For example, a common initial state effect is the multiple scattering of partons prior to the hard scattering event, which results in a significant additional transverse momentum of the final state hadrons (the “Cronin effect” [@cronin]). These effects can be identified through the use of collisions like d+Au collisions at RHIC, where the final state effects are expected to be small. In a similar way, our understanding of the partonic level cross section $\frac{d\sigma}{d\hat{t}}(ab\rightarrow cd)$ can be explored through a comparison between proton-proton collisions and pQCD calculations. In using jets as a probe of the QCD medium, this can be thought of as “calibrating” the production of these probes in the initial phase of these collisions. Final state effects, primarily energy loss as the parton traverses the medium, can be thought of as effectively modifying the fragemtation function $D_{h/c}(z_{c},Q^{2})$, which determines the statistical probability of obtaining a hadron $h$ from the fragmentation of a parton $c$, where the hadron has a longitudidinal momentum fraction $z_{c}$ of the inital parton. Fragmentation functions have been studied and determined in the relatively clean final states of $e^{+}e^{-}$ collisions. Final state effects such as energy loss can be expected to lower the energy scale of the fragmentation function, as well as increase the probability of finding soft hadrons resulting from gluon emission.
An important first question to ask is whether or not pQCD is applicable at all at RHIC energies. This has been addressed by both the STAR and PHENIX collaborations by comparing pQCD calculations with inclusive $\pi^{0}$ and jet spectra in proton-proton collisions [@PHENIX_pi0_pQCD; @STAR_jets_pQCD]. The good agreement between these pQCD calculations and experimental data indicates the pQCD is broadly applicable at RHIC, and that we have a theoretical handle on the initial state (at least in proton-proton collisions). In addition, the demonstration of $x_T$ scaling at RHIC energies in Au+Au collisions confirms that high-$p_T$ hadrons originate from hard scatterings in nuclear collisions as well [@PHENIX_xT].
The Nuclear Modification Factor
-------------------------------
In order to understand the effect of hot, dense matter on the production of high-$p_T$ particles, we construct a ratio known as the “nuclear modification factor” $R_{AA}$:
$$R_{AA}(p_{T}) = \frac{dN^{AA}/dp_{T}}{<N_{binary}>dN^{pp}/dp_{T}}.$$
The number of binary collisions for a given centrality class $<N_{binary}>$ is estimated using a Glauber model of the nuclear overlap. This overlap is a function of the centrality of the collision, defined as a fraction of the total inelastic cross section. For example, the “0-10%” most central collisions are the 10% of collisions, averaged over impact parameter, that have the most nuclear overlap. In Au+Au collisions at [$\sqrt{s_{NN}}$ = 200 GeV ]{}$<N_{binary}>\sim 900$. In the niave limit that a nucleus-nucleus collision can be though of as nothing more than a superposition of independent nucleon-nucleon collisions, the nuclear modification faction at high-$p_{T}$ should be unity.
The measured nuclear modification factor for $\pi^{0}$ and $\eta$ mesons, as well as direct photons from the PHENIX collaboration is shown in Figure \[raa\]. There is a significant suppression of high transverse momentum $\pi^{0}$ and $\eta$ mesons observed in Au+Au collisions, consistent with substantial energy loss of the scattered partons. A similar suppression is observed for charged particles as well as data taken at [$\sqrt{s_{NN}}$ = 130 GeV ]{}[@phenix_highpt]. In contrast, direct photons show no suppression. This is because the photon contains no net color charge, and therefore is able to escape the medium with interacting strongly. The fact that the $R_{AA}$ for direct photons is $\sim$ 1 is another indication that initial state effects are well under control in Au+Au collisions, and that the initial rate of hard scatterings is properly “calibrated” by proton+proton measurements.
Models of energy loss that incorporate the expansion of the system indicate the energy loss in the matter created in heavy-ion collisions at RHIC may be as much as fifteen times larger than the energy loss of a comparable parton in ordinary nuclear matter. Since the energy loss is proportional to the gluon density, this implies the gluon density created in a heavy ion collision is more than an order of magnitude larger than in cold nuclei. However, the two leading models for the energy loss of light quarks both different approaches to model the energy loss by gluon radiation. The BDMPS model [@salgado] characterizes the energy loss by a transport parameter, $\hat{q}$, which is the average momentum transfer squared imparted to the parton per unit length of the medium transversed. Formally, this parameter enters via an integral over the spectrum of emitted gluons, which is related to the medium density of color charges. The value of this parameter, averaged over medium expansion, that best fits the data is $\hat{q}$ between 5 and 15 $GeV^{2}/fm $ (see Figure \[fig:salgado\]). In the GLV model [@GLV] the energy loss is characterized by the initial medium density of gluons $dN_{G}/dy$. The observed suppression in Au+Au collisions is best fit by a value $dN_{G}/dy \simeq 1000$ gluons per unit rapidity. It is tempting to attempt to interpret these model parameters as directly related to the properties of the medium. This interpretation is called into question, however, by the fact that the authors of these models do not agree on the correspondence between $\hat{q}$ and $dN_{G}/dy$. For example, one of of the authors of the GLV model claims the correspondence [@GLV_corr]:
$$\frac{dN_{G}}{dy} = 900 \Leftrightarrow \overline{\hat{q}} = 0.35 - 0.85 GeV^{2}/fm$$
(where the $\overline{\hat{q}}$ average is done over a 1+1D Bjorken expansion) while the the correspondence claimed by a BDMPS author is [@BDMPS_corr]:
$$\overline{\hat{q}} = 5 GeV^{2}/fm \Leftrightarrow \frac{dN_{G}}{dy} = 900.$$
While both models provide a reasonable description of the data, the disagreement between the two approaches makes it questionable to interpret either the BDPMS opacity $\overline{\hat{q}}$ of GLV initial gluon density $dN_{G}/dy$ as measurements of the physics properties of the QCD medium created at RHIC.
In addition to uncertainties in the interpretation of the model calculations, the fact that both models rely on radiative energy loss only has recently been called into question [@collision_pap1], as will be discussed in the next section. Finally, it is important to note that the large suppression observed in high-$p_T$ hadrons implies a very large energy loss for light quarks. This in turn implies that the energy loss measurements we make with light quarks are biased to the surface of the produced medium, and are not necessarily probing the medium in bulk. Ideally, we would like a set of differential probes that are sensitive to the strongly interacting matter at different effective depths, as shown in Figure \[fig:wicks1\].
The simple model calculations in Figure \[fig:wicks1\] make apparant that differential probes of the medium are required, and this aspect can be achieved by combining light quark energy loss measurements with measurements of the energy loss of charm and botto quarks. A combination of such probes will allow better characterization of the QCD matter in bulk, and minimize the surface bias of the light quark data.
Heavy Quarks
============
The measurement of the production of heavy quarks is an exciting probe of heavy-ion collisions and a strong focus of the future of the RHIC physics program. The energy loss of heavy quarks is suppressed relative to light quarks through interference of the forward gluon emission amplitudes, resulting in a modified gluon emission spectrum $dI/d\omega$ [@annu_rev_bdmps]:
$$\omega \frac{dI}{d\omega}|_{HEAVY} = \frac{\omega \frac{dI}{d\omega}|_{LIGHT}}{1+(\frac{m_Q}{E_Q})\frac{1}{\theta^{2}}}$$
(the so-called “dead cone” effect). However, in order to characterize the suppression of heavy quarks in nucleus-nucleus collisions an overall knowledge of the total yield in proton-proton collisions at RHIC energies is required. At the present time the measurement of heavy quarks at high-$p_T$ is difficult at RHIC because neither STAR nor PHENIX can identify the displaced vertex associated with the decay of charm mesons. The production of charm quarks is inferred from the measurement of non-photonic electrons, as described in the next section.
Inclusive Electrons
-------------------
The PHENIX and STAR collaborations have measured the production of single inclusive electrons and positrons from non-photonic sources using a meson cocktail subtraction [@STAR_npe; @PHENIX_new_npe]. The PHENIX experiment cross-checks the cocktail results with a set of special runs with an additional photon converter surrounding the beam pipe. Because of the low-mass design of the detector, the PHENIX experiment has a substantially higher signal to background than STAR at high-$p_T$. The remaining $e^{+}$ and $e^{-}$ sample is assumed to originate from the semileptonic decay of heavy quarks. The transverse momentum distribution of electrons and positrons from non-photonic sources from PHENIX and STAR is shown in Figure \[fig:phenix\_vs\_star\] for proton-proton collisions at [$\sqrt{s_{NN}}$ = 200 GeV ]{}. It is important to note that, at the present time, there is a substantial disagreement between the two experiments with the STAR data roughly a factor of two to three larger than the PHENIX data at high-$p_T$. Until this difference is resolved it will be difficult to properly evaluate the suppression of heavy quarks in heavy-ion collisions.
One interesting comparison that can be made is to compare the non-photonic electron measurements with QCD FONLL (Fixed Order Next To Leading Logartithm) calculations, in much the same way we evaluated the applicability of pQCD at RHIC in the beginning of this article. These calculations show a reasonable agreement for $J/\Psi$ production in CDF, but the production of $D^0$ mesons is lower than the data by roughly a factor two [@CDF_fonll]. Figure \[fig:electrons\] shows a comparison between the PHENIX data and FONLL calculations, while the same is shown for the STAR data in Figure \[fig:star\_electrons\]. It is interesting to note that while the PHENIX data shows a similar behavior to the CDF $D^0$ comparison, namely the data is larger than the FONLL calculation by a factor of $\sim$1.5, the STAR data is a factor of $\sim$5 greater than the calculation. In both cases, however, the calculations seem to be in fair agreement with the shape of the data as a function of $p_T$.
The transverse momentum distribution of electrons and positrons from non-photonic sources from PHENIX and STAR is shown in Figure \[fig:wicks2\] for 10% Au+Au collisions at [$\sqrt{s_{NN}}$ = 200 GeV ]{}. Note that both the PHENIX and STAR data for $R_{AA}$ were calculated using earlier proton-proton reference data. This data is shown compared to model calculations of the energy loss of heavy quarks from radiative processes only, and a combination of gluon radiation and elastic collisional losses combined with pathlength fluctuations due to collsion geometry. One key point that Figure \[fig:wicks2\] makes apparant is that collisional losses were underestimated in models of light quark energy loss, and a re-examination of these models is called for. Even so, the range of theoretical values for $R_{AA}$ for heavy quarks is higher than the data from both STAR and PHENIX, and only consistent with model calculations that include only charm quarks. This observation, however, may be purely coincidental and if taken seriously must explain the lower than expected production of bottom quarks at RHIC energies.
The energy loss of heavy quarks is, at present, an intriguing puzzle in the RHIC data. While further refinement of existing measurements by STAR and PHENIX will eventually resolve the discrepancies in the non-photonic electron data, additional detector upgrades will be required before charm and bottom quark contributions can be separated and studied independently. Both experiments have upgrades planned (the PHENIX VTX and STAR Heavy Flavor Tracker) that will add additional high-resolution tracking very close to the beam pipe. These upgrade detectors, combined with expected improvements in luminosity between RHIC I and RHIC II, will enable high statistics studies of the propagation of both charm and bottom quarks in the QCD medium of heavy-ion collisions.
Conclusions
===========
This initial work at RHIC studying the suppression of high-$p_T$ hadrons in ultrarelativistic heavy-ion collisions is one of the key pieces of evidence to demonstrate that the matter created in these collisions is something fundamentally new, and very exciting. Model calculations of the suppression based on the radiative losses of light quarks as they traverse a colored medium have demonstarted good agreement with the data for different colliding systems, centralities and energies. However, they have significant shortcomings. The leading model calculations - BDMPS and GLV - differ in the greatly when trying to draw a correspondence between the parameters of these models, which calls into question the wisdom of interpreting these parameters as representative of physical properties of the medium. In addition, it now apprears that collisional losses cannot be fully neglected even for light quarks, and further refinement of the energy loss models, at least at low-$p_T$, is necessary. Finally, studies of light quark energy loss are inherently biased towards the surface of the created medium because their interactions with the QCD matter are so strong.
Studies of heavy quark energy loss offer a less strongly interaction probe of the medium, mainly because of the suppression of forward gluon radiation due to the “dead cone” effect. Using the charm and bottom quarks as probes will allow an investigation of the medium along a greater pathlength and can potentially yield additional information about the character of the QCD matter. Early studies at RHIC have made use of the detection of non-photonic electrons to study the production of heavy quarks, but current measurements from the STAR and PHENIX collaborations are in disagreement about non-photonic electron measurements in proton-proton collisions. It will be difficult to draw conclusions about suppression in heavy ion collisions until this disagreement is resolved. In the future, detector upgrade to PHENIX and STAR and luminosity upgrades at RHIC will allow systematic studies of heavy quark energy loss and the explicit separation of charm and bottom quark contributions, completing a full set of differential parton probes that can be used to study the hot, dense matter created in ultrarelativistic heavy ion collisions.
[**Acknowledgements**]{}
The author would like to thank the organizers for the invitation to participate in this conference. This work was supported by a grant from the US Department of Energy DE-FG02-92ER40692.
[99]{} For an introduction to the first three years of RHIC physics, see the RHIC “white papers”: Nuclear Physics A Volume 757, Issues 1-2 , 8 August 2005, Pages 1-27 (BRAHMS), Pages 28-101 (PHOBOS), Pages 102-183 (STAR) and Pages 184-283 (PHENIX). See the talk of Olga Barranikova in this conference. J. W. Cronin [*et al.*]{}, Phys. Rev. [**D11**]{}, 3105 (1975). F. Simon, arXiv:hep-ex/0604028. S. Adler [*et al.*]{}, Phys. Rev. Lett. [**91**]{}, 241803 (2003). S. Adler [*et al.*]{}, Phys. Rev. C [**69**]{}, 034910 (2004). K. Adcox [*et. al*]{}, Phys. Rev. Lett [**88**]{}, 022301 (2002). Eskola, Honkanen, Salgado and Weidemann, Nucl. Phys. A [**747**]{}, 511 (2005). M. Gyulassy, P. Levai and I. Vitev, Phys. Lett. B [**594**]{}, 371 (2001). I. Vitev, Hard Probes 2006. U. Wiedemann, Hard Probes 2006. M. G. Mustafa, Phys. Rev. C [**72**]{}, 014905 (2005). Wicks [*et. al*]{}, arXiv:nucl-th/0512076. R, Baier, D. Schiff, and B. G. Zakharov, Annu. Rev. Nucl. Part. Sci. 2000. 50:37-69 Phys. Rev. Lett. [**91**]{}, 241804 (2003). Phys. Rev. Lett. [**96**]{}, 032011 (2006). A. Adare [*et al.*]{}, arXiv:hep-ex/0609010. Abelev [*et. al*]{}, arXiv:nucl-ex/060712.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Given a sequence of regular planar domains converging in the sense of kernel, we prove that the corresponding Green’s functions converge uniformly on the complex sphere, provided the limit domain is also regular, and the connectivity is uniformly bounded.'
address:
- 'School of mathematical sciences, Shanghai Jiao Tong University, 800 Dongchuan RD, Shanghai 200240, China'
- '215 Carnegie, Mathematics Department, Syracuse University, Syracuse, NY 13244, USA'
author:
- Sergei Kalmykov
- 'Leonid V. Kovalev'
bibliography:
- 'amsrefs.bib'
title: 'Uniform convergence of Green’s functions'
---
Introduction
============
Convergence theorems for holomorphic maps, harmonic functions, and Green’s functions usually concern their uniform convergence on compact subsets of a domain (e.g., [@Duren; @Pommerenke]). However, their uniform convergence on all of the domain is also of interest. A classical result on the uniform convergence of normalized conformal maps $f_n\colon \mathbb D\to\mathbb C$ is a 1922 theorem of Radó [@Rado], which related it to the Frechét distance between boundary curves. Radó theorem has recently become a useful tool in the theory of Schramm-Loewner evolution [@CamiaNewman; @CamiaNewman2]; it was also extended to quasiconformal maps [@Palka].
In this note we study the uniform convergence of Green’s functions of domains on the complex sphere and in $\mathbb R^n$. Under the Dirichlet boundary condition, Green’s function of a regular domain can be continuously extended by $0$ outside of the domain, which makes it possible to discuss uniform convergence on the entire space. We will prove that uniform convergence holds under the kernel convergence of planar domains, provided their connectivity is uniformly bounded (Theorem \[multiplyconnected\]). In higher dimensions, uniform convergence may fail even for domains homeomorphic to a ball (Example \[examplespace\]).
Preliminaries
=============
All domains we consider are assumed to be regular for the Dirichlet boundary problem. These can be characterized as the domains $\Omega\subset \overline{\mathbb C}$ such that for every $w\in \Omega$ there exists *Green’s function* $g_\Omega(\cdot, w)$ such that
- $g_\Omega(\cdot, w)$ is continuous on $\overline{\mathbb C}\setminus \{w\}$ and is zero on $\overline{\mathbb C}\setminus \Omega$;
- $g_\Omega(\cdot, w)$ is harmonic in $\Omega\setminus \{w\}$;
- $g_\Omega(z,w) = -\log|z-w| + O(1)$ as $z\to w$ if $w$ is finite;
- $g_\Omega(z,w)=\log|z|+O(1)$, as $z\rightarrow w$ if $w=\infty$.
As for higher dimensions, the [*Green’s function*]{} $g(\cdot,w)$, of a domain $\Omega\subset \mathbb{R}^n$, $n>2$ with the pole at $w\in \Omega$ is defined by the following conditions.
- $g_\Omega(\cdot, w)$ is continuous on $\mathbb{R}^n \setminus \{w\}$ and is zero on ${\mathbb R}^n\setminus \Omega$;
- $g_\Omega(\cdot, w)$ is harmonic in $\Omega\setminus \{w\}$;
- $g_\Omega(z,w) = |z-w|^{2-n} + O(1)$ as $z\to w$.
We write $g=g_\Omega$ and $g_n=g_{\Omega_n}$.
Recall that monotone convergence of domains (that is, $\Omega_n\subset \Omega_{n+1}$ or $\Omega_n\supset \Omega_{n+1}$ for all $n$) implies pointwise convergence of their Green’s functions; moreover this convergence is uniform on compact subsets of the limit domain [@MR1814344 p. 94]. Such convergence properties are relevant for the studies of the asymptotic behavior of capacities of planar sets and condensers [@Dubinin]. A concept more general than monotone convergence is kernel convergence, which we define next.
Let $w\in\mathbb{C}$ be given and let $\Omega_n\subset\overline{\mathbb C}$ be domains such that $w$ is an interior point of $\bigcap_{n=1}^\infty \Omega_n$. Following [@Pommerenke p. 13], we say that $$\Omega_n \rightarrow \Omega \ \ \text{as} \ \ n\rightarrow \infty \ \ \text{with respect to } w$$ in the sense of [*kernel convergence*]{} if
- $\Omega$ is a domain such that $w\in \Omega$ and some neighborhood of every $z\in \Omega$ lies in $\Omega_n$ for large $n$;
- for $z\in \partial \Omega$ there exist $z_n\in \partial \Omega_n$ such that $z_n\rightarrow z$ as $n\rightarrow \infty$.
An equivalent definition is found in [@Duren p. 77] and [@MR0247039 p. 54]. According to it, $\Omega$ is the kernel of $\{\Omega_n\}$ if it is the maximal domain containing $w$ such that every compact subset of $\Omega$ belongs to all but finitely many of the domains $\Omega_n$. The convergence to $\Omega$ in the sense of kernel requires that every subsequence of $\{\Omega_n\}$ also has the same kernel $\Omega$.
The definition of kernel convergence is also applicable to the convergence of domains in $\mathbb R^n$ for $n>2$.
\[oneside\] If $\Omega_n\to \Omega$ in the sense of kernel, then $\sup_{z\ne w} (g(z,w) - g_n(z,w)) \to 0$ for every $w\in\Omega$.
Let $M_n=\sup_{z\ne w} (g(z,w) - g_n(z,w))$. Since $\Omega_n$ is a regular domain, we have $g_n(z,w)=0$ for $z\in \partial \Omega_n$. Hence $M_n\ge 0$.
Given $\epsilon>0$, let $U=\{z\in\Omega\colon g(z,w)> \epsilon \}$. Since $\overline{U}$ is a compact subset of $\Omega$, there exists $N$ such that $\overline{U}\subset \Omega_n$ for all $n\ge N$. Applying the maximum principle to $g-g_n -\epsilon$ on the set $U$ yields $g-g_n\le \epsilon$ on $U$, since $g=\epsilon$ on $\partial U$. On the complement of $U$, the inequality $g-g_n\le \epsilon$ holds by virtue of $g \le \epsilon$. This proves $M_n\le \epsilon$.
Finally, let us introduce the notation $D(a,r)$ for the open disk of radius $r$ and center $a$, and $C(a,r) = \partial D(a,r)$. As a special case, $\mathbb D=D(0,1)$.
Simply connected domains
========================
In this section we consider the important special case of simply-connected domains in $\mathbb C$. The proof of this case is different from the proof of Theorem \[multiplyconnected\] concerning the multiply connected domains.
\[simplyconnected\] Suppose that simply-connected regular domains $\Omega_n\subset\mathbb C$ converge in the sense of kernel to a regular simply-connected domain $\Omega$. Then for every $w\in\Omega$ Green’s functions $g_{n}(\cdot, w)$ converge to $g(\cdot, w)$ uniformly on $\mathbb C\setminus \{w\}$.
Suppose the uniform convergence fails. Using Lemma \[oneside\] and passing to a subsequence, we may assume there exist $\epsilon>0$ such that $\sup_{\mathbb C\setminus \{w\}}(g_n-g)\ge \epsilon$ for all $n$. The function $g_n-g$ is bounded and harmonic on the set $U = \mathbb C\setminus (\partial \Omega_n\cup\partial \Omega )$ after the removable singularity at $w$ is eliminated. Since $g_n-g \le 0$ on $\partial \Omega_n$, it follows from the maximum principle that $g_n-g$ attains its maximum on $\partial\Omega$.
Pick $z_n\in\partial\Omega$ such that $g_n(z_n)\ge \epsilon$.
We claim that the sequence $\{z_n\}$ is bounded. Indeed, the distances $\operatorname{dist}(w,\partial\Omega_n)$ must be bounded by some constant $d$, for otherwise the limit domain $\Omega$ would be $\mathbb C$, which does not have Green’s function. Let $h$ be Green’s function of the domain $\mathbb C\setminus (-\infty, -d]$ with the pole at $0$. A theorem of Krzyż [@Krzyz] (see also [@Baernstein]) asserts that Green’s function increases under circular symmetrization of the domain, which implies $g_n(z_n) \le h(|z_n-w|)$. Since $h(t)\to 0$ as $t\to\infty$, the fact that $g_n(z_n)\ge \epsilon$ implies that $\{z_n\}$ is a bounded sequence.
By compactness, passing to a subsequence, we may assume that $z_n\to z\in\mathbb C$. Let $\phi_n\colon \mathbb D\to \Omega_n$ be a conformal map such that $\phi_n(0)=w$. For every $\zeta\in \mathbb D$ the Koebe $1/4$-theorem and the distortion theorem yield $$\label{Koebe14}
\operatorname{dist}(\phi_n(\zeta), \partial\Omega_n) \ge
\frac14 |\phi_n'(\zeta)|(1-|\zeta|) \ge \frac{(1-|\zeta|)^2}{4(1+|\zeta|)^3}
\ge \frac{(1-|\zeta|)^2}{32}$$ Note also that $g_n(\phi_n(\zeta)) = -\log|\zeta| $.
When $1-|\zeta|\le 1/2$, we have $-\log|\zeta |\le 2(1-|\zeta|)$. If a point $z=\phi_n(\zeta)$ satisfies $\operatorname{dist}(z, \partial\Omega_n) \le 1/128$, then implies $1-|\zeta|\le \sqrt{32 \operatorname{dist}(z,\partial\Omega_n)}\le 1/2$, hence $$g_n(z) = -\log|\zeta| \le 2(1-|\zeta|) \le \sqrt{128 \operatorname{dist}(z, \partial\Omega_n)}$$
Therefore, if $z_n\to z$ is a sequence such that $g_n(z)\ge \epsilon$ for all $n$, then $\operatorname{dist}(z,\partial \Omega_n) \ge c>0$ for all sufficiently large $n$. This contradicts the definition of kernel convergence, since $z\in\partial\Omega$.
Multiply connected domains
==========================
Here we consider multiply connected domains on the Riemann sphere $\overline{\mathbb{C}}$, subject to a uniform upper bound on their connectivity.
\[multiplyconnected\] Suppose that regular domains $\Omega_n\subset \overline{\mathbb C}$ converge in the sense of kernel to a regular domain $\Omega$. If there exists a constant $N$ such that each domain $\Omega_n$ has at most $N$ boundary components, then for every $w\in\Omega$ Green’s functions $g_{n}(\cdot, w)$ converge to $g(\cdot, w)$ uniformly on $\mathbb C\setminus \{w\}$.
\[Greenbound\] For every $\epsilon>0$ there exists $\delta \in (0,1)$ with the following property. If $\Omega\subset \overline{\mathbb C}$ is a domain containing distinct finite points $z, w$, and there is $\rho\le |z-w|/2$ such that some connected component of $\mathbb C\setminus \Omega$ meets both $C(z,\rho)$ and $C(z,\delta \rho)$, then $g_\Omega(z,w)<\epsilon$.
This is a consequence of Theorem 2.2 in [@Totik]. Indeed, let $\Gamma$ be a component of $\mathbb C\setminus \Omega$ mentioned in the statement. After applying a Möbius transformation that sends $z$ to $0$ and $w$ to $\infty$ and rescaling the resulting domain, $\Gamma$ is transformed to a compact connected set $E$ that meets both the unit circle $C(0,1)$ and the circle $C(0, r)$ with $r\sim \delta$. In particular, we have $r\le 1/2$ when $\delta$ is small enough. In this setting, Theorem 2.2 [@Totik] implies that $g_{\mathbb C\setminus E}(0,\infty) \le C\sqrt{r}$ with an absolute constant $C$. This proves the lemma, since $r$ can be made arbitrarily small by choosing $\delta$ small.
Suppose that the uniform convergence fails. As in the proof of Theorem \[simplyconnected\], we have $\epsilon>0$ and a sequence $z_n\in\partial\Omega$ such that $g_n(z_n)\ge \epsilon$. Passing to a subsequence, we get $z_n\to z_0\in \partial \Omega$, the convergence being with respect to the spherical metric on $\overline{\mathbb C}$. Applying a Möbius transformation, we can make sure that $z_0$ and $w$ are finite points in $\mathbb C$.
Let $d=|z_0-w|$. Pick a small $\lambda\in (0,1)$ and consider the concentric annuli $$A_j = \{z \colon d \lambda^{j+2} < |z-z_0|< d\lambda^{j}\}, \quad j\in\mathbb N,$$ as well as their counterparts centered at $z_n$, $$A_j(n) = \{z \colon d \lambda^{j+2} < |z-z_n|< d\lambda^{j}\}, \quad j\in\mathbb N.$$ Let $J$ be the set of indices $j\in\mathbb N$ such that $A_j$ is not contained in $\Omega$.
**Claim 1**: The set $J$ has at most $3N$ elements, where $N$ is as in the statement of the theorem.
Suppose to the contrary that $J$ contains a subset $J'$ of cardinality $3N+1$. By Lemma \[Greenbound\] no connected component of $\Omega_n^c$ meets more than one of the circles $\{z \colon |z-z_n| = d\lambda^{j}\}$, provided $\lambda$ is small enough. It follows that a component of $\Omega_n^c$ can intersect at most three of the annuli $A_j(n)$. Therefore, for each $n$ there exists $j_n\in J'$ such that $A_{j_n}(n)\subset \Omega_n$. Since $J'$ is finite, there exists a subsequence of the domains $\Omega_n$ for which $j_n$ is the same index $j'\in J'$.
We claim $A_{j'}\subset \Omega$. Indeed, otherwise there is a compact subset $K\subset A_{j'}$ which is not contained in $\Omega$. But $K\subset A_{j'}(n)$ for all sufficiently large $n$, which implies $K\subset \Omega_n$ for all such $n$. By the definition of kernel convergence, $K\subset \Omega$ and thus $A_{j'}\subset\Omega$. Since the containment $A_{j'}\subset\Omega$ contradicts the definition of $J$, Claim 1 is proved.
Since $J$ is finite, $z_0$ is an isolated boundary point of $\Omega$, a contradiction to the regularity of domain $\Omega$.
Counterexamples
===============
The first example is standard; it demonstrates that without assuming the regularity of limit domain one cannot hope for uniform convergence even if the domains are nested.
\[nonregularlimit\] The regular domains $\Omega_n= \{z : 1/n < |z|< 1\}$ converge in the sense of kernel to the punctured unit disk $\Omega = \mathbb D\setminus\{0\}$. Since Green’s function of $\Omega$ is the same as the one of $\mathbb D$, the convergence $g_n(z, 1/2)\to g(z,1/2)$ fails to be uniform in $\Omega$.
Our second example shows that a uniform bound on the number of boundary components in Theorem \[multiplyconnected\] cannot be omitted.
\[exampleplane\] There exists a sequence of regular domains $\Omega_n\subset\mathbb C$ that converge in the sense of kernel to the unit disk $\mathbb D$ but their Green functions do not converge to $g_{\mathbb D}$ even pointwise.
To construct $\Omega_n$, let $A_n$ be a finite $(1/n)$-net in the set $D(0,2) \setminus \mathbb D$. The domain $D(0,2)\setminus A_n$ is not regular, and its Green’s function is the same as of $D(0,2)$. For sufficiently small $r>0$ the closed $r$-neighborhood of $A_n$ consists of $n$ disjoint disks contained in $D(0,2)\setminus \overline{\mathbb D}$. Removing these disks from $D(0,2)$ we obtain a regular domain $\Omega(n,r)$.
As $r\to 0$, Green’s function of $\Omega(n,r)$ converges pointwise to Green’s function of $D(0,2)$. Choose $r$ small enough so that $$g_{\Omega(n,r)}(1/2, 0) \ge g_{D(0,2)}(1/2, 0) - \frac1n$$ and let this domain be $\Omega_n$.
By construction, $\Omega_n\to\mathbb D$ in the sense of kernel convergence. However, $$g_{\Omega_n}(1/2, 0) \to g_{D(0,2)}(1/2, 0) > g_{\mathbb D}(1/2,0)$$ as claimed.
The number of boundary components of $\Omega_n$ in the above example grows indefinitely, as it must according to Theorem \[multiplyconnected\]. However, in higher dimensions such examples can be constructed with domains homeomorphic to a ball.
\[examplespace\] For $d\ge 3$, there exists a sequence of regular domains $\Omega_n\subset\mathbb R^d$ that converge in the sense of kernel to the unit ball $\mathbb B$ but their Green functions do not converge to $g_{\mathbb B}$ even pointwise. Furthermore, each $\Omega_n$ can be chosen to be homeomorphic to $\mathbb B$.
Let $A_n$ be a finite $(1/n)$-net in the set $B(0,2) \setminus \mathbb B$. Pick a simple $C^\infty$-smooth curve $\Gamma$ that passes through every point of $A_n$ and is contained in $B(0,2) \setminus \mathbb B$ except for one endpoint, which lies on $\partial B(0,2)$. Removing $\Gamma$ from $B(0,2)$ does not affect Green’s function since the curve has codimension at least $2$ and is therefore a polar set (Theorems 5.14 and 5.18 in [@HaymanKennedy]).
For sufficiently small $r>0$ the closed $r$-neighborhood of $\Gamma$ is a tubular neighborhood with no self-intersections. Removing it from $B(0,2)$ disks we obtain a regular domain $\Omega(n,r)$ that is homeomorphic to $B(0,2)$. As in Example \[exampleplane\], $r$ can be made small enough so that as $n\to\infty$, Green’s function converges to $g_{B(0,2)}$ on compact subsets of $\mathbb B$, rather than to $g_{\mathbb B}$. In the sense of kernel, the domains converge to $\mathbb B$, however.
| {
"pile_set_name": "ArXiv"
} |
---
author:
- 'Raf Cluckers, Julia Gordon and Immanuel Halupczok'
title: 'Motivic functions, integrability, and uniform in $p$ bounds for orbital integrals'
---
Introduction
============
The purpose of this short article is two-fold: to announce the new results related to analytic properties of motivic (exponential) functions, which we hope could be widely applicable; and to summarize the recent applications of these results in harmonic analysis on $p$-adic groups.
[Concretely]{}, motivic (exponential) functions are complex-valued functions on $p$-adic manifolds defined uniformly in $p$ by means of a rich yet sufficiently manageable language of logic. The word “exponential” refers to the situation when additive characters of the local field (leading to exponential sums) are included in the set of ingredients one is allowed to use when making such a function. The class of motivic exponential functions has two key features that make it particularly useful. On the one hand, these functions can be described in a field-independent manner, and thus are an ideal tool for transferring results between different local fields. On the other hand, the class of motivic (exponential) functions is closed under integration with respect to parameters, and thanks to this property one can show that many functions naturally arising in harmonic analysis on $p$-adic groups belong to this class.
The project of using motivic integration in harmonic analysis on $p$-adic groups was initiated by T.C. Hales in 1998, and [goes]{} under the emerging name of “motivic harmonic analysis”. One of the goals of this article is to collect results about “definability” of various objects arising in representation theory of $p$-adic groups and the applications in harmonic analysis that follow from such definability, summarizing the state [of the art]{} of this project.
In the following sections, we [summarize]{} most of the definitions we use that relate to (motivic) functions, motivic measures, etc. However, we do not [include]{} here any definitions related to harmonic analysis on reductive groups, as they are widely available in other sources, for example, in the beautiful article [@kottwitz:clay].
[**Acknowledgment.**]{} This article summarizes the results of a project carried out over several years. It certainly would not have happened without the influence of Tom Hales and the prior work of Jonathan Korman and Jyotsna Diwadkar. It is a pleasure to thank many people who have helped us at various stages of the project, (and those who collaborated on it in earlier stages), in particular, Jan Denef, François Loeser, Fiona Murnaghan, Clifton Cunningham, Loren Spice, Sug-Woo Shin, Nicolas Templier, Tasho Statev-Kaletha, William Casselman, Gopal Prasad, and Lance Robson. I.H. was supported by the SFB 878 of the Deutsche Forschungsgemeinschaft; J.G. was supported by NSERC; R.C. was supported in part by the European Research Council under the European Community’s Seventh Framework Programme (FP7/2007-2013) with ERC Grant Agreement nr. 246903 NMNAG, by the Labex CEMPI (ANR-11-LABX-0007-01), and by the Fund for Scientific Research of Flanders, Belgium (grant G.0415.10).
What is a motivic (exponential) function?
=========================================
Informally, motivic functions are built from definable functions in the Denef-Pas language and motivic *exponential* functions are built from motivic functions, additive characters on local fields, and definable functions serving as arguments of the characters. As such, they are given independently of the local field and can be interpreted in any local field and with any additive character. Let us recall the definition of the Denef-Pas language first, and proceed to motivic (exponential) functions afterwards, where we [slightly]{} simplify the terminology and notation of [@CGH-1].
Denef-Pas language {#sub:DP}
------------------
The Denef-Pas language is a first order language of logic designed for working with valued fields. Formulas in this language will allow us to uniformly [handle]{} sets and functions for all [local]{} fields. We start by defining two sublanguages of the language of Denef-Pas: the language of rings and the Presburger language.
### The language of rings
Apart from the symbols for variables $x_1, \dots, x_n,\dots$ and the usual logical symbols equality ‘$=$’, parentheses ‘$($’, ‘$)$’, the quantifiers ‘$\exists$’, ‘$\forall$’, and the logical operations conjunction ‘$\wedge$’, negation ‘$\neg$’, disjunction ‘$\vee$’, the language of rings consists of the following symbols:
- constants ‘$0$’, ‘$1$’;
- binary functions ‘$\times$’, ‘$+$’.
A (first-order) formula in the language of rings is any syntactically correct formula built out of these symbols. One usually omits the words ‘first order’.
If a formula in the language of rings has $n$ free (i.e. unquantified) variables then it defines a subset of $R^n$ for any ring $R$. For example the formula “$\exists x_2\, (x_2\times x_1 = 1)$” defines the set of units $R^\times$ in any ring $R$; note that quantifiers [(by convention)]{} always run over the ring in question. Note also that quantifier-free formulas in the language of rings define constructible sets (in the sense of algebraic geometry).
### Presburger language {#Pres}
A formula in Presburger’s language is built out of variables running over ${{\mathbb Z}}$, the logical symbols (as above) and symbols ‘$+$’, ‘$\le$’, ‘$0$’, ‘$1$’, and for each $d=2,3,4,\dots$, a symbol ‘$\equiv_d$’ to denote the binary relation $x\equiv y \pmod{d}$. Note the absence of the symbol for multiplication.
Since multiplication is not allowed, sets and functions defined by formulas in the Presburger language are in fact very basic, cf. [@CPres] or [@Presburger]. For example, definable subsets of the line $\ZZ$ are finite unions of arithmetic progressions (in positive or negative direction) and points. [By [@Presburger], quantifiers are never needed to describe Presburger sets, so]{} all definable sets are of simple form.
### Denef-Pas language {#DP}
The Denef-Pas language is a three-sorted language in the sense that its formulas speak about three different “sorts” of elements: those of the valued field, of the residue field and of the value group (which will always be ${{\mathbb Z}}$ in our setting). Each variable in such a formula only runs over the elements of one of the sorts, i.e., there are three disjoint sets of symbols for the variables of the different sorts. For a formula to be syntactically correct, one has to pay attention to the sorts when composing functions and plugging them into relations.
In addition to the variables and the logical symbols, the formulas use the following symbols.
- In the valued field sort: the language of rings.
- In the residue field sort: the language of rings.
- In the ${{\mathbb Z}}$-sort: the Presburger language.
- a symbol ${\mathrm{ord}}(\cdot)$ for the valuation map from the nonzero elements of the valued field sort to the ${{\mathbb Z}}$-sort, and [a symbol]{} ${\overline{\mathrm{ac}}}(\cdot)$ for the so-called angular component, which is a [multiplicative]{} function from the valued field sort to the residue field sort (more about this function below).
A formula in this language can be interpreted in any discretely valued field $F$ which comes with a uniformizing element $\varpi$, by letting the variables range over $F$, [over its]{} residue field $k_F$, and [over]{} ${{\mathbb Z}}$, respectively, depending on the sort they belong to; ${\mathrm{ord}}$ is the valuation map (defined on $F^\times$ and such that ${\mathrm{ord}}(\varpi)=1$), and ${\overline{\mathrm{ac}}}$ is defined as follows. If $x$ is a unit in ${\mathcal{O}}_F$ (that is, ${\mathrm{ord}}(x)=0$), then ${\overline{\mathrm{ac}}}(x)$ is the residue of $x$ modulo $\varpi$ (thus, an element of the residue field). For all other nonzero $x$, one puts ${\overline{\mathrm{ac}}}(x) := \varpi^{-{\mathrm{ord}}(x)}x \bmod (\varpi)$; thus, ${\overline{\mathrm{ac}}}(x)$ is the first non-zero coefficient of the $\varpi$-adic expansion of $x$. Finally, we define ${\overline{\mathrm{ac}}}(0)=0$.
Thus, a formula $\varphi$ in this language with $n$ free valued-field variables, $m$ free residue-field variables, and $r$ free ${{\mathbb Z}}$-variables gives naturally, for each discretely valued field $F$, a subset $\varphi(F)$ of $F^n\times {k}_F^m\times {{\mathbb Z}}^r$: namely, $\varphi(F)$ is the set of all the tuples where the interpretation of $\varphi$ in $F$ is “true” (where a variable appearing inside a quantifier runs over either $F$, $k_F$, or $\ZZ$, respectively, depending on its sort).
### Adding constants to the language {#sect:ldpo}
In many situations one needs to work with geometric objects defined over some fixed base number field, or over its ring of integers ${\mathcal{O}}$. In such situations, on top of the symbols for the constants that are already present in the above language (like $0$ and $1$), we will add to the Denef-Pas language all elements of ${\mathcal{O}}[[t]]$ as extra symbols for constants in the valued field sort. We denote the resulting language by ${{\mathcal L}_{{\mathcal{O}}}}$.
Given a complete discretely valued field $F$ [which]{} is an algebra over ${\mathcal{O}}$ via [a chosen]{} algebra homomorphism $\iota:{\mathcal{O}}\to F$, [with]{} a chosen uniformizer $\varpi$ of the valuation ring ${\mathcal{O}}_F$ of $F$, one can interpret the formulas in ${{\mathcal L}_{{\mathcal{O}}}}$ as described in the previous subsection, where the new constants from ${\mathcal{O}}$ are interpreted as elements of $F$ by using $\iota$, the constant symbol $t$ is interpreted as the uniformizer $\varpi$, and thus, by the completeness of $F$, elements of ${\mathcal{O}}[[t]]$ can be naturally interpreted in $F$ as well.
\[AO\] Let ${{\mathcal C}}_{\mathcal{O}}$ be the collection of all triples $(F, \iota , \varpi)$, where $F$ is a non-Archimedean local field which allows at least one ring homomorphism from ${\mathcal{O}}$ to $F$, the map $\iota:{\mathcal{O}}\to F$ is such a ring homomorphism, and $\varpi$ is a uniformizer for $F$. Let ${{\mathcal A}}_{\mathcal{O}}$ be the collection of triples $(F, \iota , \varpi)$ in ${{\mathcal C}}_{\mathcal{O}}$ with $F$ of characteristic zero, and let ${{\mathcal B}}_{\mathcal{O}}$ be the collection of the triples $(F, {\iota} , \varpi)$ such that $F$ has positive characteristic.
Given an integer $M$, let ${{\mathcal C}}_{{\mathcal{O}}, M}$ be the collection of $(F,{\iota},\varpi)$ in ${{\mathcal C}}_{\mathcal{O}}$ such that the residue field of $F$ has characteristic larger than $M$, and similarly for ${{\mathcal A}}_{{\mathcal{O}}, M}$ and ${{\mathcal B}}_{{\mathcal{O}}, M}$.
Since our results and proofs are independent of the choices of the map $\iota$ and the uniformizer $\varpi$, we will often just write $F\in {{\mathcal C}}_{\mathcal{O}}$, instead of naming the whole triple. For any $F\in {{\mathcal C}}_{{\mathcal{O}}}$, write ${\mathcal{O}}_F$ for the valuation ring of $F$, $k_F$ for its residue field, and $q_F$ for the cardinality of $k_F$.
Definable sets and motivic functions {#subsub:functions}
------------------------------------
[The ${{\mathcal L}_{{\mathcal{O}}}}$-formulas introduced in the previous section allow us to obtain a]{} field-independent notion of subsets of $F^n\times k_F^m\times \ZZ^r$ for $F\in {{\mathcal C}}_{\mathcal{O}}$.
\[defset\] A collection $X = (X_F)_{F}$ of subsets $X_F\subset F^n\times {k}_F^m\times {{\mathbb Z}}^r$ for $F$ in ${{\mathcal C}}_{\mathcal{O}}$ is called a *a definable set* if there is an ${{\mathcal L}_{{\mathcal{O}}}}$-formula $\varphi$ such that $X_F = \varphi(F)$ for each $F$, where $\varphi(F)$ is as explained at the end of §\[DP\].
[By Definition \[defset\]]{}, a “definable set” is actually a collection of sets indexed by $F\in {{\mathcal C}}_{\mathcal{O}}$; such practice is not uncommon in model theory and has its analogues in algebraic geometry. [A particularly simple definable set is $(F^n\times k_F^m\times \ZZ^r)_F$; we introduce the simplified notation]{} ${\rm VF}^n\times {\rm RF}^m\times \ZZ^r$, where ${\rm{VF}}$ stands for valued field and ${\rm{RF}}$ for residue field. [We apply the typical set-theoretical notation to definable sets $X, Y$, [e.g.]{}, $X \subset Y$ (if $X_F \subset Y_F$ for each $F \in {{\mathcal C}}_{\mathcal{O}}$), $X \times Y$, and so on.]{}
For definable sets $X$ and $Y$, a collection $f = (f_F)_F$ of functions $f_F:X_F\to Y_F$ for $F\in{{\mathcal C}}_{\mathcal{O}}$ is called a definable function and denoted by $f:X\to Y$ if the collection of graphs of the $f_F$ is a definable set.
There are precise quantifier elimination statements for [the]{} Denef-Pas language, which lead to the study of the geometry of definable sets and functions for $F$ in ${{\mathcal C}}_{{\mathcal{O}},M}$, where often $M$ is large and depends on the data defining the set. To give an overview of these results is beyond of the scope of this survey, but the reader may consult [@CCL:metric], [@CCL:lipschitz], [@CH:; @Lipschitz] and others.
We now come to motivic functions, for which definable functions are the building blocks as mentioned above.
\[motfun\] Let $X = (X_F)_F$ be a definable set. A collection $f = (f_F)_F$ of functions $f_F:X_F\to\CC$ is called *a motivic function* on $X$ if and only if there exist integers $N$, $N'$, and $N''$, such that, for all $F\in {{\mathcal C}}_{\mathcal{O}}$, $$f_F(x)=\sum_{i=1}^N q_F^{\alpha_{iF}(x)} ( \# (Y_{iF})_x ) \big( \prod_{j=1}^{N'} \beta_{ijF}(x) \big) \big( \prod_{\ell=1}^{N''} \frac{1}{1-q_F^{a_{i\ell}}} \big), \mbox{ for } x\in X_F,$$ for some
- nonzero integers $a_{i\ell}$,
- definable functions $\alpha_{i}:X\to \ZZ$ and $\beta_{ij}:X\to \ZZ$,
- definable sets $Y_i\subset X\times {\rm RF}^{r_i}$,
where, for $x\in X_F$, $(Y_{iF})_x$ is the finite set $\{y\in k_F^{r_i}\mid (x,y)\in Y_{iF}\}$.
Motivic exponential functions {#subsub:expfunctions}
-----------------------------
For any local field $F$, let ${\mathscr{D}}_F$ be the set of the additive characters on $F$ that are trivial on [the maximal ideal $\mathfrak m_F$]{} and nontrivial on ${\mathcal{O}}_F$.
For $\Lambda \in {\mathscr{D}}_F$, we will write $\bar\Lambda$ for the induced additive character on $k_F$.
\[expfun\] Let $X = (X_F)_F$ be a definable set. A collection $f = (f_{F,\Lambda})_{F,\Lambda}$ of functions $f_{F,\Lambda}:X_F\to\CC$ for $F\in {{\mathcal C}}_{\mathcal{O}}$ and $\Lambda\in {\mathscr{D}}_F$ in is called *a motivic exponential function* on $X$ if there exist integers $N>0$ and $r_i\geq 0$, motivic functions $f_i=(f_{iF})$, definable sets $Y_i\subset X\times {\rm RF}^{r_i}$ and definable functions $g_i:Y_i\to {\rm VF}$ and $e_i:{Y_i}\to {\rm RF}$ for $i=1,\ldots,N$, such that for all $F\in {{\mathcal C}}_{\mathcal{O}}$ and all $\Lambda\in {\mathscr{D}}_{F}$ $$f_{F,\Lambda}(x)=\sum_{i=1}^N f_{iF}(x)\Big( \sum_{y \in (Y_{iF})_x}\Lambda\big(g_i(x,y)\big) \cdot \bar \Lambda\big(e_i(x,y)\big)\Big)\mbox{ for all } x\in X_F.$$
[Compared to Definition \[motfun\], the counting operation $\#$ has been replaced by taking exponential sums, which makes the motivic exponential functions a richer class than the motivic functions.]{} Motivic (exponential) functions form classes of functions which are stable under integrating some of the variables out, see Theorem \[thm:mot.int.\] below. In fact, the search for this very property led to their definition.
Families
--------
[There are natural notions of “uniform families” of the objects introduced in Definitions \[defset\], \[motfun\] and \[expfun\]. We fix the following terminology.]{}
\[deffam\] Suppose that $A$ is a definable set. A [*family of definable sets*]{} with parameter in $A$ is a definable set $X\subset A\times Y$ with some definable set $Y$. We denote such a family by $(X_a)_{a \in A}$. For each $F\in {{\mathcal C}}_{\mathcal{O}}$ and $a\in A_F$, the sets $X_{F,a}:= \{y\in Y_F\mid (a,y)\in X_F \}$ are called the family members of the family $X_F = (X_{F,a})_{a\in A_F}$.
Similarly, for $(X_a)_{a \in A}$ as above, an (exponential) motivic, resp. definable, function $f$ on $X$ can be considered as a *[family $(f_a)_{a \in A}$]{} of (exponential) motivic*, resp. *definable functions* with parameter in $A$. Its members are functions $f_{F,a}$ or $f_{F,\Lambda,a}$ for $F\in {{\mathcal C}}_{\mathcal{O}}$, $\Lambda\in {\mathscr{D}}_F$ and $a\in A_F$.
Whenever we call a specific function $f_{F}$ motivic, we implicitly think of $F$ as varying and we mean that there exists an $M > 0$ and a motivic function [$g$ with $f_F = g_F$]{} for all $F \in C_M$. Likewise, we think of $F$ as varying whenever a specific set $X_{F}$ or function $f_{F}$ is called definable. All the results presented in this article will only be valid for sufficiently big residue characteristic. In particular, for any of the uniform objects $X$ we introduced in Defitions \[defset\], \[motfun\], \[expfun\], we are actually only interested in $(X_F)_{F \in {{\mathcal C}}_{{\mathcal{O}},M}}$ for $M$ sufficiently big.
Measure and integration
-----------------------
To integrate a motivic function $f$ on a definable set $X$, we need a uniformly given family of measures on each $X_F$. For $X = {\rm VF}$, we put the Haar measure on $X_F = F$ so that ${\mathcal{O}}_F$ has measure $1$; on $k_F$ and on $\ZZ$, we use the counting measure. To obtain measures on arbitrary definable sets $X$, we use “definable volume forms”, as follows.
In a nutshell, a definable volume form $\omega$ on $X$ is a volume form defined by means of definable functions on definable coordinate charts, see [@cluckers-loeser-nicaise]. Such a definable volume form specializes to a volume form $\omega_F$ on $X_F$ for every $F$ [with sufficiently large residue characteristic; more precisely, a definable volume form specializes to a top degree differential form for every $F$, but its non-vanishing can be assured only when the residue characteristic of $F$ is sufficiently large.]{}
Any $F$-analytic subvariety of $F^n$, say, everywhere of equal dimension, together with an $F$-analytic volume form, carries a natural measure associated to the volume form, cf. [@Bour]. This carries over to the motivic setting, i.e., a volume form $\omega_F$ as above induces a measure on $X_F$. By a “motivic measure”, we mean a family of measures on each $X_F$ that arises in this way from a definable volume form.
Any algebraic volume form defined over ${\mathcal{O}}$ on a variety $V$ over ${\mathcal{O}}$ gives rise to a definable volume form on $V(F)$ for [$F\in {{\mathcal C}}_{{\mathcal{O}}, M}$ for a sufficiently large constant $M$]{}, in the above sense. However, often one deals with volume forms on, say, orbits of elements of a group defined over a local field, or other volume forms defined over a local field, for which it is not always easy to obtain that they can be defined in a way which does not depend on the particular local field.
The classes of motivic (exponential) functions are natural to work with for the purposes of integration, as shown by the following theorem, which generalizes a similar result of [@cluckers-loeser:fourier] to the class of functions summable in the classical sense of $L^1$-integrability.
[@CGH-1]\*[Theorem 4.3.1]{}\[thm:mot.int.\] Let $f$ be a motivic function, resp. a motivic exponential function, on $X\times Y$ for some definable sets $X$ and $Y$, with $Y$ equipped with a motivic measure $\mu$. Then there exist a motivic function, resp. a motivic exponential function, $g$ on $X$ and an integer $M>0$ such that for each $F\in {{\mathcal C}}_{{\mathcal{O}},M}$, each $\Lambda\in {\mathscr{D}}_F$ and for each $x\in X_F$ one has $$g_F(x) = \int_{y\in Y_F}f_F(x,y)\,d\mu_F, \mbox{ resp. } g_{F,\Lambda}(x) = \int_{y\in Y_F}f_{F,\Lambda}(x,y)\,d\mu_F,$$ whenever the function $Y_F\to\CC:y\mapsto f_F(x,y)$, resp. $y\mapsto f_{F,\Lambda}(x,y)$ lies in $L^1(Y_F, \mu_F)$.
New transfer principles for motivic exponential functions {#sec:trpr}
=========================================================
Theorem \[thm:mot.int.\] leads to the transfer of integrability and transfer of boundedness principles, which we quote from [@CGH-1]. (For simplicity, we quote the version without parameters, that is, we take the parameter space to be a point in [@CGH-1]\*[Theorem 4.4.1]{} and [@CGH-1]\*[Theorem 4.4.2]{}).
[@CGH-1]\*[Theorem 4.4.1]{}\[thm:ti\] Let $f$ be a motivic exponential function on ${\rm VF}^n$. Then there exists $M>0$, such that for the fields $F\in{{\mathcal C}}_{{\mathcal{O}}, M}$, the truth of the statement that $f_{F,\Lambda}$ is (locally) integrable for all $\Lambda \in {\mathscr{D}}_F$ depends only on the isomorphism class of the residue field of $F$.
[@CGH-1]\*[Theorem 4.4.2]{}\[thm:bounded\] Let $f$ be a motivic exponential function on ${\rm VF}^n$. Then, for some $M>0$, for the fields $F\in{{\mathcal C}}_{{\mathcal{O}}, M}$, the truth of the statement that $f_{F, \Lambda}$ is (locally) bounded for all $\Lambda \in {\mathscr{D}}_F$ depends only on the isomorphism class of the residue field of $F$.
In the versions with parameters of these results, [$f$]{} is a family of motivic exponential functions with parameter in a definable set $A$. In that case, each of the two theorems states that there exists a motivic exponential function $g$ on $A$ such that the [statement under consideration]{} is true for $f_{F,\Lambda,a}$ iff $g_{F,\Lambda}(a) = 0$, that is, we relate the *locus of integrability*, resp. the *locus of boundedness* of the motivic (exponential) function $f$ to the zero locus of another motivic (exponential) function $g$.
New “uniform boundedness” results for motivic functions {#sec:ub}
=======================================================
It turns out that the field-independent description of motivic functions leads not only to transfer principles, but also to uniform (in the residue characteristic of the field) bounds for those motivic functions whose specializations are known to be bounded for individual local fields (for now, for motivic functions without the exponential). We quote the following theorems from [@S-T].
\[thm:presburger-fam\][@S-T]\*[Theorem B.6]{} Let $f$ be a motivic function on $W \times {{\mathbb Z}}^n$, where $W$ is a definable set. Then there exist integers $a,b$ and $M$ such that for all $F\in {{\mathcal C}}_{{\mathcal{O}}, M}$ the following holds.
If there exists any (set-theoretical) function $\alpha:\ZZ^n\to{{\mathbb R}}$ such that $$| f_F(w,\lambda) |_{{{\mathbb R}}} \le \alpha(\lambda) \mbox{ on } W_F \times {{\mathbb Z}}^n,$$ then one actually has $$| f_F(w,\lambda) |_{{{\mathbb R}}} \le q_F^{a+b \|\lambda\|} \mbox{ on } W_F \times {{\mathbb Z}}^n,$$ where [$\|\lambda\| = \sum_{i=1}^n |\lambda_i|$]{}, and where $|\cdot|_{{\mathbb R}}$ is the norm on ${{\mathbb R}}$.
We observe that in the case with $n=0$, the theorem yields that if a motivic function $f$ on $W$ is such that $f_F$ is bounded on $W_F$ for each $F\in {{\mathcal C}}_{{\mathcal{O}}, M}$, then the bound for $|f_F|_{{\mathbb R}}$ can be taken to be $q_F^a$ for some $a\geq 0$, uniformly in $F$ with large residue characteristic.
The following statement allows one to transfer bounds which are known [for local fields of]{} characteristic zero to local fields of positive characteristic, and vice versa.
\[thm:transfer-fam\][@S-T]\*[Theorem B.7]{} Let $f$ be a motivic function on $W \times {{\mathbb Z}}^n$, where $W$ is a definable set, and let $a$ and $b$ be integers. Then there exists $M$ such that, for any $F\in {{\mathcal C}}_{{\mathcal{O}}, M}$, whether the statement $$\label{transfer}
f_F(w,\lambda)\le q_F^{a+b \|\lambda\|} \mbox{ for all } (w,\lambda) \in W_F \times {{\mathbb Z}}^n$$ holds or not, only depends on the isomorphism class of the residue field of $F$.
A few words about proofs.
-------------------------
The common strategy in the proofs of Theorems [\[thm:mot.int.\]]{} – \[thm:transfer-fam\] (assuming the set-up of motivic integration developed in [@cluckers-loeser] and [@cluckers-loeser:fourier]) is the following. The first step is to reduce the problem from arbitrary motivic exponential functions $f$ to motivic exponential functions $g_i$ of a simpler form, by writing $f$ as a sum $\sum_i g_i$. For some of the results (e.g. to obtain a precise description of the locus of integrability or boundedness of $f$), one needs to prove that not too much cancellation can occur in this sum; this is in fact one of the main difficulties in the proofs of Theorems \[thm:ti\] and \[thm:bounded\].
In the next step, we eliminate all the valued-field variables, possibly at the cost of introducing more residue-field and $\ZZ$-valued variables, which uses the Cell Decomposition Theorem. Once we have a motivic function that depends only on the residue-field and value-group variables, we again break it into a sum of simpler terms, and then get rid of the residue-field variables. Finally, the question is reduced to the study of functions $h : \ZZ^r \to {{\mathbb R}}$, which are, up to a finite partition of $\ZZ^r$ into sets [which are]{} definable in the Presburger language (see Section \[Pres\]), sums of products of linear functions in $x \in \ZZ^r$ and of powers of $q_F$, where the power also depends linearly on $x$. For most $x$, the behaviour of such a sum $h$ is governed by a dominant term, and for a single term, each of the theorems is easy to prove. To control the locus where no single term of $h$ is dominant, we use the powerful result by Wilkie about “o-minimality of ${{\mathbb R}}_{\mathrm{exp}}$”.
Which objects arising in harmonic analysis are definable?
=========================================================
In the remainder of the article, when we say “definable”, we mean, definable in the Denef-Pas language $ {{{\mathcal L}_{{{\mathbb Z}}}}}$ (introduced in Section \[sect:ldpo\]). If one wishes to think specifically of reductive grouse defined over a given global field ${\mathbf{F}}$ with the ring of integers ${\mathcal{O}}$ and its completions, then one can use the language ${{\mathcal L}_{{\mathcal{O}}}}$ instead. We write ${{\mathcal A}}_M$, ${{\mathcal B}}_M$, and ${{\mathcal C}}_M$ for ${{\mathcal A}}_{{{\mathbb Z}}, M}$, ${{\mathcal B}}_{{{\mathbb Z}}, M}$ and ${{\mathcal C}}_{{{\mathbb Z}}, M}$, respectively (see Definition \[AO\]).
Connected reductive groups
--------------------------
We recall that a split connected reductive group over a local field is determined by its root datum $\Psi=(X_\ast, \Phi, X^\ast, \Phi^\vee)$. By an *absolute* root datum of ${\mathbf{G}}$ we will mean the root datum of ${\mathbf{G}}$ over the separable closure of the given local field; it is a quadruple of this form.
By an unramified root datum we mean a root datum of an unramified reductive group over a local field $F$, i.e. a quintuple $\xi=(X_\ast, \Phi, X^\ast, \Phi^\vee, \theta)$, where $\theta$ is the action of the Frobenius element of $F^{\mathrm{ur}}/F$ on the first four components of $\xi$.
Given an unramified root datum $\xi$, it is shown in [@cluckers-hales-loeser]\*[§4]{} that a connected reductive group ${\mathbf{G}}(F)$ with root datum $\xi$ is “definable using parameters”, for all local fields $F\in {{\mathcal C}}_M$, where $M$ is determined by the root datum. This result is extended to all connected reductive groups (not necessarily unramified) in [@CGH-2]\*[§3.1]{}, [@S-T]\*[B.4.3]{}. More precisely, by “definable using parameters”, we mean that ${\mathbf{G}}(F)$ is a member of a family (in the sense of Definition \[deffam\]) of definable connected reductive groups with parameter in a suitable definable set $Z$. This parameter encodes information about the extension over which ${\mathbf{G}}$ splits (see [@CGH-2]\*[§3.1]{} for details). In the remainder of this article, every definable or motivic object in ${\mathbf{G}}(F)$ is parametrized by $Z$, and all results are uniform in the family.
Haar measure and Moy-Prasad filtrations
---------------------------------------
If ${\mathbf{G}}$ is unramified, the canonical measure on ${\mathbf{G}}(F)$ defined by Gross [@gross:motive] is motivic for $F\in {{\mathcal C}}_M$ (with $M$ depending only on the root datum that determines ${\mathbf{G}}$), see [@cluckers-hales-loeser]\*[§7.1]{}. For a root datum defining a ramified group ${\mathbf{G}}$, [to the best of our knowledge, this is presently not known]{}. [However, in [@S-T]\*[Appendix B]{} we prove]{} that there is a constant $c$ that depends only on the root datum, such that Gross’ measure can be renormalized by a factor between $q^{-c}$ and $q^c$ [( [possibly]{} depending on $F$)]{} to give a motivic measure on ${\mathbf{G}}(F)$, where $q=p^r$ is the cardinality of the residue field of $F$, and $p$ is assumed large enough so that the group is tamely ramified.
In fact, this question is closely related to whether Moy-Prasad filtration subgroups ${\mathbf{G}}(F)_{x,r}$ are definable (here $x$ is a point in the building of ${\mathbf{G}}(F)$, and $r\ge 0$). This question has been answered in a number of important special cases. Namely, if ${\mathbf{G}}$ is unramified over $F$, and $x$ is a hyper-special point, then we prove in [@CGH-2]\*[§3]{} that ${\mathbf{G}}_{x, r}$ is definable for all $r\ge 0$. The same proof (which relies on the split case and taking Galois-fixed points) works in the case when ${\mathbf{G}}$ is ramified over $F$, $x$ is a special point, and $r>0$, but fails for $r=0$ since Galois descent no longer works. This is exactly the reason that the question whether the canonical measure is motivic is open in the ramified case. We also prove in [@CGH-2]\*[§3]{} that for $x$ an optimal point in the alcove (in the sense of the definition by Adler and DeBacker, [@adler-debacker:bt-lie]), ${\mathbf{G}}(F)_{x,r}$ is definable for $r>0$, and for all $r\ge 0$ if ${\mathbf{G}}$ is unramified over $F$, and therefore the $G$-domain ${\mathbf{G}}(F)_r$ is definable. We also prove the corresponding results for the Lie algebra. [We note that the first results on definability of ${\mathbf{G}}(F)_{x,0}$ in some special cases appeared in [@diwadkar:thesis].]{}
Orbital integrals and their Fourier transforms
----------------------------------------------
There are two ways to think of an orbital integral: as a distribution on the space $C_c^\infty({\mathbf{G}}(F))$ of locally constant compactly supported functions on ${\mathbf{G}}(F)$ (or, respectively, on the Lie algebra), or to fix a test function $f$ on ${\mathbf{G}}(F)$ (respectively, on ${\mathfrak{g}}(F)$), and consider the orbital integral ${\mathcal{O}}_\gamma(f)$ (respectively, ${\mathcal{O}}_X(f)$) as a function on ${\mathbf{G}}(F)$ (respectively, on the Lie algebra ${\mathfrak{g}}(F)$).
The earliest results concerning definability of orbital integrals use the second approach: in [@cunningham-hales:good], C. Cunningham and T.C. Hales prove that with $f$ fixed to be the characteristic function of ${\mathfrak{g}}({\mathcal{O}}_F)$ and [if]{} ${\mathbf{G}}$ [is]{} unramified, ${\mathcal{O}}_X$ is a motivic function on the set of so-called *good* regular elements in ${\mathfrak{g}}(F)$. In fact, they prove a slightly more refined result, giving explicitly the residue field parameters that control the behaviour of this function. In [@hales:orbital_motivic], T.C. Hales proves that orbital integrals are motivic “on average”. Both of these papers pre-date the most general notion of a motivic function, and because of this, extra complications arise from having to explicitly track the parameters needed to define orbital integrals.
[In order to be able to use]{} the first approach to orbital integrals (as distributions), we [define a notion of a “motivic distribution” (without this name, it first appears in this context in [@cluckers-cunningham-gordon-spice]\*[§3]{})]{}. We say that a distribution $\Phi$ is motivic if [it is, in fact, a collection of distributions defined for each $F\in {{\mathcal C}}_{M'}$ for some $M'>0$, and]{} for any [suitable]{} family $(f_a)_{a\in S}$ of motivic test functions indexed by a parameter $a$ varying over a definable set $S$, there exists a motivic function [$g$]{} on $S$ [and an [$M \ge M'$]{}]{} such that $\Phi(f_{a,F})=g(a)$ for all $F \in {{\mathcal C}}_M$ and all [$a \in S_F$,]{} [where by “suitable” we mean that $f_{F,a}$ is in the domain of $\Phi$ for all $F\in {{\mathcal C}}_M$ and $a\in S_F$.]{}
The most general current results on orbital integrals are:
1. Cluckers, Hales and Loeser prove in [@cluckers-hales-loeser] that for ${\mathbf{G}}$ – unramified over $F$, and with $f$ fixed to be the characteristic function of ${\mathfrak{g}}({\mathcal{O}}_F)$, $X\mapsto {\mathcal{O}}_X(f)$ is a motivic function on the set of regular semisimple elements in ${\mathfrak{g}}(F)$. In fact, their method shows also that the orbital [integrals of regular semisimple elements $X\in {\mathfrak{g}}(F)$ form a family of]{} motivic distribution in the sense defined above, [indexed by $X$]{}.
2. Still under the assumption that ${\mathbf{G}}$ is unramified, in [@S-T]\*[§B.6]{}, we prove that with a suitable normalization of measures, non-regular orbital integrals [${\mathcal{O}}_\gamma$]{} of tame elements (note that all elements are tame when the residue characteristic $p$ is large enough), form a family of motivic distributions (i.e., they also depend on $\gamma\in {\mathbf{G}}(F)$ in a definable way, thus showing that the orbital integrals are motivic in the sense of both approaches).
If the centralizer $C_G(\gamma)$ of $\gamma$ splits over an unramified extension, our normalization of the measure coincides with the canonical measure on the orbit that one gets from the canonical measure on $C_G(\gamma)$; if $C_G(\gamma)$ is ramified, the motivic and canonical measures might differ by a constant not exceeding a fixed power of $q$. This result should easily generalize to the case when $G$ is ramified as well (though one more constant from the measure on $G$ would appear).
We also note that even though we state the result for orbital integrals on the group, since this is what is required for the estimates that are the main goal of [@S-T]\*[Appendix B]{}, in the course of the proof we essentially prove the analogous statement for the Lie algebra as well.
3. It follows from the results of McNinch [@mcninch:nilpotent] and Jantzen [@jantzen] that the set of nilpotent elements $\mathcal N$ in the Lie algebra ${\mathfrak{g}}(F)$ is definable, cf [@CGH-2]\*[§4.1]{}. We prove in [@CGH-2]\*[§4]{} that nilpotent orbital integrals (on the Lie algebra) [form a family of]{} motivic distributions, [indexed by the set of nilpotent elements $\mathcal N$]{}.
4. On the Lie algebra, one can consider the Fourier transform $\widehat{{\mathcal{O}}}_X(f)={\mathcal{O}}_X(\widehat f)$ of an orbital integral. By a theorem of Harish-Chandra in the characteristic zero case, and of Huntsinger [@adler-debacker:mk-theory] in positive characteristic, this distribution is represented by a locally constant function supported on the set of regular elements ${\mathfrak{g}}(F)^{{\mathrm{reg}}}$. We will denote this function by $\widehat{\mu}_X(Y)$. In [@CGH-2]\*[Theorem 5.7, Theorem 5.11]{} we prove that:
1. For nilpotent elements $X$, $\widehat \mu_X(Y)$ is a motivic function defined on $\mathcal N\times {\mathfrak{g}}$.
2. For regular semisimple elements $X$, $\widehat \mu_X(Y)$ is a motivic function defined on $\mathcal {\mathfrak{g}}^{\mathrm{reg}}\times {\mathfrak{g}}$.
Classification of nilpotent orbits, and Shalika germs
-----------------------------------------------------
It is well-known that when the residue characteristic of the field $F$ is large enough, there are finitely many nilpotent orbits in ${\mathfrak{g}}(K)$. Using Waldspurger’s parametrization of nilpotent orbits for classical Lie algebras [@waldspurger:nilpotent]\*[§I.6]{}, it is possible to prove that these orbits are parametrized by the residue-field points of a definable set. This approach is carried out in detail for odd orthogonal groups in [@diwadkar:o_n]. Alternatively, one can use DeBacker’s parametrization of nilpotent orbits that uses Bruhat-Tits theory. This approach is demonstrated for the Lie algebra of type $G_2$ (as well as for ${\mathfrak{sl}}_n$) in J. Diwadkar’s thesis [@diwadkar:thesis].
Using both parametrizations of nilpotent orbits, and the explicit matching between them due to M. Nevins, [@nevins:param], L. Robson proves in [@lance:thesis] that for ${\mathfrak{sp}}_{2n}$, Shalika germs are motivic functions, up to a “motivic constant”. More precisely, [one of the main results of [@gordon-robson] is]{}:
Let ${\mathfrak{g}}={\mathfrak{sp}}_{2n}$. Let $\mathcal N$ be the set of nilpotent elements in ${\mathfrak{g}}$. Then
1. There exists a definable set $\mathcal E\subset {\rm RF}^n$ (that is, having only residue-field variables), and a definable function $h:\mathcal N\to \mathcal E$, such that for every $d\in \mathcal E$, $h^{-1}(d)$ is an adjoint orbit, and each orbit appears exactly once as the fibre of $h$.
2. There exist a finite collection of integers ${(a_i)}_{i=0}^{r}$, a motivic function $\Gamma$ on $\mathcal E\times{\mathfrak{g}}^{{\mathrm{reg}}}$, and a constant $M>0$, such that for $F\in {{\mathcal C}}_M$, for every $d\in {\mathcal E}_F$, the function $\left(q^{a_0}\prod_{i=1}^r \frac 1{1-q^{a_i}}\right)^{ {-1}}\Gamma_F(d, \cdot)$ is a representative of the Shalika germ on ${\mathfrak{g}}^{{\mathrm{reg}}}$ corresponding to $d$.
In a different context of groups over ${{\mathbb C}}\llp t\rrp $, a motivic interpretation of the sub-regular germ is given in E. Lawes’ thesis, [@lawes:thesis].
Transfer factors and other ingredients of the Fundamental Lemma
---------------------------------------------------------------
In [@cluckers-hales-loeser], the authors prove that transfer factors, weights, and other ingredients of the Fundamental Lemma are motivic functions.
Characters
----------
So far, the only results on Harish-Chandra characters of representations have been restricted to the cases where the character can be related to an orbital integral. One of the fundamental problems is that the parametrization of supercuspidal representations via $K$-types (and the classification of the $K$-types due to J.-K. Yu and Ju-Lee Kim) uses multiplicative characters of the field, which cannot be easily incorporated into the motivic integration framework. Because of this, at present we do not have a way of talking about representations in a completely field-independent way. There are two partial results on characters of some representations in a suitable neighbourhood of the identity, where Murnaghan-Kirillov theory holds.
1. The main result of [@gordon:depth0] can be expressed in modern terms as: *the character of a depth-zero representation of $\mathrm{Sp}_{2n}$ or $\mathrm{SO}_{2n+1}$ is a motivic distribution (in the sense defined above) on the set of topologically unipotent elements*.
2. For positive-depth so-called *toral, very supercuspidal* representations, it is shown in [@cluckers-cunningham-gordon-spice], that up to a constant, their Harish-Chandra characters are motivic exponential functions on a neighbourhood of the identity on which Murnaghan-Kirillov theory works, and an explicit parameter space (over the residue field) is constructed for these representations up to crude equivalence, where we identify two representations if their characters on this neighbourhood of the identity are the same.
However, thanks to Harish-Chandra’s local character expansion, extended to an expansion near an arbitrary tame semisimple element by J. Adler and J. Korman [@adler-korman:loc-char-exp], one can prove results about characters, using motivic integration, even without knowing that they are themselves motivic functions. We give an example of such an application in the next section.
We note that proving that the coefficients of Harish-Chandra’s local character expansion are motivic is a question of the same order of difficulty as the question about the characters themselves, and so it is presently far from known.
Local integrability results in positive characteristic
======================================================
Following Harish-Chandra (and using the modification by R. Kottwitz [@S-T]\*[Appendix A]{} for non-regular elements), for a semisimple element $\gamma\in {\mathbf{G}}(F)$, we define the discriminant $$D^G(\gamma):=
\prod_{\alpha\in \Phi\atop{ \alpha(\gamma)\neq 1}}|1-\alpha(\gamma)|, \quad \text{and}
\quad D(\gamma): =\prod_{\alpha\in \Phi}|1-\alpha(\gamma)|,$$ where $\Phi$ is the root system of ${\mathbf{G}}$.
On the Lie algebra, define $D(X)=\prod_{\alpha\in \Phi}|\alpha(X)|$. Following [@kottwitz:clay], we call a function $h$ defined on ${\mathfrak{g}}(F)^{\mathrm{reg}}$ *nice*, if it satisfies the following conditions:
- when extended by zero to all of ${\mathfrak{g}}({F})$, it is locally integrable, and
- the function $D(X)^{1/2}h(X)$ is locally bounded on ${\mathfrak{g}}({F})$.
Similarly, call a function on ${\mathbf{G}}({F})^{\mathrm{reg}}$ *nice*, if it satisfies the same conditions on ${\mathbf{G}}({F})$, with $D(X)$ replaced by its group version $D(\gamma)$.
Combining the theorems of Harish-Chandra in characteristic zero with the Transfer Principles of §\[sec:trpr\] above, in [@CGH-2] we obtain
\[thm:orb int loc int\]
1. For a connected reductive group ${\mathbf{G}}$ (defined via specifying a root datum), there exists a constant $M_{\mathbf{G}}>0$ that depends only on the absolute root datum of ${\mathbf{G}}$, such that for every ${F}\in {{\mathcal B}}_{M_{\mathbf{G}}}$, and every nilpotent orbit ${\mathcal{O}}$ in ${\mathfrak{g}}({F})$, the function $\widehat \mu_{{\mathcal{O}}}$ is a nice function on ${\mathfrak{g}}({F})$.
2. There exists a constant $M_{\mathbf{G}}^{\mathrm{reg}}>0$ that depends only on the absolute root datum of ${\mathbf{G}}$, such that for ${F}\in {{\mathcal B}}_{M_{\mathbf{G}}^{\mathrm{reg}}}$, for every $X\in {\mathfrak{g}}({F})^{\mathrm{reg}}$, the function $\widehat \mu_X$ is a nice function on ${\mathfrak{g}}({F})$.
Local character expansion, [which is crucial for our next result]{}, in large positive characteristic is proved by DeBacker [@debacker:homogeneity] near the identity, and by Adler-Korman [@adler-korman:loc-char-exp] near a general tame semisimple element. These results require an additional hypothethis on the existence of the so-called mock exponential map ${\mathbf{e}}$, discussed in [@adler-korman:loc-char-exp], which we will not quote here; we note only that for classical groups one can take ${\mathbf{e}}$ to be the Cayley transform.
\[thm:char\] Given an absolute root datum $\Psi$, there exists a constant $M>0$, such that for every $F\in {{\mathcal B}}_{M}$, every connected reductive group ${\mathbf{G}}$ defined over $F$ with the absolute root datum $\Psi$, and every admissible representation $\pi$ of ${{\mathbf{G}}}(F)$, the Harish-Chandra character $\theta_{\pi}$ is a nice function on ${\mathbf{G}}(F)$, provided that mock exponential map ${\mathbf{e}}$ exists for ${\mathbf{G}}(F)$.
Finally, Theorem \[thm:orb int loc int\] also implies (thanks to a result of DeBacker) that Fourier transforms of general invariant distributions on ${\mathfrak{g}}(F)$ with support bounded modulo conjugation are represented by nice functions in a neighbourhood of the origin.
Uniform bounds for orbital integrals
====================================
In order to state the next result we need to introduce the notation for a family of test functions, namely, the generators of the spherical Hecke algebra. Assume that ${\mathbf{G}}$ is an unramified reductive group with root datum $\xi$. Let $K={\underline G}({{\mathcal{O}}_F})$ be a hyperspecial maximal compact subgroup (where ${\underline G}$ is the smooth model associated with a hyperspecial point by Bruhat-Tits theory).
We have a Borel subgroup $B=TU$, and let $A$ be the maximal $F$-split torus in $T$. For $\lambda\in X_\ast(A)$, let $\tau_\lambda^G$ be the characteristic function of the double coset $K\lambda(\varpi)K$.
In [@S-T]\*[Appendix B]{}, we prove, using Theorem \[thm:presburger-fam\],
\[thm:main2\] Consider an unramified root datum $\xi$. Then there exist constants $M>0$, $a_\xi$ and $b_\xi$ that depend only on $\xi$, such that for each non-Archimedean local field $F$ with residue characteristic at least $M$, the following holds. Let ${\mathbf{G}}$ be a connected reductive algebraic group over $F$ with the root datum $\xi$. [Let $A$ be a maximal $F$-split torus in ${\mathbf{G}}$, and let $\tau_\lambda^G$ be as above.]{} Then for all [$\lambda\in X_\ast(A)$]{} with $\|\lambda\|\le \kappa$, $$|{{\bf \mathrm{O}}}_\gamma(\tau_\lambda^G)|\le {q_F}^{a_\xi+b_\xi\kappa} D^G(\gamma)^{-1/2}$$ [for all semisimple elements $\gamma\in {\mathbf{G}}(F)$.]{}
It is sometimes useful to reformulate theorems of this type to make an assertion about a group defined over a global field. This is possible because given a connected reductive group over a global field ${\mathbf{F}}$, there are finitely many possibilities for the root data of [${\mathbf{G}}_v={\mathbf{G}}\otimes_{{\mathbf{F}}}{\mathbf{F}}_v$]{}, as $v$ runs over the set of finite places of ${\mathbf{F}}$, and [${\mathbf{F}}_v$ denotes the completion of ${\mathbf{F}}$ at $v$]{}. Here we give an example of one such reformulation, which proved to be useful in [@S-T] ([we also refer to §7 of loc.cit. for a more detailed explanation of the setting]{}).
\[thm:main\] Let ${\mathbf{G}}$ be a connected reductive algebraic group over ${\mathbf{F}}$, [with a maximal split torus $A$]{}. There exist constants $a_G$ and $b_G$ that depend only on the global integral model of ${\mathbf{G}}$ such that [for all but finitely many places $v$]{}, for all $\lambda\in X_\ast(A_v)$ with $\|\lambda\|\le \kappa$, $$|{{\bf \mathrm{O}}}_\gamma(\tau_\lambda^G)|\le q_v^{a_G+b_G\kappa} D^G(\gamma)^{-1/2}$$ for all semisimple elements $\gamma\in {\mathbf{G}}({\mathbf{F}}_v)$, where $q_v$ is the cardinality of the residue field of ${\mathbf{F}}_v$.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We derive the von Kármán-Howarth equation for a full three dimensional incompressible two-fluid plasma. In the long-time limit and for very large Reynolds numbers we obtain the equivalent of the hydrodynamic “four-fifth” law. This exact law predicts the scaling of the third-order two-point correlation functions, and puts a strong constraint on the plasma turbulent dynamics. Finally, we derive a simple expression for the 4/5 law in terms of third-order structure functions, which is appropriate for comparison with in-situ measurements in the solar wind at different spatial ranges.'
author:
- 'N. Andrés$^{1,2}$'
- 'P.D. Mininni$^{2,3}$'
- 'P. Dmitruk$^{2,3}$'
- 'D.O. Gómez$^{1,2}$'
bibliography:
- 'cites.bib'
title: 'von Kármán-Howarth equation for three-dimensional two-fluid plasmas'
---
Introduction {#intro}
============
In the paradigmatic case of incompressible hydrodynamic turbulence, @vkh1938 obtained an evolution equation for the second-order correlation tensor under the assumptions of isotropy and homogeneity. This is the so-called von Kármán-Howarth (vKH) equation, and is one of the cornerstones of turbulence theory. The equation, which relates the time evolution of the second-order correlation velocity tensor to the divergence of the third-order correlation velocity tensor, has been extensively studied in the literature [@R1940; @H1975; @MY1975; @F1995]. One of its most important corollaries is the so-called “four-fifths” law. Assuming the existence of an inertial energy range for very large Reynolds number, it predicts a linear scaling of the longitudinal two-point third-order velocity structure function with the distance between the two points. This scaling puts a strong constraint on the dynamics of fully developed turbulence. In particular, one crucial consequence is that the increment of the velocity field $\delta u$ between two points separated by $r$ is proportional to $r^{1/3}$, which in Fourier space leads to the famous Kolmogorov spectrum $E(k)\sim k^{-5/3}$ for the energy [@K1941a].
Multiple attempts have been made to extend these results to turbulent plasmas. Chandrasekhar [@Ch1951] derived vKH equations in the one-fluid incompressible magnetohydrodynamic (MHD) approximation. Using Elsässer variables, @P1998a [@P1998b] derived the equations for third-order structure functions and for correlation functions, and assuming full isotropy (i.e., including mirror symmetry), homogeneity, and equipartition between kinetic and magnetic energy, they obtained the corresponding $4/5$ law. In the absence of mirror-symmetry the structure of the second-order correlation tensors is more complex. In particular, @P2003 reported an exact equation for homogeneous and isotropic MHD turbulent flows with non-zero helicity. In the large Reynolds number and long-time limit, the authors recovered linear scaling for the third-order correlation tensors.
At spatial scales approaching the ion skin-depth (although still larger than the electron skin-depth) the MHD model is no longer appropriate. It can be replaced by the Hall-MHD (HMHD) model, which is an extension to MHD including the Hall current and the electron pressure (see, e.g., [@K1973]). The HMHD model has been extensively studied in recent years, both analytically [@T1983; @MY1998; @Mi2002; @S2003] and numerically [@Mi2003; @Ma2003; @D2006]. @Ga2008 derived the vKH equation for the three-dimensional (3D) incompressible HMHD equations in the non-helical case, i.e., when mirror-symmetry is considered, and obtained exact scaling laws for the third-order correlation tensors. [However, the HMHD model also has limitations to describe a plasma. First, tt ignores the inertia of the electrons, which plays a role in the nonlinear dynamics at small scales. Second, it misses kinetic plasma dissipation mechanisms such as Landau damping, cyclotron-resonant damping, and other mode-coupling dissipative processes. As a result, the HMHD model can only be expected to hold down to scales, of the order of the ion-skin depth.]{}
[Interestingly, many of the results for a turbulent flow described above can be derived independently of the mechanism of dissipation. Here we use this fact to]{} derive the vKH equation for the 3D incompressible two-fluid equations (vKH-TF) describing a fully ionized hydrogen plasma, extending all previous MHD and HMHD results, which can be regarded as particular cases in the proper asymptotic limits. As a result, we obtain scaling predictions for a turbulent plasma while retaining the whole dynamics of both the electron and ion flows throughout all the relevant spatial scales. [These predictions should be valid down to the largest scale in which a plasma dissipation mechanism is present. The derivation is done considering in particular applications in space plasmas. We start by considering some of the implications of the “four-fifths” law for turbulence in hydrodynamics and in MHD, followed by the derivation of the vKH-TF equation, and its relevance to the description of turbulent space plasmas.]{}
The importance of the 4/5 law {#4/5lawHD}
=============================
[As mentioned in the Introduction, the vKH equation [@vkh1938] can be used as the starting point to obtain several essential results in turbulence theory. In particular, the hydrodynamic 4/5 law [@K1941b] states that in the limit of infinite Reynolds number, the longitudinal two-point third-order velocity structure function, $$B_{\parallel\parallel\parallel}^{vvv}(\textbf{r}) =
\left<\{[\textbf{v}(\textbf{x}+\textbf{r})-
\textbf{v}(\textbf{x})]\cdot\textbf{r}\}^3\right> ,
\label{eq:struc}$$ evaluated for spatial increments **r**, is given in terms of the spatial increment itself and of the mean energy dissipation per unit of mass $\varepsilon$ by $$\label{4/5hd}
B_{\parallel\parallel\parallel}^{vvv}(\textbf{r}) = -\frac{4}{5}\varepsilon r ,$$ where $r=|\textbf{r}|$, and the angular bracket in Eq. (\[eq:struc\]) denotes ensemble average. The prefactor $4/5$ in Eq. (\[eq:struc\]) gives its name to this exact result.]{}
[It is interesting that this result is independent of the dissipation mechanisms acting in the flow, and it only requires the existence of some dissipation at sufficiently small scales to get rid off all the power injected at large scales. As such, this law links information that is accessible from flow measurements at intermediate scales (the third-order moment of the velocity structure function, associated to the energy flux transported among scales by non-linear effects), to small scale quantities such as the rate of energy dissipation. As a result, this exact law has been used to identify the range of scales in which dissipation is negligible (the so-called “inertial range” of scales in a turbulent flow, for which the exact law holds), to meassure energy dissipation rates, and to define a Reynolds number without precise knowledge of the flow viscosity or of the microscopic dissipation mechanisms (see, e.g., [@WEY2007] and references therein).]{}
[The generalization of Eq. to the MHD case [@P1998a; @P1998b] relates longitudinal two-point third-order structure functions of the velocity and magnetic fields, to the spatial increment and to the energy dissipation rate. Although the prefactor changes, the law is still known as “four-fifths”. The validity of the MHD law has been subjected to several numerical testings (see, e.g., [@MI2009F; @Bo2009; @W2010; @Y2012]). Among several important results, the 4/5 law in MHD has been used to measure the energy cascade rate at large scales of the solar wind [@SV2007; @Sa2008; @M2008], to estimate the Reynolds number of solar wind turbulence [@WEY2007], to investigate large-scale solar wind models [@M1999; @MB2008], to predict the decay of MHD turbulence [@Da2009; @W2012], and to determine scaling exponents in measurements and in simulations using the so-called Extended Self-Similarity (ESS) hypothesis [@R1993; @G1997]. This last example highlights the importance of having an exact law: the ESS method uses the 4/5 law to calibrate structure functions, decreasing uncertainties and improving the analysis of the measurements.]{}
[The generalization of the vKH and the corresponding 4/5 law for a plasma would thus be useful not only to study scaling laws at large and intermediate turbulent plasma scales, but also to discern at what scale dissipation becomes relevant, and therefore the scaling would become invalid. In the solar wind and in space plasmas, there is a debate on whether scaling laws observed at high frequencies correspond to the inertial or dissipative ranges (see, e.g., [@M2011]). A derivation of exact laws for two-fluid equations for a fully ionized plasma is therefore a decisive step to elucidate this point, which should includes previous MHD and HMHD results as asymptotic limits. Next, we define the correlation tensors used to derive the 3D vKH-TF equation, and the two-fluid equations used to describe the plasma, and proceed to derive the 3D vKH-TF equation and the $4/5$ law.]{}
Correlation tensors {#def}
===================
The properties and structure of second-order correlation tensors are extensively discussed in [@ORM1997]. A derivation of the general form of homogeneous third-order correlation tensors can be found in [@P2003]. For completeness, we briefly present the main results of tensorial algebra needed to obtain the vKH-TF equation. Given two solenoidal vector fields **a**(**x**) and **b**(**x**), we define the second-order correlation tensor as $$\begin{aligned}
{R}_{ij}^{ab}(\textbf{x},\textbf{x}')=\left<a_i(\textbf{x})b_j(\textbf{x}')\right> = \left<a_i b_j' \right>,\end{aligned}$$ where **x**’$=$**x**+**r**, and the angular bracket denotes ensemble average. Spatial homogeneity implies that all regions of space are similar so far as statistical properties are concerned, which suggests that the results of averaging over a large number of realizations can be obtained equally well by averaging over a large region of space for one realization [@Ba1953]. Therefore, under this assumption, the second-order correlation tensors depend only on the relative separation **r** and is independent of its origin.
The most general expression for a solenoidal second-order correlation tensor is [@Ba1953] $$\begin{aligned}
\frac{{R^{ab}_{ij}(\textbf{r})}}{c_{ab}}=f^{ab}(r)\delta_{ij}+\frac{r}{2}\partial_rf^{ab}(r)P_{ij}+\epsilon_{ijk}\frac{r_k}{r}\tilde{f}^{ab}(r) ,\end{aligned}$$ where $c_{ab}=\bar{a}\bar{b}$, $\bar{a}$ is the root mean square value of **a**, and $c_{ab}f^{ab}(r)$ is the (dimensionless) longitudinal (along **r**) scalar correlation function. In addition, we have introduced a new pseudo-scalar function $\tilde{f}^{ab}(r)$ which is related to the not mirror-symmetric (or helical) part of the tensor. This pseudo-scalar function is, in general, not zero and related to the helicity of the fields involved. Finally, $P_{ij}(r)=\delta_{ij}- r_ir_j/r^2$ is the projector into the subspace of incompressible flows.
Third-order correlation tensors needed to derive the vKH-TF equation are of the form $$\begin{aligned}
S_{ikj}^{abc}(\textbf{x},\textbf{x}')=\left<a_i(\textbf{x})b_k(\textbf{x})c_j(\textbf{x}'))\right> = \left< a_i b_k c_j' \right> ,\end{aligned}$$ where **c**(**x**) is another vector field. These tensors satisfy $$\begin{aligned}
S_{ikj}^{abc}(\textbf{r}) & = & s_{11}(r)r_j\delta_{ik}+s_{12}(r)r_k\delta_{ij}+s_{13}(r)r_i\delta_{kj} \nonumber \\
&&+ s_3(r)r_ir_kr_j + \tilde{s}_{23}(r)r_i\epsilon_{kjm}r_m \nonumber \\ &&+ \tilde{s}_{22}(r)r_k\epsilon_{jim}r_m + \tilde{s}_{21}(r)r_j\epsilon_{ikm}r_m ,\end{aligned}$$ where the seven scalar functions are even functions of $r$, as for the second-order correlation tensor. Note that we omitted the superscript $abc$ for the sake of simplicity. The incompressibility condition leads to relationships between these seven functions and reduces the problem to only four generating functions (see [@P2003]). If the first two indices (i.e., $i$ and $k$) are symmetric, the tensor is generated only by one scalar and one pseudo-scalar function (the latter related to helicity [@Ch1996; @G2000; @K2003]).
Finally, the second and third order structure functions are defined in terms of the increments of the vectors as $$\begin{aligned}
\label{sf2}
{B^{ab}_{ij}(\textbf{r})} &=& \left<(a_i'-a_i)(b_j'-b_j)\right> , \\ \label{sf3}
{B^{abc}_{ikj}(\textbf{r})} &=& \left<(a_i'-a_i)(b_k'-b_k)(c_j'-c_j)\right> .\end{aligned}$$ Using homogeneity the following relations between structure functions and correlation tensors can be derived $$\begin{aligned}
\label{str}
{B^{ab}_{ij}(\textbf{r})} &=& 2R_{ij}^{ab}(0) - {R^{ab}_{ij}(\textbf{r})} - {R^{ab}_{ij}(-\textbf{r})} , \\
{B^{abc}_{ijk}(\textbf{r})} &=& \big[{S^{abc}_{ijk}(\textbf{r})}-{S^{abc}_{ijk}(-\textbf{r})}\big]\nonumber +\big[{S^{acb}_{ikj}(\textbf{r})} \\ && -{S^{acb}_{ikj}(-\textbf{r})}\big] + \big[{S^{bca}_{jki}(\textbf{r})}-{S^{bca}_{jki}(-\textbf{r})}\big] .\end{aligned}$$
Two fluid equations {#model}
===================
The equations of motion for a quasi-neutral incompressible plasma of ions and electrons with mass $m_{i,e}$, charge $\pm e$, densities $n_{i}=n_{e}=n$, pressures $p^{(i,e)}$, and respective velocities $\textbf{v}$ and $\textbf{u}$ are [@A2014a; @A2014b; @A2016] $$\begin{aligned}
m_en\frac{d \textbf{u}}{dt} &= -en\left(\textbf{E}+\frac{1}{c}\textbf{u}\times\textbf{B}\right)-\boldsymbol\nabla p^{(e)} + \mu^{(e)} \nabla^2\textbf{u}- \textbf{D} , \label{eq:1} \\
m_in\frac{d \textbf{v}}{dt} &= en\left(\textbf{E}+\frac{1}{c}\textbf{v}\times\textbf{B}\right)-\boldsymbol\nabla p^{(i)} + \mu^{(i)} \nabla^2\textbf{v}+ \textbf{D} , \label{eq:2} \\
\textbf{J} &= \frac{c}{4\pi}\boldsymbol\nabla\times\textbf{B}={en}(\textbf{v}-\textbf{u}) .\end{aligned}$$ Here $d/dt = \partial/\partial t + {\bf u} \cdot \nabla$ is the total derivative, $\textbf{B}$ and $\textbf{E}$ are the magnetic and electric fields, $\textbf{J}$ is the electric current density, $c$ is the speed of light, $\mu^{(i,e)}$ are viscosities, and $\textbf{D}$ is the rate of momentum gained by ions due to collisions with electrons, and assumed proportional to the relative speed between species, $\textbf{D} = - nm_i\nu_{ie}(\textbf{v} - \textbf{u})$, where $\nu_{ie}$ is the collisional frequency of ions against electrons. Incompressibility implies $\boldsymbol\nabla\cdot\textbf{u} = 0 = \boldsymbol\nabla\cdot\textbf{v}$. [Note that these equations neglect kinetic plasma dissipation mechanisms due to either electrons or ions. As mentioned in Sec. \[4/5lawHD\], the 4/5 law will hold for scales larger than any of the scales associated with these dissipation mechanisms, and as such, it should be independent of the actual damping mechanism acting in the plasma. The vKH-TF equation, on the other hand, will have explicit dissipation terms, and these terms may differ if other damping mechanisms are considered.]{}
Equations (\[eq:1\]) and (\[eq:2\]) can be written in dimensionless form in terms of a typical length $L_0$, the particle density $n$, a typical velocity $v_A=B_0/(4\pi nM)^{1/2}$ (the Alfvén velocity, where $B_0$ is a typical value of ${\bf B}$, and $M\equiv m_i+m_e$), and with the electric field in units of $E_0 = v_AB_0/c$, $$\begin{aligned}
\label{dlesse}
\mu\frac{d \textbf{u}}{dt} &= -\frac{1}{\lambda}(\textbf{E}+\textbf{u}\times\textbf{B})-\boldsymbol\nabla p^{(e)} +\nu^{(e)}\nabla^2\textbf{u}-\frac{\textbf{d}}{\lambda} , \\
\label{dlessp}
(1-\mu)\frac{d \textbf{v}}{dt} &= \frac{1}{\lambda}(\textbf{E}+\textbf{v}\times\textbf{B})-\boldsymbol\nabla p^{(i)}+\nu^{(i)}\nabla^2\textbf{v}+\frac{\textbf{d}}{\lambda} , \\
\label{dlesso}
\textbf{J} &= \frac{1}{\lambda}(\textbf{v}-\textbf{u}) .\end{aligned}$$ where we have introduced the dimensionless parameters $\mu\equiv m_e/M$ and $\lambda\equiv c/\omega_{M}L_0$, where $\omega_{M}=(4\pi e^2n/M)^{1/2}$ has the form of a plasma frequency for a particle of mass $M$. The dimensionless momentum exchange rate is $\textbf{d} = -\eta\textbf{J}$, where $\eta = m_ic^2\nu_{ie}/(4\pi e^2nv_A L_0)$ is the (dimensionless) electric resistivity. Dimensionless ion and electron inertial lengths can be defined in terms of their corresponding plasma frequencies $\omega_{i,e}=(4\pi e^2n/m_{i,e})^{1/2}$ simply as $\lambda_{i,e}\equiv c/\omega_{i,e}L_0$, and their expressions in terms of $\mu$ and $\lambda$ are simply $\lambda_i=(1-\mu)^{1/2}\lambda$ and $\lambda_e=\mu^{1/2}\lambda$. Note that in the limit of electron inertia equal to zero, we obtain $\omega_{M}=\omega_{i}$, and therefore $\lambda = \lambda_i = c/\omega_{i}L_0$ reduces to the usual Hall parameter.
To obtain a hydrodynamic description of the two-fluid plasma, we can write **u** and **v** in terms of two vector fields (see [@A2014b]): the hydrodynamic velocity $\textbf{U} = (1-\mu)\textbf{v}+\mu\textbf{u}$, and $\textbf{J}$ as given by Eq. . From these two fields, it is trivial to obtain **u** and **v** as $\textbf{u}=\textbf{U} - (1-\mu)\lambda\textbf{J}$ and $\textbf{v}=\textbf{U} + \mu\lambda\textbf{J}$. We will now see that this hydrodynamic description is useful to obtain the vKH-TF equation.
Results for a two-fluid plasma {#results}
==============================
The von Kármán-Howarth equation {#3dvkh}
-------------------------------
The $i$-th component of eqs. and evaluated at point **x** is $$\begin{aligned}
\label{motione_i}
{} && \mu \partial_tu_i = - \mu u_k\partial_k u_i - \frac{1}{{\lambda}}\big[E_i+\epsilon_{ilm}U_lB_m \nonumber \\ {} && -(1-\mu){\lambda}\epsilon_{ilm}J_lB_m\big] - \partial_i p^{(e)} + \nu^{(e)}\partial_{kk}^2u_i + \frac{\eta}{{\lambda}}J_i , \\ \label{motionp_i}
{} && (1-\mu) \partial_tv_i = - (1 - \mu) v_k\partial_{r_{k}} v_i + \frac{1}{{\lambda}}\big[E_i+\epsilon_{ilm}U_lB_m \nonumber \\ {} && + \mu{\lambda}\epsilon_{ilm}J_lB_m\big] - \partial_i p^{(i)} + \nu^{(i)}\partial_{kk}^2v_i - \frac{\eta}{{\lambda}}J_i .\end{aligned}$$
The time evolution equation for the second-order correlation tensor ${R^{uu}_{ij}(\textbf{r})}$ results from Eq. after multiplying by $u_j'= u_j(\textbf{x}+\textbf{r})$, and adding the $j$-th component of Eq. at point $\textbf{x}'$ multiplied by $u_i$. The time evolution of ${R^{vv}_{ij}(\textbf{r})}$ is obtained by performing similar operations on Eq. . The end result is $$\begin{aligned}
\label{motione_ij}
\partial_t\big[ &\mu& {R^{uu}_{ij}(\textbf{r})} \big] = \partial_{r_{k}}\big\{\mu\big[{S^{uuu}_{ikj}(\textbf{r})}-{S^{uuu}_{jki}(-\textbf{r})} \big] \nonumber \\
{} &-& (1-\mu)\big[{S^{BBu}_{ikj}(\textbf{r})}-{S^{BBu}_{jki}(-\textbf{r})} \big]\big\} - \big[{R^{Eu}_{ij}(\textbf{r})} \nonumber \\
{} &+&{R^{uE}_{ij}(\textbf{r})}\big]/{\lambda}\nonumber - \big[\epsilon_{ilm}{S^{UBu}_{lmj}(\textbf{r})} + \epsilon_{jlm}{S^{UBu}_{lmi}(-\textbf{r})}\big]/{\lambda}\nonumber \\
{} &+& 2\partial^2_{r_{k}r_{k}}\big[\nu^{(e)}{R^{uu}_{ij}(\textbf{r})}\big]+\eta\big[{R^{Ju}_{ij}(\textbf{r})}+{R^{uJ}_{ij}(\textbf{r})}\big]/{\lambda}\\
\label{motionp_ij}
\partial_t\big[ ( &1& -\mu){R^{vv}_{ij}(\textbf{r})} \big] = \partial_{r_{k}}\big\{(1-\mu)\big[{S^{vvv}_{ikj}(\textbf{r})}-{S^{vvv}_{jki}(-\textbf{r})} \big] \nonumber \\
{} &-& \mu\big[{S^{BBv}_{ikj}(\textbf{r})}-{S^{BBv}_{jki}(-\textbf{r})} \big]\big\} + \big[{R^{Ev}_{ij}(\textbf{r})} +{R^{uE}_{ij}(\textbf{r})}\big]/{\lambda}\nonumber \\
{} &+& \big[\epsilon_{ilm}{S^{UBv}_{lmj}(\textbf{r})} + \epsilon_{jlm}{S^{UBv}_{lmi}(-\textbf{r})}\big]/{\lambda}\nonumber \\
{} &+& 2\partial^2_{r_{k}r_{k}}\big[\nu^{(e)}{R^{vv}_{ij}(\textbf{r})}\big]-\eta\big[{R^{Jv}_{ij}(\textbf{r})}+{R^{vJ}_{ij}(\textbf{r})}\big]/{\lambda}\end{aligned}$$ where we have used the divergence-free condition for the fields $\textbf{u},\textbf{v}$, and $\textbf{B}$, $\partial_{r_k} \left< . \right>=\partial_k' \left< . \right>=-\partial_k \left< . \right>$ from homogeneity, and the relation $$\begin{aligned}
{S^{abc}_{ijk}(\textbf{r})}=\left<a_i(\textbf{x})b_j(\textbf{x})c_k(\textbf{x}+\textbf{r})\right>=\nonumber\\\left<a_i(\textbf{x}+\textbf{r})b_j(\textbf{x}+\textbf{r})c_k(\textbf{x})\right>={S^{abc}_{ijk}(-\textbf{r})} .\end{aligned}$$ Note that the gradient terms vanish because of isotropy [@Ba1953]. Adding eqs. and we obtain $$\begin{aligned}
\label{genera}
{} \partial_t \big[ &\mu& {R^{uu}_{ij}(\textbf{r})} + (1-\mu){R^{vv}_{ij}(\textbf{r})} \big] - \big[{R^{EJ}_{ij}(\textbf{r})}+{R^{JE}_{ij}(\textbf{r})}\big] \nonumber \\
{} &=& \partial_{r_{k}}\big\{\mu\big[{S^{uuu}_{ikj}(\textbf{r})}-{S^{uuu}_{jki}(-\textbf{r})} \big] + (1-\mu)\big[{S^{vvv}_{ikj}(\textbf{r})} \nonumber \\
{} &-& {S^{vvv}_{jki}(-\textbf{r})} \big]
-(1-\mu)\big[{S^{BBu}_{ikj}(\textbf{r})}-{S^{BBu}_{jki}(-\textbf{r})} \big] \nonumber \\
{} &-& \mu\big[{S^{BBv}_{ikj}(\textbf{r})}-{S^{BBv}_{jki}(-\textbf{r})} \big]\big\} + \big[\epsilon_{ilm}{S^{UBJ}_{lmj}(\textbf{r})} \nonumber \\
{} &+& \epsilon_{jlm}{S^{UBJ}_{lmi}(-\textbf{r})}\big] + 2\partial^2_{r_{k}r_{k}}\big[\nu^{(e)}{R^{uu}_{ij}(\textbf{r})}+ \nu^{(i)}{R^{vv}_{ij}(\textbf{r})}\big] \nonumber \\
{} &-& 2 \eta{R^{JJ}_{ij}(\textbf{r})}.\end{aligned}$$ The pseudo-tensor ${S^{UBJ}_{lmi}(\textbf{r})}$ on the r.h.s of Eq. can be expressed in terms of the derivative of the proper tensor ${S^{UBB}_{lmi}(\textbf{r})}$, since $$\begin{aligned}
\epsilon_{ilm}{S^{UBJ}_{jmi}(\textbf{r})} = \epsilon_{ilm}\epsilon_{jpq}\partial_{r_p}{S^{UBB}_{lmq}(\textbf{r})} = \partial_{r_p}{T^{UBB}_{ipj}(\textbf{r})} .\end{aligned}$$ We can write this tensor using the velocity of the species $${T^{UBB}_{ipj}(\textbf{r})} = \mu{T^{uBB}_{ipj}(\textbf{r})} + (1-\mu){S^{vBB}_{ipj}(\textbf{r})}.$$ Finally, noting that ${R^{JJ}_{ij}(\textbf{r})} = -\epsilon_{ipq}\epsilon_{jrs}\partial^2_{r_{p}r_{r}}{R^{BB}_{qs}(\textbf{r})}$, Eq. can be written as $$\begin{aligned}
\label{general}
\partial_t\big[ &\mu& {R^{uu}_{ij}(\textbf{r})} + (1-\mu){R^{vv}_{ij}(\textbf{r})}\big] - \big[{R^{EJ}_{ij}(\textbf{r})}+{R^{JE}_{ij}(\textbf{r})}\big] \nonumber \\
{} &=& \partial_{r_{k}}\big\{\mu\big[{S^{uuu}_{ikj}(\textbf{r})}-{S^{uuu}_{jki}(-\textbf{r})} \big] + (1-\mu)\big[{S^{vvv}_{ikj}(\textbf{r})} \nonumber \\
{} &-& {S^{vvv}_{jki}(-\textbf{r})} \big] -(1-\mu)\big[{S^{BBu}_{ikj}(\textbf{r})}-{S^{BBu}_{jki}(-\textbf{r})} \big] \nonumber \\
{} &-& \mu\big[{S^{BBv}_{ikj}(\textbf{r})}-{S^{BBv}_{jki}(-\textbf{r})} \big] + \mu\big[{T^{uBB}_{ikj}(\textbf{r})}\nonumber \\
{} &+& {T^{uBB}_{ikj}(-\textbf{r})}\big]+(1-\mu)\big[{T^{vBB}_{ikj}(\textbf{r})}+{T^{vBB}_{ikj}(-\textbf{r})}\big]\big\} \nonumber \\
{} &+& 2\partial^2_{r_{k}r_{k}}\big[\nu^{(e)}{R^{uu}_{ij}(\textbf{r})}+\nu^{(i)}{R^{vv}_{ij}(\textbf{r})} \big] + \nonumber \\ &+& 2\eta\epsilon_{ipq}\epsilon_{jrs}\partial^2_{r_{p}r_{r}}{R^{BB}_{qs}(\textbf{r})}.\end{aligned}$$ Equation is an exact law, valid even for anisotropic turbulence [@M1981; @P2008; @C2009]. This is our first main result.
We can now take the trace of Eq. , $$\begin{aligned}
\label{vkh_parallel1}
\partial_t\big[ &\mu& {R^{uu}_{ii}(\textbf{r})} + (1-\mu){R^{vv}_{ii}(\textbf{r})} + {R^{BB}_{ii}(\textbf{r})} \big] \nonumber \\
{} &=& \partial_{r_{k}}\big\{2\mu{S^{uuu}_{iki}(\textbf{r})} + 2(1-\mu){S^{vvv}_{iki}(\textbf{r})} \nonumber \\ {} &-& 2(1-\mu){S^{BBu}_{iki}(\textbf{r})} -2\mu{S^{BBv}_{iki}(\textbf{r})} + \mu\big[{T^{uBB}_{iki}(\textbf{r})} \nonumber \\
{} &+& {T^{uBB}_{iki}(-\textbf{r})}\big]+(1-\mu)\big[{T^{vBB}_{iki}(\textbf{r})}+{T^{vBB}_{iki}(-\textbf{r})}\big]\big\} \nonumber \\
{} &+& 2\partial^2_{r_{k}r_{k}}\big[\nu^{(e)}{R^{uu}_{ii}(\textbf{r})}+\nu^{(i)}{R^{vv}_{ii}(\textbf{r})}+\eta{R^{BB}_{ii}(\textbf{r})}\big],\end{aligned}$$ where, for $a$ equal to $u$ or $v$, $$\begin{aligned}
{T^{aBB}_{iki}(\textbf{r})} &=&\epsilon_{ilm}\epsilon_{ikq}{S^{aBB}_{lmq}(\textbf{r})} = (\delta_{lk}\delta_{mq}-\delta_{lq}\delta_{mk}){S^{aBB}_{lmq}(\textbf{r})} \nonumber \\
{} &=& {S^{aBB}_{kqq}(\textbf{r})}-{S^{aBB}_{qkq}(\textbf{r})}.\end{aligned}$$ From homogeneity, we have noted that $$\begin{aligned}
\label{mag}
{R^{EJ}_{ii}(\textbf{r})}+{R^{JE}_{ii}(\textbf{r})} &=& -\partial_t{R^{BB}_{ii}(\textbf{r})}.\end{aligned}$$ In the asymptotic limit of $\textbf r \rightarrow 0$, Eq. corresponds to twice the time variation of the total magnetic energy.
Now, assuming isotropy, we introduce the explicit form of the second- and third-order correlation tensors, $$\begin{aligned}
{} && \frac{{R^{aa}_{ii}(\textbf{r})}}{c_{aa}} = (3+r\partial_r)f^{aa}(r) , \\
{} && \frac{{S^{aab}_{iki}(\textbf{r})}}{c_{aab}} = \frac{4k^{aab} +r\partial_rk^{aab}}{2r}r_k= \text{S}^{aab}(r)r_k ,\end{aligned}$$ $$\begin{aligned}
\frac{{S^{abb}_{kii}(\textbf{r})}-{S^{abb}_{iki}(\textbf{r})}}{c_{abb}} &=& \frac{2k^{abb}+r\partial_rk^{abb}-4q^{abb}}{r}r_k \nonumber \\
{} &=& \text{T}^{abb}(r)r_k ,\end{aligned}$$ where $$\begin{aligned}
f^{aa}(r) &=& R_{\parallel\parallel}^{aa}/c_{aa} , \,\,\,\,
k^{aab}(r) = {S^{aab}_{\parallel\parallel\parallel}(\textbf{r})}/c_{aab} , \nonumber \\
q^{abb}(r) &=& {S^{abb}_{\perp\parallel\perp}(\textbf{r})}/c_{abb} , \nonumber\end{aligned}$$ are the longitudinal and transversal correlations of the fields $a,b=u,v$ and $B$, and where $c_{aa}=\bar{a}\bar{a}$ and $c_{abb}=\bar{a}\bar{b}\bar{b}$. By longitudinal ($\parallel$) and transversal ($\perp$), we mean in the direction along and perpendicular to the displacement vector **r**, respectively. Using these expressions and after some manipulation, Eq. reduces to $$\begin{aligned}
\label{vkh_parallel2}
( 3 &+& r\partial_r)\partial_t\big[\mu c_{uu}f^{uu} + (1-\mu) c_{vv}f^{vv} + c_{BB}f^{BB}\big] \nonumber \\
{} &=& 2(3+r\partial_r)\big\{\mu c_{uuu} S^{uuu} + (1-\mu) c_{vvv} S^{vvv} \nonumber \\
{} &-& (1-\mu) c_{BBu} S^{BBu} - \mu c_{BBv} S^{BBv} + \mu c_{uBB} T^{uBB} \nonumber \\
{} &+& (1-\mu) T^{vBB}
+2\partial_r\big[r^4\partial_r \big(\nu^{(e)}c_{uu}f^{uu}+\nu^{(i)}c_{vv}f^{vv} \nonumber \\
{} &+& \eta c_{BB}f^{BB}\big)\big]/r^4\big\}\end{aligned}$$ where we have used identities for isotropic turbulence [@Ba1953] $$\begin{aligned}
\partial_{kk} &= \partial^2_{rr} + \frac{2}{r}\partial_r ,\\
\big(\partial_{rr}^2+\frac{2}{r}\partial_r\big)(3+r\partial_r) &= (3+r\partial_r)\frac{1}{r^4}\partial_r(r^4\partial_r).\end{aligned}$$ A first integral of Eq. is $$\begin{aligned}
\label{vkh}
\partial_t\big[&\mu& c_{uu}f^{uu} + (1-\mu) c_{vv}f^{vv} + c_{BB}f^{BB} \big]\nonumber \\
{} &=& 2\big[\mu c_{uuu} S^{uuu} + (1-\mu) c_{vvv} S^{vvv} \nonumber \\
{} &-& (1-\mu) c_{BBu} S^{BBu} - \mu c_{BBv} S^{BBv} + \mu c_{uBB} T^{uBB} \nonumber \\
{} &+& (1-\mu) T^{vBB} \big] +2\partial_r\big[r^4\partial_r \big(\nu^{(e)}c_{uu}f^{uu} \nonumber \\
{} &+& \nu^{(i)}c_{vv}f^{vv}+\eta c_{BB}f^{BB})\big]/r^4.\end{aligned}$$ This exact relation is the von Kármán-Howarth equation for an incompressible two-fluid plasma, which is the second main result of the present paper.
The 4/5 law {#4/5law}
-----------
The vKH-TF Eq. can be writen in terms of the structure functions using the relation $$\begin{aligned}
{R^{aa}_{\parallel\parallel}(\textbf{r})}=\left<a_{\parallel}a_{\parallel}\right>-{B^{aa}_{\parallel\parallel}(\textbf{r})}/2,\end{aligned}$$ where we have used Eq. . Therefore $$\begin{aligned}
\partial_t\bigg[ &\mu& \left<u^2_{\parallel}\right>+(1-\mu)\left< v^2_{\parallel}\right>+\left<B^2_{\parallel}\right>\bigg] \nonumber \\
{} &-& \partial_t\big[\mu B^{uu}_{\parallel\parallel}+(1-\mu)B^{vv}_{\parallel\parallel}+B^{BB}_{\parallel\parallel}\big] \nonumber \\
{} &=& 2\big[ \mu c_{uuu} S^{uuu} + (1-\mu) c_{vvv} S^{vvv} - (1-\mu) c_{BBu} S^{BBu} \nonumber \\
{} &-& \mu c_{BBv} S^{BBv} + \mu c_{uBB} T^{uBB} + (1-\mu) T^{vBB}\big] \nonumber \\
{} &+& 2\nu^{(e)} r^{-4} \partial_r\left[r^4\partial_r\left(\left<u^2_{\parallel}\right>-B_{\parallel\parallel}^{uu}/2\right)\right] \nonumber \\
{} &+& 2\nu^{(i)} r^{-4}\partial_r\left[r^4\partial_r\left(\left<v^2_{\parallel}\right>-B_{\parallel\parallel}^{vv}/2\right)\right]\nonumber \\
{} &+& 2\eta r^{-4}\partial_r\left[r^4\partial_r\left(\left<B^2_{\parallel}\right>-B_{\parallel\parallel}^{BB}/2\right) \right] ,\end{aligned}$$ where we identify the mean energy dissipation rate per unit mass $\varepsilon_T$ for isotropic turbulence as $$\partial_t\left(\mu\left<u^2_{\parallel}\right>+(1-\mu)\left< v^2_{\parallel}\right>+\left<B^2_{\parallel}\right>\right) = -\frac{2}{3}\varepsilon_T.$$
Hereafter we adopt the usual assumption for fully developed turbulence [@F1995], i.e., we use the long-time limit in which a stationary regime is reached for sufficiently large Reynolds numbers, and in which $\varepsilon_T$ tends to a finite positive value. Therefore, at the inertial range $$\begin{aligned}
\label{pre4/5}
-\frac{2}{3}r\varepsilon_T &= \mu c_{uuu} (4k^{uuu} +r\partial_rk^{uuu}) \nonumber \\
{} &+ (1-\mu) c_{vvv} (4k^{vvv} +r\partial_rk^{vvv}) \nonumber \\
{} &- (1-\mu) c_{BBu} (4k^{BBu} +r\partial_rk^{BBu}) \nonumber \\
{} &- \mu c_{BBv} (4k^{BBv} +r\partial_rk^{BBv}) \nonumber \\
{} &+ 8\mu c_{uBB} (k^{uBB}+r\partial_rk^{uBB}/2-2q^{uBB}) \nonumber \\
{} &+ 8(1-\mu)c_{vBB} (k^{vBB}+r\partial_rk^{vBB}/2-2q^{vBB}).\end{aligned}$$ Equation can also be written in terms of the third-order correlation tensors as $$\begin{aligned}
\label{almost4/5}
-\frac{1}{6}r\varepsilon_T =\mu(S_{\parallel\perp\perp}^{uuu}+\frac{1}{2}S_{\parallel\parallel\parallel}^{uuu}) + (1-\mu) (S_{\parallel\perp\perp}^{vvv}+\frac{1}{2}S_{\parallel\parallel\parallel}^{vvv}) \nonumber \\ - (1-\mu) (S_{\parallel\perp\perp}^{BBu}+\frac{1}{2}S_{\parallel\parallel\parallel}^{BBu}) - \mu (S_{\parallel\perp\perp}^{BBv}+\frac{1}{2}S_{\parallel\parallel\parallel}^{BBv}) \nonumber \\ + \mu (S_{\parallel\perp\perp}^{uBB}-S_{\perp\parallel\perp}^{uBB}) + (1-\mu) (S_{\parallel\perp\perp}^{vBB}-S_{\perp\parallel\perp}^{vBB})\end{aligned}$$ where $\perp=2$ or 3 (no summation on $\perp$). Finally, using eqs. , and the incompressibility condition (${S^{abb}_{\perp\perp\parallel}(\textbf{r})}=-{B^{baa}_{\parallel\parallel\parallel}(\textbf{r})}/2$), we write Eq. as a function of the third-order structure functions $$\begin{aligned}
\label{4/5}
-\frac{4}{3}r\varepsilon_T = \bigg[{B^{vvv}_{\parallel ii}(\textbf{r})}+{B^{vBB}_{\parallel ii}(\textbf{r})}-{B^{BvB}_{\parallel ii}(\textbf{r})}-{B^{BuB}_{\parallel ii}(\textbf{r})} \bigg] \nonumber \\+ \mu\bigg[{B^{uuu}_{\parallel ii}(\textbf{r})}+{B^{uBB}_{\parallel ii}(\textbf{r})}-{B^{vvv}_{\parallel ii}(\textbf{r})}-{B^{vBB}_{\parallel ii}(\textbf{r})}\bigg].\end{aligned}$$ This equation is the $4/5$ law for the two-fluid plasma, and is the third main result of the present paper.
Discussion {#disc}
----------
Equations , , and give respectively: (1) an exact relation for the correlation functions of anisotropic turbulence (e.g., in the presence of a guide field) in a two-species incompressible plasma, (2) the vKH equation for isotropic turbulence, and (3) the $4/5$ law for the scaling of the flux in the inertial range. At the large scales, i.e., when $\lambda\rightarrow0$ and $\mu\rightarrow0$, we recover the MHD results [@Ch1951; @P1998a; @P1998b], and the hydrodynamic result when the magnetic field is taken equal to zero. When only $\mu\rightarrow0$ we obtain the HMHD case, previously studied by @Ga2008. However, unlike the results in @Ga2008, here we also express the results in terms of structure functions. [Also, it is important to emphasize that our expressions are written in terms of the velocity of each species in the plasma, which allow easier comparison with observations in space physics and with data from numerical simulations]{}.
The exact Eq. implies a scaling law for the third-order structure functions in the large Reynolds numbers and long-time limits. This law imposes correlations between the basic fields **u**, **v** and **B**, putting a strong constraint on the plasma turbulent dynamics. Finally, the equivalent expression involving only third-order structure functions is appropriate for comparisons with [*in-situ*]{} measurements in the solar wind at different spatial ranges (such as those in [@S2009]), or with laboratory plasmas.
Under certain assumptions, eqs. and also provide predictions for the scaling of the energy spectrum in a turbulent plasma. At the largest scales, using Eq. and assuming energy equipartition, we recover the well known result $\delta B\sim r^{1/3}$, which corresponds to the Kolmogorov spectrum for the total (kinetic plus magnetic) energy $E(k)\sim k^{-5/3}$. At intermediate scales (assuming $\mu=0$ and $r>\lambda$), and using the fact that the electron velocity $\delta u$ is proportional to $\delta B/r$, Eq. leads to $\delta B\sim r^{2/3}$, which corresponds to a magnetic energy spectrum $E_B(k)\sim k^{-7/3}$. Finally, at the smallest scales ($r<\lambda_e$) where the terms proportional to $\mu$ in Eq. become dominant, a new scaling for the energy inertial range emerges. At these scales, using $\delta u\sim \delta B/r$, we obtain $\delta B\sim r^{4/3}$, and therefore $E_{B}(k)\sim k^{-11/3}$. This scaling was recently obtained from dimensional arguments and observed in numerical simulations in [@A2014b]. It is also remarkable that these three different scalings are compatible with previous theoretical and numerical results [@M2010; @G2008; @Mi2007; @Mi2003; @Bi1997; @MG2010] and with solar wind observations [@SV2007; @MB2008; @S2009; @S2010; @S2011; @A2009].
Regarding the intermediate HMHD range, it is worth mentioning that an energy spectrum with a slope equal to $-7/3$ is incompatible with the assumption of an asymptotic separation of scales between the forced scales and the dissipative (assumed small) scales, as for a spectrum steeper than $-2$ dissipation peaks at large scales (i.e., in the HMHD range, at wavenumber $k\sim 1/\lambda$). However, at scales below the electron skin depth (where the HMHD model is no longer appropriate), the total energy spectrum (dominated by electron kinetic energy) is $E_k\sim k^{-5/3}$ [@A2014b]. Thus, as long as the HMHD range in the plasma is not too broad [(and as long as kinetic plasma damping mechanisms can be neglected at those scales)]{}, the scalings obtained above should hold. The dissipation anomaly for steep spectra has been studied before in the literature, for instance, in regularized MHD models [@P2006; @H2002].
Conclusions {#conclus}
===========
[We derived the von Kármán-Howarth equation, and its corresponding 4/5 law, for a 3D incompressible two-fluid plasma model. In particular, the derived vKH-TF equation can be written compactly in the usual form found for hydrodynamics [@R1940] as $$\frac{\partial Q}{\partial t} = - \frac{\partial T}{\partial r} +
\frac{\partial}{\partial r} \left[ \frac{1}{r^4}
\frac{\partial}{\partial r}r^4 D(Q)\right] ,$$ where $Q$ is a function associated to the second-order correlation tensors of the fields, $T$ to the third-order correlation tensors, and $D(Q)$ takes into account all dissipative effects. The structure of the functions $Q$ and $T$ \[see., e.g., Eq. (\[general\]) valid in the more general anisotropic case, and Eq. (\[vkh\]) valid in the isotropic case\] is independent of the mechanism of dissipation acting on the plasma. However, the structure of the function $D(Q)$ depends on the dissipation mechanism, and should therefore be expected to change if kinetic plama dissipation mechanisms (e.g, Landau damping, cyclotron-resonant damping, or other mode-coupling dissipative processes) are present.]{}
[To obtain the generalization of the 4/5 law for a two-fluid plasma \[Eq. (\[4/5\])\] we adopted the usual assumption of fully developed turbulence, where an asymptotic regime is expected to be reached for sufficiently large Reynolds numbers. This allows for the existence of a range of scales for which dissipation mechanisms are negligible (i.e., $D\approx 0$ at those scales), and as a result $\partial_t Q$ can be associated to the mean energy dissipation rate per unit mass. Note that this is independent of the particular dissipation mechanisms in the plasma, as long as they act at sufficiently small scales, and as long as they allow for an asymptotic regime to be reached for sufficiently large scale separation. Another equivalent approach to derive the 4/5 law is to consider that in the turbulent steady regime, for which $\partial_tQ\approx0$, the dissipative term must be equal to the mean power injected into the system per unit mass, which is equal to $-\varepsilon_T$. Therefore, regardless of the particular mechanisms of dissipation present in the plasma (and in particular, in the solar wind), Eq. (\[4/5\]) provides an exact law which should hold as long as the energy injection rate balances the energy dissipation rate, and for all scales larger than the scale at which collisional or kinetic plasma dissipation mechanisms become dominant. The length of this scale will depend on specific properties of the plasma considered. As an example, in the solar wind at 1 AU dissipation mechanisms seem to become relevant for frequencies of $0.5$ Hz [@L1998], although there is a debate on whether this scale is indeed dissipative or inertial [@M2011].]{}
[The 4/5 law for a two-fluid plasma thus includes the effect of ion and electron inertia in the scaling of turbulence, and generalizes previous results obtained for MHD and HMHD. For scales larger than dissipative scales, it implies a specific scaling for structure functions of the velocity of each species and of the magnetic field, which can be related to the energy dissipation rate at the smallest scales. Therefore, it provides a way to test if a range of scales in a plasma is inertial or dissipative: if scaling laws observed in that range satisfy Eq. (\[4/5\]), then the scaling is the result of a turbulent cascade. It also provides a way to calibrate observations (e.g., higher order structure functions) against the 4/5 law, as often done when the ESS hypothesis are used to analyze data. In the range of scales in which it holds, it implies scaling laws for the energy spectrum at scales larger and smaller than the ion skin depth, as discussed in Sec. \[disc\]. Equation (\[4/5\]) gives a way to quantify the total energy dissipation rate per unit mass in a plasma from observations at scales larger than the kinetic scales at which the dissipation mechanisms become dominant. Finally, since the expressions are given in terms of the velocity field of each species, the procedure used here can be extended to consider multi-species plasmas in a straightforward fashion.]{}
In the study of turbulent flows, exact laws provide an essential tool to analyze data and understand non-linear cascades. Over the last years, the sustained increase in the spatial and temporal resolution of space missions such as Cluster (ESA) or the new NASA MMS (Magnetospheric MultiScale) mission has opened the possibility to study a number of small-scale plasma phenomena as never before. The exact laws derived here allow investigation of the nature of turbulent magnetic field fluctuations at a broad range of scales in space plasmas, and will be indispensable to understand the nature of turbulence at the smallest scales in the solar wind.
Acknowledgments {#acknowledgments .unnumbered}
===============
We acknowledge support from grants No. PIP 11220090100825, UBACyT 20020110200359 and 20020100100315, and PICT 2011-0454 and 2011-1529.
| {
"pile_set_name": "ArXiv"
} |
---
author:
- 'E.V. Luschevskaya'
- 'O.V. Teryaev'
- 'D.Yu. Golubkov'
- 'O.V. Solovjeva'
- 'R.A.Ishkuvatov'
title: 'Tensor polarizability of the vector mesons from $SU(3)$ lattice gauge theory'
---
Introduction {#intro}
============
The influence of strong magnetic fields on quark-hadron matter represents rich and full of surprising effects area of science. These fields could exist in the Early Universe, they can influence the physics of neutron stars, and lead to non-trivial effects in non-central heavy-ion collisions in terrestrial laboratories. Behaviour of the hadron energy in external magnetic field may provide information about the particle internal structure. In strong magnetic field the hadronic wave function deforms. This deformation is defined both by QED and QCD interactions inside the hadron. The magnetic polarizability and hyperpolarizabilities are quantities describing the response of the hadron to the external magnetic field, which we have explored in our previous work [@2017]. Apparently, this non-linear response arises solely due to the strong QCD interaction binding quarks together.
The magnetic field effect on the hadronic energy and structure was explored by theoretical models [@Simonov:2013; @Simonov; @Cho:2015; @Taya:2015; @Kawaguchi:2016; @Hattori:2016; @Gubler:2015] as well as lattice calculations [@Martinelli:1982; @Liu:2015; @Luschevskaya:2015a; @Savage:2015; @Luschevskaya:2015b; @Bali:2015; @Luschevskaya:2016]. The polarizabilities of hadrons were also investigated by analytical methods using dispersion relations in [@baldin; @Filkov:2006], the magnetic moments were studied in [@Samsonov; @Aliev; @Djukanovic; @Lee; @Owen].
Below we discuss the calculations in lattice gauge theory with chiral invariant Dirac operator without dynamical quarks. Our method enables to calculate the hadronic energies for different spin projections on the magnetic field axis. Magnetic polarizabilities and moments can be extracted as the fit parameters from the magnetic field value dependence of the energy. The values of the magnetic polarizabilities depend on the meson spin projection on the field axis. The physical meaning of this phenomenon is related to different deformations of a hadron in the various space direction. Therefore, the presence of a magnetic field creates a kind of anisotropy in space, which can lead to tensor polarization (alignment) of the vector meson and, after its decay, to dileptonic asymmetry in collisions of heavy ions. Dilepton anisotropy [@Teryaev] is an important physical characteristic which is sensitive to different channels of particle’s decay and can be utilized to disentangle contributions of some channels. Dilepton asymmetries provide information about the evolution of quark-gluon plasma in non-central heavy-ion collisions [@Baym:2017gzx]. Here we study the influence of magnetic field on the tensor polarization of $\rho$ mesons which was detected in angular distribution of their decay products. This requires a more accurate calculation of polarizability, so, we repeat we repeat, improve and extend our previous analysis [@2017; @Luschevskaya:2015a].
Technical details of the simulations {#sec-1}
====================================
We used ensembles of statistically independent $SU(3)$ gauge field configurations. For the generation of these configurations the improved Lüscher-Weisz action [@Luscher:1985] was used: $$S=\beta_{{\mbox{imp}}} \sum_{{\mbox{pl}}} S_{{\mbox{pl}}}-\frac{\beta_{{\mbox{imp}}}}{20 u^2_0}\sum_{{\mbox{rt}}}S_{{\mbox{rt}}},$$ where $S_{{\mbox{pl}},{\mbox{rt}}}=(1/3){\mbox{Tr}}(1-U_{{\mbox{pl}},{\mbox{rt}}})$ is the plaquette and rectangular loop terms respectively, $u_0 = (W_ {1 \times 1})^{1/4} $ is defined by the relation $ W_{1 \times 1} = \langle (1/2) {\rm Tr} U_{{\mbox{pl}}} \rangle$ calculated at zero temperature [@Bornyakov:2005].
The calculations were carried out in $SU(3)$ lattice gauge theory without dynamical quarks. We consider the lattice volumes $N_t \times N_s^3 = 18^4$ and $20^4$ and a set of lattice spacings $a=\{0.095, 0.105, 0.115\}\, {\mbox{fm}}$. In Table \[tbl:sim\_det\] we show the lattice volume $N_t\times N_s^3$, the lattice spacing $a$, the corresponding $\beta_{{\mbox{imp}}}$ values and the number of configurations.
$N_t \times N_s^3 $ $ \beta_{{\mbox{imp}}}$ $a, \ {\mbox{fm}}$ $ N_{{\mbox{conf}}} $
--------------------- ------------------------- -------------------- -----------------------
$18^4$ 8.20 0.115 300
$18^4$ 8.30 0.105 400
$18^4$ 8.45 0.095 200
$20^4$ 8.20 0.115 300
\[tbl:sim\_det\]
In order to construct meson correlation function, we used quark propagators, which can be extracted as a series of eigenfunctions and eigenvalues of discrete lattice version of the Dirac operator.
$$D^{-1}(x,y)=\sum_{k<M}\frac{\psi_k(x) \psi^{\dagger}_k(y)}{i \lambda_k+m},
\label{lattice:propagator}$$
where $ M = $ 50 is used for the calculations.
The next step is the numerical solution of the Dirac equation and finding the eigenfunctions $\psi_k$ and the eigenvalues $\lambda_k$ of the Dirac operator for a quark located in an external gauge field $A_{\mu}$. $$D \psi_k=i \lambda_k \psi_k, \ \ D=\gamma^{\mu} (\partial_{\mu}-iA_{\mu})
\label{Dirac}$$ In these calculations we used the Neuberger overlap operator [@Neuberger:1997]. This operator enables to consider the limit of massless quarks without breaking the chiral symmetry and can be written in the following form $$D_{{\mbox{ov}}}=\frac{\rho}{a}\left( 1+D_W/\sqrt{D^{\dagger}_W D_W} \right),
\label{overlap}$$ where $D_W = M-\rho/a $ is the Wilson-Dirac operator with the negative mass parameter $\rho/a$, $a$ is the lattice spacing, $M$ is the Wilson term. The sign function ${\mbox{sign}}(H_W)$ of the Hermitian Wilson-Dirac operator $H_W$ is determined by the following relation $$\gamma_5 {\mbox{sign}}(H_W)= D_W/ \sqrt{(D_W)^{\dagger} D_W}.
\label{sign_function}$$ It is calculated using the min-max method of approximation by polynomials.
The fermionic fields satisfy the periodic boundary conditions in space and the antiperiodic boundary conditions in time. We investigate the behaviour of the ground state energy of the meson in a gauge field, which is the sum of the gluonic field and the external constant magnetic field. The magnetic field interacts only with quarks, so the magnetic field was added only into the Dirac operator. A constant magnetic field $B$ is directed along the z-axis. $$A_{\mu \, ij}\rightarrow A_{\mu \, ij} + A_{\mu}^{B} \delta_{ij},
\label{exchange}$$ where $$A^B_{\mu}(x)=\frac{B}{2} (x_1 \delta_{\mu,2}-x_2\delta_{\mu,1}).$$
The quantization of the magnetic field on a torus was originally discussed in [@Hooft:1979; @Zainuddin:1989; @Chen:1996]. In order to satisfy the boundary conditions for fermions, the magnetic field has to be quantized on the lattice [@Al-Hashimi:2009] $$qB = \frac{2 \pi k} {(aN_s) ^ 2}, \ \ k \in \mathbb {Z},
\label{quantization}$$ where $ q = - 1/3 \, e $, $aN_s$ is the lattice space extension. The calculation of the eigenfunctions and eigenvalues of the Dirac operator allows us to find the value of the two-point correlation function by means of which we determine the energy of the ground state.
Meson correlation functions
===========================
We calculate the following correlation functions $$\langle\psi^{\dagger}(x) \gamma_i \psi(x) \psi^{\dagger}(y) \gamma_j \psi(y)\rangle_A,
\label{observables}$$ where $ \gamma_i, \gamma_j $ are the Dirac gamma matrices with the Lorentz indexes $ i, j = 1,2, 3 $ $ x = (\textbf {n} a, n_ta) $ and $ y = (\textbf{n}^{\prime} a, n^{\prime}_t a) $ are the coordinates on the lattice.
The spatial coordinates on the lattice are written as follows: $ \textbf{n}, \textbf{n}^{\prime} \in \Lambda_3 = \{(n_1, n_2, n_3) | n_i = 0,1, ..., N -1 \}$, $n_t, n^{\prime}_t$ are the numbers of sites in the time direction. In the Euclidean space $\psi^{\dagger} = \bar{\psi}$. The correlators of the $\rho^{\pm}$ mesons are computed with the use of the relation $$\langle \bar{\psi}_{d,u}(x) \gamma_i \psi_{u,d}(x) \bar{\psi}_{u,d}(y) \gamma_j \psi_{d,u}(y) \rangle_A=-{\mbox{Tr}}[\gamma_iD^{-1}_{u,d}(x,y)\gamma_jD^{-1}_{d,u}(y,x)],
\label{lattice:correlator}$$ where $D^{-1}_d$ and $D^{-1}_u$ are the propagators of the $d$ and $u$ quarks in the coordinate space. We consider the isospin symmetry case, so at zero magnetic field $D^{-1}_d=D^{-1}_u$, but at non-zero field this equality is not fulfilled.
Since the $u$ and $d$ quarks interact differently with the field, the correlator of the $\rho^0$ meson is represented by the sum $$\langle \bar{\psi}_{d}(x) \gamma_i \psi_{d}(x) \bar{\psi}_{d}(y) \gamma_j \psi_{d}(y) + \bar{\psi}_{u} (x)\gamma_i \psi_{u}(x) \bar{\psi}_{u}(y)\gamma_j \psi_{u}(y) \rangle_A=
\label{lattice:correlator_rho0}$$ $$-{\mbox{Tr}}[\gamma_iD^{-1}_{d}(x,y)\gamma_jD^{-1}_{d}(y,x)]-{\mbox{Tr}}[\gamma_iD^{-1}_{u}(x,y)\gamma_jD^{-1}_{u}(y,x)].$$
We numerically perform Fourier transform of the correlators from the coordinate space to the momentum space. In order to find the ground-state energy, we consider mesons with zero spatial momentum.
The ground-state energy of a meson with a definite spin projection on the magnetic field axis is defined by the covariant density matrix. We express the spin density matrix in terms of transverse ($e_x=(0,1,0,0), e_y=(0,0,1,0)$) and longitudinal ($e_z=(0,0,0,1)$) polarization vectors. Then the energies of the $\rho^{\pm}$ and $\rho^0$ mesons with the spin projection $s_z=0$ are obtained from equations and respectively after Fourier transform, where $i,j=3$.
The combinations of the correlators $$C(s_z=\pm 1)=\langle O_1 (t)\bar{O}_1 (0)\rangle_A+\langle O_2 (t)\bar{O}_2 (0)\rangle_A \pm i(\langle O_1 (t)\bar{O}_2 (0)\rangle_A-\langle O_2 (t)\bar{O}_1 (0)\rangle_A)
\label{eq:CVV1}$$ give the energies of vector mesons with the spin projections equal to $+1$ and $-1$ on the field axis, where $O_1=\psi^{\dagger}_{d,u}(x) \gamma_1 \psi_{u,d}(x),\, O_2=\psi_{d,u}^{\dagger}(x) \gamma_2 \psi_{u,d}(x)$ are the interpolation operators of the $\rho^{\pm}$ mesons. The interpolation operators for the $\rho^0$ case are constructed similarly to taking into account .
The correlation function can be expanded in a series over the eigenstates of the Hamiltonian $\widehat{H}$\
$$\langle O_i(t) \bar{O}_j(0)\rangle_T=\frac{1}{Z}\sum_{m,n}\langle m|e^{-(T-t)\widehat{H}} \widehat{O}_i|n \rangle \langle n|e^{-t\widehat{H}} \widehat{O}_j^{\dagger}|m \rangle=
\label{series}$$ $$= \frac{1}{Z}\sum_{m,n}e^{-(T-t)E_m}\langle m|\widehat{O}_i|n \rangle e^{-tE_n}\langle n| \widehat{O}_j^{\dagger}|m \rangle ,$$ where $i,j=1,2,3$, $E_m$ and $E_n$ are the energies of the excited states with the numbers $m$ and $n$, and $Z=\sum_n \langle n|e^{-T\widehat{H}} |n \rangle=\sum_n e^{-TE_n}$ is the partition function.
In expression we take out the factor $e^{-TE_0}$, as a result, we obtain the following relation\
$$\langle O_i(t) \bar{O}_j(0)\rangle_T=\frac{\sum_{m,n}e^{-(T-t)\Delta E_m}\langle m|\widehat{O}_i|n \rangle e^{-t \Delta E_n}\langle n| \widehat{O}_j^{\dagger}|m \rangle}{1+e^{-T \Delta E_1}+e^{-T \Delta E_2}+...},
\label{series2}$$ where $\Delta E_n=E_n-E_0$. In the thermodynamical limit of the theory $T\rightarrow\infty$ from we obtain $$\langle O_i(t) \bar{O}_j(0)\rangle_{T\rightarrow \infty}= \sum_{n} \langle 0|\widehat{O}_i|n \rangle \langle n| \widehat{O}_j^{\dagger}|0 \rangle e^{-t E_n}.$$ In our case we got the following expression\
$$C(n_t)=\langle \psi^{\dagger}(\textbf{0},n_t) \gamma_i \psi(\textbf{0},n_t) \psi^{\dagger}(\textbf{0},0) \gamma_j \psi(\textbf{0},0)\rangle_A =$$ $$\sum_k\langle 0|\widehat{O}_i|n \rangle \langle n|\widehat{O}^{\dagger}_{j}|0 \rangle e^{-n_t a E_n}.
\label{sum}$$ One can see from that at-large $n_t$ the main contribution comes from the ground state. Taking into account periodic boundary conditions on the lattice we obtain the final formula $$C_{fit}(n_t)=A_0 e^{-n_t a E_0} + A_0 e^{-(N_T-n_t) a E_0}=$$ $$2A_0 e^{-N_T a E_0/2} \cosh ((\frac{N_T}{2}-n_t) a E_0),
\label{coshfit}$$ where $A_0$ is the constant, $E_0$ is the ground state energy. We use this formula for fitting the correlation functions obtained from the lattice propagators. From this fits the ground state energies can be obtained.
Calculation of $m_{eff}$ and its statistical errors
===================================================
The $\rho$-meson effective masses $m_{eff}$ were extracted from $\chi^2$-fits to the standard asymptotic parametrization , with fit parameters being $m_{eff} = aE_0$ and $A_0$. The effective mass can be also found from the following equation: $$\frac{C(n_t)}{C(n_t+1)}=\frac{\cosh(m_{eff}(n_t-N_T/2))}{\cosh(m_{eff}(n_t+1-N_T/2))},
\label{effmass}$$ in which the boundary conditions on the lattice are taken into account. We have checked that the results for the energies coincide with the results obtained from , taking into account the correlation matrix between the adjacent points of the plateau.
In order to control the stability of resulting $m_{eff}$ values, i.e. to determine the $m_{eff}(n_t)$ plateau, the fits were performed in four ranges of $n_t = \left[\frac{N_T}{2}-k, \frac{N_T}{2}+k\right], \,\, k=1..4$, where the fit quality could be satisfactory, $\chi^2/d.o.f. \sim 1$. Despite that the fit quality was often acceptable already at $k \simeq 4$, the values of $m_{eff}$ at $k = 4$ could still be systematically overestimated.
Usually, the $m_{eff}$ plateau was reached at $k=3$, i.e. $n_t=6\div12 \, (7\div13)$ for lattice volume $18^4 (20^4)$. In some cases a narrower $n_t$ range was conservatively chosen after a visual inspection of the fits to ensure that $m_{eff}$ was well within the plateau. A typical graph of $m_{eff}(\rho^0)$ depending on the fit range $n_t$ is shown in Fig. \[Fig:m\_eff\_example\] for the lattice volume $18^4$, the lattice spacing $a=0.115\ {\mbox{fm}}$, spin projections $|s_z|=0,1$ on the magnetic field axis and two values of the magnetic field.
The uncertainties attributed to the resulting $\rho$-meson effective masses $m_{eff}$ are statistical-only, as determined by the fit in the chosen $n_t$ range.
Magnetic polarizabilities of the $\rho^{\pm}$ mesons {#sec-5}
====================================================
In the magnetic field the energy levels of the point-like charged particle are described by the following dependency [@Kharzeevbook] $$E^2=p_z^2+(2n+1)|qB|-gsqB+m^2,
\label{eqLL}$$ where $p_z$ is the momentum in the magnetic field direction, $n$ is the principal quantum number, $q$ is the particle electric charge, $g$ is the g-factor, $s$ and $m$ are the particle spin and mass respectively.
As we discussed earlier [@2017] the magnetic field can affect the internal structure of vector mesons if it is sufficiently strong. This influence is characterized by non-zero magnetic polarizabilities and hyperpolarizabilities which depend on the projection of the meson spin $s_z$ on the direction of the magnetic field. It means that the response of the fermionic currents inside the meson is defined by the mutual orientation of the quarks and the external magnetic field. These phenomena are interesting and can give a contribution to the polarization of the emitted charged particles in the strong magnetic field which we discuss below.
According to the parity conservation for the spin projection $s_z=0$ the energy squared gets corrections from the non-linear terms of even powers in a magnetic field. For the spin projections $s_z=+1$ and $s_z=-1$ the energy squared contains the terms of both even and odd powers in the field. We found the contributions of these terms depend on the interval of magnetic fields considered. So, for the lattices considered the correction of the third power term to the lowest energy sub-level with $qs_z=+1$ is not larger than $20\%$ and compatible with errors at $eB\in [0,1.2]\, {\mbox{GeV}}^2$. It may be seen from Fig. 9 presented in our previous work [@2017]. Similarly, a correction of the fourth power does not give a significant contribution to the square of the energy for the case $s_z=0$.
Therefore, we obtain the dipole magnetic polarizability for the spin projection $s_z=0$ from the fit of lattice data by the following relation $$E^2_{s_z=0} =|eB|+m^2 -4 \pi m \beta_m (eB)^2
\label{eq:betas0}$$ at $eB\in[0,1.2]\ {\mbox{GeV}}^2$, where $eB$ is the magnetic field in ${\mbox{GeV}}^2$, $\beta_m$ is the dipole magnetic polarizability and fit parameter, $m$ is the fit parameter also. The lattice results together with the fitting curves are shown in Fig. \[Fig:rho:beta\_s0\] which shows the energy squared increasing with the magnetic field value. At $eB\sim 0.2\div0.3\, {\mbox{GeV}}^2$ we observe a hump which leads to increase of the errors in determination of $\beta_m(s_z=0)$ value. It can be a result of lattice spacing and lattice volume effects or have some physical explanation. We cannot perform a rigorous analysis due to high errors of the calculations, but we note that this hump is absent for the lattice with smaller lattice spacing $a=0.095\ {\mbox{fm}}$. Therefore it is more probably a lattice spacing artefact.
The values of the $\beta_m$ with the values of $\chi^2/n_{d.o.f.}$ and lattice parameters are shown in Table \[Table\_s0\].
$V$ $m_{\pi}({\mbox{MeV}})$ $a({\mbox{fm}})$ $\beta_m({\mbox{GeV}}^{-3})$ $\chi^2/d.o.f.$
-------- ------------------------- ------------------ ------------------------------ -----------------
$18^4$ $574 \pm 7$ $0.105$ $0.03\pm 0.01$ $6.90$
$18^4$ $395 \pm 6$ $0.115$ $0.028\pm 0.006$ $0.53$
$18^4$ $541 \pm 3$ $0.115$ $0.027\pm 0.004$ $1.25$
: The magnetic dipole polarizability $\beta_m$ of the charged $\rho$ meson with spin projection $s_z=0$ obtained from the fits of the lattice data sets. The pion masses are represented in the second column, the lattice spacings in the third column, the values of $ \chi^2/d.o.f$ are shown in the last one.[]{data-label="Table_s0"}
The behaviour of the energy squared for the case $qs_z=+1$ (which corresponds to $\rho^{-}$ at $s_z=-1$ and $\rho^{+}$ at $s_z=+1$) can be described by the dependency $$E^2_{qs_z=+1} =|eB| -g(eB) +m^2 -4 \pi m \beta_m (eB)^2
\label{eq:betasp1}$$ at $eB\in[0,1.2]\ {\mbox{GeV}}^2$. The magnetic dipole polarizability was obtained from the fit of lattice data by the relation , where $m$, $g$ and $\beta_m$ are the fit parameters.
We represent this energy component in Fig.\[Fig:rho:beta\_s1\] for the lattice volume $18^4$, lattice spacings $0.105\ {\mbox{fm}}$, $0.115\ {\mbox{fm}}$ and for the lattice volume $20^4$, lattice spacing $0.115\ {\mbox{fm}}$. The values of pion mass are also shown.
The values of the magnetic dipole polarizability are represented in Table \[Table\_s1\]. These results are obtained from the 3-parameter fit in comparison with the results of our previous paper (see Table 3 and Fig.9), where we fixed the $g$-factor value and used 2-parametric fit. The $\beta_m$ values are in a good agreement with our previous ones within the error range.
$V$ $m_{\pi}({\mbox{MeV}})$ $a({\mbox{fm}})$ $g$-factor $\beta_m({\mbox{GeV}}^{-3})$ $\chi^2/d.o.f.$
-------- ------------------------- ------------------ ----------------- ------------------------------ -----------------
$18^4$ $574 \pm 7$ $0.105$ $2.48 \pm 0.19$ $-0.049\pm 0.010$ $2.66$
$18^4$ $541 \pm 3$ $0.115$ $2.26 \pm 0.14$ $-0.041\pm 0.006$ $2.32$
$20^4$ $535 \pm 4$ $0.115$ $2.19\pm 0.12$ $-0.044\pm0.006$ $1.48$
$18^4$ $395 \pm 6$ $0.115$ $2.12\pm 0.13 $ $ -0.039\pm 0.006$ $1.49$
: The magnetic dipole moment and the magnetic dipole polarizability of the charged $\rho$ meson with $qs_z=+1$ for the lattice spacings $0.105\ {\mbox{fm}}$, $0.115\ {\mbox{fm}}$, the lattice volume $18^4$, various pion masses and for the lattice spacing $0.115\ {\mbox{fm}}$, the lattice volume $20^4$ and the pion mass $m_{\pi}=535(4)\ {\mbox{MeV}}$ with their errors and $ \chi^2/d.o.f$ values. The results were obtained with the use of 3-parametric fit at $eB\in[0,1.2]$ shown in Fig. \[Fig:rho:beta\_s1\].[]{data-label="Table_s1"}
In Fig.\[Fig:rho:allspins\] we represent the energy squared for different meson spin projections on the field axis for the lattice volume $18^4$, the lattice spacing $a=0.115\ {\mbox{fm}}$ and the pion mass equal to $541\ {\mbox{MeV}}$. The energy errors for the case $qs_z=-1$ are sufficiently high. For the case $qs_z=-1$ the fit was using the formula $$E^2_{qs_z=-1} =|eB| +g(eB) +m^2 -4 \pi m \beta_m (eB)^2\,.
\label{eq:betasm1}$$
To estimate the contribution of the magnetic dipole polarizability to the lepton asymmetry it is needed to know the dipole magnetic polarizability for the $s_z=+1$ and $s_z=-1$ cases. However, due to big statistical errors the magnetic dipole polarizability cannot be obtained from the energy dependence at $s_z=+1$ if we consider $\rho^-$ or from $s_z=-1$ if we consider $\rho^+$. Nevertheless, the parity conservation demands the equality of the magnetic dipole polarizabilities for $s_z=+1$ and $s_z=-1$. Then it follows from and $$E^2_{qs_z=-1}-E^2_{qs_z=+1}=2g(eB)\,.
\label{eq:diff}$$
In order to make sure that the relation is satisfied for our data we represent the value of $(E^2_{qs_z=-1}-E^2_{qs_z=+1})/2$ versus the magnetic field value in Fig. \[Fig:rho:gfactor\]. This quantity increases linearly with the magnetic field. The values of the $g$-factor obtained from these fits are equal to $2.4\pm0.1$ for the lattice spacing $a=0.105\ {\mbox{fm}}$ and $2.40\pm0.04$ for the spacing $a=0.115\ {\mbox{fm}}$. The errors of the $g$-factor determination are underestimated because the effective mass plateau is very noisy for the $qs_z=-1$. We have determined the $g$-factor from the lowest energy sub-level in our previous work [@2017].
Magnetic polarizability of the $\rho^0$ meson {#sec-6}
=============================================
We have discussed the magnetic polarizabilities of the neutral $\rho$ mesons previously in [@Luschevskaya:2015a], where we considered the toy model with one type of quarks. In an external magnetic field the u- and d-quarks couple differently to the magnetic field. It has to be taken into account when we calculate the physical observables. We repeat the analysis represented in [@Luschevskaya:2015a] with the difference that we find the energies from the correlation function and increase the statistics.
In the relativistic case the energy of the neutral $\rho$ meson with spin projection $s_z=0$ is described by the following dependency $$E^2 =m^2 -4 \pi m \beta_m (eB)^2-4 \pi m \beta_m^{h2} (eB)^4-4 \pi m \beta_m^{h4} (eB)^6-4 \pi m \beta_m^{h6} (eB)^8-...\,,
\label{eq:beta_rho0_s0}$$ where $\beta_m^{h2}$, $\beta_m^{h4}$ and $\beta_m^{h6}$ are the various magnetic hyperpolarizabilities of higher orders, $m$ is the mass of the meson at zero field.
In Fig. \[Fig:rho:s0uudd\] we represent the energy squared of the $\rho^0$ meson with $s_z=0$ versus the field squared for various lattices and pion masses at $(eB)^2\in[0:0.5]\, {\mbox{GeV}}^4$. The lattice data are depicted by points. We observe the energy decrease rapidly for all the lattice data. It also strongly depends on the lattice spacing, the lattice volume, and the pion mass. The lines are the fits of these data obtained with the use of formula when we also include terms $\sim (eB)^{10}$ and $\sim (eB)^{12}$. For the lattice with spacing $a=0.105\ {\mbox{fm}}$ the fits were performed at $(eB)^2\in[0:1.7]\, {\mbox{GeV}}^4$, for the other lattices we use $(eB)^2\in[0:1.5]\, {\mbox{GeV}}^4$. The terms of higher powers of field begin to contribute significantly at low fields. So, it is difficult to extract the magnetic polarizability, because the quantization condition imposes a limitation on the minimal field value. We do not perform an extrapolation to the chiral limit since we are interested only in the qualitative predictions at this stage of investigations.
In Table \[Table:rho0:beta\_s0\] one can find the magnetic polarizability $\beta_m$ and hyperpolarizability $\beta_m^{2h}$ obtained from the fits. The lattice volume $V$, the lattice spacing $a$, the pion mass $m_{\pi}$, the interval of fields selected for the fitting procedure and $\chi^2/n.d.o.f.$ are also shown. The results agree with each other within the errors.
$V$ $a({\mbox{fm}})$ $m_{\pi}({\mbox{MeV}})$ $\beta_m({\mbox{GeV}}^{-3})$ $\beta_m^{2h}({\mbox{GeV}}^{-7})$ $n$ $(eB)^2({\mbox{GeV}}^4$) $\chi^2/d.o.f.$
-------- ------------------ ------------------------- ------------------------------ ----------------------------------- ------ -------------------------- -----------------
$18^4$ $0.105$ $574 \pm 7$ $ 0.66\pm 0.16$ $ -2.51 \pm 0.98 $ $10$ $[0:1.7]$ $1.04 $
$18^4$ $0.115$ $541 \pm 3$ $ 0.90\pm0.16$ $ -5.11 \pm 1.59 $ $12$ $[0:1.5]$ $ 2.46 $
$20^4$ $0.115$ $535 \pm 4$ $ 0.95\pm0.15$ $ -5.78 \pm 1.60 $ $12$ $[0:1.5]$ $2.63 $
$18^4$ $0.115$ $395 \pm 6$ $ 0.98\pm 0.30 $ $ -5.79 \pm 2.74 $ $12$ $[0:1.5]$ $ 3.32 $
: The value of the magnetic dipole polarizability $\beta_m$ and the magnetic hyperpolarizability $\beta_m^{2h}$ for the $\rho^0$ with spin projection $s_z=0$ are shown for the various lattice volumes $V$, lattice spacings $a$ and pion masses $m_{\pi}$. The degree of polynomial $n$, field range used for the fitting and $\chi^2/d.o.f$ values are represented in columns sixth to eighth correspondingly.[]{data-label="Table:rho0:beta_s0"}
At $eB\in[0:1.2]\ {\mbox{GeV}}^2$ for the spin projection $|s_z|=1$ on the field axis the energy squared of the neutral vector meson can be described by the following relation: $$E^2 =m^2 -4 \pi m \beta_m (eB)^2-4 \pi m \beta_m^{h1} (eB)^3\,.
\label{eq:beta_rho0_s1}$$ In Fig. \[Fig:rho:s1uudd\] the energy squared is shown for the $\rho^0$ meson with the spin projection $|s_z|=1$, the energies of the neutral vector meson for the $s_z=+1$ and $s_z=-1$ coincide due to the conservation of C-parity. The magnetic polarizability $\beta_m$ is obtained from the fit of the lattice data by formula for the lattices with spacings $0.105\ {\mbox{fm}}$ and $0.115\ {\mbox{fm}}$, where $m$, $\beta_m$ and $\beta_m^{h1}$ are the fit parameters. The lattice data for $a=0.084\ {\mbox{fm}}$ and $a=0.095\ {\mbox{fm}}$ do not allow to extract statistically significant $\beta_m$ values, but we show them to check the lattice volume and lattice spacing effects. The $\beta_m$ values with the errors and other parameters are shown in Table \[Table:rho0:beta\_s1\]. The results agree with each other within the errors.
$V$ $a({\mbox{fm}})$ $m_{\pi}({\mbox{MeV}})$ $\beta_m({\mbox{GeV}}^{-3})$ $\beta_m^{1h}({\mbox{GeV}}^{-5})$ $eB,\ {\mbox{GeV}}^2$ $\chi^2/d.o.f.$
-------- ------------------ ------------------------- ------------------------------ ----------------------------------- ----------------------- -----------------
$18^4$ $0.105$ $574 \pm 7$ $ -0.10 \pm 0.02 $ $ 0.07 \pm 0.02$ $ [0:1.1] $ $0.47$
$18^4$ $0.115$ $541 \pm 3$ $ -0.07 \pm 0.02 $ $ 0.03 \pm 0.03$ $ [0:1.1] $ $0.87$
$20^4$ $0.115$ $535 \pm 4$ $ -0.10 \pm 0.02$ $0.06 \pm 0.03$ $ [0:1.1] $ $1.54$
$18^4$ $0.115$ $395 \pm 6$ $ -0.11\pm 0.03 $ $ 0.08 \pm 0.03$ $ [0:1.1] $ $0.65 $
: The value of the magnetic dipole polarizability $\beta_m$ and the magnetic hyperpolarizability $\beta_m^{1h}$ are shown for $\rho^0$ with $|s_z|=1$, for the lattice volumes $18^4$ and $20^4$, the lattice spacings $0.105\ {\mbox{fm}}$ and $0.115\ {\mbox{fm}}$ and various pion masses. The fourth and the last columns contain the intervals of the magnetic field used for the fit and $\chi^2/d.o.f.$ correspondingly.[]{data-label="Table:rho0:beta_s1"}
Tensor magnetic polarizability {#sec-7}
==============================
In the previous sections we have found that the energy of the $\rho^0$ meson with $s_z=\pm 1$ increases versus the magnetic field value, while the energy of the $\rho^0$ meson with $s_z=0$ diminishes quickly versus the magnetic field value. For the $\rho^{\pm}$ the energy decreases for the case $qs_z=+1$ and increases for the $qs_z=0$ and $qs_z=-1$ cases.
Such energy behaviour has some physical consequences. The small energy is more profitable than the high energy, so the longitudinal polarization of the $\rho^0$ meson corresponding to $s_z=0$ has to dominate in the collisions. Therefore, in the non-central heavy-ion collision the magnetic field favours longitudinal polarization of the $\rho^0$ mesons.
The dilepton asymmetries in non-central heavy ion collisions depend on the energy behaviour of the vector mesons. According to the vector dominance principle the vector mesons can directly convert to virtual photons. In turn, the electromagnetic decay of virtual photons is one of the main sources of dilepton production in heavy-ion collisions. When dileptons are produced in non-central heavy-ion collisions, their anisotropy depends on the response of the spin structure of intermediate resonances, such as $\rho^0$ to the magnetic field of the collision. Therefore, it is of particular interest to identify and distinguish between various sources of dileptons emitted from non-central heavy-ion collisions.
The shape of the dilepton distribution is characterized by the following differential cross section: $$\frac{d\sigma}{dM^2 d\cos\theta}=A(M^2)(1+B\cos^2 \theta),
\label{cross_sec}$$ where $M^2=(p_1+p_2)^2$ is the energy of the lepton pair in their rest frame, $p_1$ and $p_2$ are the four-momenta of the leptons, $\theta$ is the angle between the momenta of the virtual photon and the lepton. The asymmetry coefficient $B$ is defined by the polarization of the virtual photons produced in the collisions: $$B =\frac{\gamma_{\perp}-\gamma_{\parallel}}{\gamma_{\perp}+\gamma_{\parallel}},
\label{asymm}$$ where the $\gamma_{\perp,\parallel}$ are the contributions of the transverse and longitudinal polarizations of the virtual intermediate photon.
Our lattice calculations make possible to get the estimation of the asymmetry factor for such processes. For the vector particle in Cartesian basis the polarization tensor has the following form $$P_{ij}=\frac{3}{2}\langle s_i s_j+s_j s_i \rangle -2 \delta_{ij}.$$ If $w_{s_z=+1}$, $w_{s_z=-1} $ and $w_{s_z=0}$ are the probabilities that the $\rho$ meson has a spin projection on the field direction equal to $+1$, $-1$ and $0$ correspondingly, then the value of the component $P_{33}$ can be represented in terms of these probabilities in the following form: $$P_{33}=w_{s_z=+1}+w_{s_z=-1}-2w_{s_z=0}=\frac{N_{s_z=+1}+N_{s_z=-1}-2N_{s_z=0}}{N_{s_z=+1}+N_{s_z=-1}+N_{s_z=0}},
\label{asymmetry_coef}$$ where $N_{s_z=+1}$, $N_{s_z=-1}$ and $N_{s_z=0}$ are the numbers of particles with different spin projections.
As $w_{s_z=+1}+w_{s_z=-1}+w_{s_z=0}=1$, therefore, $P_{33}=1-3w_{s_z=0}$ and $-2 \leq P_{33}\leq 1$. The tensor polarizability describes the effect of the magnetic field for the spin states of the $\rho$ meson, in turn, the spin states and tensor polarization is revealed in its decay to lepton pair, see [@Teryaev]. In a strong magnetic field the spin of the particle tends to align along the field direction, but the non-zero temperature leads to the spin-flipping.
The differential cross section for the decay of $\rho$ meson to the lepton pair may be written as $$\frac{d\sigma}{dM^2 d\cos\theta}=N(M^2)(1+\frac{1}{4}P_{33}(3\cos^2 \theta-1)).
\label{cross_sec_2}$$
Comparing with one can see that $$B=\frac{3P_{33}}{4-P_{33}}.$$ For the transversely polarized $\rho$ meson $B=1$ and for the longitudinally polarized $B=-1$.
We also introduce the tensor polarizability $$\beta_t=\frac{\beta_{s_z=+1}+\beta_{s_z=-1}-2\beta_{s_z=0}}{\beta_{s_z=+1}+\beta_{s_z=-1}+\beta_{s_z=0}},$$ which is the measure of the magnetic field effect on a vector meson, in particular for high magnetic fields and temperature $P\sim \beta_t$.
We calculate $\beta_{t}$ on the lattice taking into account the equality $\beta_{s_z=+1}=\beta_{s_z=-1}$. The results are presented in Table \[betarho0\] for the neutral $\rho$-meson and in Table \[betarho\] for the charged $\rho$-meson. The large negative values of $\beta_t$ suggest the dominating longitudinal polarization of the $\rho^0$-meson. The dileptons are mainly emitted in the directions perpendicular to the magnetic field axis. This is a convincing result, as we clearly see from Figure \[Fig:rho:s0uudd\] and \[Fig:rho:s1uudd\],that the energy of the state with $s_z=0$ decreases, and the energy of $|s_z|=1$ increases.
It was found previously that for soft dileptons the longitudinal polarization also dominates [@Polikarpov]. This result was obtained by comparing the formation of soft dileptons with a nonzero component of the conductivity of the strongly interacting matter, parallel to the external magnetic field [@Luschevskaya].
$V$ $a({\mbox{fm}})$ $m_{\pi}({\mbox{MeV}})$ $\beta_t$
-------- ------------------ ------------------------- -----------------
$18^4$ $0.105$ $574 \pm 7$ $-3.3\pm 0.6 $
$18^4$ $0.115$ $541 \pm 3$ $-2.6\pm 0.2$
$20^4$ $0.115$ $535 \pm 4$ $-2.8 \pm 0.3 $
$18^4$ $0.115$ $395 \pm 6$ $-2.9\pm 0.5$
: The tensor polarizability $\beta_t$ of the $\rho^{0}$ meson is shown in the last column for the lattice volume $V$, the lattice spacing $a$ and the pion mass $m_{\pi}$.[]{data-label="betarho0"}
$V$ $a({\mbox{fm}})$ $m_{\pi}({\mbox{MeV}})$ $\beta_t$
-------- ------------------ ------------------------- --------------
$18^4$ $0.105$ $574 \pm 7$ $2.3\pm 0.7$
$18^4$ $0.115$ $541 \pm 3$ $2.5\pm 0.5$
$18^4$ $0.115$ $395 \pm 6$ $2.7\pm 0.7$
: The tensor polarizability $\beta_t$ of the $\rho^{\pm}$ is represented for various lattices and pion masses.[]{data-label="betarho"}
Conclusion
==========
In this paper the calculation of the magnetic dipole and tensor polarizabilities have been presented for the charged and neutral $\rho$ mesons on the lattice volume. We perform a thorough analysis of the behaviour of the effective mass plateau depending on the number of points used for the fitting by hyperbolic cosine function.
In Section \[sec-5\] the magnetic dipole polarizability and $g$-factor for the charged mesons have been represented. We include the additional sets of data into consideration and increase statistics with the lattice spacing $a=0.105\ {\mbox{fm}}$ and volume $18^4$. This allows one to obtain the more statistically significant values of the magnetic dipole polarizabilities for the spin projection $s_z=0$, see Table \[Table:rho0:beta\_s0\]. In contrast to the previous results [@2017], we have calculated the dipole magnetic polarizability for the case $qs_z=1$ from the 3-parametric fit , where the $g$-factor is a free parameter of the fit. Obviously, we could find the $g$-factor values at smaller field interval and use these values for the $\beta_m$ determination, but the purpose of this work was to obtain the qualitative predictions.
In Section \[sec-6\] the dipole magnetic polarizability have been extracted for the case of the neutral $\rho$ meson. The magnetic dipole polarizability was calculated, taking into account that the $u$ and $d$-quarks couple differently to the magnetic field. In [@Luschevskaya:2015a] the magnetic dipole polarizability of $\rho^0$ was calculated in the toy model with only one type of quark. Section \[sec-7\] is devoted to the discussion of the polarization of the dileptons which result from the decays of $\rho$ mesons. We have found that the longitudinal polarization of the $\rho^0$ mesons dominates in the collisions, because the low energy is more preferential than high energy. Therefore the dileptons occurring due to decays of the $\rho^0$ mesons will be emitted perpendicular to the direction of the magnetic field. This result reinforces the previous results obtained in [@Luschevskaya], i.e. the nonzero conductivity of the quark-hadronic matter in a strong magnetic field. We introduce new characteristics of the meson magnetic properties - the tensor palarizability. This quantity has been suggested to be related to the coefficient of asymmetry in the differential cross section for the dilepton production.
The authors are grateful to FAIR-ITEP supercomputer center where these numerical calculations were performed. This work is completely supported by the grant from the Russian Science Foundation (project number 16-12-10059).
[99]{}
E.V. Luschevskaya, O.E. Solovjeva, O.V. Teryaev, [*Determination of the properties of vectors mesons in external magnetic field by quenched SU(3) lattice QCD, JHEP*]{} [**9**]{} (2017) 142
M.A. Andreichikov, B.O. Kerbikov, V.D. Orlovsky and Yu.A. Simonov, [*Meson Spectrum in Strong Magnetic Fields, Phys.Rev. D*]{} **87** (2013) 094029 \[arXiv: 1304.2533\]
V.D. Orlovsky and Yu.A. Simonov, [*Nambu-Goldstone mesons in strong magnetic field, JHEP*]{} [**1309**]{} (2013) 136 \[arXiv:1306.2232\]
S. Cho, K. Hattori, S.H. Lee, K. Morita and Sho Ozaki, [*Charmonium Spectroscopy in Strong Magnetic Fields by QCD Sum Rules: S-Wave Ground States, Phys.Rev. D*]{} **91** (2015) 045025 \[arXiv:1411.7675\]
H. Taya, [*Hadron Masses in Strong Magnetic Fields, Phys.Rev. D*]{} **92** (2015) 014038 \[arXiv:1412.6877\]
M. Kawaguchi and S. Matsuzaki,[*Vector Meson Masses from Hidden Local Symmetry in Constant Magnetic Field, Phys.Rev. D*]{} **93** (2016) 125027 \[arXiv:1511.06990\]
K. Hattori, T. Kojo and N. Su, [*Mesons in strong magnetic fields: (I) General analyses, Nucl.Phys. A*]{} [**951**]{} (2016) 1 \[arXiv:1512.07361\]
Ph. Gubler, K. Hattori, S.H. Lee, M. Oka, S. Ozaki and K. Suzuki, [*D mesons in a magnetic field, Phys.Rev. D*]{} **93** (2016) 054026 \[arXiv:1512.08864\]
G. Martinelli, G. Parisi, R. Petronzio and F. Rapuano, [*The proton and neutron magnetic moments in lattice QCD, Phys. Lett.*]{} [**116B**]{} (1982) 434
H. Liu, L. Yu and M. Huang, [*Charged and neutral vector meson under magnetic field, Phys. Rev. D*]{} [**91**]{} (2015) 014017 \[arXiv:1408.1318\]
E.V. Luschevskaya, O.E. Solovjeva, O.E. Kochetkov and O.V. Teryaev, [*Magnetic polarizabilities of light mesons in SU(3) lattice gauge theory, Nucl. Phys. B*]{} [**898**]{} (2015) 627 \[arXiv: 1411.4284\]
S.R. Beane, E. Chang, W. Detmold, K. Orginos, A. Parreno, M.J. Savage and B.C. Tiburzi, [*Ab initio calculation of the $np\rightarrow d\gamma$ radiative capture process, Phys.Rev.Lett.*]{} [**115**]{} (2015) 132001 \[arXiv:1505.02422\]
E.V. Luschevskaya, O.E. Kochetkov, O.V. Teryaev and O.E. Solovjeva, [*$\pi^{\pm}$ and $\rho^{0,\pm}$ mesons in a strong magnetic field on the lattice, JETP Letters*]{} [**101**]{} no.10 (2015) 674
G. Bali, B.B. Brandt, G. Endrodi and B. Glaessle, [*QCD spectroscopy and quark mass renormalization in external magnetic fields with Wilson fermions,*]{} The 33rd International Symposium on Lattice Field Theory, [*PoS LATTICE2015*]{} [**265**]{} (2016) \[arXiv:1510.03899\]
E.V. Luschevskaya, O.E. Solovjeva and O.V. Teryaev, [*Magnetic polarizability of pion, Phys.Lett.B*]{} [**761**]{} (2016) 393
A.M. Baldin, [*Polarizability of nucleons, Nucl. Phys.*]{} **18** (1960) 310
L.V.Fil’kov and V.L.Kashevarov, [*Determination of $\pi^+$-meson polarizabilities from $\gamma \gamma\rightarrow \pi^+ \pi^-$ process, Phys. Rev. C*]{} **73** (2006) 035210 \[nucl-th/0512047\]
A. Samsonov, [*Magnetic moment of the rho-meson in QCD sum rules: perturbative corrections, JHEP*]{} [ **0312**]{} (2003) 061 \[hep-ph/0308065\]
T.M. Aliev, A. Özpineci and M. Savcı, [*Magnetic and quadrupole moments of light spin-1 mesons in light cone QCD sum rules, Phys. Lett. B*]{} [**678**]{} (2009) 470
D. Djukanovic, E. Epelbaum, J. Gegelia and U.-G. Meissner, [*The magnetic moment of the $\rho$-meson, Phys. Lett. B*]{} [**730**]{} (2014) 115 \[arXiv:1309.3991\]
F.X. Lee, S. Moerschbacher and W. Wilcox, [*Magnetic moments of vector, axial, and tensor mesons in lattice QCD, Phys. Rev. D*]{} [**78**]{} (2008) 094502
B. Owen, W. Kamleh, D. Leinweber, B. Menadue, and S. Mahbub, [*Light Meson Form Factors at near Physical Masses, Phys. Rev. D*]{} [**91**]{} (2015) 074503 \[arXiv:1501.02561\]
E.L. Bratkovskaya, O.V. Teryaev, V.D. Toneev, [*Anisotropy of dilepton emission from nuclear collisions, Physics Letters B*]{} [**348**]{} (1995) 283
G. Baym, T. Hatsuda and M. Strickland, [*Structure of virtual photon polarization in ultrarelativistic heavy-ion collisions, Nucl. Phys. A*]{} [**967**]{} (2017) 712
M. Lüscher and P. Weisz, [*On-shell improved lattice gauge theories, Commun. Math. Phys.*]{} [**97**]{} (1985) 59
V.G. Bornyakov, E.-M. Ilgenfritz and M. Müller-Preussker, [*Universality check of Abelian Monopoles, Phys. Rev. D*]{} [**72**]{} (2005) 054511 \[hep-lat/0507021\]
H. Neuberger, [*Exactly massless quarks on the lattice, Phys. Lett. B*]{} [**417**]{} (1998) 141 \[hep-lat/9707022\]
G.’t Hooft, [*A property of electric and magnetic flux in non-abelian gauge theories, Nucl. Phys. B*]{} [**153**]{} (1979) 141
H. Zainuddin, [*Group-theoretic quantization of a particle on a torus in a constant magnetic field*]{}, [Phys. Rev. D]{} [**40**]{} (1989) 636
G.-H. Chen, [*Degeneracy of Landau levels and quantum qroup $sl_q(2)$, Phys. Rev. B*]{} [**53**]{} (1996) 9540
M.H. Al-Hashimi and U.J. Wiese, [*Discrete Accidental Symmetry for a Particle in a Constant Magnetic Field on a Torus, Annals Phys.*]{} [**324**]{} (2009) 343 \[arXiv: 0807.0630\]
Dmitri E. Kharzeev, Karl Landsteiner, Andreas Schmitt, and Ho-Ung Yee, [*Strongly interacting matter in magnetic fields: an overview, Lecture Notes in Physics Volume*]{} [**871**]{} (2013)
P.V. Buividovich, M.I. Polikarpov, O.V. Teryaev, [*Lattice studies of magnetic phenomena in heavy-ion collisions, Lecture Notes in Physics*]{} [**871**]{} (2013) 377
P.V. Buividovich, M.N. Chernodub, D.E. Kharzeev, T.K. Kalaydzhyan, E.V. Luschevskaya, M.I. Polikarpov, [*Magnetic-Field-Induced insulator-conductor transition in SU(2) quenched lattice gauge theory, Phys. Rev. Lett.*]{} [**105**]{} (2010) 132001
| {
"pile_set_name": "ArXiv"
} |
---
abstract: |
We consider the following Keller-Segel system of degenerate type: $$\label{eqn;KS}
\tag{KS} \left\{
\begin{array}{llll}
& \displaystyle \frac{\partial u}{\partial t} =
\frac{\partial}{\partial x}
\Big( \frac{\partial u^m}{\partial x}
- u^{q-1} \cdot \frac{\partial v}{\partial x} \Big),
& x \in {{\rm I\kern-.22em R}}, \ t>0, \nonumber \\
& \displaystyle 0 =
\frac{\partial^2 v}{\partial x^2} - \gamma v + u,
& x \in {{\rm I\kern-.22em R}}, \ t>0, \nonumber \\
& u(x,0) = u_0(x), \quad
& x \in {{\rm I\kern-.22em R}},
\end{array}
\right.$$ where $m>1, \ \gamma > 0, \ q \ge 2m$. We shall first construct a weak solution $u(x,t)$ of (KS) such that $u^{m-1}$ is Lipschitz continuous and such that $\displaystyle u^{m-1+\delta}$ for $\delta>0$ is of class $C^1$ with respect to the space variable $x$. As a by-product, we prove the property of finite speed of propagation of a weak solution $u(x,t)$ of (KS), [*i.e.,*]{} that a weak solution $u(x,t)$ of (KS) has a compact support in $x$ for all $t>0$ if the initial data $u_0(x)$ has a compact support in ${{\rm I\kern-.22em R}}$. We also give both upper and lower bounds of the interface of the weak solution $u$ of (KS).
author:
- |
\
[ Department of Mathematics, Tsuda University,]{}\
[2-1-1, Tsuda-chou, Kodaira-shi, Tokyo, 187-8577, [Japan]{}]{},\
[[email protected]]{}
title: |
Finite speed of propagation in 1-D degenerate\
Keller-Segel system
---
Introduction
============
We consider the following Keller-Segel system of degenerate type: $$\label{eqn;KS}
\tag{KS} \left\{
\begin{array}{llll}
& \displaystyle \frac{\partial u}{\partial t} =
\frac{\partial}{\partial x}
\Big( \frac{\partial u^m}{\partial x}
- u^{q-1} \cdot \frac{\partial v}{\partial x} \Big),
& x \in {{\rm I\kern-.22em R}}, \ t>0, \nonumber \\
& \displaystyle
0 =
\frac{\partial^2 v}{\partial x^2} - \gamma v + u,
& x \in {{\rm I\kern-.22em R}}, \ t>0, \nonumber \\
& u(x,0) = u_0(x), \quad
& x \in {{\rm I\kern-.22em R}},
\end{array}
\right.$$ where $m>1, \ \gamma > 0, \ q \ge 2m$. The initial data $u_0$ is a non-negative function and in $L^{1} \cap L^{\infty}({{\rm I\kern-.22em R}})$ with $u_0^m \in H^1({{\rm I\kern-.22em R}}). $ This equation is often called as the Keller-Segel model describing the motion of the chemotaxis molds. (see [*e.g.,*]{} [@ChPe].)
The aim of this paper is to construct a weak solution $u(x,t)$ of (KS) such that $u^{m-1}$ is Lipschitz continuous and such that $\displaystyle u^{m-1+\delta}$ for $\delta>0$ is of class $C^1$ with respect to the space variable $x$. The regularity property whether $u^{m-1}$ is Lipschitz continuous or of class $C^1$ plays an important role for the investigation of the behaviour of the interface to the solution $u$ of (KS). Our result shows that the power $m-1$ to $u$ exhibits the borderline behaviour between Lipschitz continuity and $C^1$-regularity. Indeed, as a by-product of Lipschitz continuity for $u^{m-1}$, we prove that a weak solution $u(x,t)$ of (KS) possesses the property of finite speed of propagation [*i.e.,*]{} that a weak solution $u(x,t)$ of (KS) has a compact support in $x$ for all $t>0$ if the initial data $u_0(x)$ has a compact support in ${{\rm I\kern-.22em R}}$.
Similar results have been obtained for the porous medium equation: $$\tag{PME}
\left\{
\begin{array}{llll}
\displaystyle
\frac{\partial U}{\partial t}
& = &
\displaystyle
\frac{\partial^2 U^m}{\partial x^2},
\hspace{2cm} & x \in {{\rm I\kern-.22em R}}, \ t>0,
\nonumber \\
U(x,0) & = & U_0(x),
\hspace{2cm} & x \in {{\rm I\kern-.22em R}}.
\end{array}
\right.$$ It is known that the comparison principle gives both upper and lower bounds of all solutions $U$ to (PME) by means of the [*Barenblatt solution*]{} $V_B$ which is an exact solution of (PME). Hence the property of finite speed of propagation of $U$ is a direct consequence of the explicit form of $V_B$ since supp $V_B(\cdot,t)$ is compact in ${{\rm I\kern-.22em R}}$ for all time $t$.
Our purpose is to prove the property of finite speed of propagation for (KS) to which the comparison principle is not available. To this end, one makes use of the notion of the domain of dependence which is useful for the proof of uniqueness of solutions to the linear wave equations. For instance, the half-cone like region $D_T$ defined by $$\begin{aligned}
D_T
& := &
\Big\{ (x,t)
; \
-c t +a \le x \le c t +b, \quad 0 \le t <T
\Big\},
\quad a<b, \ c>0\end{aligned}$$ makes it possible to prove that the solution of the linear wave equation with the propagation speed $c$ vanishes on $D_T$ for the initial data $u_0$ such that $u_0(x) \equiv 0$ on $I \equiv [a,b]$.
To deal with (KS), we generalize such an idea, and consider the [*curved*]{} half-cone like region. Indeed, suppose that $u_0(x)=0$ on $I$. Then our curved half-cone like region $D_T$ with respect to $I$ can be expressed by $$\begin{aligned}
\label{region-DT}
D_T
& := &
\Big\{ (x,t)
; \
\xi(t) \le x \le \Xi(t), \quad 0 \le t <T
\Big\},\end{aligned}$$ where $\xi(t)$ and $\Xi(t)$ are the solutions of the following initial value problems: $$\tag{IE} \left\{
\begin{array}{lll}
\xi^{'}(t)
& = & - \frac{\partial}{\partial x}
\Big(\frac{m}{m-1} u^{m-1} \Big)(\xi(t),t)
+ u^{q-2}
\cdot \frac{\partial v}{\partial x} (\xi(t),t), \quad \xi(0)=a,
\label{int;1} \\
\Xi^{'}(t)
& = & - \frac{\partial}{\partial x}
\Big(\frac{m}{m-1} u^{m-1} \Big)(\Xi(t),t)
+ u^{q-2}
\cdot \frac{\partial v}{\partial x} (\Xi(t),t), \quad \Xi(0)=b.
\end{array}
\right.$$ Unfortunately, Lipschitz continuity of $u^{m-1}$ is too weak to ensure the existence of solutions $\{\xi(t), \Xi(t)\}$ to (IE). Hence we need to regularize $u$ by $u_{\varepsilon}$ with small parameter $\varepsilon>0$, and deal with the approximating solutions $\{\xi_{\varepsilon}(t), \Xi_{\varepsilon}(t)\}$ which correspond to (IE) with $u$ replaced by $u_{\varepsilon}$. It is shown that Lipschitz continuity of $u^{m-1}$ guarantees the existence of uniform limit $\{\xi(t), \Xi(t)\}$ on $0 \le t \le T$ of $\{\xi_{\varepsilon}(t), \Xi_{\varepsilon}(t)\}$ as $\varepsilon \to 0$. Then we see that $u(x,t)=0$ on $D_T$.\
Our definition of a weak solution to (KS) now reads:\
[**Definition 1**]{} [*Let $m,\gamma$ and $q$ be constants as $m>1, \ \gamma> 0, \ q \ge 2.
$ Let $u_0$ be a non-negative function in ${{\rm I\kern-.22em R}}$ with $u_0 \in L^{1} \cap L^{\infty}({{\rm I\kern-.22em R}})$ and $u_0^m \in H^1({{\rm I\kern-.22em R}})$. A pair of non-negative functions $(u,v)$ defined in ${{\rm I\kern-.22em R}}\times [0,T)$ is said to be a weak solution of [(KS)]{} on $[0,T)$ if\
[i)]{} $u \in L^{\infty} (0,T; L^2({{\rm I\kern-.22em R}})),$ $u^m \in L^2(0,T;H^1({{\rm I\kern-.22em R}}))
,$\
[ii)]{} $
v \in L^{\infty}(0,T; H^2({{\rm I\kern-.22em R}})),
$\
[iii)]{} $(u,v)$ satisfies (KS) in the sense of distributions: [*i.e.,*]{} $$\begin{aligned}
&& \int_0^T \int_{{{\rm I\kern-.22em R}}}
\left(\partial_x u^m \cdot \partial_x \varphi
- u^{q-1} \partial_x v \cdot \partial_x \varphi
- u \cdot \partial_t \varphi \right) \ dxdt
\ = \ \int_{{{\rm I\kern-.22em R}}} u_0(x) \cdot \varphi (x,0) \ dx, \\
&&
\hspace{6.5cm}
{\it for \ all \ functions} \ \varphi \in C_0^{\infty}({{\rm I\kern-.22em R}}\times [0,T)),
\\
&& \\
&&
-\partial_x^2 v + \gamma v - u =0
\qquad {\it for \ a.a.} \ (x,t) \ {\it in } \ {{\rm I\kern-.22em R}}\times (0,T).\end{aligned}$$* ]{}\
Concerning the local-in-time existence of weak solutions to (KS), the following result can be shown by a slight modification of argument developed by the author [@Sblow-uq-2005 Theorem 1.1].
\[[local existence of weak solution and its $L^{\infty}$ uniform bound]{} \] \[pr;local existence\] \
Let $m>1,\ \gamma > 0, \ q \ge 2$. Suppose that the initial data $u_0$ is non-negative everywhere. Then, ${\rm (KS)}$ has a non-negative weak solution [($u, v$)]{} on $[0,T_0)$ with $
T_0 = \Big( \| u_{0} \|_{L^{\infty}({{\rm I\kern-.22em R}})} + 2 \Big)^{-q}.
$ Moreover, $u(t)$ satisfies the following [*a priori*]{} estimate $$\begin{aligned}
\label{by}
\| u(t) \|_{L^{\infty}({{\rm I\kern-.22em R}})}
& \le & \| u_0 \|_{L^{\infty}({{\rm I\kern-.22em R}})} + 2
\ \ \quad {\rm for \ all} \ t \in [0,T_0). \end{aligned}$$
\
[**Remark 1.**]{} Concerning the global-in-time existence of weak solutions to (KS), the author and Kunii [@KS-SuKu] obtained the following result: Let $m,\gamma,q$ and the initial data $u_0$ be as in Definition 1. In the case $q<m+2$, there exists a weak solution $u$ of (KS) on $[0,\infty).$ On the other hand, in the case $q \ge m+2$, the weak solution $u$ of (KS) on $[0,\infty)$ can be constructed provided $\|u_0\|_{L^{\frac{q-m}{2}}({{\rm I\kern-.22em R}})}$ is sufficiently small.\
Now, we construct a weak solution $u$ of (KS) with some additional regularity for the velocity potential $u^{m-1}$.
\[thm:lip\] Let $m>1, \gamma > 0$ and $q \ge 2m$. Let the initial data $u_0$ be as in Definition 1. In addition, we assume that $u_0^{m-1}$ is Lipschitz continuous in ${{\rm I\kern-.22em R}}$. Then, the weak solution $u$ of (KS) on $[0,T_0)$ given by Proposition \[pr;local existence\] has the following additional properties (i) and (ii):\
(i) $u^{m-1}(x,t)$ is Lipschitz continuous with respect to $x$ for all $0 \le t < T_0$ with the estimate $$\begin{aligned}
\label{pr-adj}
\sup_{0<t<T_0}\| \partial_x u^{m-1}(t) \|_{L^{\infty}({{\rm I\kern-.22em R}})}
& \le & C,\end{aligned}$$ where $C=C(m,\gamma,q,u_0)$.\
(ii) For every $\delta>0$, $u^{m-1+\delta}(\cdot,t)$ is a $C^1$-function with respect to $x$ for all $0<t<T_0$ with the property that $\partial_x u^{m-1+\delta}(x,t)=0$ at the points $(x,t) \in {{\rm I\kern-.22em R}}\times
(0,T_0)$ such that $u(x,t)=0$. Furthermore, in the case of $1<m<2$, we have $\partial_x u(x,t)=0$ at the same points $(x,t)$ as above.
[**Remark 2.**]{} (i) By the fundamental inequality $$\begin{aligned}
\label{holder-est}
&& |u(x,t)-u(y,t)| \ \le \
\left\{
\begin{array}{lcl}
\frac{2^{1/m-1}}{m-1}
\|u\|_{L^{\infty}(Q_{T_0})}^{2-m}|u^{m-1}(x,t)-u^{m-1}(y,t)|, & \ 1<m<2,& \\
|u^{m-1}(x,t)-u^{m-1}(y,t)|^{\frac{1}{m-1}}, & \ m \ge 2 &
\end{array}
\right.\end{aligned}$$ for all $x,y \in {{\rm I\kern-.22em R}}, \ 0<t<T_0,$ we have by Theorem \[thm:lip\] that for every $0<t<T_0$, $u(\cdot,t)$ is a Hölder continuous function in ${{\rm I\kern-.22em R}}$ with the exponent $\mu=\min\{1,\frac{1}{m-1}\}$.\
(ii) For (PME), it is well-known that $\partial_x U^{m-1}(\cdot,t)$ becomes a discontinuous function in ${{\rm I\kern-.22em R}}$ after some definite time $t$. Our result in Theorem \[thm:lip\] makes it clear that continuity in $x$ of $\partial_x u^p(x,t)$ is guaranteed for all $p>m-1$ and all $0<t<T_0$. It seems to be an interesting question whether $\partial_x
u^{m-1}(\cdot,t)$ is really discontinuous in ${{\rm I\kern-.22em R}}$ or not.\
(iii) The hypothesis $q \ge 2m$ seems to be redundant. Indeed, such restriction on $q$ stems from choice of the transformation $\psi$ in (\[PSIAT\]) which may have a certain freedom to apply the Bernstein method to the uniform estimate of $\partial_x u^{m-1}$ for $\varepsilon$. It should be noted that in the case $\gamma=0$, we can relax this restriction to $q \ge m+1$. See (\[JJ\]) below.\
\
By (\[pr-adj\]) in Theorem \[thm:lip\], we can construct a pair of continuous functions $\xi(t)$ and $\Xi(t)$ on $[0,T_0)$ such that the region $D_{T_0}$ defined by (\[region-DT\]) belongs to the interior of the domain surrounded by the interface of $u$, which leads us to the property of finite speed of propagation to (KS).
\[thm;finite speed\] Let $m>1, \ \gamma > 0$ and $q \ge 2m$. Let the initial data $u_0$ be as in Definition 1. In addition, we assume that $u_0(x)=0$ on some interval $I \equiv [a,b]$ and that $u_0^{m-1}$ is Lipschitz continuous in ${{\rm I\kern-.22em R}}$. Suppose that $u$ is the weak solution of (KS) on $[0,T_0)$ given by Theorem \[thm:lip\]. Then, there exists a pair $\{\xi(t),\Xi(t)\}$ of continuous functions on $[0,T_0)$ with the following properties (i) and (ii):\
(i) $\xi, \Xi \in W^{1,\infty}(0,T_0)$ with $\xi(0)=a, \ \Xi(0)=b$;\
(ii) $u(x,t)=0$ for $\xi(t) \le x \le \Xi(t), \ 0 \le t<T_0$.
[**Remark 3.**]{} (i) Concerning (PME), the interface of $U$ can be explicitly determined by the solutions $\hat{\xi}(t)$ and $\hat{\Xi}(t)$ of the following initial value problems: $$\left\{
\begin{array}{lll}
\hat{\xi}^{'}(t)
& = & - \frac{\partial}{\partial x}
\Big(\frac{m}{m-1} U^{m-1} \Big)(\hat{\xi}(t),t),
\quad
\hat{\xi}(0)=a,
\nonumber \\
\hat{\Xi}^{'}(t)
& = & - \frac{\partial}{\partial x}
\Big(\frac{m}{m-1} U^{m-1} \Big)(\hat{\Xi}(t),t),
\quad \hat{\Xi}(0)=b.
\nonumber
\end{array}
\right.$$ Indeed, by the comparison principle Knerr [@Kn] showed that if $U_0(x) = 0$ on some interval $I=[a,b]$ and $U_0(x)>0$ on $I^c={{\rm I\kern-.22em R}}\backslash I$, then it holds that $U(x,t) = 0$ for $\hat{\xi}(t) \le x \le \hat{\Xi}(t)$ and $U(x,t)>0$ for $x < \hat{\xi}(t)$ and $x > \hat{\Xi}(t)$ for all $0 \le t <\infty$. We call such $\hat{\xi}(t)$ and $\hat{\Xi}(t)$ the [*interface*]{} of (PME).\
(ii) Compared with (PME), it is not clear whether (IE) determines the exact interface of (KS) to which the comparison principle is not available. However, if $\xi_1(t)$ and $\Xi_1(t)$ are the interface of (KS), [*i.e.,*]{} that $\xi_1(t)$ and $\Xi_1(t)$ have the property that $$u(x,t) = 0 {\rm \ \ in} \ I_{t} := [\xi_1(t),\Xi_1(t)]
\quad {\rm and} \quad u(x,t)>0 \ {\rm in \ some \ neighbourhood \ outside \
of } \ I_{t}$$ for all $0 \le t <T_0$, then we can see that $\xi(t)$ and $\Xi(t)$ given by Theorem \[thm;finite speed\] satisfy the estimates $$\xi_1(t) \le \xi(t), \qquad \Xi(t) \le \Xi_1(t) \quad {\rm for \ all} \
0 \le t < T_0.$$ Hence our result may be regarded as an estimate of the maximum and the minimum of the interface of (KS). Other observations were done by Mimura-Nagai [@MN] and Bonami-Hilhorst-Logak-Mimura [@B-H-L-M-2].\
This paper is organized as follows. In Section 2, we shall first recall the approximating problem (KS)$_{\varepsilon}$ of (KS) introduced by [@KS-SuKu]. Our main purpose is devoted to the derivation of uniform gradient bound with respect to $\varepsilon>0$ of the approximating velocity potential $w_{\varepsilon}=\frac{m}{m-1} u_{\varepsilon}^{m-1}$, where $u_{\varepsilon}$ is the smooth solution of (KS)$_{\varepsilon}$. Bernstein’s method plays an important role to obtain our uniform estimate. (see [*e.g.,*]{} [@Li].) Then in Section 3, by the standard compactness argument, we shall prove the Lipschitz continuity of the velocity potential $w=\frac{m}{m-1}u^{m-1}$ for the weak solution $u$ of (KS). It is expected that $\partial_x w(x,t)$ becomes a discontinuous function in $x$ after some finite time $t$. However, we shall show that for $p>m-1$, $\partial_x u^p(x,t)$ is, in fact, a continuous function in ${{\rm I\kern-.22em R}}$ for all $t \in [0,T_0)$. Section 4 is devoted to the construction of continuous curves $\xi(t)$ and $\Xi(t)$ such that $u(x,t)=0$ on $D_{T_0}$ defined by (\[region-DT\]), which implies the property of the finite speed of propagation for (KS).\
We will use the simplified notations:\
1) $\partial_t = \frac{\partial}{\partial t}, \ $ $\partial_x = \frac{\partial}{\partial x}, \ $ $\partial^2_{x} = \frac{\partial^2}{\partial x^2}, \ $ $\partial^3_{x} = \frac{\partial^3}{\partial x^3}, \ $\
2) $\| \cdot \|_{L^r} = \| \cdot \|_{L^r({{\rm I\kern-.22em R}})},
(1 \le r \le \infty), \
\int \cdot \ dx:= \int_{{{\rm I\kern-.22em R}}} \cdot \ dx,
$\
3) $Q_T:= {{\rm I\kern-.22em R}}\times (0,T)$, \
4) When the weak derivatives $\partial_x u, \partial_x^2 u$ and $\partial_t u$ are in $L^p(Q_T)$ for some $p \ge 1$, we say that $u \in W^{2,1}_p(Q_T)$, [*i.e.*]{}, $$\begin{aligned}
W^{2,1}_p(Q_T)
& := & \Big\{u \in L^p(0,T;W^{2,p}({{\rm I\kern-.22em R}})) \cap
W^{1,p}(0,T;L^p({{\rm I\kern-.22em R}})) ; \\
&& \quad
\| u \|_{W^{2,1}_p(Q_T)}
:= \|u\|_{L^p(Q_T)}+\|\partial_x u\|_{L^p(Q_T)}
+ \|\partial_x^2u\|_{L^p(Q_T)}
+ \|\partial_t u\|_{L^p(Q_T)} < \infty
\Big\} .\end{aligned}$$
[**Approximating Problem**]{} {#3}
=============================
In order to justify the formal arguments, we introduce the following approximating equations of (KS): \[sect; time local\] $${\rm (KS)_{\varepsilon}}
~\left\{\begin{array}{rclll}
\begin{displaystyle}
\partial_t u_{\varepsilon} \end{displaystyle}(x,t)
& = &
\partial_x
\Big(\partial_x
(u_{\varepsilon} + \varepsilon)^m
- (u_{\varepsilon}+\varepsilon)^{q-2} u_{\varepsilon} \cdot
\partial_x v_{\varepsilon} \Big) ,
& (x,t) \in {{\rm I\kern-.22em R}}\times (0,T),
\vspace{2mm} \\
0
& = &
\partial_x^2 v_{\varepsilon} - \gamma v_{\varepsilon}
+ u_{\varepsilon},
& (x,t) \in {{\rm I\kern-.22em R}}\times (0,T),
\vspace{2mm} \\
u_{\varepsilon}(x,0)
& = & u_{0\varepsilon}(x), \
& x \in {{\rm I\kern-.22em R}},
\end{array}
\right.
\hspace{3cm}$$ where $\varepsilon>0$ is a positive parameter.\
Let us introduce the following assumption on the initial data $u_{0\varepsilon}$ with $\varepsilon>0$.
(A.1)
: $u_{0\varepsilon} \ge 0$ for all $x \in {{\rm I\kern-.22em R}}$ and $u_{0\varepsilon} \in W^{2,p}({{\rm I\kern-.22em R}})$ with $$\begin{aligned}
\sup_{0<\varepsilon<1}
\| u_{0\varepsilon}\|_{L^p({{\rm I\kern-.22em R}})}
\le \|u_0\|_{L^p({{\rm I\kern-.22em R}})}
&&
{\rm for \ all} \ p \in [1,\infty],
\nonumber \\
\|u_{0\varepsilon} - u_0 \|_{L^p({{\rm I\kern-.22em R}})} \to 0 \quad {\rm as}
\ \varepsilon \to 0
&&
{\rm for \ all} \ p \in [1,\infty). \end{aligned}$$
(A.2)
: $u_{0\varepsilon} \in W^{1,2}({{\rm I\kern-.22em R}})$ with $
\sup_{0<\varepsilon<1}
\| \partial_x u_{0\varepsilon} \|_{L^2({{\rm I\kern-.22em R}})}
\le \|\partial_x u_0 \|_{L^2({{\rm I\kern-.22em R}})}.
$
We call ($u_{\varepsilon}, v_{\varepsilon}$) a [*strong solution*]{} of (KS)$_{\varepsilon}$ if it belongs to $W^{2,1}_p \times W^{2,1}_p(Q_T)$ for some $p \ge 1$ and (KS)$_{\varepsilon}$ is satisfied almost everywhere.\
For the strong solution, we consider the case $p=3$ and introduce the space ${\bf W}(Q_T)$ defined by $$\begin{aligned}
\label{def-wqt}
{\bf W}(Q_T)
& := &
W^{2,1}_{3}
\times
W^{2,1}_{3}(Q_T).\end{aligned}$$ In [@Sblow-uq-2005]–[@KS-SuKu], the following proposition concerning the existence of the strong solution was proved :
\[pr;apto\] ([local existence of approximating solution]{}) Let $m \ge 1, \ \gamma > 0, \ q \ge 2$. We take $T_0:=(\|u_0\|_{L^{\infty}({{\rm I\kern-.22em R}})}+2)^{-q}$. Then, for every $\varepsilon>0$ and every initial data $u_{0\varepsilon}$ satisfying the hypothesis [(A.1)]{}, ${\rm (KS)_{\varepsilon}}$ has a unique non-negative strong solution [($u_{\varepsilon}, v_{\varepsilon}$)]{} in ${\bf W}(Q_{T_0})$. Moreover, $u_{\varepsilon}(t)$ satisfies the following [*a priori*]{} estimate $$\begin{aligned}
\label{estK0}
\| u_{\varepsilon}(t) \|_{L^{\infty}({{\rm I\kern-.22em R}})}
& \le & \| u_0 \|_{L^{\infty}({{\rm I\kern-.22em R}})} + 2
\ \quad {\rm for \ all} \ t \in [0,T_0) \
{\rm and \ all} \ \varepsilon \in (0,1].\end{aligned}$$
[**Remark 4.**]{} (i) It should be noted that the time interval $[0,T_0)$ of the existence of the strong solution $(u_{\varepsilon},v_{\varepsilon})$ can be taken uniformly with respect to $\varepsilon>0$.\
(ii) The weak solution $(u,v)$ of (KS) on $[0,T_0)$ given by Proposition \[pr;local existence\] can be constructed as the weak limit of $(u_{\varepsilon},v_{\varepsilon})$ as $\varepsilon \to 0$, where $(u_{\varepsilon},v_{\varepsilon})$ is the strong solution in Proposition \[pr;apto\]. More precisely, by choosing a subsequence of $(u_{\varepsilon},v_{\varepsilon})$ which we denote by $(u_{\varepsilon},v_{\varepsilon})$ itself for simplicity, we have $$\begin{aligned}
u_{\varepsilon}
\ \ \rightharpoonup \ \ u \qquad && {\rm weakly-star}
\ {\rm in} \ L^{\infty}(0,T_0; L^2({{\rm I\kern-.22em R}})),
\nonumber \\
u_{\varepsilon}^m
\ \ \rightarrow \ \ u^m \qquad && {\rm weakly } \ {\rm in} \
L^2(0,T_0; H^1({{\rm I\kern-.22em R}})) \ {\rm and \ strongly \ in} \
C([0,T_0); L_{loc}^2({{\rm I\kern-.22em R}})),
\nonumber \\
v_{\varepsilon}
\ \ \rightharpoonup \ \ v \qquad &&
{\rm weakly-star} \ {\rm in} \
L^{\infty}(0,T_0; H^2({{\rm I\kern-.22em R}}))\end{aligned}$$ as $\varepsilon \to 0.$ In what follows, we assume that the sequence of approximating solutions ($u_{\varepsilon},v_{\varepsilon}$) satisfies the above convergence.\
(iii) The strong solution $(u_{\varepsilon},v_{\varepsilon}) \in {\bf
W}(Q_{T_0})$ is more regular. Indeed, for every $\varepsilon>0$, it can be shown that $
u_{\varepsilon}, v_{\varepsilon} \in C^{\infty}({{\rm I\kern-.22em R}}\times (0,T_0)).
$\
The following lemma gives the gradient estimate for the velocity potential $u^{m-1}$.
\[lip\] Let $m>1, \ \gamma > 0$ and $q \ge 2m$. Let the initial data $u_0$ be as in Definition 1. For every $\varepsilon>0$, we take $u_{0\varepsilon}$ so that the hypothesis [(A.1)–(A.2)]{} are satisfied. In addition, we assume that $u_{0\varepsilon}^{m-1}$ is Lipschitz continuous in ${{\rm I\kern-.22em R}}$. Then the strong solution $u_{\varepsilon}$ of (KS)$_{\varepsilon}$ on $[0,T_0)$ given by Proposition \[pr;apto\] has the following property: $$\begin{aligned}
\label{adj}
\sup_{0<\varepsilon<1}
\Big(
\sup_{0<t<T_0}\| \partial_x (u_{\varepsilon}+\varepsilon)^{m-1} \|_{L^{\infty}({{\rm I\kern-.22em R}})}
\Big)
& \le & C,\end{aligned}$$ where $C=C(m,\gamma,q,u_0)$.
[*Proof of Lemma \[lip\].*]{} For the sake of simplicity, we denote $(u_{\varepsilon}, v_{\varepsilon})$ by $(u,v)$. To treat the velocity potential, let us define $w := \frac{m}{m-1}(u+\varepsilon)^{m-1}$. Multiplying the first equation of (KS)$_{\varepsilon}$ by $m(u+\varepsilon)^{m-2}$ and then rewriting the resultant identity in terms of $w$, we have $$\begin{aligned}
\label{w}
\partial_t w
& = &
(m-1)w \cdot \partial_{x}^2 w
+ |\partial_x w|^2
- \Big( (q-2) (u+\varepsilon)^{q-3} u
+ (u + \varepsilon)^{q-2}
\Big) \cdot \partial_x v
\cdot \partial_x w
\nonumber \\
&& -(m-1)(u+\varepsilon)^{q-3}u
\cdot \partial_x^2 v
\cdot w. \end{aligned}$$ Now we apply Bernstein’s method. Introducing the convex transformation $\psi: \bar{w} \to w$, determined below (\[PSIAT\]), we rewrite the identity (\[w\]) by means of $\bar{w}=\psi^{-1}(w)$ in the following form: $$\begin{aligned}
\label{apri nabla; bar u..}
\partial_t \bar{w}
& = & (m-1) \psi \cdot
\Big(\frac{\psi^{''}}{\psi^{'}} |\partial_x \bar{w}|^2
+ \partial_x^2 \bar{w} \Big)
+
\psi^{'} |\partial_x \bar{w}|^2
\nonumber \\
&& \quad -
\Big((q-2)(u+\varepsilon)^{q-3} u+ (u+\varepsilon)^{q-2} \Big)
\cdot \partial_x v \cdot \partial_x \bar{w}
- (m-1) \frac{\psi}{\psi^{'}} \cdot
(u+\varepsilon)^{q-3} u \cdot \partial_x^2 v. \end{aligned}$$ We note that $$\begin{aligned}
&& \psi(\bar{w})
\ = \ w \ = \ \frac{m}{m-1} (u+\varepsilon)^{m-1},
\quad
\psi^{'}(\bar{w}) \cdot \partial_x \bar{w}
\ = \ m(u+\varepsilon)^{m-2} \partial_x u,
\label{psi;w-1} \\
&& (u+\varepsilon)^{q-i} \cdot \partial_x u
\ = \ \frac{1}{m} \cdot \psi^{'} \cdot (u+\varepsilon)^{q-m-i+2}
\cdot \partial_x \bar{w}
\quad {\rm for} \ i=3,4.
\label{psi;w-2} \end{aligned}$$ Differentiating both sides of (\[apri nabla; bar u..\]) with respect to $x$, we obtain from (\[psi;w-1\]) and (\[psi;w-2\]) that $$\begin{aligned}
\label{bar u 1..}
\partial_t \partial_{x} \bar{w}
& = & (m-1)
\psi'(\bar{w}) \cdot
\Big(\frac{\psi^{''}(\bar{w})}{\psi^{'}(\bar{w})}
|\partial_x \bar{w}|^2
+ \partial_x^2 \bar{w}
\Big) \cdot \partial_x \bar{w}
\nonumber \\
&& + \Big(
(m-1) \psi(\bar{w}) \cdot
\Big( \frac{ \psi''(\bar{w})}{\psi'(\bar{w})} \Big)^{'}
+ \psi^{''}(\bar{w})
\Big)
\cdot (\partial_x \bar{w})^3
\nonumber \\
&& + \ 2 \Big(
(m-1) \psi(\bar{w}) \cdot \frac{\psi''(\bar{w})}{\psi'(\bar{w})}
+ \psi^{'}(\bar{w})
\Big)
\cdot \partial_x \bar{w}
\cdot \partial_x^2 \bar{w}
\ + \ (m-1) \psi(\bar{w})
\cdot \partial_x^3 \bar{w}
\nonumber \\
&& - \ (q-2)(q-3) \cdot \frac{1}{m} \cdot \psi^{'}(\bar{w})
\cdot (u+\varepsilon)^{q-m-2} u
\cdot \partial_x v \cdot (\partial_x \bar{w})^2
\nonumber \\
&& - \ 2(q-2) \cdot \frac{1}{m} \cdot \psi^{'}(\bar{w})
\cdot (u+\varepsilon)^{q-m-1}
\cdot \partial_x v \cdot (\partial_x \bar{w})^2
\nonumber \\
&& - \ (q-2) \cdot \frac{m-1}{m} \cdot \psi(\bar{w})
\cdot (u+\varepsilon)^{q-m-2} u
\cdot \partial_x^2 v
\cdot \partial_x \bar{w}
\nonumber \\
&& - \ (q-2) \cdot \frac{m-1}{m} \cdot \psi(\bar{w})
\cdot (u+\varepsilon)^{q-m-2} u
\cdot \partial_x v \cdot
\partial_x^2 \bar{w}
\nonumber \\
&& - \ \frac{m-1}{m} \cdot \psi(\bar{w})
\cdot (u+\varepsilon)^{q-m-1}
\cdot \partial_x^2 v \cdot \partial_x \bar{w}
\nonumber \\
&& - \ \frac{m-1}{m} \cdot \psi(\bar{w})
\cdot (u+\varepsilon)^{q-m-1}
\cdot \partial_x v
\cdot \partial_x^2 \bar{w}
\nonumber \\
&& - \ \frac{(m-1)^2}{m}
\Big( \frac{ \psi'(\bar{w})}{\psi(\bar{w})} \Big)^{'}
\cdot \psi \cdot (u+\varepsilon)^{q-m-2} u
\cdot \partial_x^2 v \cdot \partial_x \bar{w}
\nonumber \\
&& - \ \frac{m-1}{m} \cdot
\psi(\bar{w}) \cdot
(u +\varepsilon)^{q-m-2}
\Big((q-2) u +\varepsilon
\Big) \cdot \partial_x^2 v \cdot \partial_x \bar{w}
\nonumber \\
&& - \ \frac{(m-1)^2}{m} \cdot \frac{ \psi(\bar{w})^2}{\psi'(\bar{w})}
\cdot (u +\varepsilon)^{q-m-2} u
\cdot \partial_x^3 v.\end{aligned}$$ Let us put $U:=|\partial_x \bar{w}|^2$, then the following identities hold. $$\begin{aligned}
\label{der;k1..}
\partial_x \bar{w} \cdot \partial_x^2 \bar{w}
= \frac12 \partial_x U,
\qquad
\partial_x \bar{w} \cdot \partial_x^3 \bar{w}
& = &
\frac{1}{2} \partial_x^2 U - (\partial_x^2 \bar{w})^2. \end{aligned}$$ Multiplying (\[bar u 1..\]) by $\partial_x \bar{w}$ and using (\[der;k1..\]), the resultant equation in terms of $U$ reads: $$\begin{aligned}
\label{U..}
\frac{1}{2} \cdot \partial_t U
& = & \Big((m-1) \psi \big( \frac{\psi^{''}}{\psi^{'}} \big)^{'}
+ m\psi^{''}
\Big) U^2
+ \Big(
(m-1) \cdot \psi \cdot \frac{\psi^{''}}{\psi^{'}}
+ \frac{m+1}{2} \psi^{'}
\Big) \partial_x \bar{w} \cdot \partial_x U
\nonumber \\
&&
+ \ (m-1) \psi
\cdot \Big(
\frac12 \partial_x^2 U - (\partial_x^2 \bar{w})^2
\Big)
\nonumber \\
&& - \ (q-2) \cdot \frac{1}{m} \cdot \psi^{'}(\bar{w})
\cdot (u+\varepsilon)^{q-m-2} \cdot
\Big((q-1)u+2\varepsilon \Big)
\cdot \partial_x v
\cdot \partial_x \bar{w} \cdot U
\nonumber \\
&& - \ (q-2) \cdot \frac{m-1}{m} \cdot \psi(\bar{w})
\cdot (u+\varepsilon)^{q-m-2} u
\cdot \partial_x^2 v
\cdot U
\nonumber \\
&& - \ (q-2) \cdot \frac{m-1}{m} \cdot \psi(\bar{w})
\cdot (u+\varepsilon)^{q-m-2} u
\cdot \partial_x v \cdot
\frac{1}{2} \partial_x U
\nonumber \end{aligned}$$ $$\begin{aligned}
&& - \ \frac{m-1}{m} \cdot \psi(\bar{w})
\cdot (u+\varepsilon)^{q-m-1}
\cdot \partial_x^2 v \cdot U
\nonumber \\
&& - \ \frac{m-1}{m} \cdot \psi(\bar{w})
\cdot (u+\varepsilon)^{q-m-1}
\cdot \partial_x v \cdot \frac{1}{2} \partial_x U
\nonumber \\
&& - \ \frac{(m-1)^2}{m}
\Big( \frac{ \psi'(\bar{w})}{\psi(\bar{w})} \Big)^{'}
\cdot \psi \cdot (u+\varepsilon)^{q-m-2} u
\cdot \partial_x^2 v \cdot U
\nonumber \\
&& - \ \frac{m-1}{m} \cdot
\psi(\bar{w}) \cdot
(u +\varepsilon)^{q-m-2}
\Big((q-2) u +\varepsilon
\Big) \cdot \partial_x^2 v \cdot U
\nonumber \\
&& - \ \frac{(m-1)^2}{m} \cdot \frac{ \psi(\bar{w})^2}{\psi'(\bar{w})}
\cdot (u +\varepsilon)^{q-m-2} u
\cdot \partial_x^3 v \cdot \partial_x \bar{w}.\end{aligned}$$ We consider a sequence $\{\eta_k(x)\}_{k=-\infty}^{\infty}$ of cut-off functions such that $$\begin{aligned}
&& {\rm supp} \ \eta_k = \{x \in {{\rm I\kern-.22em R}}; \ -2+k \le x \le 2+k \},
\label{ino-oi}
\\
&& \eta_k(x) = 1 \qquad {\rm for} \ -1+k \le x \le 1+k,
\label{ino-oi-222}\end{aligned}$$ with $$\begin{aligned}
\label{opsi}
&& |\partial_x \eta_k(x)|
\le c_1 (\eta_k(x))^{\frac{3}{4}},
\quad
-c_2 \eta_k(x)
\ \le \
\partial_x^2 \eta_k(x)
\ \le \ c_3
\qquad {\rm for} \ x \in {{\rm I\kern-.22em R}},\end{aligned}$$ where $c_1, c_2$ and $c_3$ are positive constants independent of $k$. In Remark 5 below, we give an example of such $\{\eta_k\}_{k=-\infty}^{\infty}$.\
Multiplication of (\[U..\]) by $\eta_k$ yields $$\begin{aligned}
\label{U..ij}
\frac{1}{2} \cdot \partial_t (\eta_k U)
& = &
\Big((m-1) \psi \big( \frac{\psi^{''}}{\psi^{'}} \big)^{'}
+ m\psi^{''}
\Big) \eta_k U^2
\ - \ (m-1) \psi
(\partial_x^2 \bar{w})^2 \eta_k
\nonumber \\
&& \ + \ I \cdot \partial_x^2(\eta_k U)
\ + \ J \cdot \partial_x(\eta_k U)
\ + \ R_k,\end{aligned}$$ where $$\begin{aligned}
I & := & \frac{m-1}{2} \psi,
\nonumber \\
J & := & \Big(
(m-1) \cdot \psi \cdot \frac{\psi^{''}}{\psi^{'}}
+ \frac{m+1}{2} \psi^{'}
\Big) \partial_x \bar{w}
\ - \ (m-1) \psi \cdot \partial_x \eta_k
\nonumber \\
&& - \frac{m-1}{2m} \cdot \psi (u+\varepsilon)^{q-m-2}\Big( (q-1)u+\varepsilon \Big)
\partial_x v,\end{aligned}$$ and $R_k$ is regarded as the remainder term defined by $$\begin{aligned}
\label{RK}
R_k & := &
\sum_{j=1}^7 R_k^{(j)}\end{aligned}$$ with $$\begin{aligned}
R_k^{(1)} & := & - \Big(
(m-1) \cdot \psi \cdot \frac{\psi^{''}}{\psi^{'}}
+ \frac{m+1}{2} \psi^{'}
\Big) \partial_x \eta_k \cdot \partial_x \bar{w} \cdot U
\nonumber \\
&&
- \ \frac{q-2}{m} \cdot \psi^{'}(\bar{w})
\cdot (u+\varepsilon)^{q-m-2} \cdot
\Big((q-1)u+2\varepsilon \Big)
\cdot \partial_x v \cdot \eta_k
\cdot \partial_x \bar{w} \cdot U,
\nonumber \\
R_k^{(2)} & := &
(m-1) \cdot \psi \cdot (\partial_x \eta_k)^2 U,
\nonumber \\
R_k^{(3)} & := & \frac{m-1}{2m} \cdot \psi(\bar{w})
\cdot (u+\varepsilon)^{q-m-2} \Big( (q-1)u + \varepsilon \Big)
\cdot \partial_x v \cdot \partial_x \eta_k \cdot U,
\nonumber \\
R_k^{(4)} & := &
- \ \frac{m-1}{m} \cdot \psi(\bar{w})
\cdot (u+\varepsilon)^{q-m-2} \Big( (2q-3)u + 2\varepsilon \Big)
\cdot \partial_x^2 v \cdot \eta_k
U,
\nonumber \\
R_k^{(5)} & := &
- \ \frac{m-1}{2} \cdot \psi \cdot \partial_x^2 \eta_k \cdot U,
\nonumber \\
R_k^{(6)} & := &
- \ \frac{(m-1)^2}{m} \cdot
\Big( \frac{ \psi'(\bar{w})}{\psi(\bar{w})} \Big)^{'}
\cdot \psi \cdot (u+\varepsilon)^{q-m-2} u
\cdot \partial_x^2 v \cdot \eta_k U,
\nonumber \\
R_k^{(7)} & := &
- \ \frac{(m-1)^2}{m} \cdot \frac{ \psi(\bar{w})^2}{\psi'(\bar{w})}
\cdot (u +\varepsilon)^{q-m-2} u
\cdot \partial_x^3 v \cdot \partial_x \bar{w} \cdot \eta_k.\end{aligned}$$ Now we choose the transformation $\psi(r)$ by $$\begin{aligned}
\label{PSIAT}
\psi(r) := \frac{m}{m-1} (\|u_0\|_{L^{\infty}({{\rm I\kern-.22em R}})}+2+\varepsilon)^{m-1}
\cdot \frac{r}{3}(4-r),
\qquad 0 \le r \le 1.\end{aligned}$$ Then we observe that the coefficient of the first term $\eta_k U^2$ of the right-hand side in (\[U..ij\]) is negative, in particular $$\begin{aligned}
\label{psi;negative}
\Big((m-1) \psi \big( \frac{\psi^{''}}{\psi^{'}} \big)^{'}
+ m\psi^{''}
\Big) \eta_k U^2
& \le & - M \cdot \eta_k U^2
\qquad
{\rm for} \ (x,t) \in {{\rm I\kern-.22em R}}\times [0,T_0),\end{aligned}$$ where $ \displaystyle M:=\frac{m(11m - 3)}{12(m-1)} >0$.\
Indeed, since Proposition \[pr;apto\] states that $$\frac{m}{m-1} \varepsilon^{m-1}
\le w(x,t)
\le \frac{m}{m-1}
(\|u_0\|_{L^{\infty}({{\rm I\kern-.22em R}})}+2+\varepsilon)^{m-1} \ =: \ L$$ holds for all $(x,t) \in {{\rm I\kern-.22em R}}\times [0,T_0)$, the definition $\psi(\bar{w})=w$ yields $$\begin{aligned}
\label{re;barw}
(0<) \ 2-\sqrt{4-\frac{3m\varepsilon^{m-1}}{(m-1)L}}
\le \bar{w}(x,t) \le 1,
\qquad (x,t) \in {{\rm I\kern-.22em R}}\times [0,T_0)\end{aligned}$$ for sufficiently small $\varepsilon>0.$ Moreover, by (\[re;barw\]) we have $$\begin{aligned}
&&
\frac{2L}{3}
\ \le \
\psi^{'}(\bar{w})
\ = \ \frac{2L}{3}(2-\bar{w})
\ \le \ \frac{4L}{3},
\quad
\psi^{''}(\bar{w})
\ = \ -\frac{2L}{3},
\label{w1} \\
&&
\frac{1}{2}
\ \le \
\left|
\frac{\psi^{''}}{\psi^{'}}
\right|
\ \le \ 1,
\quad
-1 \ \le \
\Big( \frac{\psi^{''}}{\psi^{'}} \Big)^{'}
\ = \ \frac{\psi^{'}\psi^{'''}- (\psi^{''})^2}{(\psi^{'})^2}
\ = \ -\Big( \frac{\psi^{''}}{\psi^{'}} \Big)^2
\ \le \ - \frac14.
\label{w2}\end{aligned}$$ Now from (\[w1\]) and (\[w2\]), we see that the left-hand side of (\[psi;negative\]) is bounded by $$\begin{aligned}
\Big((m-1) \psi \big( \frac{\psi^{''}}{\psi^{'}} \big)^{'}
+ m\psi^{''}
\Big) \cdot \eta_k U^2
& \le &
\Big(- \frac{(m-1) L}{4} - \frac{2mL}{3}
\Big) \cdot \eta_k U^2
\nonumber \\
& = & - \frac{m(11m - 3)}{12(m-1)} \cdot \eta_k U^2
= -M \cdot \eta_k U^2 <0.\end{aligned}$$ On the other hand, suppose that $\eta_k U$ attains its maximum at the point $(x_0,t_0) \in {{\rm I\kern-.22em R}}\times (0,T_0)$. Then it holds by $\eta_k \ge 0$ that $$\begin{aligned}
\label{ux,ut}
\frac{1}{2} \cdot \partial_t (\eta_k U)(x_0,t_0) \ \ge \ 0,
\quad
\partial_x^2 (\eta_k U)(x_0,t_0) \ \le \ 0,
\quad \partial_x (\eta_k U)(x_0,t_0) \ = \ 0.\end{aligned}$$ Combining (\[psi;negative\]), (\[w1\])–(\[ux,ut\]) with (\[U..ij\]), we obtain $$\begin{aligned}
\label{uiui}
M \cdot \eta_k U^2
& \le & \sum_{j=1}^7 R_k^{(j)},
\label{IJ1-J5}\end{aligned}$$ where $\{ R_k^{(j)} \}_{1 \le j \le 7}$ is given by (\[RK\]).\
We are going to estimate the seven terms $\{R_k^{(j)}\}_{1 \le j \le 7}$. To this end, firstly integrating the second equation of (KS)$_{\varepsilon}$ on $(-\infty, x)$, we have $$\begin{aligned}
\partial_x v(x,t)
= \gamma \int_{-\infty}^x v(y,t) \ dy
- \int_{-\infty}^x u(y,t) \ dy, \end{aligned}$$ which yields $$\begin{aligned}
\label{proof;th-2}
\sup_{0<t<T_0} \| \partial_x v(t)\|_{L^{\infty}({{\rm I\kern-.22em R}})}
& \le & 2\| u_0 \|_{L^1({{\rm I\kern-.22em R}})}.\end{aligned}$$ Here we have used the fact that $v(x,t) > 0$ together with $
\gamma \int_{-\infty}^{\infty} v(y,t) \ dy=
\int_{-\infty}^{\infty} u(y,t) \ dy$ for all $t \in [0,T_0)$.\
By (\[opsi\]), (\[proof;th-2\]), Young’s inequality, and the relation $\partial_x^2 v = \gamma v-u$, we have for $q \ge m+1$ that $$\begin{aligned}
\label{J1-J4}
\sum_{j=1}^4 R_k^{(j)} & \le &
C \ + \ \frac{M}{8} \cdot \eta_k U^2,\end{aligned}$$ where $C$ is a constant depending on $m,\gamma,q$ and $u_0$. By (\[opsi\]) and Young’s inequality, we have $$\begin{aligned}
\label{sumgni}
R_k^{(5)}
& \le &
\frac{c_2(m-1)}{2} \psi \cdot \eta_k U
\ \le \
C \ + \ \frac{M}{8} \cdot \eta_k U^2.\end{aligned}$$ We are now going to estimate $R_k^{(6)}$. Since $$\begin{aligned}
-\frac{2L}{3} \cdot \frac{1}{\psi}
-\Big( \frac{4L}{3} \Big)^2
\cdot \frac{1}{\psi^2}
\ \le \
\Big( \frac{\psi^{'}}{\psi} \Big)^{'}
\ = \ \frac{\psi\psi^{''}- (\psi^{'})^2}{(\psi)^2}
\ \le \ -\frac{2L}{3} \cdot \frac{1}{\psi}
\ < \ 0,
\label{w3}\end{aligned}$$ and since $\partial_x^2 v = \gamma v-u$, by the hypothesis that $q \ge 2m$, we have that $$\begin{aligned}
R_k^{(6)} & = & - \frac{(m-1)^2}{m} \cdot
\Big(\frac{ \psi'(\bar{w})}{\psi(\bar{w})} \Big)^{'} u
\cdot
\psi(\bar{w})
\cdot (u+\varepsilon)^{q-m-2}
\cdot \partial_x^2 v
\cdot \eta_k U
\nonumber \\
& \le &
\frac{(m-1)^2}{m} \cdot
\frac{2L}{3}
\cdot (u+\varepsilon)^{q-m-1}
\cdot \gamma v
\cdot \eta_k U
\nonumber \\
&& \ + \ (m-1) \cdot
\Big(\frac{4L}{3} \Big)^2
\cdot (u+\varepsilon)^{q-2m}
\cdot \gamma v
\cdot \eta_k U
\nonumber \\
& \le &
C \ + \ \frac{M}{8} \cdot \eta_k U^2.
\label{JJ}\end{aligned}$$ Indeed, since $v(x,t)$ satisfies $$\begin{aligned}
\label{semi; z}
v(x,t)
& = & \int_{{{\rm I\kern-.22em R}}} G(x-y) \cdot u(y,t) \ dy \end{aligned}$$ with the Bessel potential $G(x)$ which can be express as $$\begin{aligned}
\label{GI}
G(x) & = & \frac{1}{\sqrt{4\pi}}
\int_0^{\infty} s^{-\frac{1}{2}} \cdot e^{-\gamma s-\frac{|x|^2}{4s}} \ ds
\ = \ \frac{e^{-\sqrt{\gamma}|x|}}{2 \gamma},\end{aligned}$$ it holds that $
G \in L^p({{\rm I\kern-.22em R}})
$ for all $1 \le p \le \infty$ and that $$\begin{aligned}
\sup_{0<t<T} \| v(t) \|_{L^{\infty}({{\rm I\kern-.22em R}})}
& \le & C \sup_{0<t<T} \|u(t)\|_{L^1({{\rm I\kern-.22em R}})} \ = \ C \|u_{0\varepsilon}\|_{L^1({{\rm I\kern-.22em R}})}
\ \le \ \|u_0\|_{L^1({{\rm I\kern-.22em R}})}.\end{aligned}$$ By (\[psi;w-1\]) and (\[psi;w-2\]), it holds $$\begin{aligned}
\label{J5}
\lefteqn{
\psi \cdot (u+\varepsilon)^{q-m-2} u \cdot \partial_x^3 v
} \nonumber \\
& = &
\frac{m}{m-1} (u+\varepsilon)^{q-3} u \cdot
(\gamma \partial_x v - \partial_x u )
\nonumber \\
& \le &
\frac{m}{m-1} \cdot \gamma \cdot
(u+\varepsilon)^{q-2} \cdot 2\| u_0 \|_{L^1}
+
\frac{1}{m-1} \cdot \psi^{'}(\bar{w}) (u+\varepsilon)^{q-m-1} u\cdot
\partial_x \bar{w},\end{aligned}$$ which yields $$\begin{aligned}
\label{J5-3}
R_k^{(7)} & \le &
C \ + \ \frac{M}{8} \cdot \eta_k U^2.\end{aligned}$$ Substituting (\[J1-J4\]), (\[sumgni\]), (\[JJ\]) and (\[J5-3\]) into (\[IJ1-J5\]), we obtain $$\begin{aligned}
M \cdot \eta_k U^2
& \le & C.\end{aligned}$$ Recalling $U=|\partial_x \bar{w}|^2$, we have by (\[ino-oi\]) and the above estimate that $$\begin{aligned}
\label{Rm}
|\partial_x \bar{w}|^2 \ =: \ U
& \le &
C \ \qquad {\rm for} \ -1+k \le x \le 1+k, \quad 0 \le t <T_0, \end{aligned}$$ where $C$ is a constant independent of $\varepsilon$ and $k$. Repeating the same argument as the above for $k=0, \pm 1, \pm 2, \cdots,$ we obtain the upper bound of $|\partial_x \bar{w}|$ which is independent of $\varepsilon$ in the whole interval ${{\rm I\kern-.22em R}}$.\
We recall the definition of $w$ and $\psi({\bar{w}})$: $$\begin{aligned}
\label{relation;u-w-barw}
\frac{m}{m-1}(u_{\varepsilon}+\varepsilon)^{m-1}
& = &
w
\ = \
\psi(\bar{w}).\end{aligned}$$ Differentiating both sides of (\[relation;u-w-barw\]) with respect to $x$, we have by (\[w1\]) and (\[Rm\]) that $$\begin{aligned}
\label{fga}
\left| \partial_x (u_{\varepsilon}+\varepsilon)^{m-1} \right|
& = &
\frac{m-1}{m}
\cdot
\psi^{'}(\bar{w}) \cdot
|\partial_x \bar{w}|
\ \le \ C \quad {\rm for \ all} \ (x,t) \in {{\rm I\kern-.22em R}}\times [0,T_0),\end{aligned}$$ which yields (\[adj\]). This proves Lemma \[lip\].\
[**Remark 5.**]{} In the proof of Lemma \[lip\], we have used a sequence $\{\eta_k(x)\}_{k=-\infty}^{\infty}$ of cut-off functions with properties (\[ino-oi\])–(\[opsi\]). Taking $\eta(x)$ as $$\begin{aligned}
&&
\eta(x) = \left\{
\begin{array}{llll}
0 & \quad {\rm for} \quad x \le -2 \\
2(2+x)^4 &
\quad {\rm for} \quad -2 < x < -\frac{3}{2}, \\
1-2(x+1)^4 &
\quad {\rm for} \quad -\frac{3}{2} < x \le -1, \\
1 & \quad {\rm for} \quad -1 \le x \le 1, \\
1-2(x-1)^4 &
\quad {\rm for} \quad 1 < x \le \frac{3}{2}, \\
2(2-x)^4 &
\quad {\rm for} \quad \frac{3}{2} < x < 2, \\
0 & \quad {\rm for} \quad x \ge 2,
\end{array}
\right.\end{aligned}$$ and then defining $\eta_{k}$ by $\eta_{k}(x) := \eta(x-k)$ for $k=0, \pm 1, \pm 2, \cdots$, we see that $\{\eta_k(x)\}_{k=-\infty}^{\infty}$ has the desired properties (\[ino-oi\])–(\[opsi\]). \
Proof of Theorem \[thm:lip\] {#section;3}
============================
Let us first show that for every $t \in [0,T_0)$, $\{u_{\varepsilon}(\cdot,t)\}_{\varepsilon>0}$ is a sequence of uniformly bounded and equi-continuous functions in ${{\rm I\kern-.22em R}}$. Indeed, the uniform bound is a consequence of (\[estK0\]). By (\[estK0\]), (\[adj\]) and (\[holder-est\]) with $u$ replaced by $u_{\varepsilon}+\varepsilon$, it holds $$\begin{aligned}
|u_{\varepsilon}(x,t)-u_{\varepsilon}(y,t)|
& \le & C(\|u_0\|_{L^{\infty}}+2)^2 |x-y|^{\mu},
\quad
\mu=\min\{1,\frac{1}{m-1}\}\end{aligned}$$ for all $x,y \in {{\rm I\kern-.22em R}}, \ 0 \le t <T_0$, and all $\varepsilon>0$, where $C$ is the same constant as in (\[adj\]). This implies that $\{u_{\varepsilon}(\cdot,t)\}_{\varepsilon>0}$ is a family of equi-continuous functions in ${{\rm I\kern-.22em R}}$ for all $0 \le t <T_0$. Hence by the Ascoli-Arzela theorem, there is a subsequence of $\{u_{\varepsilon}(\cdot,t)\}_{\varepsilon>0}$, which we denoted by $\{u_{\varepsilon}(\cdot,t)\}_{\varepsilon>0}$ itself such that $$\begin{aligned}
\label{Lm;strong--sta}
u_{\varepsilon}(\cdot,t) & \longrightarrow & u(\cdot,t)
\qquad {\rm as} \ \ \varepsilon \to 0\end{aligned}$$ uniformly in every compact interval $I \subset {{\rm I\kern-.22em R}}$.\
On the other hand, by (\[adj\]) and the weakly-star compactness of $L^{\infty}(Q_{T_0})$, there exists a sequence of $\{u_{\varepsilon}\}_{\varepsilon>0}$, which we denote by $\{u_{\varepsilon}\}_{\varepsilon>0}$ itself for simplicity, and a function $\tilde{u} \in L^{\infty}(Q_{T_0})$ such that $$\begin{aligned}
\partial_x (u_{\varepsilon}+\varepsilon)^{m-1} \
\rightarrow
\tilde{u} \hspace{1cm}
{\rm weakly-star}
& \ {\rm in} \
L^{\infty}(Q_{T_0}) \end{aligned}$$ with $$\begin{aligned}
\| \tilde{u} \|_{L^{\infty}(Q_{T_0})} & \le &
\liminf_{\varepsilon \to +0}
\| \partial_x (u_{\varepsilon}+\varepsilon)^{m-1} \|_{L^{\infty}(Q_{T_0})}.\end{aligned}$$ By (\[Lm;strong–sta\]), it is easy to see that $\tilde{u}= \partial_x u^{m-1}$, which yields the desired estimate (\[pr-adj\]).\
Next, we shall show that $\partial_x u^{m-1+\delta}(\cdot,t)$ is a continuous function in ${{\rm I\kern-.22em R}}$ for all $0<t<T_0$ and for all $\delta>0$ with the additional property that $\partial_x u^{m-1+\delta}(x,t)=0$ at the point $(x,t)$ such as $u(x,t)=0$. To this aim, we follow a similar argument employed in Aronson [@Ar3]. Let $u(x_0,t_0)>0$. Then we see by the standard argument that both $\partial_x u$ and $\partial_x u^{m-1+\delta}$ with $\delta>0$ are continuous functions in a neighbourhood of $(x_0,t_0)$. Therefore, it suffices to prove that $\partial_x u^{m-1+\delta}(\cdot,t)$ is a continuous function in a neighbourhood of $x_1$ such as $u(x_1,t)=0$ with the additional property that $\partial_x u^{m-1+\delta}(x_1,t)=0$. By virtue of (\[Lm;strong–sta\]), for every $t \in [0,T_0)$ and every compact interval $I \subset {{\rm I\kern-.22em R}}$, it holds that $u_{\varepsilon}(\cdot,t) \to u(\cdot,t)$ uniformly on $I$. Therefore, by Remark 2, there exists $a_0>0$ such that $$\begin{aligned}
\label{hani}
0 \le u_{\varepsilon}(x,t)
& \le &
|u_{\varepsilon}(x,t) -u(x,t)| + |u(x,t)-u(x_1,t)| +u(x_1,t)
\nonumber \\
& \le & 2 a^\mu \end{aligned}$$ holds for all $x \in I_a(x_1):=\{x \in {{\rm I\kern-.22em R}}; |x-x_1| < a\}$ and for all $0<a \le a_0$ and for all $0<\varepsilon<1$, where $\mu:= \min \{1, \frac{1}{m-1} \}$.\
On the other hand, since we have $$\begin{aligned}
\label{sekibun}
u_{\varepsilon}^{m-1+\delta}(x,t)-u_{\varepsilon}^{m-1+\delta}(x^{\prime},t)
& = &
\frac{m-1+\delta}{m-1} \int_{x^{\prime}}^x u_{\varepsilon}^{\delta}(x,t)
\cdot
\partial_x u_{\varepsilon}^{m-1}(x,t) \ dx,\end{aligned}$$ it follows from (\[hani\]),(\[sekibun\]) and Lemma \[lip\] that $$\begin{aligned}
\label{ilet}
|u_{\varepsilon}^{m-1+\delta}(x,t)-u_{\varepsilon}^{m-1+\delta}(x^{\prime},t)|
& \le &
C (2a^{\mu})^{\delta} |x-x^{\prime}|
\qquad
{\rm for \ all} \ x,x^{\prime} \in I_a(x_1)\end{aligned}$$ and for all $0 < a \le a_0$ and for all $0<\varepsilon<1$, where $C$ depends on $m,\gamma,q,u_0$ but not on $\varepsilon$. Letting $\varepsilon \to +0$ in (\[ilet\]), we have by (\[Lm;strong–sta\]) that $$\begin{aligned}
\label{base}
\lefteqn{
|u^{m-1+\delta}(x,t)-u^{m-1+\delta}(x^{\prime},t)|
} \nonumber \\
& \le &
C (2a^{\mu})^{\delta} |x-x^{\prime}|
\quad
{\rm for \ all} \ x,x^{\prime} \in I_a(x_1)
\ {\rm and \ all} \ 0 < a \le a_0.\end{aligned}$$ Taking $x=x_1$ in (\[base\]) and then letting $x^{\prime} \to x_1$, we have $$\begin{aligned}
|\partial_x u^{m-1+\delta}(x_1,t)| & \le & C(2a^{\mu})^{\delta}, \quad 0<a \le a_0.\end{aligned}$$ Hence we have by letting $a \to 0$ that $$\begin{aligned}
\partial_x u^{m-1+\delta}(x_1,t)=0.\end{aligned}$$ Similarly, letting $x^{\prime} \to x$ in (\[base\]), we have $$\begin{aligned}
\label{base-2}
|\partial_x u^{m-1+\delta}(x,t)| & \le & C(2a^{\mu})^{\delta}
\qquad {\rm for \ all} \ 0<a\le a_0,\end{aligned}$$ which implies that $\partial_x u^{m-1+\delta}(\cdot,t)$ is continuous at $x_1$. Since $x_1$ can be taken arbitrary in such a way that $u(x_1,t)=0$, we conclude that $\partial_x u^{m-1+\delta}(\cdot,t)$ is a continuous function in ${{\rm I\kern-.22em R}}$ for all $t \in [0,T_0)$ with the additional property that $\partial_x u^{m-1+\delta}(x,t)=0$ for the point $(x,t)$ such as $u(x,t)=0$.\
The case of $1<m<2$ can be handled in a similar manner as above and we conclude that $\partial_x u(\cdot,t)$ is a continuous function in ${{\rm I\kern-.22em R}}$ for all $t \in [0,T_0)$ with the additional property that $\partial_x u(x,t)=0$ for the point $(x,t)$ such as $u(x,t)=0$. This completes the proof of Theorem \[thm:lip\].
Proof of Theorem \[thm;finite speed\] {#section;4}
======================================
Let $(u_{\varepsilon},v_{\varepsilon})$ be the unique strong solution of (KS)$_{\varepsilon}$ given by Proposition \[pr;apto\]. For a fixed $R>0$, we take $a,b>0$ such as $-R<a<b<R$ and consider the following ordinary differential equations: $${\rm (IE)_{\xi}} :\left\{
\begin{array}{llll}
& \xi_{\varepsilon}^{'}(t) =
\frac{m}{m-1} \partial_x \big( u_{\varepsilon} + \varepsilon
\big)^{m-1} (\xi_{\varepsilon}(t),t)
- \big(u_{\varepsilon}
+ \varepsilon
\big)^{q-3}
u_{\varepsilon}
\cdot \partial_x v_{\varepsilon}(\xi_{\varepsilon}(t),t),
\ \ 0 \le t<T_0,
\\
& \xi_{\varepsilon}(0) = a,
\end{array}
\right.$$ and $${\rm (IE)_{\Xi}} :\left\{
\begin{array}{llll}
& \Xi_{\varepsilon}^{'}(t) =
\frac{m}{m-1} \partial_x \big( u_{\varepsilon} + \varepsilon
\big)^{m-1} (\Xi_{\varepsilon}(t),t)
- \big(u_{\varepsilon}
+ \varepsilon
\big)^{q-3}
u_{\varepsilon}
\cdot \partial_x v_{\varepsilon}(\Xi_{\varepsilon}(t),t),
\ \ 0 \le t<T_0,
\\
& \Xi_{\varepsilon}(0) = b.
\end{array}
\right.$$ By Remark 4 (iii), (\[estK0\]), (\[adj\]) and (\[proof;th-2\]), we have $$\begin{aligned}
\label{2R}
\partial_x (u_{\varepsilon}+\varepsilon)^{m-1}
\in C^1([-2R,2R] \times [0,T_0))\end{aligned}$$ and $$\begin{aligned}
&& \sup_{0<\varepsilon<1}
\Big(
\sup_{0<t<T_0}
\{
\| \partial_x
(u_{\varepsilon}+\varepsilon)^{m-1}(\cdot,t) \|_{L^{\infty}(-2R,2R)}
+
\|
\big(u_{\varepsilon}
+ \varepsilon
\big)^{q-3}
u_{\varepsilon}
\cdot \partial_x v_{\varepsilon}(\cdot,t)
\|_{L^{\infty}(-2R,2R)}
\}
\Big)
\nonumber \\
&& \quad \le C,\end{aligned}$$ where $C=C(m,\gamma,q,u_0)$. We now chose $R>0$ large enough such that $\frac{2R}{C} > T_0$. Then, it follows from the well-known theorem on the existence and uniqueness of local solutions to the initial value problem for the ordinary differential equations that both (IE)$_{\xi}$ and (IE)$_{\Xi}$ have a unique $C^1$-solution $\xi_{\varepsilon}(t)$ and $\Xi_{\varepsilon}(t)$ on $[0,T_0)$ for all $\varepsilon>0$, respectively.\
We consider the following domain: $$\begin{aligned}
D_{\tau}
& := &
\displaystyle
\bigcup_{t \in [0,\tau]}
I_{t} \times \{t\}, \quad
I_{t} \ := \
\Big\{ x \in {{\rm I\kern-.22em R}}; \
\xi_{\varepsilon}(t) \le x \le \Xi_{\varepsilon}(t)
\Big\}
\quad \ {\rm for} \ 0 <\tau < T_0.\end{aligned}$$ By the local uniqueness of the initial value problem (IE)$_\xi$ and (IE)$_{\Xi}$, we obtain that $$\xi_{\varepsilon}(t) < \Xi_{\varepsilon}(t)
\qquad
{\rm for \ all } \
0 \le t <T_0.$$ Let us define the gradient $\overrightarrow{\nabla}$ and the vector $\textbf{F}$ on $(x,t)$ by $$\begin{aligned}
\overrightarrow{\nabla}
& := & (\partial_x, \partial_t),
\quad
\textbf{F}(x,t)
\ := \
\Big(- \partial_x (u_{\varepsilon}+\varepsilon)^m
+ (u_{\varepsilon} + \varepsilon)^{q-2} u_{\varepsilon}
\cdot \partial_x v_{\varepsilon},
\ \ u_{\varepsilon}+\varepsilon
\Big).\end{aligned}$$ Then it follows from the first equation of (KS)$_{\varepsilon}$ that $$\begin{aligned}
\label{a1}
\lefteqn{
\int_{D_{\tau}} \overrightarrow{\nabla} \cdot \textbf{F}(x,t) \ dxdt
} \nonumber \\
& = & \int_{D_{\tau}} \partial_t u_{\varepsilon}
- \partial_x \Big(
\partial_x (u_{\varepsilon} +\varepsilon)^m
- (u_{\varepsilon} + \varepsilon)^{q-2} u_{\varepsilon}
\cdot \partial_x v_{\varepsilon}
\Big) \ dxdt
\ = \ 0\end{aligned}$$ for all $0<\tau<T_0$. Taking two curves $C_1$ and $C_2$ as $$\begin{aligned}
C_1 := \{(x,t) = (\xi_{\varepsilon}(t),t) ; \ 0<t<\tau \},
\quad
C_2 := \{(x,t) = (\Xi_{\varepsilon}(t),t) ; \ 0<t<\tau \},\end{aligned}$$ we have $$\begin{aligned}
\partial D_{\tau} & = & I_0 \ \cup \ C_1 \ \cup \ C_2 \ \cup \ I_{\tau}.\end{aligned}$$ Hence, the Stokes formula gives $$\begin{aligned}
\label{a2}
\lefteqn{
0= \int_{D_{\tau}} \overrightarrow{\nabla} \cdot \textbf{F}(x,t) \ dxdt
} \nonumber \\
& = &
\int_{\partial D_{\tau}} \textbf{F}(x,t) \cdot {\bf n} \ dS
\nonumber \\
& = & \int_a^b \textbf{F}(x,0) \cdot (0,-1) \ dx
- \int_{\Xi_{\varepsilon}(\tau)}^{\xi_{\varepsilon}(\tau)}
\textbf{F}(x,t) \cdot (0,1) \ dx
+ \int_{C_1}
\textbf{F} \cdot {\bf n_1} \ dS
+ \int_{C_2}
\textbf{F} \cdot {\bf n_2} \ dS
\nonumber \\
\quad
& = & - \int_a^b (u_{0\varepsilon}+\varepsilon) \ dx
+ \int_{\xi_{\varepsilon}(\tau)}^{\Xi_{\varepsilon}(\tau)}
(u_{\varepsilon}+\varepsilon) \ dx
+ \int_{C_1}
\textbf{F} \cdot {\bf n_1} \ dS
+ \int_{C_2}
\textbf{F} \cdot {\bf n_2} \ dS,\end{aligned}$$ where ${\bf n_1}$ and ${\bf n_2}$ denote the unit outer normals to $C_1$ and $C_2$, respectively. Since $$\begin{aligned}
{\bf n_1} & = & \frac{(1, \xi_{\varepsilon}^{\prime}(t))}
{\sqrt{1+(\xi_{\varepsilon}^{\prime}(t))^2}},
\qquad
{\bf n_2} \ = \ \frac{(1, \Xi_{\varepsilon}^{\prime}(t))}
{\sqrt{1+(\Xi_{\varepsilon}^{\prime}(t))^2}},\end{aligned}$$ we have by (IE)$_{\xi}$ and (IE)$_{\Xi}$ that $$\begin{aligned}
\label{agh-1}
\textbf{F} \cdot {\bf n_1} & = & 0
\quad {\rm on} \ \ C_1,
\qquad
\textbf{F} \cdot {\bf n_2} \ = \ 0
\quad {\rm on} \ \ C_2.\end{aligned}$$ Combining (\[a1\])–(\[agh-1\]), we have $$\begin{aligned}
\label{key;ineq}
\int_{\xi_{\varepsilon}(\tau)}^{\Xi_{\varepsilon}(\tau)}
(u_{\varepsilon}(x,\tau)+\varepsilon) \ dx
& = &
\int_a^b (u_{0\varepsilon}(x)+\varepsilon) \ dx,
\qquad 0 \le \tau <T_0.\end{aligned}$$ On the other hand, we obtain from (\[proof;th-2\]), Proposition \[pr;apto\] and Lemma \[lip\] that $$\begin{aligned}
\sup_{0<\varepsilon<1} \| \xi_{\varepsilon}\|_{L^{\infty}(0,T_0)}
& \le & a+
\Big(
\frac{m}{m-1} \cdot C
+ 2(\|u_0\|_{L^{\infty}}+2)^{q-2} \cdot \| u_0 \|_{L^1}
\Big) \cdot T_0,
\label{111}
\\
\sup_{0<\varepsilon<1} \|\xi_{\varepsilon}^{'}\|_{L^{\infty}(0,T_0)}
& \le &
\frac{m}{m-1} \cdot C
+ 2(\|u_0\|_{L^{\infty}}+2)^{q-2} \cdot \| u_0 \|_{L^1}.
\label{222}\end{aligned}$$ Hence it follows by the Ascoli-Arzela theorem that there exists a subsequence of $\{\xi_{\varepsilon}(t)\}$, still denoted by $\{\xi_{\varepsilon}(t)\}_{\varepsilon>0}$, and a function $\xi \in C^{0,1}[0,T_0)$ such that $$\begin{aligned}
\label{xi}
\xi_{\varepsilon}(t) \ \to \ \xi(t) \qquad
{\rm as} \ \varepsilon \to 0
\quad {\rm uniformly \quad for \ every} \ \ t \in [0,T_0).\end{aligned}$$ Obviously, a similar argument to $\Xi_{\varepsilon}(t)$ also holds, and there exist a subsequence of $\{\Xi_{\varepsilon}(t)\}_{\varepsilon>0}$, still denoted by $\{\Xi_{\varepsilon}(t)\}$, and $\Xi \in C^{0,1}[0,T_0)$ such that $$\begin{aligned}
\label{Xi}
\Xi_{\varepsilon}(t) \ \to \ \Xi(t) \qquad
{\rm as} \ \varepsilon \to 0
\quad {\rm uniformly \quad for \ every} \ \ t \in [0,T_0).\end{aligned}$$ Since $u_0 \equiv 0$ on $[a,b]$, by letting $\varepsilon \to 0$ in (\[key;ineq\]), we have $$\begin{aligned}
\label{conv;ineq}
\int_{\xi(t)}^{\Xi(t)} u(x,t) \ dx
\ = \
\int_a^b u_0(x) \ dx \ = \ 0\end{aligned}$$ for all $0 \le t <T_0$. Indeed, we may assume $$\begin{aligned}
-2R < \xi_{\varepsilon}(t) < \Xi_{\varepsilon}(t) < 2R
\qquad
{\rm for \ all } \ \varepsilon>0, \ {\rm and \ all} \ 0 \le t <T_0,\end{aligned}$$ where $R>0$ is the same as in (\[2R\]). Hence it follows from (\[Lm;strong–sta\]), (\[xi\]), (\[Xi\]) and Proposition \[pr;local existence\] that $$\begin{aligned}
\label{conv;ineq-2}
\lefteqn{
\Big|
\int_{\xi_{\varepsilon}(t)}^{\Xi_{\varepsilon}(t)} (u_{\varepsilon}+\varepsilon) \ dx
-
\int_{\xi(t)}^{\Xi(t)} u \ dx
\Big|
} \nonumber \\
& \le &
\Big| \int_{\xi_{\varepsilon}(t)}^{\Xi_{\varepsilon}(t)}
(u_{\varepsilon} - u) \ dx \Big|
\ + \
\varepsilon \int_{\xi_{\varepsilon}(t)}^{\Xi_{\varepsilon}(t)} \ dx
\ + \
\Big| \int_{\xi_{\varepsilon}(t)}^{\Xi_{\varepsilon}(t)}
u \ dx
- \int_{\xi(t)}^{\Xi(t)}
u \ dx
\Big|
\nonumber \\
& \le &
\Big(\sup_{-2R \le x \le 2R}
|u_{\varepsilon}(x,t)-u(x,t)| + \varepsilon \Big)
(\Xi_{\varepsilon}(t)-\xi_{\varepsilon}(t))
\nonumber \\
&&
\quad \ + \
\|u\|_{L^{\infty}(Q_{T_0})}
\Big( |\Xi_{\varepsilon}(t)-\Xi(t)|
+
|\xi_{\varepsilon}(t)-\xi(t)|
\Big)
\nonumber \\
& \le &
4R \Big(\sup_{-2R \le x \le 2R}
|u_{\varepsilon}(x,t)-u(x,t)| + \varepsilon \Big)
\ + \
\Big(
\| u_0 \|_{L^{\infty}} +2
\Big)
\Big( |\Xi_{\varepsilon}(t)-\Xi(t)|
+
|\xi_{\varepsilon}(t)-\xi(t)|
\Big)
\nonumber \\
& \to & 0 \qquad {\rm as} \ \varepsilon \to +0,\end{aligned}$$ which yields (\[conv;ineq\]). Since $u$ is non-negative in ${{\rm I\kern-.22em R}}\times [0,T_0)$, we conclude from (\[conv;ineq\]) that $$\begin{aligned}
u(x,t)=0 \qquad {\rm for} \quad \xi(t) \le x \le \Xi(t), \ 0 \le t<T_0.\end{aligned}$$ This proves Theorem \[thm;finite speed\].
[9]{} [, [Regularity properties of flows through porous media: The interface]{}, [*Arch. Rational Mech. Anal.*]{}, [**37**]{} (1970), 1-10. MR 41 656. , [Regularity properties of flows through porous media: A counterexample]{}, [*SIAM J. Appl. Math.*]{}, [**19**]{} (1970), 299-307. MR 42 683. , [Regularity properties of flows through porous media]{}, [*SIAM J. Appl. Math.*]{}, [**17**]{} (1969), 461-467. , [Singular limit of a chemotaxis-growth model]{}, [*Adv. Differential Equations* ]{}, [**72**]{} (2001), 1173–1218. , [Nonlinear aspects of chemotaxis ]{}, [*Math. Biosci.*]{}, [**56**]{} (1981), 217–237. , [The porous medium equation in one dimension]{}, [*Trans. Amer. Math. Soc.*]{}, [**2**]{}, (1977), 381-415. , [Global solution to the semi-linear Keller-Segel model of parabolic-parabolic type with small data in $L^r({{\rm I\kern-.22em R}}^n)$]{}, [*submitted*]{}. , [ Keller-Segel system of parabolic-parabolic type with initial data in weak $L^{\frac{n}{2}}({{\rm I\kern-.22em R}}^n)$ and its application to the self-similar solution]{}, [*submitted*]{}. , [Strong solutions to the Keller-Segel system with the weak $L^{\frac{n}{2}}$ initial data and its application to the blow-up rate]{}, [*submitted*]{}. , [Local existence and finite time blow-up in the 2-D Keller-Segel system,]{} [*submitted*]{}. , [Finite time blow-up of the solution for the nonlinear parabolic equation of the drift diffusion type]{}, [*Diff. Integral Equations*]{}, [**16**]{}, (2003), 427-452. , [Second Order Parabolic Differential Equations]{}, World Scientific publishing Co. Pte. Ltd. (1996). , [**43**]{}, (1983), 449-464. , [Lipschitz continuous solutions of some doubly nonlinear parabolic equations]{}, [*Discrete Contin. Dynam. Systems,*]{} [**8**]{}, (2002), 647-670. , [Global existence in the sub-critical cases and finite time blow-up in the super-critical cases to degenerate Keller-Segel systems,]{} [*Differential Integral Equations,*]{} [**9**]{}, (2006), 841-876. , [Time Global Existence and Asymptotic Behavior of Solutions to Degenerate Quasi-linear Parabolic Systems of Chemotaxis]{}, [*Differential Integral Equations,*]{} [**20**]{}, (2007), 133-180. , [ Global existence and decay properties for a degenerate Keller-Segel model with a power factor in drift term, ]{} [*J. Differential Equations,*]{} [**227**]{}, (2006), 333-364. ]{}
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Many realizations of solid-state qubits involve couplings to leakage states lying outside the computational subspace, posing a threat to high-fidelity quantum gate operations. Mitigating leakage errors is especially challenging when the coupling strength is unknown, e.g., when it is caused by noise. Here we show that simple pulse sequences can be used to strongly suppress leakage errors for a qubit embedded in a three-level system. As an example, we apply our scheme to the recently proposed charge quadrupole (CQ) qubit for quantum dots. These results provide a solution to a key challenge for fault-tolerant quantum computing with solid-state elements.'
author:
- Joydip Ghosh
- 'S. N. Coppersmith'
- Mark Friesen
title: 'Pulse sequences for suppressing leakage in single-qubit gate operations'
---
Recent advances in semiconducting quantum dots make them promising candidates for universal quantum computing [@PhysRevA.57.120; @Elzerman2004; @Petta2005; @Hanson2007; @Shulman2014; @Veldhorst2015]. However, performing gate operations with high enough fidelity to support fault-tolerant error correction remains a key challenge [@Taylor2005]. Two ways that a qubit can fail to have high fidelity are (i) the qubit could decay or dephase, and (ii) quantum information could leak out of the qubit’s logical subspace into other quantum states in the physical system [@PhysRevA.71.052301; @PhysRevLett.103.110501; @West2012; @PhysRevA.88.062329; @PhysRevB.88.161303]. While several recent proposals in quantum dots have focused on suppressing dephasing from environmental noise [@Kim2014; @Eng2015], relatively little effort has gone into suppressing leakage [@PhysRevB.88.161303].
Several approaches for reducing leakage errors have been developed for superconducting qubits, including analytic pulse shaping [@PhysRevA.87.022309; @PhysRevLett.103.110501] and optimal quantum control [@0953-2048-27-1-014001; @PhysRevLett.114.200502]. [It is also known that leakage errors are, in principle, suppressible for a system-bath model with a composite sequence of a large number of pulses [@PhysRevA.71.052301].]{} However, suppressing leakage errors below the fault-tolerant threshold under experimental conditions for quantum dot qubits is challenging, because of the need to apply smoothly varying [short]{} control pulses [@Barnes2015] [within a time much less than the coherence times of the system]{}, and also because the qubits experience fluctuations that vary in time and strength [@PhysRevB.94.045435]. For semiconducting systems, these constraints preclude using analytically derived pulse shapes, since these require knowing the noise strength as an input parameter [@PhysRevLett.103.110501], or quantum control strategies that rely on optimizing an unrealistically large number of control parameters [@PhysRevLett.114.200502; @0953-2048-27-1-014001].
In this work, we develop a simple and experimentally feasible protocol based on composite pulses that suppresses leakage errors in a three-level quantum system where the leakage state is coupled to one of the logical states with an unknown but static coupling strength (referred to as the quasistatic noise approximation [@PhysRevB.94.045435]). Whereas the leakage error scales quadratically with noise amplitude in conventional pulsed-gate schemes for this model, our protocol significantly improves the gate fidelities by eliminating computational errors in the logical subspace up to fourth order in the noise amplitude and leakage errors up to sixth order.
In a basis comprised of two logical states and one leakage state, the model Hamiltonian is given by \[eq:ModelHamiltonian\] H=H\_[z]{} + H\_[x]{} + H\_[leak]{}, with &&H\_[z]{}=(
[ccc]{} 1 & 0 & 0\
0 & -1 & 0\
0 & 0 & -
), H\_[x]{}=g(
[ccc]{} 0 & 1 & 0\
1 & 0 & 0\
0 & 0 & 0
)\
&&H\_[leak]{}=(
[ccc]{} 0 & 0 & 0\
0 & 0 & 1\
0 & 1 & 0
) , where $\epsilon_{\rm q}$ and $g$ are the independent control parameters for rotations about the $z$- and $x$-axes of the Bloch sphere in the logical subspace, $\xi$ denotes the unknown coupling between the leakage state and one of the logical states, and $\zeta$ denotes the (scaled) leakage state energy in the absence of coupling. While the Hamiltonian (\[eq:ModelHamiltonian\]) provides a very general description of a 2-level system coupled to a leakage state [@PhysRevA.88.052330; @PhysRevLett.103.110501], we focus here on the semiconducting charge quadrupole (CQ) qubit [@Friesen:2016aa], for which the logical states have different charge distributions but the same center of mass, yielding a qubit that is inherently protected from the predominant type of noise in this system: uniform electric field fluctuations.
The CQ qubit is formed in three adjacent semiconducting quantum dots sharing a single electron [@Friesen:2016aa]. In the localized charge basis $\{{\left|100\right\rangle},{\left|010\right\rangle},{\left|001\right\rangle}\}$, where the basis states denote the electron being in the $1^{\rm st}$, $2^{\rm nd}$ or the $3^{\rm rd}$ dot respectively, the Hamiltonian is given by H\_[CQ]{}=(
[ccc]{} \_[d]{} & t\_[A]{} & 0\
t\_[A]{} & \_[q]{} & t\_[B]{}\
0 & t\_[B]{} & -\_[d]{}
)+, where $U_{1,2,3}$ are the on-site potentials for the three dots, $t_{A,B}$ are tunnel couplings between adjacent dots, and $\epsilon_{\rm d}=(U_{1}-U_{3})/2$ and $\epsilon_{\rm q}=U_{2}-(U_{1}+U_{3})/2$ denote the dipolar and quadrupolar detuning parameters, respectively. We now define a new set of basis states [@Friesen:2016aa] [|C]{}=[|010]{}, [|E]{}=, [|L]{}=, \[eq:basisDef\] where ${\left|C\right\rangle}$ and ${\left|E\right\rangle}$ correspond to logical states and ${\left|L\right\rangle}$ denotes the leakage state. In this basis, the Hamiltonian becomes \[eq:Hcqcel\] \_[CQ]{}=(
[ccc]{} & &\
& - & \_[d]{}\
& \_[d]{} & -
) , where we have neglected a term proportional to the identity. Tunnel couplings and detuning parameters can be independently tuned in quantum dots [@PhysRevLett.116.110402; @PhysRevLett.116.116801], and we set them here to $t_{\rm A}=t_{\rm B}$ and $\epsilon_{\rm d}=0$. In this case, it can be shown that $\tilde{H}_{\rm CQ}$ forms a decoherence-free subspace (DFS) with respect to uniform electric field fluctuations [@Friesen:2016aa]. Fluctuations of the dipolar detuning parameter $\epsilon_{\rm d}$ break this DFS by coupling the states ${\left|E\right\rangle}$ and ${\left|L\right\rangle}$.
![Energy eigenvalues of the CQ qubit, as functions of (a) $\epsilon_{\rm q}$ (with $g/h=0$ and $\delta\epsilon_{\rm d}/h=0.3$ GHz) and (b) $g$ (with $\epsilon_{\rm q}/h=0$ and $\delta\epsilon_{\rm d}/h=0.3$ GHz). The vertical dashed lines denote the idle point ($\epsilon_{\rm q}=g=0$) for the CQ qubit. The insets (which use the same time axis) show the bang-bang pulse profiles for $\epsilon_{\rm q}$ and $g$ that implement the $\mathcal{R}_{\rm zxz}$ composite gate \[see Eq. (\[eq:RzxzDefinition\])\]. These pulses generate $z$-rotations in (a), and $x$-rotations in (b).[]{data-label="fig:energyLevels"}](energyLevels.pdf){width="3.3in"}
Under the conditions described above, $\tilde{H}_{\rm CQ}$ reduces to Eq. (\[eq:ModelHamiltonian\]), where $\zeta=1$, $\xi=\epsilon_{\rm d}$, $g=(t_{\rm A}+t_{\rm B})/\sqrt{2}$. [For the remainder of this paper, we will adopt Eq. (\[eq:Hcqcel\]) with $t_{\rm A}=t_{\rm B}$ as our model Hamiltonian. We restrict our analysis to the special case $\zeta=1$, as appropriate for CQ qubits, although it is also possible to generalize our approach to the case $\zeta \neq 1$ by considering longer pulse sequences that are more challenging to implement, experimentally.]{} Figure \[fig:energyLevels\] shows the energy levels of $\tilde{H}_{\rm CQ}$ for typical device parameters, as a function of $\epsilon_{\rm q}$ and $g$. Here, the vertical dashed lines denote the idle point for qubit gate operations. A rotation about the $z$-axis of the Bloch sphere can now be performed by pulsing $\epsilon_{\rm q}$ (with $g=0$) from the idle point to a region with $|\epsilon_{\rm q}| \gg 0$, as shown in Fig. \[fig:energyLevels\](a). Here the large energy gap between logical states ${\left|C\right\rangle}$ and ${\left|E\right\rangle}$ generates the phase required for the rotation. Similarly, a rotation about the $x$-axis of the Bloch sphere is performed by pulsing $g$ (with $\epsilon_{\rm q}=0$) from the idle point, as shown in Fig. \[fig:energyLevels\](b).
The predominant noise in this system arises from the motion of trapped charge in the semiconductor or dielectric materials [@Dial2013; @Wu19082014; @BrandurPreprint]. We model it as $\epsilon_{\rm q}=\bar{\epsilon}_{\rm q}+\delta\epsilon_{\rm q}$ and $\epsilon_{\rm d}=\bar{\epsilon}_{\rm d}+\delta\epsilon_{\rm d}$, where $\bar{\epsilon}_{\rm q,d}$ are the average quadrupolar and dipolar detuning control parameters (here, $\bar \epsilon_\text{d}=0$), and $\delta\epsilon_{\rm q,d}$ are the corresponding fluctuations. In [@Friesen:2016aa] it is argued that, while $\delta\epsilon_{\rm d}\gg \delta\epsilon_{\rm q}$, the effects of $\delta\epsilon_{\rm d}$ are largely suppressed by the DFS. Here, we focus on the residual effects of $\delta\epsilon_{\rm d}$, which primariliy cause leakage. Experimentally, it is known that the noise spectrum of $\delta\epsilon_{\rm d}$ is dominated by low-frequency fluctuations, which are slow compared to gate operations [@Dial2013; @Kawakami2016]. We therefore assume that while $\delta\epsilon_{\rm d}$ is variable, its value remains constant during a given gate operation [@PhysRevB.94.045435]. [In this work, we neglect the less dominant sources of noise, such as high-frequency components and control jitters for voltage and timing.]{}
We now show how leakage due to $\delta\epsilon_{\rm d}$ can be suppressed for arbitrary single-qubit rotations. We first consider the “bang-bang" limit, in which the control pulses switch instantaneously between two values, and we analytically determine the probabilities for leakage and computational errors. Later, we will relax the bang-bang constraint and numerically compute the fidelities of single-qubit rotations with smoothly varying pulse profiles, which are more realistic for controlling semiconducting qubits.
The unitary operators for noisy $z$- and $x$-rotations are U\_[z]{}(\_[q]{},\_[d]{},)&=&e\^[-i/\_]{}\
U\_[x]{}(g,\_[d]{},)&=&e\^[-i/2g]{} , \[eq:UzUx\] for arbitrary rotation angles $\varphi$ and $\theta$. For bang-bang gates, these angles are related to the corresponding gate times as $\varphi=t_z(\epsilon_\text{q}/\hbar)$ and $\theta=t_x(2g/\hbar)$. Hence, $\varphi$ must have the same sign as $\epsilon_\text{q}$, and $\theta$ must be positive, since $g> 0$ for most solid-state devices, including quantum dots. Our approach will be to compose $U_\text{z}$ and $U_\text{x}$ gates to obtain arbitrary single-qubit rotations in the logical subspace as a function of $\epsilon_\text{q}$, $g$, $\varphi$, and $\theta$, and then determine the conditions that should be imposed on these parameters to suppress the leading order leakage terms in the composed gates.
If $z$- and $x$-rotations are performed using “bare" $U_z$ and $U_x$ operations, i.e. with no gate sequence, it can be shown that leakage error, and hence the computational error in the logical subspace, both scale as $(\delta\epsilon_{\rm d}/g)^{2}$ [@Friesen:2016aa]. In [@Supp] we show that it is impossible to achieve better results by composing just two bare gates. We therefore focus here on compositions of three bare gates, of the form \[eq:RzxzDefinition\] \_[zxz]{}(,)=U\_[z]{}(\_[q]{},\_[d]{},)U\_[x]{}(g,\_[d]{},)U\_[z]{}(-\_[q]{},\_[d]{},). A schematic diagram of a bang-bang pulse sequence for this $\mathcal{R}_{\rm zxz}$ operation is shown in the insets of Fig. \[fig:energyLevels\].
We now expand $\mathcal{R}_{\rm zxz}$ about $\delta\epsilon_{\rm d}=0$ and note that the first order terms of $\delta\epsilon_{\rm d}$ vanish from the Taylor expansion if we set [@Supp] \_[q]{}=[-()]{}. \[eq:zxzCondition\] This condition, along with the conditions on $\varphi$, $\theta$, and $g$, are all satisfied when $2\pi < \theta < 4\pi$, which means that $U_x$ rotations are between 1 and 2 cycles around the Bloch sphere.
We note that $\mathcal{R}_{\rm zxz}(\theta,\varphi)$ corresponds to a rotation of angle $\theta$ about an arbitrary axis $\hat{n}=\cos(\varphi/2)\hat{x}+\sin(\varphi/2)\hat{y}$ in the $x$-$y$ plane of the logical subspace (modulo an irrelevant overall phase). To estimate analytically how the computational and leakage errors depend on the noise $\delta\epsilon_{\rm d}/g$, we evaluate the higher order terms in the Taylor expansion of $\mathcal{R}_{\rm zxz}(\theta,\varphi)$ [@Supp], observing that the error probability within the logical subspace $|\mathcal{R}_{\rm zxz}({\delta\epsilon_{\rm d} \neq 0})-\mathcal{R}_{\rm zxz}({\delta\epsilon_{\rm d}=0})|^2$ scales as $(\delta\epsilon_{\rm d}/g)^{4}$, while the probability of leakage errors $P_{LC,LE}=\left|{\left\langle L \right|}\mathcal{R}_{\rm zxz}{\left|C,E\right\rangle}\right|^{2}$ scales as $(\delta\epsilon_{\rm d}/g)^{6}$. [Therefore, for a typical experimental value of $\delta\epsilon_{\rm d}/g \sim 10^{-1}$ [@BrandurPreprint], the error in the computational subspace is $\sim 10^{-4}$ and $P_{LC,LE} \sim 10^{-6}$.]{} This represents a remarkable improvement over the bare gates, [for which the leakage error is of order $10^{-2}$ when $\delta\epsilon_{\rm d}/g \sim 10^{-1}$ (For a detailed comparison of leakage errors in simple and composite pulse sequences, see [@Supp])]{}.
Finally, we construct a completely general and arbitrary rotation on the Bloch sphere. Since the effective rotation axis for an $\mathcal{R}_{\rm zxz}$ gate lies anywhere in the $x$-$y$ plane, an arbitrary rotation requires just two steps [@2013arXiv1303.0297S]: $\mathcal{R}_{\rm zxz}(\theta_2,\varphi_2)\mathcal{R}_{\rm zxz}(\theta_1,\varphi_1)$, comprising 6 bare gates. Although there are penalties associated with longer gate sequences, such as reduced gate speeds and accumulation of errors, the sequences described here improve the overall scaling of errors with respect to the noise amplitude. Hence, they are always beneficial if $\delta\epsilon_\text{d}$ is small enough.
While an $\mathcal{R}_{\rm zxz}$ gate produces arbitrary rotations in the $x$-$y$ plane, Eq. (\[eq:zxzCondition\]) indicates that the control parameter $g$ ($\epsilon_\text{q}$) diverges in the limit $\theta\rightarrow 2\pi$ ($4\pi$). This implies that an identity gate cannot be implemented using a single $\mathcal{R}_{\rm zxz}$ operation. A high-fidelity identity gate is an essential ingredient for fault-tolerant quantum computing because most qubits remain idle during most of the quantum error-correction cycle. We cannot use the “idle gate," corresponding to $\epsilon_\text{q}=g=0$, as an identity gate because $\delta\epsilon_\text{d}\neq 0$ causes errors. It is possible to obtain an identity gate from the sequence $[\mathcal{R}_{\rm zxz}(\pi,\varphi)]^2$, which comprises 6 bare gates; however it is interesting to ask whether shorter identity sequences exist.
We now present a viable identity sequence, $\mathcal{R}_{I}$, that uses just 4 bare gates: Following the same procedure used to construct $\mathcal{R}_{\rm zxz}$, we consider the sequence \[eq:Identity\] \_[I]{}=\^2. $\mathcal{R}_{I}$ is clearly an identity operation when $\delta\epsilon_{\rm d}=0$. To address the case $\delta\epsilon_{\rm d} \neq 0$, we perform a Taylor expansion of $\mathcal{R}_{I}$ about $\delta\epsilon_{\rm d}=0$ [@Supp], noting that the first-order term in $\delta\epsilon_{\rm d}$ vanishes due to the special form of the sequence, even without imposing special constraints on $\epsilon_q$ and $g$. The Taylor expansion also indicates that the probability of computational errors, given by $\left|{\left\langle C \right|}\mathcal{R}_{I}{\left|E\right\rangle}\right|^{2} = \left|{\left\langle E \right|}\mathcal{R}_{I}{\left|C\right\rangle}\right|^{2}$, scales as $(\delta\epsilon_{\rm d}/g)^{4}$, while the probability of leakage errors, given by $\left|{\left\langle L \right|}\mathcal{R}_{I}{\left|C\right\rangle}\right|^{2} = \left|{\left\langle L \right|}\mathcal{R}_{I}{\left|E\right\rangle}\right|^{2}$, scales as $(\delta\epsilon_{\rm d}/g)^{6}$.
To understand why our pulse sequences work, we now describe a method for simultaneously visualizing the evolution of the logical and leakage states on a pair of coupled Bloch spheres. An arbitrary state ${\left|\Psi\right\rangle}$ of a three-level quantum system can be expressed in the $\left\{{\left|C\right\rangle}, {\left|E\right\rangle}, {\left|L\right\rangle}\right\}$ basis using four angle variables: [|]{}=
(/2)(/2)\
(/2)(/2)e\^[i]{}\
(/2)e\^[i]{}
. When $\vartheta=0$, ${\left|\Psi\right\rangle}$ maps onto ${\left|\psi_{g}\right\rangle}=\cos(\chi/2){\left|C\right\rangle}+\sin(\chi/2)e^{i\varrho}{\left|E\right\rangle}$, comprising a two-level system. Similarly, when $\chi=0$, ${\left|\Psi\right\rangle}$ maps onto ${\left|\psi_{r}\right\rangle}=\cos(\vartheta/2){\left|C\right\rangle}+\sin(\vartheta/2)e^{i\varsigma}{\left|L\right\rangle}$, comprising a different two-level system. The full mapping ${\left|\Psi\right\rangle}\rightarrow \{{\left|\psi_g\right\rangle},{\left|\psi_r\right\rangle}\}$ is bijective, and the states ${\left|\psi_{g}\right\rangle}$ and ${\left|\psi_{r}\right\rangle}$ may alternatively be represented as unit vectors on a pair of Bloch spheres that we label ‘green’ and ‘red,’ respectively. The angles $\vartheta$, $\chi$, $\varrho$, and $\varsigma$ here correspond to the usual polar and azimuthal angles of the respective Bloch spheres.
In the absence of leakage, the state vector of the red sphere points to the north pole and the state vector of the green sphere describes the logical qubit in the usual way. When leakage occurs, the state vector of the red sphere deviates from north, with its latitude describing the amplitude of the leakage state, and its azimuth describing the relative phase of ${\left|L\right\rangle}$ with respect to ${\left|C\right\rangle}$. The green sphere still describes the logical qubit, with its state vector renormalized to lie on the unit sphere. For quantum systems with multiple leakage states, we could extend this geometrical representation to include multiple Bloch spheres, with one sphere describing the logical states, and one additional sphere for each leakage state. Below, we restrict our analysis to a single leakage state and compute the time-dependent trajectories of the green and red Bloch spheres for the $\mathcal{R}_{\rm zxz}$ pulse sequence.
\[fig:Evolution\] ![A numerically optimized gate operation. (a) The pulse sequence for an $X_{-\pi/2}$ gate, assuming rounded pulse profiles with $50$ ps rise times. Here, the optimization parameters are the gate times $t_z$ and $t_x$, and the peak value of $\epsilon_\text{q}$. The peak value of $g/h=3$ GHz is fixed. (b) Leakage probabilities as a function of time. The final leakage error is of order $10^{-10}$. (c) Time evolution of the logical and leakage states on green and red “Bloch spheres" respectively, [for the smooth pulse shown in (a).]{} Here, the color of the state vectors indicates the time. The transitions between $z$-, $x$-, and $z$-rotations are indicated by symbols, as consistent with (a).[]{data-label="fig:Evolution"}](Evolution.pdf "fig:"){width="3.2in"}
Our analytical results for pulse sequences were previously obtained in the bang-bang limit. However in charge qubit experiments, the combination of high-frequency filtering and short gate times ($\sim 0.1$-1 ns) means that pulses will experience significant rounding. To account for this, we now consider pulse sequences with smooth profiles and finite rise times. Under such conditions, constraints on the pulse shape, such as Eq. (\[eq:zxzCondition\]), must be modified. To do this, we keep our previous sequences, but vary the control parameters and gate times. The parameters are chosen to maximize the process fidelity, defined as [@Pedersen2007; @PhysRevA.81.052340] F= . \[eq:FidelityDef\] Such projections do not generally yield unitary operators; however Eq. (\[eq:FidelityDef\]) takes this into account, and captures both computational and leakage errors.
A typical result from our gate optimization procedure is shown in Fig. \[fig:Evolution\]. Here, we consider an effective $X_{-\pi/2}$ rotation, obtained from the 3-step sequence defined in Fig. \[fig:Evolution\](a), which is analogous to the bang-bang sequence $\mathcal{R}_{\rm zxz}(2\pi+\pi/2,2\pi)$. We adopt several fixed device parameters, including error-function pulse profiles with rise times of 50 ps [@Supp], a peak tunnel coupling of $g/h=3$ GHz, and a typical experimental noise value of $\delta \epsilon_{d}/h = 0.3$ GHz [@BrandurPreprint]. We then numerically optimize the gate times $t_z$ and $t_x$ defined in the figure, and the peak value of $\epsilon_q$, to maximize the fidelity. Figures \[fig:Evolution\](b) and \[fig:Evolution\](c) show the time evolution of the resulting gate for the initial state $\cos(\pi/10){\left|C\right\rangle}+\sin(\pi/10){\left|E\right\rangle}$. Although there is significant leakage in the middle of the operation, the final leakage probability is strongly suppressed. From Eq. (\[eq:FidelityDef\]), the final infidelity is found to be $1-F=2.46\times10^{-8}$, while the total leakage errors corresponding to ${\left|C\right\rangle}$ and ${\left|E\right\rangle}$ states are $1.5\times10^{-10}$ and $1.2\times10^{-10}$ respectively. [These results are obtained for the smooth pulse sequence shown in Fig. \[fig:Evolution\](a), and represent orders of magnitude of improvement in the infidelity compared to the bang-bang pulse under the constraint (\[eq:zxzCondition\]). This improvement can be attributed to the numerical optimization procedure. Indeed, by relaxing constraint (\[eq:zxzCondition\]), while retaining the three-pulse construction, the dynamics can be modified to cancel out higher order noise effects, for either bang-bang or smooth pulses.]{}
The benefit of the two-Bloch-sphere representation is apparent in Fig. \[fig:Evolution\](c), where it provides a qualitative understanding of how our pulse sequence suppresses leakage errors. Starting from the north pole of the red sphere, the leakage probability grows slowly during the first $z$-rotation. As it grows, the leakage state accumulates phase, causing the state vector to spiral outward and downward. The special constraint on the control parameters, analogous to Eq. (\[eq:zxzCondition\]), causes the leakage phase to be inverted during the $x$-rotation, similar to a Hahn echo. The final $z$-rotation then yields a time evolution that is opposite of the initial evolution (a reverse spiral), which eventually passes very close to the north pole.
In conclusion, we have developed special pulse sequences for a three-level system, comprised of two logical states and a leakage state, which suppress the leakage probability and enable arbitrary, high-fidelity, single-qubit rotations. We consider a general model, in which only one of the logical states is coupled to the leakage state. A crucial feature of this method is that it requires no knowledge of the coupling to the leakage state. The method is perfectly suited for the quantum dot charge quadrupole (CQ) qubit, whose leakage coupling is caused by charge noise. Taking the CQ as our model system we have shown that computational errors in the gate operation arise at fourth order in the noise amplitude, while leakage errors arise at sixth order. This represents a substantial improvement over conventional pulse schemes, for which errors arise at second order in the noise amplitude. [While it is possible to design high-fidelity quantum gates for quantum dot qubits using optimal control schemes [@0953-2048-27-1-014001; @PhysRevLett.114.200502], we emphasize that our approach yields high gate fidelities without knowing the amplitude of the noise, which sets it apart from those schemes. We also note that a Hahn echo sequence, which has the same level of complexity as our sequence, has recently been employed successfully to improve the fidelities of charge dipole qubits [@Kim2015]. This suggests that our scheme could also yield high-fidelity gate operations in charge quadrupole qubits.]{} To understand why our pulse sequence works, we have introduced an intuitive geometrical mapping of the three-level system, in which the logical and leakage states are represented on two coupled Bloch spheres. In this way, the cancellation of leakage errors appears similar to a Hahn echo. Our proposal therefore provides an important step towards fault-tolerant quantum computation in quantum dots.
We thank Mark Eriksson and Brandur Thorgrimsson for many helpful discussions. The authors would like to acknowledge support from the Vannevar Bush Faculty Fellowship program sponsored by the Basic Research Office of the Assistant Secretary of Defense for Research and Engineering and funded by the Office of Naval Research through grant N00014-15-1-0029. This work was also supported in part by Army Research Office Grant W911NF-12-0607 and National Science Foundation (NSF) Grant PHY-1104660.
[32]{}ifxundefined \[1\][ ifx[\#1]{} ]{}ifnum \[1\][ \#1firstoftwo secondoftwo ]{}ifx \[1\][ \#1firstoftwo secondoftwo ]{}““\#1””@noop \[0\][secondoftwo]{}sanitize@url \[0\][‘\
12‘\$12 ‘&12‘\#12‘12‘\_12‘%12]{}@startlink\[1\]@endlink\[0\]@bib@innerbibempty [****, ()](http://link.aps.org/doi/10.1103/PhysRevA.57.120) [****, ()](http://dx.doi.org/10.1038/nature02693) [****, ()](http://science.sciencemag.org/content/309/5744/2180) [****, ()](http://link.aps.org/doi/10.1103/RevModPhys.79.1217) [****, ()](http://dx.doi.org/10.1038/ncomms6156) [****, ()](http://dx.doi.org/10.1038/nature15263) [****, ()](http://dx.doi.org/10.1038/nphys174) [****, ()](http://link.aps.org/doi/10.1103/PhysRevA.71.052301) [****, ()](http://link.aps.org/doi/10.1103/PhysRevLett.103.110501) [****, ()](http://stacks.iop.org/1367-2630/14/i=8/a=083002) [****, ()](http://link.aps.org/doi/10.1103/PhysRevA.88.062329) [****, ()](http://link.aps.org/doi/10.1103/PhysRevB.88.161303) [****, ()](http://dx.doi.org/10.1038/nature13407) [****, ()](http://advances.sciencemag.org/content/1/4/e1500214) [****, ()](http://link.aps.org/doi/10.1103/PhysRevA.87.022309) [****, ()](http://stacks.iop.org/0953-2048/27/i=1/a=014001) [****, ()](http://link.aps.org/doi/10.1103/PhysRevLett.114.200502) [****, ()](http://dx.doi.org/10.1038/srep12685) [****, ()](http://link.aps.org/doi/10.1103/PhysRevB.94.045435) [****, ()](\doibase 10.1103/PhysRevA.88.052330) [****, ()](https://www.nature.com/articles/ncomms15923) [****, ()](http://link.aps.org/doi/10.1103/PhysRevLett.116.110402) [****, ()](http://link.aps.org/doi/10.1103/PhysRevLett.116.116801) [****, ()](http://link.aps.org/doi/10.1103/PhysRevLett.110.146804) [****, ()](\doibase 10.1073/pnas.1412230111) @noop [ ()]{}, [****, ()](http://www.pnas.org/content/113/42/11738.abstract) @noop @noop [ ()]{}, [****, ()](http://www.sciencedirect.com/science/article/pii/S0375960107003271) [****, ()](http://link.aps.org/doi/10.1103/PhysRevA.81.052340) [****, ()](http://dx.doi.org/10.1038/nnano.2014.336)
Supplemental Material for “Pulse sequences for suppressing leakage in single-qubit gate operations"
===================================================================================================
In these Supplemental Materials we first show that two-pulse sequences cannot correct leakage errors, which leads us to study three-pulse sequences. We then provide details of the derivations of internal errors produced by our gate protocols, as well as the pulse shapes used in our simulations. Finally, we present a comparison of gate process fidelities between simple and composite pulse sequences.
Demonstration that a composition of two pulses cannot correct leakage errors
----------------------------------------------------------------------------
As mentioned in the main text, the leakage error for bare $x$- and $z$-rotations implemented in a charge quadrupole qubit scales as $(\delta\epsilon_{\rm d}/g)^2$. Here we show that composing two such pulses is also insufficient for suppressing the leading-order leakage errors.
We consider the following two composed operations, U\_[zx]{}(\_[q]{},g,\_[d]{},,)&=& U\_[z]{}(\_[q]{},\_[d]{},)U\_[x]{}(g,\_[d]{},)\
U\_[xz]{}(\_[q]{},g,\_[d]{},,)&=& U\_[x]{}(g,\_[d]{},)U\_[z]{}(\_[q]{},\_[d]{},), \[eq:UzUx\] where the bare gates are defined as U\_[z]{}(\_[q]{},\_[d]{},)&=&e\^[-i]{}\
U\_[x]{}(g,\_[d]{},)&=&e\^[-i]{}, \[eq:UzUx\] and the various parameters and expressions are defined the same as in the main text.
First we consider $U_{\rm zx}$. Occupation of the leakage state can be expressed in terms of the matrix elements ${\left\langle L \right|}U_{\rm zx}{\left|C\right\rangle}$ and ${\left\langle L \right|}U_{\rm zx}{\left|E\right\rangle}$. Expanding these two expressions about $\delta\epsilon_{\rm d}=0$ yields leading-order terms proportional to \[eq:UZXCond\] 2\_[q]{} \^2()+g () \_[q]{} ()+g (), respectively. To suppress the leakage errors to leading order, we must simultaneously set both of these expressions to zero. It is easy to show that the only solutions to this problem involve “null" rotations, for which either $\epsilon_\text{q}=0$ or $g=0$. However, these do not represent valid solutions, because $U_\text{zx}$ then reduces to one of the bare rotations, $U_\text{z}$ or $U_\text{x}$, which provide no leakage protection.
Applying the same considerations to $U_\text{xz}$ yields the constraints \[eq:UXZCond\] ()-1=\_[q]{} ()+g =0, which also do not allow viable solutions. This proves that it is impossible to suppress the leading-order effects of leakage using compositions of just two bare rotations.
[One can also attempt to compose two bare $z$- or $x$-rotations. However, it can be explicitly shown that leading order leakage error cannot be suppressed for either of these pulses under realistic conditions.]{} We therefore turn to compositions of three bare rotations.
Analysis of errors for $\mathcal{R}_{\rm zxz}(\theta,\varphi)$ {#sec:Rzxz}
--------------------------------------------------------------
In this section we consider the three-pulse sequence $\mathcal{R}_{\rm zxz}(\theta,\varphi)$ and determine the requirements for achieving noise suppression in this case.
A Taylor expansion of $\mathcal{R}_{\rm zxz}(\theta,\varphi)$ about $\delta\epsilon_{\rm d}=0$ yields
\[eq:RzxzExpanded\] \_[zxz]{}(,)= (
[ccc]{} () & -ie\^[-]{} () & 0\
-ie\^ () & () & 0\
0 & 0 & 1\
)-(+2()) (
[ccc]{} 0 & 0 & e\^[-]{} ()\
0 & 0 & i()\
e\^ () & i() & 0\
) ,
up to $O[\delta\epsilon_\text{d}/g]$. This $O[\delta\epsilon_\text{d}/g]$ term clearly causes leakage.
We can make the $O[\delta\epsilon_\text{d}/g]$ term vanish by choosing control parameters such that \_[q]{}=[-()]{}. \[eq:SzxzCondition\] When this condition is satisfied, the next-leading-order terms in the Taylor expansion become \[eq:RzxzExpanded\] \_[zxz]{}(,)= (
[ccc]{} () & -ie\^[-]{} () & 0\
-ie\^ () & () & 0\
0 & 0 & 1\
) + (
[ccc]{} () & i e\^[-]{}() & 0\
i e\^() & () & 0\
0 & 0 & (,)\
)()\^[2]{}+(), where $$\begin{gathered}
\mathcal{A}(\theta)=2\sin\left(\frac{\theta}{4}\right)\left[\sin\left(\frac{\theta}{4}\right)-\frac{\theta}{4}\cos\left(\frac{\theta}{4}\right)\right], \nonumber \\
\mathcal{B}(\theta)=\sin\left(\frac{\theta}{2}\right)-\frac{\theta}{4}\cos\left(\frac{\theta}{2}\right)-\tan\left(\frac{\theta}{4}\right).\end{gathered}$$ From Eq. (\[eq:RzxzExpanded\]), we see that $\mathcal{R}_{\rm zxz}\bigr\rvert_{\delta\epsilon_{\rm d} \neq 0}-\mathcal{R}_{\rm zxz}\bigr\rvert_{\delta\epsilon_{\rm d}=0} \sim \left(\delta\epsilon_{\rm d}/g\right)^{2}$, which implies that the error probabilities within the computational subspace scale as $\sim \left(\delta\epsilon_{\rm d}/g\right)^{4}$. However at this order, the leakage probabilities, $P_{LC}=\left|{\left\langle L \right|}\mathcal{R}_{\rm zxz}{\left|C\right\rangle}\right|^{2}$ and $P_{LE}=\left|{\left\langle L \right|}\mathcal{R}_{\rm zxz}{\left|E\right\rangle}\right|^{2} $, both vanish. Expanding $\mathcal{R}_{\rm zxz}(\theta,\varphi)$ to even higher order in $(\delta\epsilon_\text{d}/g)$ yields $$\begin{gathered}
P_{LC} = \left(\frac{\delta\epsilon_{\rm d}}{g}\right)^{6}\left|\frac{1}{3}\sin^{2}\left(\frac{\theta}{4}\right)-\frac{\theta}{4}\tan\left(\frac{\theta}{4}\right)+\frac{2}{3}\tan^{2}\left(\frac{\theta}{4}\right)\right|^2 + \mathcal{O}\left(\frac{\delta\epsilon_{\rm d}}{g}\right)^{10}\nonumber \\
P_{LE} = \left(\frac{\delta\epsilon_{\rm d}}{g}\right)^{6}\left|\frac{2}{3}\tan\left(\frac{\theta}{4}\right)-\frac{\theta}{4}+\frac{1}{6}\sin\left(\frac{\theta}{2}\right)\right|^2+ \mathcal{O}\left(\frac{\delta\epsilon_{\rm d}}{g}\right)^{10}.\end{gathered}$$
Analysis of errors for the Identity gate {#sec:identity}
----------------------------------------
Here we discuss the computational and leakage error probabilities of the Identity gate that we have defined as \[eq:Identity\] \_[I]{}=\^2. Taylor expanding $\mathcal{R}_{I}$ about $\delta\epsilon_{\rm d}=0$ gives \[eq:IdentityExpanded\] \_[I]{}=\_[33]{}-i(1+)(
[ccc]{} 0 & 1 & 0\
1 & 0 & 0\
0 & 0 & 0\
)()\^2-i(1+)(
[ccc]{} 0 & 0 &\
0 & 0 & 1\
& 1 & 0\
)()\^3+ ()\^[4]{}, where $\mathbb{1}_{3\times3}$ is the $3\times3$ identity matrix. It is interesting to note that no $O[\delta\epsilon_\text{d}/g]$ term occurs in this expansion, due to the special form of Eq. (\[eq:Identity\]). The $O[(\delta\epsilon_\text{d}/g)^2]$ term cannot be made to vanish because this would require $\epsilon_{\rm q}<0$ (since $g>0$); the condition $\epsilon_{\rm q}<0$ is unphysical here, since it would require a negative gate time. The resulting computational errors in $\mathcal{R}_{I}$ and $\mathcal{R}_{\rm zxz}$ are therefore both $O[\left(\delta\epsilon_{\rm d}/g\right)^{4}]$. For $\mathcal{R}_{I}$, we obtain the following leading-order results for the computational and leakage error probabilities: $$\begin{gathered}
P_{EC}=\left|{\left\langle E \right|}\mathcal{R}_{I}{\left|C\right\rangle}\right|^{2} = \left|{\left\langle C \right|}\mathcal{R}_{I}{\left|E\right\rangle}\right|^{2} = \left(\pi+\frac{4\pi{g}}{\epsilon_{\rm q}}\right)^2\left(\frac{\delta\epsilon_{\rm d}}{g}\right)^4+\mathcal{O}\left(\frac{\delta\epsilon_{\rm d}}{g}\right)^{8}, \nonumber \\
P_{LC}=\left|{\left\langle L \right|}\mathcal{R}_{I}{\left|C\right\rangle}\right|^{2} = \left(\frac{\pi{g}}{\epsilon_{\rm q}}\right)^{2}\left(\pi+\frac{4\pi{g}}{\epsilon_{\rm q}}\right)^2\left(\frac{\delta\epsilon_{\rm d}}{g}\right)^6 +\mathcal{O}\left(\frac{\delta\epsilon_{\rm d}}{g}\right)^{10} , \nonumber \\
P_{LE}=\left|{\left\langle L \right|}\mathcal{R}_{I}{\left|E\right\rangle}\right|^{2} = \left(\pi+\frac{4\pi{g}}{\epsilon_{\rm q}}\right)^2\left(\frac{\delta\epsilon_{\rm d}}{g}\right)^6+\mathcal{O}\left(\frac{\delta\epsilon_{\rm d}}{g}\right)^{10}.
\label{eq:PerrorRI}\end{gathered}$$ Here we note that the limits $\epsilon_\text{q}$ or $g\rightarrow 0$ appear to cause the error probabilities to diverge. However these limits would require infinite gate times in Eq. (\[eq:Identity\]). Therefore they do not correspond to valid gates.
Error function pulse profile {#sec:pulse}
----------------------------
Control pulses from an arbitrary waveform generator (AWG) are smoothed out, intentionally or unintentionally, before they reach the physical qubit inside the dilution refrigerator. In our numerical simulations, we model these smooth pulse shapes as error functions [@PhysRevA.87.022309]. If the initial and final times of the switch are $t_1$ and $t_2$, and the corresponding frequencies are $f_1$ and $f_2$, then we define the switching function as \[eq:erf\] f(t)=+, where, \[z\]=\_[-z]{}\^[z]{}e\^[-k\^[2]{}]{}dk. For typical experiments in quantum dots, the switching or rise time is of order $t_2-t_1=50$ ps, which is the value we use in our simulations.
Comparison between single and composite pulse approaches {#sec:errorComp}
--------------------------------------------------------
![[Comparison of the infidelities of $X_{-\pi/2}$ gates as a function of dipolar fluctuation $\delta\epsilon_{\rm d}$ for the composite- and single-pulse approaches. Black-filled circles show the data points that are numerically computed for optimal pulses corresponding to the given $\delta\epsilon_{\rm d}$ and $\delta\epsilon_{\rm q}$. The vertical solid black line indicates a typical value of $\delta\epsilon_{\rm d}$ that is consistent with recent experiments [@Mi_2017]. The dashed lines (note that the red dashed line falls on top of the red solid line) denote infidelity for $\delta\epsilon_{\rm q}=\delta\epsilon_{\rm d}/50$ and the solid lines denote the same for $\delta\epsilon_{\rm q}=0$. The plot demonstrates that the composite pulses yield a large decrease in infidelity over a broad range of amplitudes of the dipolar noise.]{}[]{data-label="fig:infidelityComp"}](errorComp.pdf){width="0.7\linewidth"}
[Here we compare our composite pulse approach against the single-pulse approach to designing quantum gates for CQ qubit under experimental conditions. In order to obtain a fair comparison, we optimize three pulse parameters for both approaches by varying $\delta\epsilon_{\rm d}$ over a range of experimentally realistic values. Fig. [\[fig:infidelityComp\]]{} shows the results of an infidelity calculation for $X_{-\pi/2}$ gates as a function of $\delta\epsilon_{\rm d}$, based on Eq. (10) of the main text. These results indicate that in the limit $\delta\epsilon_{\rm q} \rightarrow 0$, the infidelity obtained with our composite pulse sequence is orders of magnitude better than that obtained using an optimized pulse in the single-pulse approach. For $\delta\epsilon_{\rm q} > 0$, we find that the infidelity of the gate operation using the composite pulse sequence is $\delta\epsilon_{\rm q}$-limited, as indicated by the dashed blue line in Fig. [\[fig:infidelityComp\]]{}. However, as discussed in ref. [@Friesen:2016aa], the ratio $\delta\epsilon_{\rm q}/\delta\epsilon_{\rm d} \approx d/R$ can be very small, where $d$ is the interdot spacing and $R$ is the characteristic distance between the qubit and the charge fluctuators that cause $\delta\epsilon_{\rm d}$. Hence, $\delta\epsilon_{\rm q}$ can be systematically reduced with advanced fabrication techniques that, in turn, reduce the infidelity of quantum gates performed using our approach. This is one of the greatest advantages of using our composite pulse sequence for CQ qubits.]{}
[3]{}ifxundefined \[1\][ ifx[\#1]{} ]{}ifnum \[1\][ \#1firstoftwo secondoftwo ]{}ifx \[1\][ \#1firstoftwo secondoftwo ]{}““\#1””@noop \[0\][secondoftwo]{}sanitize@url \[0\][‘\
12‘\$12 ‘&12‘\#12‘12‘\_12‘%12]{}@startlink\[1\]@endlink\[0\]@bib@innerbibempty [****, ()](http://link.aps.org/doi/10.1103/PhysRevA.87.022309) [****, ()](\doibase
10.1063/1.4974536) [****, ()](https://www.nature.com/articles/ncomms15923)
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We study the tolerance to congestion failures in communication networks with scale-free topology. The traffic load carried by each damaged element in the network must be partly or totally redistributed among the remaining elements. Overloaded elements might fail on their turn, triggering the occurrence of failure cascades able to isolate large parts of the network. We find a critical traffic load above which the probability of massive traffic congestions destroying the network communication capabilities is finite.'
author:
- Yamir Moreno
- 'Romualdo Pastor-Satorras'
- 'Alexei V[á]{}zquez'
- Alessandro Vespignani
title: 'Critical load and congestion instabilities in scale-free networks'
---
Complex heterogeneous connectivity patterns have been recently identified in several natural and technological networks [@strogatz; @bara02; @doro]. The Internet and the World-Wide-Web (WWW) networks, where nodes represent routers or web pages and edges physical connections or hyper-links, appear to have a topology characterized by the presence of “hubs” with many connections to peripherical nodes. Empirical evidence recently collected shows that this distinctive feature finds its statistical characterization in the presence of heavy-tailed degree distributions [@bara00; @fal; @broder00; @gov00; @gcalda; @gov01; @caida; @romu01]. In the Internet, for instance, the statistical analysis reveals that the degree distribution $P(k)$, defined as the probability that any node has $k$ links to other nodes, is well approximated by a power-law behavior $P(k)\sim k^{-\gamma}$, with $\gamma\approx 2.2$ [@fal; @gov00; @romu01]. This makes the Internet a capital example of the recently identified class of scale-free (SF) networks [@bara02]. The statistical physics approach has been proved to be a valuable tool for the study of complex networks, and several interesting results concerning dynamical processes taking place on complex networks have been recently reported. In particular, the absence of the percolation [@newman00; @havlin01] and epidemic [@pv01a; @virusreview] thresholds in SF networks has a large impact because of its potential practical implications. The absence of the percolating threshold, indeed, prompts to an exceptional tolerance to random damages [@barabasi00]. This is a property that assumes a great importance in communication networks, guaranteeing the connectivity capabilities of the system.
Percolation properties of SF networks refer only to the static topological connectivity properties [@newman00; @havlin01]. On the other hand, in the Internet and other communications networks, many instabilities are due to traffic load congestions [@labo1; @labo2]. The traffic load carried on the failing nodes or connections is automatically diverted to alternative paths on the networks and instabilities can spread from node to node by an avalanche of traffic congestions and overloads. For instance, route flaps have led to the transient loss of connectivity for large portions of the Internet. These instabilities are thus of a dynamical nature and depend on how information is routed and distributed in the network. The models proposed so far, however, deal with regular structures [@jap1; @jap2; @watts02] and do not take into account the complex topology of SF networks.
In this paper, we propose a simple model aimed at the study of failure cascades generated by the redistribution of traffic load by congested links or nodes in SF networks. We find that the system behavior depends on the average traffic load imposed to the network. Above a critical value of the average traffic load, a single failure has a finite probability of triggering a congestion avalanche affecting a macroscopic part of the network. The present analysis thus reveals the existence of a transition from a free phase to a congested one as a function of the amount of traffic carried by the network. Contrary to what happens for the static percolation transition [@havlin01; @newman00], loaded SF networks exhibit a finite threshold above which the system can develop macroscopic instabilities with respect to small damages if we consider the dynamics of the traffic carried on top of them. The results provided here represent a first step towards a more complete modeling of traffic instabilities in real communication networks.
In order to include the degree fluctuations of SF networks we shall use in the following the Barabási-Albert model [@bar99]. This is a stochastic growth model in which one starts from a small number $m_0$ of nodes and at each time step a new node is introduced. The new node is connected preferentially to $m$ old ones (for the simulations we used $m=3$) with a probability $\Pi(k_i)=k_i/\sum_j k_j$ proportional to the degrees $k_i$ of the nodes. The repeated iteration of this scheme gives as a result a complex network with a topological structure characterized by a power-law degree distribution $P(k)=2m^2k^{-\gamma}$ with $\gamma=3$ and average degree $\langle k \rangle=2m$. In principle, one might also consider a more general class of complex networks with variable power-law degree distributions [@bara02; @doro].
To simulate the flow of data packets on SF networks, taking into account the load redistribution in case of damages, we need to specify the initial state of the network; i.e. the load of traffic flowing through each link. An estimate of such load, assuming that the routing takes place following the minimum path, is given by the total number of shortest paths between any two nodes in the network that pass through the node $i$. This magnitude is called betweenness or load [@new01; @goh01] and has been recently studied in SF networks. This property of the network can also be defined in terms of links. In real systems, however, the amount of traffic carried by each link is a fluctuating quantity that depends on many variables such as number of users, routing agreements, and available bandwidth. For this reason, we associate to each link connecting the nodes $i$ and $j$ of the network a load $\ell_{i,j}$ drawn from a probability distribution that specifies the initial traffic load of the system. For simplicity, we have considered a uniform distribution $U(\ell)$ for $0< \ell < 1$, taking the form $$U(\ell) = \left\{
\begin{array}{cll}
\frac{1}{2 \langle \ell \rangle }, & \ell \in [0, 2 \langle \ell \rangle ] &
\mathrm{if} \, \langle \ell \rangle \leq 0.5\\ &&\\
\frac{1}{2(1- \langle \ell \rangle) }, & \ell \in [2 \langle \ell \rangle -1 ,1
] & \mathrm{if} \,\langle \ell \rangle \geq 0.5
\end{array} \right..$$ This uniform distribution implies that the minimum initial load carried by a link is bounded by a nonzero value for an average load $\langle l\rangle > 0.5$, which means that there will be no links with load smaller than this lower bound. The results reported in this paper were obtained with the initial distribution $U(\ell)$. In order to test the universality of the critical behavior, we have also considered the distribution $F(\ell)= (\langle \ell \rangle^{-1} -1) (1-\ell)^{(\langle \ell
\rangle^{-1} -2)}$, $\ell \in [0, 1 ]$, which allows the existence of links with a very small load, irrespective of the average load $\langle\ell\rangle$ flowing through the system. Both distributions $U(\ell)$ and $F(\ell)$ yield the same qualitative results. Along with the load, we associate to each link the same capacity $C$ that, without loss of generality, we fix equal to one. This choice can be considered as a first approximation since the actual difference between line bandwidths in communication networks can be large. In this perspective, we consider as the most important source of heterogeneity the flow of different amounts of load through the network.
The dynamics of the model is defined by a simple threshold process. A link is selected at random and overloaded by raising its traffic. When the load carried by a link is $\ell_{i,j}>C$, i.e. when it exceeds the link’s capacity, the link is considered congested and the load it carries is diverted among its (not overloaded) neighboring links. This amounts to consider that the time scale of the local congestion is greater than the time scale characterizing the reorganization of the routing procedure. The redistribution of the load on its turn might provoke that other links become overloaded, thus triggering a cascade of failures. We have explored two physically different settings of the load redistribution rule. The first consists on equally distributing the load of a congested link among the non-congested neighboring links. We refer to it as the deterministic redistribution rule, respectively. The second case will be called random redistribution because when a link is overloaded, a random amount of load is redistributed to each of the remaining working links in its neighborhood. Finally we note that in the rare event in which the congested link has no active neighbors, its load can be equally shared among all the remaining working lines of the network or just be considered as lost from the network. This amounts to a conserved or dissipative redistribution rule. Many physical systems display criticality only when energy is conserved [@jen98; @marro]. In distributed networks such as the Internet, however, it is common to discard packets if there is not a route available at the moment. As we shall see in the following, the results do not depend qualitatively on the conserved nature of the traffic load.
We have performed large-scale numerical simulations by applying repeatedly the rules stated above on BA networks. The sizes of the networks used in the simulations range from $N=5 \times 10^3$ nodes ($15 \times
10^3$ links) to $N=10^5$ nodes ($3 \times 10^5$ links). All numerical results have been obtained by averaging over 10 different networks and, at least, 100 different realizations of the initial load distribution.
In order to inspect the occurrence of dynamic instabilities, we construct the phase diagram of the system. The order parameter can be identified as the probability $P_\mathcal{G}$ of having a giant component $\mathcal{G}$ of connected nodes with size of the order of the network size. The giant component is defined as the largest component of the network made by nodes connected by active links, after the system has reached a stable state (when $\ell_{i,j} < C$ for all $i$ and $j$). The existence of a giant component implies that a macroscopic part of the network is still functional. If the giant component of the network is zero, the communication capabilities of the network are destroyed and a congestion of the order of the system size builds up. It is worth noticing that although we have determined the giant component size in terms of nodes and the dynamical rules of the model are expressed in terms of links, the results are completely equivalent since a connected node is defined as a node with at least one active link.
In Fig. \[figure1\] we plot the order parameter $P_\mathcal{G}$ as a function of the average load $\langle\ell\rangle$. At low values of the average load, the network always reaches a stable state in which the number of isolated nodes is very small and with probability $P_\mathcal{G}=1$ the network has a giant component of connected nodes of the order of the system size. When increasing the load imposed on the network, the system starts to develop instabilities. In particular, above a critical load $\langle\ell\rangle_c^I$, whose value depends on the model considered, with a finite probability the system evolves to a congested state without giant component of connected nodes; i.e. the largest set of connected active nodes has a density of order $N^{-1}$. This implies a probability of having a giant component $P_\mathcal{G}<1$, which is decreasing as the load is progressively increased. At an average load $\langle\ell\rangle_c^{II}\simeq 0.82$ we get that $P_\mathcal{G}=0$, signalling that, with probability one, any instability will propagate until the complete fragmentation of the network. It is worth remarking that this scenario is rather different from the percolation one in which the probability of having a giant component is abruptly dropping from one to zero at the transition point. Here, the probability decays continuously to zero and we have a wide region of $\langle\ell\rangle$ where the initial instability can trigger a destructive congestion with probability $1-P_\mathcal{G}$. Fig. \[figure2\] illustrates the probability $p(S)$ that the isolated network has a size $S$ in the case that no giant component of connected nodes has survived. The distribution is rather peaked also at relatively small values of $\langle \ell \rangle$, almost affecting the totality of the network.
The phase diagram obtained in Fig. \[figure1\] points out that the value of the average load at which $P_\mathcal{G}=0$ is relatively high. On the other hand, the value at which $P_\mathcal{G}$ is appreciably smaller than one is well below the theoretical capacity of the network measured as the capacity $C=1$ of the individual links ($\langle\ell\rangle_c^I\simeq 0.15 - 0.3$, see Fig. \[figure1\]). This evidence defines a wide region of load values in which there is a small but finite probability that a small instability propagates through the system and could explain why in real communication networks congestions of diverse degree can be encountered from time to time. In Fig. \[figure1\] we also report curves obtained for the different dynamical rules defined previously and it is interesting to remark that the highest level of stability has been obtained for the deterministic and conservative version of the model ($\langle\ell\rangle_c^I\simeq
0.3$). Moreover, the dynamics of the transition does not depend on the distribution of initial loads and the network size $N$.
Another way to shed light on the congestion dynamics is to inspect the process of generation of the congestion instability. We define the size $s$ of a congestion burst or avalanche as the total number of simultaneously overloaded links. The cumulative distribution $P(s)$ of avalanches of size larger than $s$ for several values of the average load imposed on the network and four different system sizes have been plotted in Fig. \[figure3\] for the random and dissipative definition of the present model. The main plot is in log-linear scale, so that a straight line corresponds to a power law of the form $p(s) \sim s^{-1}$ for the probability of observing an avalanche of size $s$. Power-laws with exponent $-1$ have been found for several characteristic features of Internet traffic such as latency times, queue lengths, and congestion lengths [@jap1; @mag; @will]. In the figure we focus on region close to the stable region $\langle \ell \rangle\simeq
0.20$, which means that the power-law behavior extends to values far from the instability transition. This fact confirms that it is not necessary that the network operates very close to a critical point in order to observe power-laws in the distribution of several quantities. The inset in Fig. \[figure3\] shows that the cumulative size of overloaded links also scales with the system size, the scaling dynamics, however, remaining the same. This may help understand why power-law distributions observed in real communication networks have been measured for different network sizes, i.e. both for local networks and for networks that extend to a very large scale.
In summary, we have introduced a simple threshold model aimed at the description of instabilities due to load congestion that takes into account the topological properties of SF networks. The results obtained point out that the network can freely handle traffic up to some critical average load $\langle\ell\rangle_c^I$. Above this level the network faces partial congestions that start to build up local bottlenecks in various places and small instabilities might trigger macroscopic outages with a finite probability. Above a critical load value $\langle\ell\rangle_c^{II}\simeq 0.82$ any small instability leads to the whole network collapse. In the intermediate region of network load, the number of simultaneous line casualties follows a power-law resembling what has been observed in experimental studies of the Internet. We hope that our work will provide hints for accurate modeling of the Internet and the WWW large-scale traffic behavior.
This work has been partially supported by the European Commission - Fet Open project COSIN IST-2001-33555. R.P.-S. acknowledges financial support from the Ministerio de Ciencia y Tecnología (Spain).
[99]{}
S. H. Strogatz, Nature (London) [**410**]{}, 268 (2001).
R. Albert and A.-L. Barabási, Rev. Mod. Phys. [**74**]{}, 47 (2002). S. N. Dorogovtsev and J. F. F. Mendes, Adv. Phys. [**51**]{}, 1079 (2002). A.-L. Barabási, R. Albert and H. Jeong, Physica A [**281**]{}, 69 (2000).
M. Faloutsos, P. Faloutsos, and C. Faloutsos, Comp. Com. Rev. [**29**]{}, 251 (2000).
A. Broder, R. Kumar, F. Maghoul, P. Raghavan, S. Rajagopalan, R. Stata, A. Tomkins and J. Wiener, Comp. Networks [**33**]{}, 309 (2000).
R. Govindan and H. Tangmunarunkit, in [*Proceedings of IEEE INFOCOM 2000*]{}, Tel Aviv, Israel (Ieee, Piscataway, N.J. 2000).
G. Caldarelli, R. Marchetti and L. Pietronero, Europhys. Lett. [**52**]{}, 386 (2000).
Q. Chen, H. Chang, R. Govindan, S. Jamin, S.J. Shenker and W. Willinger, INFOCOM 2002. Twenty-First Annual Joint Conference of the IEEE Computer and Communications Societies. Proceedings. IEEE , Volume: 2 , 2002
A. Broido and K. C. Claffy, in SPIE International symposium on Convergence of IT and Communication, (Denver, CO, 2001).
R.Pastor-Satorras, A. Vázquez, and A. Vespignani, Phys. Rev. Lett. [**87**]{}, 258701 (2001); A. Vázquez, R.Pastor-Satorras, and A. Vespignani, Phys. Rev. E [**65**]{}, 066130 (2002).
R. Cohen, K. Erez, D. ben-Avraham, and S. Havlin, [*Phys. Rev. Lett.*]{} [**85**]{}, 4626 (2000);
D. S. Callaway, M. E. J. Newman, S. H. Strogatz, and D. J. Watts, [*Phys. Rev. Lett.*]{} [**85**]{}, 5468 (2000).
R. Pastor-Satorras, and A. Vespignani, Phys. Rev. Lett. [**86**]{}, 3200 (2001).
R. Pastor-Satorras, and A. Vespignani, [*Handbook of Graphs and Networks: From the Genome to the Internet*]{}, eds. S. Bornholdt and H.G. Schuster (Wiley-VCH, Berlin, 2002).
R. Albert, H. Jeong, and A.-L. Barabási, Nature [**406**]{}, 378 (2000).
C. Labovitz, G.R. Malan and F. Jahanian, INFOCOM ’99. Eighteenth Annual Joint Conference of the IEEE Computer and Communications Societies. Proceedings. IEEE, Volume: 1, (1999).
C. Labovitz, A. Ahuja, F. Jahanian, Fault-Tolerant Computing, 1999. Digest of Papers. Twenty-Ninth Annual International Symposium on , 278, (1999)
M. Takayasu, H. Takayasu, and T. Sato, Physica A [**233**]{}, 924 (1996); A. Y. Tretyakov, H. Takayasu, and M. Takayasu, Physica A [**253**]{}, 315 (1998).
T. Ohira, R. Sawatari, Phys. Rev. E [**58**]{}, 193 (1998); S. Valverde, and R. V. Solé, Physica A, to appear (2002).
D.J. Watts, Proc. Nat. Acad. Sci. USA,[**99**]{}, 5766 (2002).
A.-L. Barabási, and R. Albert, Science [**286**]{}, 509 (1999); A.-L. Barabási, R. Albert, and H. Jeong, Physica A [**272**]{}, 173 (1999).
M. J. Newman, Phys. Rev. E. [**64**]{}, 016132 (2001).
K.-I. Goh, B. Kahng, and D. Kim, Phys. Rev. Lett. [**87**]{}, 278701 (2001).
H. J. Jensen, [*Self-Organized Criticality*]{} (Cambridge University Press, Cambridge, England, 1998) and references therein.
J. Marro and R. Dickman, [*Nonequilibrium phase transitions in lattice models*]{} (Cambridge University Press, Cambridge, 1999).
M. O. Magnasco, [*The thunder of distant Net storms*]{}, preprint nlin.AO/0010051 (2000).
W. Willinger, R. Govindan, S. Jamin, V. Paxson and S. Shenker, Proc. Nat. Acad. Sci. USA [**99**]{} 2573 (2002).
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We present a list of optimized damping range parameters $s_R$ to be used with the Tkatchenko-Scheffler van der Waals dispersion-correction scheme \[Phys. Rev. Lett. 102, 073005 (2009)\]. The optimal $s_R$ are obtained for seven popular generalized-gradient approximation exchange-correlation density functionals: PBE, RPBE, revPBE, PBEsol, BLYP, AM05 and PW91. The optimization is carried out in the standard way by minimizing the mean absolute error of the S22 test set, where the reference interaction energies are taken from coupled-cluster calculations. With the optimized range parameters, we assess the impact of van der Waals corrections on the ability of these functionals to accurately describe structural and thermodynamic properties of liquid water: radial distribution functions, self-diffusion coefficients and standard molar entropies.'
author:
- 'Miguel A. Caro'
title: 'Parametrization of the Tkatchenko-Scheffler dispersion correction scheme for popular exchange-correlation density functionals: effect on the description of liquid water'
---
Introduction
============
The Tkatchenko-Scheffler (TS) method has emerged in recent years as one of the most popular dispersion-correction schemes in density functional theory (DFT) calculations [@tkatchenko_2009]. The reason is that TS is a post-processing scheme, and can therefore be applied on top of DFT electron densities without significantly adding to the overall execution time of the different algorithms, where by far the most expensive part corresponds to the self-consistent calculation of the Kohn-Sham orbitals. It was shown by Marom [*et al*.]{} [@marom_2011] that TS performs better than similar approaches for a wide variety of weekly interacting molecular systems.
The TS approach involves the optimization of an adjustable parameter, $s_R$, which determines the range at which the dispersion interactions begin to become important. The value of $s_R$ depends on the intrinsic ability of the underlying exchange-correlation (XC) density functional to correctly describe van der Waals interactions. An optimal value of $s_R$ is therefore functional-dependent. Tkatchenko and Scheffler originally optimized $s_R$ [@tkatchenko_2009] for the PBE functional [@perdew_1996] by minimizing the error in the interaction energies predicted with the PBE+TS approach compared with highly accurate coupled-cluster calculations \[CCSD(T)\] of the S22 test set by Jurečka [*et al*.]{} [@jurecka_2007]. They obtained $s_R = 0.94$. Later on, Marom [[*et al*.]{}]{} optimized $s_R$ for a number of hybrids and meta-generalized-gradient approximation (meta-GGA) XC functionals [@marom_2011]. Agrawal [[*et al*.]{}]{} carried out follow-up work on the performance of TS corrections for hybrid functionals, including the role of the range-separation screening parameter, finding it to have little correlation with the optimization of $s_R$ [@agrawal_2013]. However, optimal $s_R$ have not yet been estimated for many popular GGA functionals, which are routinely used preferentially over more advanced functionals because of computational advantages. $s_R$ is not transferable across XC functionals, and using for instance the PBE value with a different GGA may lead to unacceptable errors in the calculated dispersion corrections.
Optimization of $s_R$ for several XC functionals
================================================
In this paper we optimize $s_R$ for some of the GGAs most widely used across the computational chemistry and physics communities: PBE (for benchmark with previous results), RPBE [@hammer_1999], revPBE [@zhang_1998], PBEsol [@perdew_2008], BLYP [@becke_1988; @lee_1988; @gill_1992], AM05 [@armiento_2005] and PW91 [@perdew_1992]. We use the GPAW DFT suite [@enkovaara_2010] in conjunction with the Atomic Simulation Environment (ASE) [@larsen_2017], which offer an extremely flexible Python-based environment to carry out the present calculations (a sample Python script is presented at the end of this paper that allows to optimize $s_R$ for any XC functional available from the LibXC library [@marques_2012] and compatible with GPAW). We used the GPAW grid mode with 0.18 [Å]{} spacing and PAW potentials [@bloechl_1994; @kresse_1999]. We generated the PAW setups for the PBEsol, BLYP, AM05 and PW91 functionals using the GPAW setup generation tool. The DFT calculations were carried out in fixed boundary conditions within orthogonal boxes where at least 4 [Å]{} of vacuum was allowed between the atoms and the box boundaries.
![Mean signed difference (MSD) and mean absolute error (MAE) of the TS method applied to different GGAs over a wide range of $s_R$ values.[]{data-label="01"}](01.pdf)
![Fine detail of [Fig. \[01\]]{} in the region where the errors are minimized.[]{data-label="02"}](02.pdf)
XC functional $s_R$ MAE (meV)
--------------- ------- -----------
PBE 0.940 16
RPBE 0.590 26
revPBE 0.585 29
PBEsol 1.055 30
BLYP 0.625 25
AM05 0.840 15
PW91 0.965 20
: Optimal $s_R$ fitted for each XC functional (resolved to nearest 0.005) and MAE calculated for the S22 test set. The calculations for each functional were run at the optimal value of $s_R$ for that functional.
\[03\]
The interaction energies are defined as the difference between the total energy of the interacting system (1+2) and the sum of the energies of each of the two isolated molecules (1 and 2): $$\begin{aligned}
E_i^\text{int} = E_i^{1+2} - E_i^1 - E_i^2,\end{aligned}$$ where $i$ labels each of the 22 molecular systems in the S22 set. The mean absolute error (MAE) and mean signed difference (MSD) between the interaction energies calculated with GPAW and those from the CCSD(T) S22 test set computed for each XC functional over a wide range of $s_R$ values are shown in [Fig. \[01\]]{}. One can observe the same overall trends regardless of the GGA used: the error increases rapidly as the value of $s_R$ is reduced and slowly as it is increased. It can also be observed how the MAE for each functional has its minimum at a different position, emphasizing the fact that $s_R$ values are not transferable across XC functionals and must be carefully optimized on an individual basis.
Figure \[02\] shows a detailed view of MSD and MAE in the regions where they are zero and minimum, respectively. Table \[03\] summarizes the optimal $s_R$ parameter obtained for each XC functional studied and the MAE computed at said value of $s_R$. We obtain the same value for PBE as has been previously reported [@tkatchenko_2009; @marom_2011]. As discussed by Marom [*et al*.]{} [@marom_2011], a low optimal $s_R$ indicates that the underlying XC functional does a poor job at handling dispersion interactions on its own. From [Fig. \[02\]]{} and Table \[03\] we observe how revPBE, RPBE and BLYP cluster around the low $s_R$ region, while AM05, PBE and PW91 cluster at intermediate values and PBEsol lies slightly beyond them. We also observe that AM05 and PBE see their MAE reduced to a very small value ($\sim 15$ meV) with the inclusion of TS dispersion corrections, while revPBE and PBEsol retain larger errors of up to 30 meV even with the optimal $s_R$.
To get deeper insight into the different contributions to the overall errors, [Fig. \[04\]]{} shows the signed errors (SD) for each molecule in the S22 set and each functional, where the TS corrections are computed at the optimal $s_R$ value for each XC functional. Note from the figure that the description of dispersion is not at all homogeneous across molecules. For instance, PBEsol+TS, which performs worst overall, does a very good job at describing many of the benzene-containing systems. We also note how several functionals give almost identical results to each other, for instance PW91 and PBE show almost identical curves. The same is true for RPBE and revPBE.
Based on these observations, the possibility to re-optimize $s_R$ for a very specific problem by fitting to a more suitable test set remains viable. For instance, if one is interested in achieving an accurate description of a protein, $s_R$ could be optimized to describe the dispersion interactions between a set containing different aminoacids and small organic molecules. The Python script in the Appendix automatizes the calculation of dispersion errors presented in this paper, and can be used to optimize $s_R$ for a different XC functional or test set (note that one needs to generate the corresponding PAW setups needed by GPAW beforehand if they are not installed by default).
Application of dispersion corrections to describe liquid water
==============================================================
To test the possible improvement over the underlying functionals brought about by the inclusion of dispersion corrections, we looked at molecular dynamics (MD) simulations of liquid water. A more exhaustive assessment of the ability of DFT to correctly describe the properties of water is available from the recent excellent review by Gillan, Alfè and Michaelides [@gillan_2016]; here we are more concerned with identifying systematic improvements introduced by TS dispersion corrections. A detailed study along the same lines was conducted for the BLYP, PBE and revPBE functionals [@lin_2012], although hydrogen nuclei were replaced by deuterium nuclei in that study. This has the observed effect of decreasing the self-diffusion constant of the liquid. In this study, the performance of the different GGAs for predicting structural and thermodynamic properties of water at ambient conditions (300 K) and experimental density (1 g/cm$^3$) was studied with and without TS corrections. The properties that we looked at are: self-diffusion coefficient, radial distribution functions and standard molar entropy. We used periodic supercells containing 100 water molecules, pre-equilibrated with a classical potential and the Gromacs code [@abraham_2015]. We used the SPCE water model [@berendsen_1987] (allowing for vibrations of the O-H intramolecular bonds) with the OPLS force field [@jorgensen_2005]. After this, 35 ps of *ab initio* MD (AIMD) followed. The AIMD part was performed with the plane-wave based DFT code VASP [@kresse_1996]. Since PAW potentials for VASP are only supplied for the PBE and LDA functionals, we carried out the AIMD using PBE PAW potentials for all the GGAs studied here. We used 0.5 fs as time step to correctly resolve hydrogen vibrations, and set the kinetic energy cutoff for the plane-wave basis set to 300 eV. The first 15 ps of dynamics for each simulation were discarded and the last 20 ps were analyzed with our own implementation [@caro_2016; @caro_2017b] of the 2PT method [@berens_1983; @lin_2003; @lin_2010], DoSPT [^1]. Unfortunately, we could not compute the AM05 values due to convergence problems in the MD runs. The values for all the other GGAs tested and reference experimental and classical MD values are given in Table \[05\]. Radial distribution functions are shown in [Fig. \[06\]]{} and detailed information regarding RDF peak positions is given in Table \[07\].
XC functional $D$ ($10^{-9}$ m$^2$/s) $S_0$ (J/mol/K)
---------------------- ------------------------- -----------------
PBE 0.57 43.2
PBE+TS 1.09 47.8
RPBE 1.61 54.4
RPBE+TS 1.09 49.1
revPBE 1.28 51.5
revPBE+TS 0.66 45.5
PBEsol 0.40 39.1
PBEsol+TS 0.49 40.8
BLYP 2.16 58.0
BLYP+TS 1.59 54.7
PW91 0.64 42.8
PW91+TS 0.92 45.0
OPLS+SPCE (rigid) 2.72 58.8
OPLS+SPCE (flexible) 1.54 53.3
Experiment 2.41 69.92
: Standard molar entropies $S_0$ and self-diffusion coefficients $D$ calculated for different GGAs with and without TS dispersion corrections at $T=300$ K. Experimental and classical MD results are shown for comparison.
\[05\]
In Ref. [@gillan_2016], uncorrected RPBE was observed to offer the best description of liquid water among the tested GGAs. Here, we observe the same behavior for RPBE but obtain best results in all the categories, including entropy, which was not surveyed in Ref. [@gillan_2016], for BLYP. The main difference between the sources quoted in Ref. [@gillan_2016] and this work is that all MD here were carried out with regular hydrogen nuclei (protium) rather than deuterium. We also ran all the simulations at the experimental density. Heavy water has a lower self-diffusion constant than regular water, and a quick test that we ran with the classical potential we employed for pre-equilibration (SPCE+OPLS) showed that the self-diffusion constant decreases by 40% when using deuterium nuclei instead of protium nuclei.
The revPBE functional shows very similar performance to the related RPBE functional. Surprisingly, the inclusion of TS dispersion corrections worsens the agreement with experiment for both of those functionals and BLYP, while it improves it for the others. PBEsol emerges as the absolute loser in this comparison. The bad performance of PBEsol does not only affect the quantitative description of the properties of liquid water, but it also gives a qualitatively bad behavior of the water molecules during the MD. We observed significant spontaneous hydroxyl and hydronium ion formation with PBEsol: 4% of the oxygens were observed to be part of either a hydroxyl or a hydronium group at any time of the dynamics. This can be graphically observed as the lack of a clear minimum in the O-H radial distribution function in [Fig. \[06\]]{}. Therefore, we conclude that the PBEsol functional should be avoided for any simulation that involves water molecules, such as simulations of solid/water interfaces, in order to prevent unphysical results.
XC functional $r^{(1)}_\text{max}$ (Å) $g^{(1)}_\text{max}$ $r^{(1)}_\text{min}$ (Å) $g^{(1)}_\text{min}$ $r^{(2)}_\text{max}$ (Å) $g^{(2)}_\text{max}$ $r^{(2)}_\text{min}$ (Å) $g^{(2)}_\text{min}$
-------------------------- -------------------------- ---------------------- -------------------------- ---------------------- -------------------------- ---------------------- -------------------------- ----------------------
PBE 1.59 2.59 1.80 0.17 2.20 1.88 2.88 0.37
PBE+TS 1.59 2.64 1.80 0.14 2.23 1.69 2.85 0.48
RPBE 1.59 2.83 1.80 0.05 2.31 1.45 2.92 0.58
RPBE+TS 1.59 2.72 1.80 0.14 2.27 1.56 2.92 0.51
revPBE 1.59 2.80 1.80 0.07 2.27 1.52 2.92 0.54
revPBE+TS 1.59 2.66 1.80 0.17 2.23 1.64 2.85 0.47
PBEsol 1.62 2.14 1.80 0.70 2.13 2.00 2.85 0.37
PBEsol+TS 1.62 2.15 1.84 0.65 2.13 1.99 2.81 0.35
BLYP 1.59 3.07 1.80 0.08 2.27 1.41 2.92 0.66
BLYP+TS 1.59 2.94 1.80 0.11 2.23 1.49 2.92 0.64
PW91 1.59 2.47 1.80 0.20 2.20 1.89 2.88 0.37
PW91+TS 1.59 2.57 1.80 0.17 2.20 1.85 2.88 0.40
Experiment [@soper_2013] n/a n/a n/a n/a 2.43 1.34 2.97 0.69
PBE 1.01 21.03 1.22 0.01 1.62 1.87 2.49 0.03
PBE+TS 1.01 20.88 1.22 0.01 1.66 1.62 2.49 0.07
RPBE 0.97 22.19 1.22 0.00 1.80 1.19 2.49 0.11
RPBE+TS 0.97 21.20 1.22 0.00 1.84 1.67 2.41 0.07
revPBE 0.97 21.44 1.26 0.00 1.73 1.32 2.52 0.09
revPBE+TS 1.01 21.26 1.22 0.00 1.80 1.81 2.41 0.04
PBEsol 1.01 18.00 1.26 0.33 1.55 2.18 2.49 0.03
PBEsol+TS 1.01 18.23 1.26 0.29 1.55 2.12 2.41 0.03
BLYP 0.97 24.51 1.22 0.00 1.77 1.12 2.45 0.17
BLYP+TS 0.97 23.55 1.22 0.00 1.84 1.11 2.41 0.16
PW91 1.01 20.72 1.22 0.01 1.62 1.88 2.49 0.03
PW91+TS 1.01 20.81 1.22 0.01 1.62 1.84 2.49 0.04
Experiment [@soper_2013] n/a n/a n/a n/a 1.86 1.04 2.46 0.21
PBE 2.63 3.46 3.28 0.20 4.36 1.65 5.51 0.78
PBE+TS 2.63 3.05 3.28 0.37 4.32 1.46 5.58 0.78
RPBE 2.77 2.32 3.39 0.60 4.36 1.37 5.58 0.75
RPBE+TS 2.77 3.18 3.28 0.43 4.43 1.34 5.48 0.81
revPBE 2.70 2.56 3.35 0.49 4.40 1.45 5.69 0.74
revPBE+TS 2.77 3.35 3.24 0.33 4.50 1.40 5.51 0.78
PBEsol 2.56 4.13 3.24 0.18 4.29 1.64 5.55 0.75
PBEsol+TS 2.56 4.07 3.24 0.20 4.29 1.59 5.22 0.83
BLYP 2.74 2.20 3.42 0.73 4.36 1.22 5.58 0.85
BLYP+TS 2.81 2.38 3.42 0.62 4.36 1.19 5.40 0.90
PW91 2.63 3.56 3.28 0.20 4.43 1.64 5.66 0.76
PW91+TS 2.63 3.41 3.28 0.24 4.32 1.57 5.40 0.75
Experiment [@soper_2013] 2.79 2.50 3.39 0.80 4.53 1.12 5.58 0.88
\[07\]
The extremely popular PBE benefits from dispersion corrections by improved description of all the studied properties. However, the water overstructuring typical of the PBE functional cannot be completely suppressed with the inclusion of van der Waals corrections. A popular strategy, often found in the literature to tackle this issue, is to perform MD with PBE at high temperature. The also very popular BLYP functional offers by far the best description of water, in all the categories, in the absence of dispersion corrections. When adding TS corrections the description is slightly worsened, in particular the value of the self-diffusion coefficient. Unfortunately, BLYP is known to offer poor description of metallic systems [@stroppa_2008], which may limit its wider applicability.
All the functionals underestimate standard molar entropies by large factors, which vary between a best-case scenario of 17% (BLYP) and worst-case scenario of 44% (PBEsol). Similarly, the self-diffusion coefficients are also underestimated by all the tested GGAs, varying between 10% (BLYP) and 83% (PBEsol). Unfortunately, it is likely that a fully satisfactory description of liquid water cannot be achieved with any XC functional at this level of theory, although BLYP offers very good performance. With respect to dispersion corrections, it seems that the description of water is worsened by adding them for those functionals which require largest corrections, i.e., functionals which require small $s_R$ values: RPBE, revPBE and BLYP. The description of liquid water is improved by including dispersion corrections for the other functionals: PBE, PW91 and PBEsol.
Conclusion
==========
In conclusion, we have provided a list of optimal range parameters $s_R$ for different XC functionals based on minimization of the MAE in the prediction of dispersion interactions of the S22 test set. These parameters allow computations of dispersion corrections to DFT based on the Tkatchenko-Scheffler method for the following GGAs: PBE, RPBE, revPBE, PBEsol, BLYP, AM05 and PW91. We have tested the effect of dispersion corrections on the description of liquid water offered by those functionals (except for AM05), finding no systematic improvement, i.e., the improvement is strongly functional-dependent. Based on our simulations, we recommend either BLYP for organic systems (due to its limitations to correctly describe metals), or RPBE more generally, as affordable options to carry out atomistic studies of systems including water. When van der Waals interactions are expected to play a significant role, for instance in the presence of solvated organic molecules and adsorption phenomena, BLYP+TS and RPBE+TS can be good options, although the improvement over other GGAs is not so obvious as for the uncorrected cases. PBEsol was observed to perform particularly badly for liquid water, and we strongly advice against using PBEsol to simulate any system including an aqueous phase.
Computational resources for this project were provided by CSC–IT Center for Science through the Taito supercluster. The author would like to thank Michael Walter for useful comments and email correspondence, as well as being an author of the GPAW implementation of the TS scheme.
**Code for $s_R$ optimization**
===============================
The following Python script can be used to optimize the $s_R$ parameter for a given XC functional (or list of functionals) with the GPAW code:
\# This code was adapted by M.A. Caro from Michael Walter’s TS documentation on the GPAW website from \_\_future\_\_ import print\_function from ase import Atoms from ase.parallel import paropen from ase.data.s22 import data,s22 from ase.calculators.vdwcorrection import vdWTkatchenko09prl from gpaw import GPAW, FermiDirac from gpaw.cluster import Cluster from gpaw.analyse.hirshfeld import HirshfeldPartitioning from gpaw.analyse.vdwradii import vdWradii import numpy as np
h = 0.18; box = 4.
sR\_range = np.arange(0.88,1.02,0.02) xc\_list = \[“PBE”\]
for xc in xc\_list: f = paropen(’dispersion\_energies\_ disp\_en = for molecule in s22: disp\_en\[molecule\] = \[\] ss = Cluster(Atoms(data\[molecule\]\[’symbols’\], data\[molecule\]\[’positions’\])) \# Split interacting system into the separate molecules s1 = ss.find\_connected(0) s2 = ss.find\_connected(-1) assert(len(ss) == len(s1) + len(s2)) c = GPAW(xc=xc, h=h, nbands=-6, occupations=FermiDirac(width=0.1)) for s in \[s1, s2, ss\]: s.set\_calculator(c) s.minimal\_box(box, h=h) s.get\_potential\_energy() E = for sR in sR\_range: cc = vdWTkatchenko09prl(HirshfeldPartitioning(c), vdWradii(s.get\_chemical\_symbols(), xc)) cc.sR = sR s.set\_calculator(cc) E\[sR\] = s.get\_potential\_energy() disp\_en\[molecule\].append(E) \# Print for sR in sR\_range: print(“\# sR = for molecule in s22: ref = data\[molecule\]\[”interaction energy CC“\] E1 = disp\_en\[molecule\]\[0\]\[sR\] E2 = disp\_en\[molecule\]\[1\]\[sR\] E12 = disp\_en\[molecule\]\[2\]\[sR\] \# Print molecule name, predicted interaction energy and reference CCSD(T) energy print(” print(“”, file=f) f.close()
[34]{}ifxundefined \[1\][ ifx[\#1]{} ]{}ifnum \[1\][ \#1firstoftwo secondoftwo ]{}ifx \[1\][ \#1firstoftwo secondoftwo ]{}““\#1””@noop \[0\][secondoftwo]{}sanitize@url \[0\][‘\
12‘\$12 ‘&12‘\#12‘12‘\_12‘%12]{}@startlink\[1\]@endlink\[0\]@bib@innerbibempty @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [ ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [ ]{} @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} @noop [****, ()]{} and , eds., @noop [**]{} () @noop [****, ()]{} @noop [****, ()]{}
[^1]: [http://dospt.org]{}
| {
"pile_set_name": "ArXiv"
} |
---
address: |
$^{1}$Laboratoire de Physique des Solides, Université Paris-Sud, Bât 510, 91 405 Orsay (France)\
$^{2}$Laboratoire de Physique des Solides Semicristallins,\
IGA-Département de Physique, Ecole Polytechnique Fédérale de Lausanne, 1015 Lausanne (Switzerland)
author:
- 'V. Brouet$^{1}$, H. Alloul$^{1}$, T.N. Le$^{2}$, S. Garaj$^{2}$ and L. Forr[ó]{}$^{2}$'
title: |
Role of dynamic Jahn-Teller distortions in Na$_{2}$C$_{60}$ and Na$_{2}$CsC$%
_{60}$ studied by NMR
---
It has been known, almost since the discovery of alkali doped fullerenes, that A$_{4}$C$_{60}$ is insulating, while A$_{3}$C$_{60}$ is metallic and superconducting [@Haddon]. This contrasts with expectations for a rigid band filling model, as the C$_{60}$ lowest unoccupied molecular t$_{1u}$ level should form a triply degenerate band. All A$_{n}$C$_{60}$ with n6 should either be metals if the strength of the Coulomb repulsion U is small compared to the band width W or Mott insulators if U is larger. In fullerides, U/W is close to the critical ratio where a metal-insulator transition is expected [@GunnarssonPRB96], but the observation of both metals and insulators within this family is puzzling.
An attractive explanation [@Heritier; @Manini; @Gunnarsson] is that the presence of Jahn-Teller distortions (JTD) could create [*effective electronic interactions* ]{}which would modulate U, so that different compounds could be on different sides of the metal-insulator transition. More precisely, the computation of U$_{eff}$ = E(n+1)+E(n-1)-2 E(n), where n is the number of electrons per C$_{60},\ $supports this idea when the larger gain of electronic energy associated with the JTD of an evenly charged C$%
_{60}$ ball is taken into account [@Heritier; @Manini; @Gunnarsson]. It follows that in even electron systems JT effects [*add*]{} to the Coulomb repulsion to localize electrons, whereas for odd electrons systems they [*oppose* ]{}the Coulomb repulsion and favor delocalization. Hence, A$_{2n}$C$%
_{60}$ could be insulating whereas A$_{2n+1}$C$_{60}$ could become metallic despite the strong electronic correlations. To give some experimental support to this idea, a large amount of work has been devoted to the search of JTD in fullerides, but without success up to now. However, the distortion is expected to be very small and possibly dynamic, which makes it difficult to detect directly. Alternatively, recent models have suggested that the different properties between A$_{4}$C$_{60}$ (body centered tetragonal) and A$_{3}$C$_{60}$ (face cubic centered) are due to their different structures [@GunnarssonA4].
In order to sort out the relevant parameters for the physics of A$_{n}$C$%
_{60}$, we present here an NMR study of Na$_{2}$C$_{60}$ and Na$_{2}$CsC$%
_{60}$, as they have the same[* cubic*]{} structure but [*even*]{} and [*odd*]{} stoichiometries. Na$_{2}$C$_{60}$ is the only compound with n=2 known so far, but experimental studies are still limited and controverted. The first ESR studies [@Petit] concluded that its electronic properties were identical to K$_{4}$C$_{60}$ but more recently, another ESR investigation claimed that it was metallic with a metal-insulator transition at 50 K [@Kubozono]. We present the first NMR study of Na$_{2}$C$_{60}$, which allows to detect singlet-triplet excitations of JTD C$_{60}$ balls, as found in A$%
_{4}$C$_{60}$ [@ZimmerK4; @Kerkoud; @ZimmerPRB95]. This rules out a strong dependence of the electronic properties on the structure, as one expects similar properties for n = 2 and 4 due to electron-hole symmetry in the t$%
_{1u}$ band. We then extend our investigation to Na$_{2}$CsC$_{60},$ which superconducts below $T_{c}=12$ $K$ [@Tanigaki]. In the metallic phase, we evidence an anomalous contribution to the $^{13}$C and $^{133}$Cs NMR spin-lattice relaxation rate $1/T_{1}.$ As it is similar to that found in Na$%
_{2}$C$_{60}$, we assign it to the presence of Jahn-Teller distorted C$%
_{60}^{(2,4)-}.\,$This demonstrates the importance of dynamic JTD in metallic fullerides also and suggests that the metallic character of A$_{3}$C$_{60}$ could be related to an enhanced stability of C$_{60}^{(2,4)-}$.
The Na$_{2}$C$_{60}$ and Na$_{2}$CsC$_{60}$ samples were prepared by conventional solid-state reaction. Phase purity was checked by X-ray diffraction. At high $T,$ both compounds have the same structure as other A$%
_{3}$C$_{60}$ systems (face centered cubic structure (fcc) with space group Fm$\overline{3}$m) but they undergo below room $T$ an orientational ordering transition like pure C$_{60}$ (the symmetry is reduced to simple cubic (sc) with Pa$\overline{3}$ space group) [@Yldirim; @Prassides]. NMR measurements were carried out in a 7 T field and spin-lattice relaxation measurements were obtained with usual saturation recovery sequences.
Figure \[Na2T1car\] shows that $1/T_{1}$ for $^{13}$C in Na$_{2}$C$_{60}$ increases very steeply and can be modelled by an activated law $%
1/T_{1}\propto exp(-E_{a}/k_{B}T)$ with $E_{a}=140\pm 20$ meV. This contrasts with the linear $T$ dependence (the Korringa law) expected for simple metals [@hyperfin] and indicates that a gap $E_{a}$ separates the ground state from the excited states of the system. This is similar to the insulating A$_{4}$C$_{60}$ systems, where an activated behaviour dominates the relaxation with $E_{a}\simeq 50-75$ meV [@ZimmerK4; @Kerkoud; @ZimmerPRB95; @convention]. Besides this central fact, more features are evident :
i)There is a peak at 180 K which is typical of a
contribution due to the slowing down of C$_{60}$ molecular motions [@Yoshinari], as shown in more details elsewhere [@Kirchberg2000].
ii\) At low temperature, $1/T_{1}$ is enhanced with respect to the activated law. On Figure \[CarNa2Na2Cs\], it can be more clearly seen that $%
(T_{1}T)^{-1}$ tends to a constant value. If intrinsic, such a contribution could indicate a residual metallic character with a small density of states $%
n(E_{f})\simeq 1$ eV$^{-1}$spin$^{-1}$ [@hyperfin]. A very similar behaviour was in fact observed in Rb$_{4}$C$_{60},$ where an additional relaxation mechanism becomes efficient at low $T$, which was assigned to a small gap ($\simeq 20$ meV) [@Kerkoud]. This gap is easily closed by applying pressure and a growing linear contribution to the relaxation is observed with increasing pressure [*that coexists*]{} with the activated contribution [@Kerkoud]. By analogy, Na$_{2}$C$_{60}$ would be equivalent to Rb$_{4}$C$_{60}$ under an applied pressure of roughly $1$ kbar.
The relaxation behaviour does not change down to 10 K, implying that there is no transition to an ordered ground state. In particular, we do not observe any anomaly at 50 K that might be assigned to a metal-insulator transition as claimed in ref. [@Kubozono]. Hence, the ground state is non-magnetic, which is confirmed by the absence of magnetic broadening of the NMR spectra at low T [@Kirchberg2000]. This is again similar to A$_{4}$C$_{60}$.
iii\) Above 300K, $1/T_{1}$ saturates, which would only be expected for $%
T\gtrsim E_{a}/k_{B}\approx 1600\,\ K$. As the structural transition from sc to fcc takes place at 310 K, it is natural to wonder whether there is an associated change of the electronic properties, for example a smaller gap in the fcc phase. Data at higher temperatures would be needed to conclude this unambiguously.
Like 1/T$_{1}$, the ESR susceptibility $\chi _{esr\text{ }}$increases between 200 and 300 K [@Petit] and a quantitative comparison with NMR should give us further insight into the properties of the system. Therefore, we have measured $\chi _{esr\text{ }}$in our sample (see inset of Fig. \[Na2T1car\]), with results analogous to those of ref. [@Petit]. The intrinsic part $\chi _{int}\,\,$of $\chi _{esr}$ is difficult to extract precisely below 250 K as a large Curie contribution is always observed in Na$%
_{2}$C$_{60}$. From a low $T$ fit, we deduce a Curie-Weiss contribution$%
\,\,C/(T+T_{n})\,$with $T_{n}=8$ $K$ corresponding to 2.3 % impurities per C$_{60}.$ As can be seen in Fig \[Na2T1car\], the remaining part $\chi
_{int}\,$tends to a constant value of 6.10$^{-5}$ emu/mol at low $T$. Although direct conductivity measurements will be necessary to conclude about a possible weak metallicity of this compound, we note that such a Pauli contribution in $\chi _{int}$ would be consistent with the constant $%
(T_{1}T)^{-1}$ found at low $T$ that we have previously discussed. The $T$ dependent contribution $\chi \,\,$to $\chi _{int}$ can then be compared to $%
1/T_{1}$ (see right scale of Fig. \[Na2T1car\]). The simplest way to relate $1/T_{1}$ and $\chi $ is to assume that both are associated with electronic excitations characterized by a spin correlation function with an exponential decay time $\tau .$ In the limit $\omega _{e}\tau <<1$, where $%
\omega _{e}$ is the electronic Larmor frequency, one expects [@White] :
$$\frac{1}{T_{1}}=\left( \frac{A}{\hbar }\right) ^{2}\frac{\chi }{N_{A}\text{ }%
\mu _{B}^{2}}\text{ }k_{B}T\text{ }\tau$$
where $A$ is the hyperfine coupling and $N_{A}$ the Avogadro number. Fig. \[Na2T1car\] evidences the validity of this scaling between $1/T_{1}$ and $%
\chi .\;$Unfortunately, the limited $T$ range of the experiment does not allow to probe efficiently the $T$ dependence of $\tau .$ Assuming that it is constant and using $A~=~4.~10^{-20}$ erg [@hyperfin], we obtain $\tau
\simeq 8.$ $10^{-14}$ sec, which has the same order of magnitude as that found in Rb$_{4}$C$_{60}$ and is consistent with $\omega _{e}\tau <<1.$
To summarize, we conclude that [*there are strong similarities between Na*]{}$_{2}$[*C*]{}$_{60}$[* and A*]{}$_{4}$[*C*]{}$_{60}$. To describe the weakness of the metallic character of these compounds, models involving a (dynamic) JTD of the C$_{60}$ molecule that lifts the degeneracy of the t$%
_{1u}$ levels are the most likely, as they naturally yield a non-magnetic ground state. Indeed, although it could be singlet or triplet, depending on the nature of the JTD, molecular calculations indicate that the singlet state has the lowest energy [@Manini] for both n=2 and n=4. However, as Hund’s rules favor the triplet state, the two states lie close in energy and singlet-triplet like excitations would take place when one molecule goes from one distortion to the other. The 140 meV value of the gap observed in $%
1/T_{1}$ corresponds to the molecular estimate for this “spin-gap” [@Manini]. We can then attribute the relaxation to a coupling between the NMR nuclei and the thermally populated triplet states, which have a spin lifetime $\tau $. While ref. [@Manini] predicts the same spin-gap for C$%
_{60}^{2-}\,$and C$_{60}^{4-}$, the experimental difference between Na$_{2}$C$_{60}$ and A$_{4}$C$_{60}\,$could be due to their different structures, as suggested by the saturation in $1/T_{1}$ observed here at the structural transition. Alternatively, electron-hole symmetry in the t$_{1u}$ level might be broken when excitations to higher electronic levels (t$_{1g}$) are taken into account [@OBrien].
The fact that molecular properties on an energy scale of 140 meV are not smeared out by the formation in the solid of bands of typically 500 meV width requires strong electronic correlations. This was recognized by
Fabrizio [*et al.*]{}[@Fabrizio], who qualified these systems as “Mott Jahn-Teller insulators”. They emphasized that the splitting between $t_{1u}$ levels induced by the JT distortion, estimated to be 500 meV (and observed experimentally as an “optical gap” in A$_{4}$C$_{60}$ [@Fink; @Iwasa]) is too small to lead to a band insulator. Electronic correlations increase the average time spent by one electron on a C$_{60}$ ball, so that “molecular physics”, such as the JT distortions, can take place even before complete localization. This might be the case in Na$_{2}$C$_{60}$, where a residual metallic character is suggested by our low $T$ data. We now turn to the study of Na$_{2}$CsC$_{60}$, for which Fig. \[CarNa2Na2Cs\] shows $(T_{1}T)^{-1}$ compared with Na$_{2}$C$_{60}$. Below 150K, $(T_{1}T)^{-1}$ in Na$_{2}$CsC$_{60}$ is dominated by a $T$ independent contribution, in agreement with its metallic character, which is nearly suppressed in Na$_{2}$C$_{60}.\;$But at higher $T$, $(T_{1}T)^{-1}$ departs from the metallic behaviour and surprisingly, its overall behaviour is very similar to that of Na$_{2}$C$_{60}$. We want to argue here that [*this is not accidental*]{} but reveals a similar relaxation mechanism in the two compounds. We will restrict our discussion to the $sc$ phase of Na$%
_{2}$CsC$_{60}$ ($T$ 300 K), since it was recently suggested that the $fcc$ phase might be insulating [@Cegar].
The strong deviation from the Korringa law in Na$_{2}$CsC$_{60}$ had already been observed previously [@Maniwa; @Saito]. It was attributed to an increase of the density of states associated to the lattice expansion plus a peak due to molecular motions around 300 K. We present here new experimental data to refute this hypothesis. First, to avoid completely a contribution of the C$_{60}$ molecular motions to the relaxation, we have performed measurements on $^{133}$Cs which is not coupled to these motions. As can be seen on Fig. \[CsInvT1\], $(T_{1}T)^{-1}$ for $^{133}$Cs does not exhibit any molecular motion peak around 180 K. Such a peak, already discussed for Na$_{2}$C$_{60},$ is also present in the $^{13}$C relaxation of Na$_{2}$CsC$%
_{60}$ but was missed in the study of ref. [@Maniwa; @Saito]. On the other hand, as for $^{13}$C, $(T_{1}T)^{-1}$ for $^{133}$Cs does deviate above 150 K from the Korringa law, ensuring that this deviation is not due to molecular motions but to [*electronic excitations*]{}. Second, to determine an eventual $n(E_{f})$-related increase of $(T_{1}T)^{-1}$,
we have measured $\chi _{esr}$ in our sample. It does not follow a simple Pauli law, but increases slightly above 100$~$K. The increase of $(T_{1}T)^{-1}$ should scale with $\chi ^{2}$, if both were related to a variation of $n(E_{f})$ [@hyperfin]$,$ but Fig. \[CsInvT1\] shows that $(T_{1}T)^{-1}$ increases much more steeply, already at 200 K, well into the $sc$ phase. Hence, [*the increase of* ]{}$(T_{1}T)^{-1}$[* in* ]{}Na$_{2}$CsC$_{60}$[* is related to an additional relaxation channel.* ]{}It can be fitted by an activated law with $E_{a}=110\pm 5$ meV, as sketched on Fig. \[CsInvT1\].
It is then natural to propose that this additional relaxation mechanism is similar to that proposed for Na$_{2}$C$_{60}$. For a JT distorted C$%
_{60}^{3-}$, there are no possibilities for singlet-triplet transitions directly similar to the ones of a C$_{60}^{2-}$. Therefore we believe that the similarity between Na$_{2}$C$_{60}$ and Na$_{2}$CsC$_{60}$ comes directly from the presence of C$_{60}^{2-}$ in Na$_{2}$CsC$_{60}$, or C$%
_{60}^{4-}$ which are equally likely to be formed when electrons jump from ball to ball in the metal [@comment]. As the gap observed by NMR in Na$%
_{2}$C$_{60}$ is related to [*individual*]{} excitations of C$_{60}^{2-}$ and not to a band gap, similar excitations might occur in Na$_{2}$CsC$_{60}$ as well, [*if C*]{}$_{60}^{(2,4)-}$[* exist within the metal for times* ]{}$%
\tau _{pair}$[* sufficiently long compared to the spin lifetime* ]{}$\tau
\,\,$[*of an excited state*]{}. This does not imply a static charge separation in Na$_{2}$CsC$_{60}$, which is ruled out by the $^{23}$Na NMR spectra (not shown) which display one narrow line shifted from the position of Na$_{2}$C$_{60}$.
Using Eq. 1, $\tau $ can be extracted from $(T_{1}T)^{-1}$ if the singlet-triplet component in $\chi $ is known. We assume a similar $T$ dependence as that found in Na$_{2}$C$_{60}$, which scales as $\chi
~=~\alpha T~^{-1}exp~(-E_{a}/T)$ in the experimental $T$ range and the gap value $E_{a}=110$ meV determined accurately from $^{133}$Cs NMR. Such a contribution to the susceptibility can only be smaller than the increase of $%
\chi _{esr}$, which implies $\alpha \lesssim 2\ $and yields $\tau \simeq
10^{-14}\sec $. On the other hand, the time $\tau _{res}$ spent by one electron in the vicinity of one C$_{60}$, which is a lower limit for $\tau
_{pair}$, can be estimated by $\tau _{res}\simeq \hbar n(E_{f})\simeq
6.10^{-15}$ sec in Na$_{2}$CsC$_{60},$ which is only slightly shorter than $%
\tau $. Electronic correlations, Jahn-Teller effects, as well as scattering on a C$_{60}$ ball, would all increase $\tau _{res}$ compared to this simple estimate, so that we are likely in the limit $\tau _{pair}>\tau $ where the C$_{60}^{(2,4)-}$ can contribute to the NMR relaxation.
Both the gap value and $\tau $ are reduced compared with Na$_{2}$C$_{60}$. The same trend was observed in Rb$_{4}$C$_{60}$, where $\tau $ and the gap decrease with increasing pressure, [*i.e.*]{} increasing density of states [@Kerkoud]. This suggests a direct relation between $\tau $ and $%
n(E_{f})$, which raises the question of the origin of the relaxation time $%
\tau $ for triplet states. One could expect such a trend if the triplet states are relaxed by conduction electrons, although the triplet state is formed by “potential” conduction electrons, making it difficult to distinguish two different spin species. This would also predict that for higher density of states like in K$_{3}$C$_{60}$ and Rb$_{3}$C$_{60},$ the contribution of triplet states to the relaxation almost disappear. We note that although a “better” Korringa law is observed in these compounds [@Tycko], a 20%-30% increase of $(T_{1}T)^{-1}$ was still noticed between 100 K and 300 K. It could in fact have a similar origin, so that the presence of C$_{60}^{(2,4)-}$ might be a common feature of metallic A$_{3}$C$%
_{60}$ compounds.
In conclusion, we have shown that Na$_{2}$C$_{60}$ has a non-magnetic ground state and that its low energy electronic excitations are characterized by a 140 meV spin-gap. This is very similar to A$_{4}$C$_{60}$ systems and supports the “Mott JT scenario” [@Fabrizio] to describe fullerides with 2 or 4 electrons per C$_{60}$. Furthermore, we evidence very similar electronic excitations in Na$_{2}$CsC$_{60}$, coexisting with typically metallic ones. In both cases, we assign the spin-gap to singlet-triplet excitations between two JT distortions of C$_{60}^{2n-}$ balls (n = 1 or 2). Because the triplet state provides a very efficient relaxation mechanism for NMR, we could indirectly detect here for the first time the presence of dynamic JTD in a superconducting fulleride. Our study also implies that Na$%
_{2}$CsC$_{60}$ undergoes rapid charge fluctuations (on a time scale of 10$%
^{-14}$ sec) which create preferentially C$_{60}^{(2,4)-}$. We suggest that this can be rationalized as a consequence of the JTD which stabilizes evenly charged C$_{60}$.
A first indication of the role played by JTD in the electronic properties of fullerides was the observation in the metallic cubic quenched CsC$_{60}$ of[* localized C*]{}$_{60}^{2-}$ [@BrouetPRL99]. Although this large difference of charge lifetime remains to be understood, these behaviours suggest that the key feature behind the physics of cubic fullerides is due to an interplay between strong electronic interaction and “JT mediated electronic interactions”. Electronic correlations are essential for molecular excitations, such as the JTD, to exist in the solid. JT mediated electronic interactions play in turn a crucial role in determining the insulating or metallic character of a given compound, by inducing repulsive interactions in even electron systems and attractive interactions in odd electron systems, in order to promote the existence of C$_{60}^{2n-}$. The recent discovery of a way of doping continuously the C$_{60}$ t$_{1u}$ level through a field-effect device might open a new path for checking this original behaviour more systematically [@Batlogg].
We thank N. Manini and E. Tosatti for stimulating discussions and we acknowledge the financial support from the TMR program of the European Commission (“FULPROP” ERBRMRXVT970155).
R.C. Haddon [*et al.*]{}, [*Nature*]{} [**350**]{}, 320 (1991)
O. Gunnarsson, E. Koch and R. M. Martin, [*Phys. Rev. B*]{} [**54**]{}, R11 026 (1996)
W. Victoroff and M. Héritier, [*J. Phys. I France*]{} [**6**]{}, 2175 (1996)
N. Manini, E. Tosatti and A. Auerbach [*Phys. Rev. B*]{} [**49**]{}, 13008 (1994)
O. Gunnarsson, [*Phys. Rev. B*]{} [**51**]{}, 3493 (1995)
O. Gunnarsson, S.C Erwin, E. Koch and R. M. Martin [*Phys. Rev. B*]{} [**57**]{}, 2159 (1998); J. E. Han, E. Koch and O. Gunnarsson, [*Phys. Rev. Lett.*]{} [**84**]{}, 1276 (2000)
P. Petit [*et al.*]{}, [*Progress in Fullerene research*]{}, World scientific, edited by H. Kuzmany, J. Fink, M. Mehring and S. Roth, p.148 (1994)
Y. Kubozono [*et al., Phys. Rev. B*]{} [**59**]{}, 15062 (1999)
G. Zimmer, M. Helme, M. Mehring and F. Rachdi, [*Europhysics Letters*]{} [**27**]{}, 543 (1994)
R. Kerkoud [*et al.*]{}, [*J. Phys. Chem. Solids* ]{}[**57**]{}, 143 (1996)
G. Zimmer, M. Mehring, C. Goze and F. Rachdi, [*Phys. Rev. B*]{} [**52**]{}, 13 300 (1995)
K. Tanigaki [*et al.*]{}, [*Nature*]{} [**356**]{}, 419 (1992)
T. Yldirim [*et al.*]{}, [*Phys. Rev. Lett.*]{} [**71**]{}, 1383 (1993)
K. Prassides [*et al.*]{}, [*Science*]{} [**263**]{}, 950 (1994)
In a metal, one expects (T$_{1}$T)$^{-1}$ = ($\pi k_{b}/$h-.2em.2em)$A^{2}$n(E$_{f}$)$^{2}$. From the Korringa value at low T in Na$_{2}$CsC$%
_{60},$ we deduce an hyperfine coupling $A$ = 4 10$^{-20}$ erg, which lies with other estimates of $A$ in fullerides within a factor 4.
We use here E$_{a}=E_{g}$ instead of E$_{a}=E_{g}/2$, as done in [@ZimmerK4; @Kerkoud; @ZimmerPRB95] by reference to a semiconducting model.
Y. Yoshinari, H. Alloul, G. Kriza and K. Holczer, [*Phys. Rev. Lett.*]{} [**71**]{}, 2413 (1993)
V. Brouet [*et al.,* ]{}to appear in [*AIP conference proceedings (2000)*]{}
R.M. White, [*Quantum theory of Magnetism*]{}, Springer-Verlag Heidelberg New-York (1983)
O’Brien MCM, [*Phys. Rev. B*]{} [**53**]{}, 3775 (1996)
M. Fabrizio and E. Tosatti, [*Phys. Rev. B*]{} [**55**]{}, 13 465 (1997)
M. Knupfer and J. Fink, [*Phys. Rev. Lett.*]{} [**79**]{}, 2714 (1997)
Y. Iwasa and T. Kaneyasu, [*Phys. Rev. B*]{} [**51**]{}, 3678 (1995)
N. Cegar [*et al.*]{}, cond-mat/0002447
Y. Maniwa et al., [*Phys. Rev. B*]{} [**52**]{}, 7054 (1995)
T. Saito [*et al., J. of the Phys. Soc. of Japan*]{} [**64**]{}, 4513 (1995)
For simplicity, we assume here that C$_{60}^{2-}$ and C$%
_{60}^{4-}$ behave similarly, although it could be discussed following ref. [@OBrien]
R. Tycko [*et al.*]{}, [*Phys. Rev. Lett.*]{} [**68**]{}, 1912 (1992)
V. Brouet [*et al.*]{}, [*Phys. Rev. Lett.*]{} [**82**]{}, 2131 (1999)
J.H. Schon, Ch. Kloc, R.C. Haddon and B. Batlogg, [*Science*]{} [**288**]{}, 656 (2000)
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'An approximate formula connecting the true and the HBT homogeneity regions in multparticle production processes is derived. It implies that when calculating the HBT radii one should use the center of mass systems of the pairs rather than the now popular LCMS system. A discussion of several simple examples clarifies the potential and limitations of the HBT method. The even cumulants of the $\textbf{X}$-distribution, including the HBT radii, can be determined for each homogeneity region, but the relative positions of the homogeneity regions are unconstrained. This makes the HBT radii of little use for calculating quantities dependent on the interparticle interactions in coordinate space.'
author:
- |
K.Zalewski\
M.Smoluchowski Institute of Physics\
Jagellonian University, Cracow[^1]\
and\
Institute of Nuclear Physics, Cracow
title: Physics from HBT radii
---
PACS numbers 25.75.Gz, 13.65.+i\
Bose-Einstein correlations, interaction region determination.
Almost fifty years ago, in a famous paper known as GGLP [@GGL], a method of using momentum distributions for pairs of identical pions to estimate the sizes of the interaction regions, i.e. of the regions where the hadrons are produced in multiparticle production processes, has been described. Such radii, determined from momentum distributions of identical particles, have been later called, not very appropriately [@KOP], HBT radii. GGLP assumed that hadron production happens instantly and simultaneously at some time $t=0$ and that there are no correlations between the momenta of the hadrons and their production points. Considering simultaneously the momentum of a particle and its production point implies a quasiclassical approximation, but it can be made plausible [@ZAL1] that this approximation is good for heavy ion scattering and acceptable also for the other multiparticle production processes. Let us denote by $p_1$ and $p_2$ the four-momenta of the two identical pions in the pair[^2], by $x_1$ and $x_2$ the space-time positions of their production points and introduce the notation
$$\label{}
K = \frac{1}{2}(p_1+p_2);\quad q = p_1 - p_2;\quad X = \frac{1}{2}(x_1 +
x_2);\quad y = x_1 - x_2.$$
Note that $p_1^2 = p_2^2 = m^2$ implies
$$\label{onmass}
K_0 =
\frac{1}{\sqrt 2}\sqrt{E_k^2 + \frac{1}{4}\textbf{q}^2 + \sqrt{(E_k^2 + \frac{1}{4}\textbf{q}^2)^2 -
(\textbf{K}\cdot \textbf{q})^2}};\quad q_0 = {\mbox{\boldmath$\beta $}}\cdot \textbf{q},$$
where $E_k = \sqrt{\textbf{K}^2 + m^2}$ equals $K_0$ at $\textbf{q}=0$, and ${\mbox{\boldmath$\beta $}} = \frac{\textbf{K}}{K_0}$ is the velocity of the pair. GGLP found $R_{HBT}$ from the distribution of $q^2$.
Some twenty five year later Pratt [@PRA1] described a model with an exploding source. In this model, the momentum distribution of particles depends on the production point. Due to this correlation, the HBT radius differs significantly from the true radius, which is known from the input. For fixed $\textbf{K}$ the result was that, while the true radius $R$ does not depend on $|\textbf{K}|$, the HBT radius decreases from $\sqrt{\frac{2}{3}}R$ for $\textbf{K}=0$ to zero when $|\textbf{K}|$ tends to infinity. This finding was generalized by Bowler [@BOW], who pointed out that in general, whenever there are strong correlations between momenta and production points, $R_{HBT}$ measures only the region where particles of similar momentum are produced. Sinyukov [@SIN], [@AKS] associated these regions with the homogeneity regions considered in hydrodynamics and the name homogeneity region got generally accepted.
Most models in their simplest form, i.e. without final state interactions, resonance decays etc., find the $n$-particle momentum distributions for identical pions from formulae equivalent to [@KAR]
$$\label{karczm}
P(\textbf{p}_1,\ldots,\textbf{p}_n) = C_n \sum_Q \prod_{j=1}^n \rho(\textbf{p}_j;\textbf{p}_{Qj}),$$
where the summation is over all the permutations $j \rightarrow Qj$ of the indices $j=1,\ldots,n$, $C_n$ are normalization constants and $\rho(\textbf{p}_1;\textbf{p}_2)$ is some time independent, single particle density matrix. Thus $\rho(\textbf{p}_1;\textbf{p}_2)$ determines all the momentum distributions. The best [@HCS] quantum mechanical analogue of the classical phase space density is the Wigner function $W(\textbf{X},\textbf{K})$, which is a Fourier transform of $\rho(\textbf{p}_1;\textbf{p}_2)$. Following most models, we assume further that, $W(\textbf{X},\textbf{K})$ defines the HBT $\textbf{X}$-distribution for the homogeneity region corresponding to a given $\textbf{K}$.
Measuring all the momentum distributions is not enough to determine $\rho(\textbf{p}_1;\textbf{p}_2)$. As easily seen from (\[karczm\]), the observable momentum distributions do not change [@BIZ], [@ZAL2], under the transformation[^3]
$$\label{rhopri}
\rho(\textbf{p}_1;\textbf{p}_2) \rightarrow \rho_f(\textbf{p}_1;\textbf{p}_2) = e^{i\left[f(p_1)-
f(p_2)\right]}\rho(\textbf{p}_1;\textbf{p}_2),$$
where $f(p)$ is an arbitrary real-valued function of the four-vector $p$, and $p_0 = \sqrt{\textbf{p}^2+m^2}$. This transformation does not affect the HBT radii[^4] of the homogeneity regions, but it can shift and deform these regions.
An important concept introduced in [@PRA1] is the emission function $S(X,K)$. Assuming chaotic sources, i.e. no interference between particles produced at different moments of time, this is related to the density matrix in the momentum representation by the formula[^5]
$$\label{rhoemi}
\rho(\textbf{p}_1;\textbf{p}_2) = \int\!\!d^4X\;S(X,K)e^{iqX}.$$
In the spirit of the quasiclassical approximation, using the well tested (cf. e.g. [@WIH] and references given there) mass shell approximation
$$\label{masapr}
K_0 = E_k \equiv \sqrt{m^2 + \textbf{K}^2},$$
one can interpret $S(X,K)$ as the time-dependent distribution of the pairs of vectors $\{\textbf{X},\textbf{K}\}$. The time independent homogeneity region for each $\textbf{K}$ can be obtained by integrating over time
$$\label{rhotru}
p(\textbf{X}|\textbf{K}) = \int\!\!dt\;S(X,K),$$
where the notation stresses that we are interested in the distribution of $\textbf{X}$ for given $\textbf{K}$. At this point the mass shell approximation (\[masapr\]) is necessary in order to make $K_0$, and consequently $S(X;K)$ and $p(\textbf{X}|\textbf{K})$, independent of $\textbf{q}$.
Further we will call $p(\textbf{X}|\textbf{K})$ true distribution in order to distinguish it from the HBT distribution given by the Wigner function. Invariance (\[rhopri\]) means that for a given set of momentum distributions there is a variety of HBT $\textbf{X}$-distributions which all correspond to the same fit to the data. One way of choosing among them is to specify the emission function.
It was soon noticed [@PRA2] that for a given $\rho(\textbf{p}_1;\textbf{p}_2)$ there is an infinity of solutions for $S$ and that, in particular, a small long-lived source may be undistinguishable from a large short-lived source. Thus, the information about interaction regions obtained from momentum measurements is rather incomplete.
The standard relation between the density matrix and the Wigner function yields from (\[rhoemi\])
$$\label{}
W(\textbf{X},\textbf{K}) =
\int\!\!\frac{d^3q}{(2\pi)^3}\int\!\!d^4X'\;S(X',K)e^{i\textbf{q}\cdot
(\textbf{X}-\textbf{X}'+{\mbox{\boldmath$\beta $}} t)},$$
where $q_0$ has been eliminated using (\[onmass\]). We invoke now the mass shell approximation (\[masapr\]), which makes $S$ independent of $\textbf{q}$. Thus, the integration $d^3q$ gives $(2\pi)^3\delta(\textbf{X}-\textbf{X}'+{\mbox{\boldmath$\beta $}} t)$ and the integration $d^3X'$ can be performed. The result is
$$\label{rhohbt}
W(\textbf{X},\textbf{K}) = \int\!\!dt\;S(\textbf{X}+{\mbox{\boldmath$\beta $}} t,t,K).$$
The comparison of the true interaction region with the HBT one reduces to the comparison of the integrals in (\[rhotru\]) and (\[rhohbt\]).
The integrals in (\[rhotru\]) and (\[rhohbt\]), and consequently the true and HBT homogeneity regions, coincide only[^6] for ${\mbox{\boldmath$\beta $}} = \textbf{0}$. This can be achieved by working with pairs which have the same velocity ${\mbox{\boldmath$\beta $}}$ and using the reference frame where ${\mbox{\boldmath$\beta $}} = \textbf{0}$.
We summarize our findings:
- The HBT results are credible only for single homogeneity regions, i.e. at given $\textbf{K}$.
- For each homogeneity region, only the pairs with the same velocity, i.e. with the same $\textbf{K}$, are considered and one should use the reference frame, where this velocity is zero. There, the true and the HBT homogeneity regions coincide which is not the case for other frames, as for instance for the now popular (cf e.g. the review [@LPS]) LCMS frame. An additional pragmatic argument in favor of the rest frame is that the effects of the final state interactions are the simplest there (cf. e.g. [@DAP] and the references given there).
- The true radii of the homogeneity regions, more generally all the even cumulants of their $\textbf{X}$-distributions, can be obtained from the HBT analysis of the momentum distributions. For instance, when $p(\textbf{X}|\textbf{K}) =
p(-\textbf{X}|\textbf{K})$ all the odd cumulants vanish and, therefore, $p(\textbf{X}|\textbf{K})$ can be measured by the HBT method.
- The space distribution of the centers of the homogeneity regions $\langle
\textbf{X}\rangle(\textbf{K})$ is unconstrained by the HBT analysis, but once it is fixed, all the functions $p(\textbf{X}|\textbf{K})$ can be determined [@BIZ], [@ZAL2].
Let us discuss some simple examples illustrating these features of the HBT method. In order to avoid unessential complications, we will consider one space dimension and very simple emission functions constructed from $\delta$-functions and step functions. They violate the Heisenberg uncertainty relations, but all these calculations can be repeated in three dimensions using Gaussians and the results are qualitatively the same.
Let us first consider the emission function
$$\label{}
S_1(X,K;t_0,a,b,\kappa) = \delta(t - t_0)\Theta_{a,b}(\textbf{X})\delta(\textbf{K} - \kappa),$$
where $t_0$, $a < b$ and $\kappa$ are real constants. Function $\Theta_{a,b}(x)
= \frac{1}{b-a}$ for $a < x < b$ and zero outside this interval. For this emission function
$$\begin{aligned}
\label{rhomod}
\rho(\textbf{p}_1;\textbf{p}_2) &=&
\delta(\textbf{K}-\kappa)e^{-i\textbf{q}(\frac{a+b}{2} - {\mbox{\boldmath$\beta
$}} t_0)}
\frac{2\sin |\textbf{q}|\frac{b-a}{2}}{|\textbf{q}|(b-a)}\\
W(\textbf{X},\textbf{K}) &=& \Theta_{a,b}(\textbf{X}+ {\mbox{\boldmath$\beta
$}} t_0)\delta(\textbf{K}-\kappa).\end{aligned}$$
It is seen that the segment $a < \textbf{X} < b$ got shifted to $a-{\mbox{\boldmath$\beta $}} t_0 < \textbf{X} < b - {\mbox{\boldmath$\beta $}}
t_0$. Acording to (\[rhomod\]) this shift is due to the phase factor in the density matrix. Since in the formula for the two-particle correlation function only the absolute value of $\rho$ appears, experiment is blind to such shifts. This remains true also when many-particle correlation functions are measured, even when all the measurements are performed with perfect precision [@BIZ], [@ZAL2]. The fact that the position of the center of the interaction region cannot be found from the HBT analysis of the momentum distributions is, of course, well known (see e.g. [@WIH]), but one should also keep in mind that when the emission function is given the density matrix, the Wigner function and, consequently, all the HBT homogeneity regions are unambiguously defined.
As the next example we take
$$\label{}
S_2(X,K) = \frac{1}{2}\left[S_1(X,K;t_0,-a,0,\kappa_1) +
S_1(X,K;t_0,0,a,\kappa_2)\right]$$
Using the previous example, it is seen that the segment $-a < \textbf{X} < a$ is mapped onto two segments: $-a-{\mbox{\boldmath$\beta $}}_1t_0 < \textbf{X} <
-{\mbox{\boldmath$\beta $}}_1 t_0$ and $-{\mbox{\boldmath$\beta $}}_2t_0 <
\textbf{X} < a-{\mbox{\boldmath$\beta $}}_2t_0$. Let us choose
$$\label{trainv}
f(p) = -\textbf{b}\cdot \textbf{p} - \frac{1}{2}c\textbf{p}^2,$$
where $\textbf{b}$ is an arbitrary vector and $c$ an arbitrary constant, and make the transformation (\[rhopri\]). Each HBT homogeneity region gets shifted by $\textbf{b} + c\textbf{K}$. For $\kappa_1~\neq ~\kappa_2$, by a suitable choice of $c$, one can obtain any prescribed distance between the centers of the two segments. The true length of the interaction region, as seen from $S$, is $2a$. The length of the transformed interaction region, which follows just as well from the data on momentum distributions, can be any number not smaller than $a$. What is the way out? One has to invoke the homogeneity regions. For each $\textbf{K}$ separately, the length of the segment where particles with this value of $\textbf{K}$ are produced, i.e. of the homogeneity region, is reproduced correctly. The positioning of the homogeneity regions corresponding to different values of $\textbf{K}$, however, is beyond control when the momentum distributions are the only input. In three dimensions it is easy to prove [@BIZ], [@ZAL2] that by a suitable choice of function $f(p)$ in (\[rhopri\]), the positions of the centers of the homogeneity regions $\langle \textbf{X} \rangle(\textbf{K})$ can be changed into $\langle
\textbf{X} \rangle(\textbf{K}) + \textbf{g}(\textbf{K})$, where $\textbf{g}$ is an arbitrary differentiable function of $\textbf{K}$.
As an amusing example in three dimensions let us consider the models where $K_\mu \approx \lambda X_\mu$, and $\lambda$ is a constant [@CSO], [@BIZ2]. They correspond to emission functions
$$\label{}
S(X,K) = \delta^3(\textbf{X}- {\mbox{\boldmath$\beta $}} t)\overline{S}(\textbf{X},t,K),$$
where $\overline{S}$ is some function which does not affect the singularity introduced by the $\delta^3$. The corresponding Wigner function is
$$\label{}
W(\textbf{X},\textbf{K}) = \delta^3(\textbf{X})\int\!\!dt\;\overline{S}({\mbox{\boldmath$\beta $}} t,t,K).$$
Thus, the HBT interaction region reduces to one point. To be sure: such models, when properly used, are quite successful, but finding the interaction region from the density matrix in the momentum representation is their misuse.
As our last model consider the emission function
$$\label{}
S(X,K) = \Theta_{0,a}(t)\delta(\textbf{X})\delta(\textbf{K}-\kappa).$$
This corresponds to the Wigner function $$\label{}
W(\textbf{X},\textbf{K}) = \frac{1}{{\mbox{\boldmath$\beta $}}}\Theta_{0,a}(-\textbf{X}/{\mbox{\boldmath$\beta $}})
\delta(\textbf{K} - \kappa) =
\Theta_{-{\mbox{\boldmath$\beta $}a,0} }(\textbf{X})\delta(\textbf{K} - \kappa).$$ The length of the true interaction region is zero, while the length of the HBT interaction region is ${\mbox{\boldmath$\beta $}} a$. Thus, in order to get the true length from the HBT analysis one must use the reference frame where $\textbf{K} = \textbf{0}$. We conclude that, working with all the pairs which have a given velocity, one should measure their homogeneity region in their rest frame. Let us consider some further implications of our analysis:
If there are no position-momentum correlations, all the homogeneity regions, measured in the respective rest frames, are equivalent to each other and to the overall interaction region. The ambiguity (\[rhopri\]) reduces to a lack of information about the position of the center of the interaction region [@ZAL2]. Thus, in this case the HBT method works very well. One should keep in mind however that, as illustrated by our second model, the independence of the homogeneity region on $\textbf{K}$ is a necessary, but not a sufficient condition for the absence of position-momentum correlations.
If there are no interparticle interactions in coordinate space, the HBT homogeneity regions may be all we need. E.g. the entropy of a gas of noninteracting particles (in the quasiclassical approximation) depends on the accessible phase space volume. This can be calculated by integrating over coordinate space at fixed momentum and then integrating over momenta. The first integration gives just the HBT volume of the homogeneity region, though one should keep in mind the ambiguity (\[rhopri\]). If there are interparticle interactions in coordinate space, however, it may make a lot of difference whether the homogeneity regions are on top of each other, or scattered over space. Thus, the use of the HBT radii to calculate the entropy of a gas of interacting particles, or of their mean free paths, is risky.
For ${\mbox{\boldmath$\beta $}} = \textbf{0}$ the true space density $\int\!\!dt\;S(X,K)$ can be unambiguously obtained from $\rho(\textbf{p}_1,\textbf{p}_2)$ by inverse Fourier transformation. Thus, all the ambiguity in the determination of the homogeneity regions, including their absolute positions, results from (\[rhopri\]) and disappears when $f(p)$ is fixed. Conversely, when the true space density is known, it yields unambiguously $\rho(\textbf{p}_1;\textbf{p}_2)$ and thus it fixes $f(p)$.
The ambiguity in the determination of the space-time density $S$ follows from the irreversibility of transformation (\[rhoemi\]) and persists even when $f(p)$ is fixed, i.e. when $\rho(\textbf{p}_1;\textbf{p}_2)$ is known.
The author thanks A. Bialas, K.Fia[ł]{}kowski and W. Florkowski for helpful comments.
[99]{} G. Goldhaber, S. Goldhaber, W. Lee and A. Pais, *Phys. Rev.* **210**(1960)300. G.I. Kopylov and M.I Podgoretskii, *Zhur.Exp.Theor.Phys.* **69**(1975)434. K. Zalewski, hep-ph/07052961. S. Pratt *Phys. Rev. Lett.* **53**(1984)1219. M.G. Bowler, *Z. Phys.* **C29**(1985)617. Yu.M. Sinyukov, in J. Letessier, H.H. Gutbrod and J. Rafelski (eds.) *Hot Hadronic Matter Theory and Experiment*, Plenum Press, New York (1995) p. 309. S.V. Akkelin and Yu.M. Sinyukov, *Phys. Lett.* **B356**(1995)525. J. Karczmarczuk, *Nucl. Phys.* **B78**(1974)370. M. Hillery, R.F. O’Connel, M.O. Scully and E.P. Wigner, *Phys. Rep.* **106**(1984)121. A. Bia[ł]{}as and K. Zalewski, *Phys. Rev.* **D72**(2005)036009. K. Zalewski, *Phys. Rev.* **D74**(2006)114022. U.A. Wiedemann and U. Heinz, *Phys. Rep*. **319**(1999)145. S. Pratt, *Phys. Rev.* **D33**(1986)72. M.A. Lisa, S. Pratt, R. Soltz and W. Wiedemann, *Ann. Rev. Nucl. Part. Phys.* **55**(2005)357.
T. Csörgö and J.Zimańyi, *Nucl. Phys.* **A517**(1990)588. A. Bialas and K. Zalewski, *Acta Phys. Pol*. **B30**(1999)359. P. Danielewicz and S. Pratt, *Phys. Rev.* **C75**(2007)0304907.
[^1]: Address: Reymonta 4, 30 059 Krakow, Poland, e-mail: [email protected]. This work has been partly supported by the Polish Ministry of Education and Science grant 1P03B 045 29(2005-2008).
[^2]: We discuss pions for definiteness, but the discussion applies to any spin zero bosons.
[^3]: The full group of transformations includes also $\rho(\textbf{p}_1;\textbf{p}_2)\rightarrow \rho^*(\textbf{p}_1;\textbf{p}_2)$, but this corresponds to the space inversion of the interaction region and is of no interest in the present context.
[^4]: More generally, all the even cumulants of $W(\textbf{X},\textbf{K})$, at $\textbf{K}$ fixed, remain unchanged. The proof is a simple extension of the proof given for the HBT radii in [@ZAL2].
[^5]: Sometimes formulae equivalent to (\[rhoemi\]) with $\rho(\textbf{p}_1;\textbf{p}_2)$ replaced by $\rho^*(\textbf{p}_1;\textbf{p}_2)$ are used. The advantage of the present convention is that for $S(X,K) = \delta (t)g(\textbf{X},\textbf{K})$ function $g$ is just the Wigner function.
[^6]: Except for the trivial case when $S(X,K) \sim \delta(t)$.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We prove that P. Mathieu’s Open problem on constructing Gardner’s deformation for the $N{=}2$ supersymmetric $a{=}4$–Korteweg–de Vries equation has no supersymmetry-invariant solutions, whenever it is assumed that they retract to Gardner’s deformation of the scalar KdV equation under the component reduction. At the same time, we propose a two-step scheme for the recursive production of the integrals of motion for the $N{=}2$, $a{=}4$–SKdV. First, we find a new Gardner’s deformation of the Kaup–Boussinesq equation, which is contained in the bosonic limit of the super-hierarchy. This yields the recurrence relation between the Hamiltonians of the limit, whence we determine the bosonic super-Hamiltonians of the full $N{=}2$, $a{=}4$–SKdV hierarchy. Our method is applicable towards the solution of Gardner’s deformation problems for other supersymmetric KdV-type systems.'
address:
- 'Département de Mathématiques et de Statistique, Université de Montréal, C.P. 6128, succ. Centre-ville, Montréal, Québec H3C 3J7, Canada.'
- 'Mathematical Institute, University of Utrecht, P.O.Box 80.010, 3508 TA Utrecht, The Netherlands.'
- 'Department of Higher Mathematics, Ivanovo State Power University, 34Rabfakovskaya str., Ivanovo, 153003 Russia.'
- 'Department of Mathematics, Brock University, 500 Glenridge av., St. Catharines, Ontario L2S 3A1, Canada.'
author:
- 'V. Hussin'
- 'A. V. Kiselev${}^{*}$'
- 'A. O. Krutov'
- 'T. Wolf'
date: 'November 11, 2009'
---
Introduction {#introduction .unnumbered}
------------
This paper is devoted to the Korteweg–de Vries equation and its generalizations [@Gardner]. We consider completely integrable, multi-Hamiltonian evolutionary $N{=}2$ supersymmetric equations upon a scalar, complex bosonic $N{=}2$ superfield $$\label{N2SuperField}
{\boldsymbol{u}}(x,t;\theta_1,\theta_2)
=u_0(x,t)+\theta_1\cdot u_1(x,t)+\theta_2\cdot u_2(x,t)+\theta_1\theta_2
\cdot u_{12}(x,t),$$ where $\theta_1$ and $\theta_2$ are Grassmann variables satisfying $\theta_1^2=\theta_2^2=\theta_1\theta_2+\theta_2\theta_1=0$. Also, we investigate one- and two-component reductions of such four-component $N{=}2$ super-systems upon ${\boldsymbol{u}}$. In particular, we study the bosonic limits, which are obtained by the constraint $$\label{BRed}
u_1=u_2\equiv0.$$ We analyse the structures that are inherited by the limits from the full super-systems and, conversely, recover the integrability properties of the entire $N{=}2$ hierarchies from their bosonic counterparts.
We address $2^{\text{nd}}$ Open problem of [@MathieuOpen] for the $N{=}2$ supersymmetric Korteweg–de Vries equation with $a{=}4$, see [@MathieuTwo; @MathieuNew], $$\label{SKdV}
{\boldsymbol{u}}_t=-{\boldsymbol{u}}_{xxx}+3\bigl({\boldsymbol{u}}{\mathcal{D}}_1{\mathcal{D}}_2{\boldsymbol{u}}\bigr)_x
+\frac{a-1}{2}\bigl({\mathcal{D}}_1{\mathcal{D}}_2{\boldsymbol{u}}^2\bigr)_x + 3a{\boldsymbol{u}}^2{\boldsymbol{u}}_x,\qquad
{\mathcal{D}}_i=\frac{{\partial}}{{\partial}\theta_i}+\theta_i\cdot\frac{{{\mathrm d}}}{{{\mathrm d}}x}.$$ For $a{=}4$, this super-equation possesses an infinite hierarchy of bosonic Hamiltonian super-functionals $\boldsymbol{{\mathcal{H}}}^{(k)}$ whose densities $\boldsymbol{h}^{(k)}$ are integrals of motion. The problem amounts to a recursive production of such densities by using those which are already obtained. In its authentic formulation, the problem suggests finding a parametric family of super-equations ${\mathcal{E}}(\epsilon)$ upon the generating super-function $\tilde{{\boldsymbol{u}}}(\epsilon)=\sum_{k=0}^{+\infty}\boldsymbol{h}^{(k)}\cdot\epsilon^k$ for the integrals of motion such that the initial super-equation is ${\mathcal{E}}(0)$. It is further supposed that, at each $\epsilon$, the evolutionary equation ${\mathcal{E}}(\epsilon)$ exprimes a (super-)conserved current, and there is the Gardner–Miura substitution ${\mathfrak{m}}_\epsilon\colon{\mathcal{E}}(\epsilon)\to{\mathcal{E}}(0)$. Hence, expanding ${\mathfrak{m}}_\epsilon$ in $\epsilon$ and using the initial condition $\tilde{{\boldsymbol{u}}}(0)={\boldsymbol{u}}$ at $\epsilon=0$, one obtains the differential recurrence relation between the Taylor coefficients $\boldsymbol{h}^{(k)}$ of the generating function $\tilde{{\boldsymbol{u}}}$ (see [@Gardner] or [@PamukKale; @TMPh2006; @KuperIrish; @MathieuNew] and references therein for details and examples). The recurrence relations between the (super-)Hamiltonians of the hierarchy are much more informative than the usual recursion operators that propagate symmetries. In particular, the symmetries can be used to produce new explicit solutions from known ones, but the integrals of motion help to find those primary solutions.
Let us also note that, within the Lax framework of super-pseudodifferential operators, the calculation of the $(n+1)$-st residue does not take into account the $n$ residues, which are already known at smaller indices. This is why the method of Gardner’s deformations becomes highly preferrable. Indeed, there is no need to multiply any pseudodifferential operators by applying the Leibnitz rule an increasing number of times, and all the previously obtained quantities are used at each inductive step. By this argument, we understand Gardner’s deformations as the transformation in the space of the integrals of motion that maps the residues to Taylor coefficients of the generating functions $\tilde{{\boldsymbol{u}}}(\epsilon)$ and which, therefore, endows this space with the additional structure (that is, with the recurrence relations between the integrals).
Still there is a deep intrinsic relation between the Lax (or, more generally, zero-curvature) representations for integrable systems and Gardner’s deformations for them. Namely, both approaches manifest the matrix and vector field representations of the Lie algebras related to such systems [@WilsonEquiv]. Our main result is the following. Under some natural assumptions, we prove the non-existence of $N{=}2$ supersymmetry-invariant Gardner’s deformations for the bi-Hamiltonian $N{=}2$, $a{=}4$–SKdV. Still, we show that the Open problem must be addressed in a different way, and then we solve it in two steps. First, in section \[SecKBous\] we recall that the tri-Hamiltonian hierarchy for the bosonic limit of with $a{=}4$ contains the Kaup–Boussinesq equation, see [@Kaup75; @NutkuPavlov] and [@BrunelliDas94; @KuperSuperLongWaves; @PalitChowdhury96] in the context of this paper. Then in section \[SecBurg\] we construct new deformations for the Kaup–Boussinesq equation such that the Miura contraction ${\mathfrak{m}}_\epsilon$ now incorporates Gardner’s map for the KdV equation ([@Gardner], c.f. [@PamukKale; @KuperIrish]). Second, extending the Hamiltonians $H^{(k)}$ for the Kaup–Boussinesq hierarchy to the super-functionals $\boldsymbol{{\mathcal{H}}}^{(k)}$ in section \[SecHam\], we reproduce the bosonic conservation laws for with $a{=}4$. Finally, we contribute to the solution of P. Mathieu’s $3^{\text{rd}}$ Open problem [@MathieuOpen] with the description of necessary conditions upon a class of Gardner’s deformations for that reproduce its *fermionic* local conserved densities.
The standard reference in geometry of completely integrable Hamiltonian partial differential equations is [@Olver].
$N{=}2$ $a{=}4$–SKdV as bi-Hamiltonian super-extension of Kaup–Boussinesq system {#SecKBous}
================================================================================
Let us begin with the Korteweg–de Vries equation $$\label{KdV}
u_{12;t}+u_{12;xxx}+6u_{12}u_{12;x}=0.$$ Its second Hamiltonian operator, $\smash{\hat{A}_2^{\text{KdV}}} = {{\mathrm d}}^3/{{\mathrm d}}x^3+4u_{12}\,{{\mathrm d}}/{{\mathrm d}}x
+2u_{12;x}$, which relates to the functional $H^{(2)}_{\text{KdV}}=-\tfrac{1}{2}\int u_{12}^2\,{{\mathrm d}}x$, can be extended[^1] in the $(2\mid2)$-graded field setup to the parity-preserving Hamiltonian operator [@MathieuTwo], $$\label{SecondHam4x4}
\hat{P}_2=\begin{pmatrix}
-\tfrac{{{\mathrm d}}}{{{\mathrm d}}x} & -u_2 & u_1 & 2u_0\tfrac{{{\mathrm d}}}{{{\mathrm d}}x}+2u_{0;x} \\
-u_2 & \bigl(\tfrac{{{\mathrm d}}}{{{\mathrm d}}x}\bigr)^2+u_{12} & -2u_0\tfrac{{{\mathrm d}}}{{{\mathrm d}}x}-u_{0;x} & 3u_1\tfrac{{{\mathrm d}}}{{{\mathrm d}}x}+2u_{1;x} \\
u_1 & 2u_0\tfrac{{{\mathrm d}}}{{{\mathrm d}}x}+u_{0;x} & \bigl(\tfrac{{{\mathrm d}}}{{{\mathrm d}}x}\bigr)^2+u_{12}\vphantom{\Bigr)} & 3u_2\tfrac{{{\mathrm d}}}{{{\mathrm d}}x}+2u_{2;x} \\
2u_0\tfrac{{{\mathrm d}}}{{{\mathrm d}}x} & -3u_1\tfrac{{{\mathrm d}}}{{{\mathrm d}}x}-u_{1;x} & -3u_2\tfrac{{{\mathrm d}}}{{{\mathrm d}}x}-u_{2;x} & \underline{\bigl(\tfrac{{{\mathrm d}}}{{{\mathrm d}}x}\bigr)^3+4u_{12}\tfrac{{{\mathrm d}}}{{{\mathrm d}}x}+2u_{12;x}}
\end{pmatrix}.$$ Here the fields $u_0$ and $u_{12}$ are bosonic, $u_1$ and $u_2$ are fermionic together with their derivatives w.r.t. $x$. Likewise, the components $\psi_0\simeq\delta{\mathcal{H}}/\delta u_0$ and $\psi_{12}\simeq\delta{\mathcal{H}}/\delta u_{12}$ of the arguments $\vec{\psi}={}^t\bigl(\psi_0,\psi_1,\psi_2,\psi_{12}\bigr)$ of are even-graded and $\psi_1$, $\psi_2$ are odd-graded. The operator is unique in the class of Hamiltonian total differential operators that merge to scalar $N{=}2$ super-operators which are local in ${\mathcal{D}}_i$ and whose coefficients depend on the super-field ${\boldsymbol{u}}$ and its super-derivatives, see below. The operator determines the $N{=}2$ classical super-conformal algebra [@ChaichianKulish87]. Conversely, the Poisson bracket given by reduces to the second Poisson bracket for , whenever one sets equal to zero the fields $u_0$, $u_1$, and $u_2$ both in the coefficients of and in all Hamiltonians; the operator $\hat{A}_2^{\text{KdV}}$ is underlined in . By construction, P. Mathieu’s extensions of the Korteweg–de Vries equation are determined by the operator and the bosonic Hamiltonian functional $$\label{HamComponents}
{\mathcal{H}}^{(2)}=\int\Bigl[u_0u_{0;xx}\underline{-u_{12}^2}+u_1u_{1;x}+u_2u_{2;x}+
a\cdot\bigl(u_0^2u_{12}-2u_0u_1u_2\bigr)\Bigr]\,{{\mathrm d}}x,$$ which incorporates $H^{(2)}_{\text{KdV}}$ as the underlined term; similar to , the Hamiltonian will be realized by as the bosonic $N{=}2$ super-Hamiltonian. Now we have that $$u_{i;t}=\bigl(\hat{P}_2\bigr)_{ij}\bigl(\delta{\mathcal{H}}^{(2)}/\delta u_j\bigr),
\qquad i,j\in\{0,1,2,12\}.
$$This yields the system
\[SKdVComponents\] $$\begin{aligned}
u_{0;t}&=-u_{0;xxx}+\bigl(a u_0^3
-(a+2)u_0u_{12}+(a-1)u_1u_2\bigr)_x,
\label{GetmKdV}\\
u_{1;t}&=-u_{1;xxx}+\bigl(\phantom{+}(a+2)u_0u_{2;x}+(a-1)u_{0;x}u_2
-3u_1u_{12}+3a u_0^2u_1 \bigr)_x,\\
u_{2;t}&=-u_{2;xxx}+\bigl(-(a+2)u_0u_{1;x}-(a-1)u_{0;x}u_1
-3u_2u_{12}+3a u_0^2u_2 \bigr)_x,\\
\underline{u_{12;t}}&=\underline{-u_{12;xxx}-6u_{12}u_{12;x}}
+3au_{0;x}u_{0;xx}+(a+2)u_0u_{0;xxx}\notag\\
{}&{}\qquad{}+3u_1u_{1;xx}+3u_2u_{2;xx}
+3a\bigl(u_0^2u_{12} -2u_0u_1u_2\bigr)_x.\label{GetKdV}\end{aligned}$$
Obviously, it retracts to , which we underline in , under the reduction $u_0=0$, $u_1=u_2=0$. At all $a\in{\mathbb{R}}$, the Hamiltonian equals $$\label{BalanceHam}
\boldsymbol{{\mathcal{H}}}^{(2)}=\int\bigl(
\boldsymbol{u}{\mathcal{D}}_1{\mathcal{D}}_2(\boldsymbol{u})+\tfrac{a}{3}\boldsymbol{u}^3
\bigr){{\mathrm d}}\boldsymbol{\theta}{{\mathrm d}}x,\qquad
\text{where ${{\mathrm d}}\boldsymbol{\theta}={{\mathrm d}}\theta_1{{\mathrm d}}\theta_2$.}$$ Likewise, the structure , which is independent of $a$, produces the $N{=}2$ super-operator $$\label{SecondHamN=2}
\hat{\boldsymbol{P}}_2=
{\mathcal{D}}_1{\mathcal{D}}_2\tfrac{{{\mathrm d}}}{{{\mathrm d}}x}+2{\boldsymbol{u}}\tfrac{{{\mathrm d}}}{{{\mathrm d}}x}-{\mathcal{D}}_1({\boldsymbol{u}}){\mathcal{D}}_1-{\mathcal{D}}_2({\boldsymbol{u}}){\mathcal{D}}_2+2{\boldsymbol{u}}_x.$$ Thus we recover P. Mathieu’s super-equations [@MathieuNew], which are Hamiltonian with respect to and the functional : ${\boldsymbol{u}}_t=\hat{\boldsymbol{P}}_2\bigl(\tfrac{\delta}{\delta{\boldsymbol{u}}}
(\boldsymbol{{\mathcal{H}}}_2)\bigr)$. In component notation, super-equations are . The assumption that, for a given $a$, the super-system admits infinitely many integrals of motion yields the triplet $a\in\{-2,1,4\}$, see [@MathieuNew]. The same values of $a$ are exhibited by the Painlevé analysis for $N{=}2$ super-equations , see [@BourqueMathieu]. The three systems have the common second Poisson structure, which is given by , but the three ‘junior’ first Hamiltonian operators $\hat{\boldsymbol{P}}_1$ for them do not coincide [@KerstenBiHamA1; @MathieuTwo; @MathieuNew]. Moreover, system with $a{=}4$ is radically different from the other two, both from the Hamiltonian and Lax viewpoints.
The $N{=}2$ supersymmetric hierarchy of P. Mathieu’s $a{=}4$ Korteweg–de Vries equation is bi-Hamiltonian with respect to the local super-operator and the junior Hamiltonian operator[^2] $\hat{\boldsymbol{P}}{\mathstrut}_1^{a=4}={{{\mathrm d}}}/{{{\mathrm d}}x}$, which is obtained from $\hat{\boldsymbol{P}}{\mathstrut}_2^{a=4}$ by the shift ${\boldsymbol{u}}\mapsto{\boldsymbol{u}}+\boldsymbol{\lambda}$ of the super-field ${\boldsymbol{u}}$, see : $$\hat{\boldsymbol{P}}{\mathstrut}_1^{a=4}=\frac{{{\mathrm d}}}{{{\mathrm d}}x}=\frac{1}{2}\cdot
\frac{{{\mathrm d}}}{{{\mathrm d}}\boldsymbol{\lambda}}{\Bigr|}_{\boldsymbol{\lambda}=0}
\hat{\boldsymbol{P}}{\mathstrut}_2^{a=4}{\Bigr|}_{{\boldsymbol{u}}+\boldsymbol{\lambda}}.$$ The two operators are Poisson compatible and generate the tower of *nonlocal* higher structures $\hat{\boldsymbol{P}}_{k+2}=\bigl(\hat{\boldsymbol{P}}_2\circ\hat{\boldsymbol{P}}_1^{-1}\bigr)^{k}\circ\hat{\boldsymbol{P}}_2$, $k\geq1$, for the $N{=}2$, $a{=}4$–SKdV hierarchy, see [@YKSMagri; @JKGolovko2008]. Although $\hat{\boldsymbol{P}}_3$ is nonlocal (c.f. [@PalitChowdhury96]), its bosonic limits under yield the *local* third Hamiltonian structure $\hat{A}_{2}$ for the Kaup–Boussinesq equation, which determines the evolution along the second time $t_2\equiv\xi$ in the bosonic limit of the $N{=}2$, $a{=}4$–SKdV hierarchy see Proposition on p. .
The Kaup–Boussinesq system [@Kaup75] arising here is equivalent to the Kaup–Broer system (the difference amounts to notation). A bi-Hamiltonian $N{=}2$ super-extension of the latter is known from [@KuperSuperLongWaves]. A tri-Hamiltonian two-fermion $N{=}1$ super-extension of the Kaup-Broer system was constructed in [@BrunelliDas94] such that in the bosonic limit the three known Hamiltonian structures for the initial system are recovered. At the same time, a boson-fermion $N{=}1$ super-extension of the Kaup–Broer equation with two local and the nonlocal third Hamiltonian structures was derived in [@PalitChowdhury96]; seemingly, the latter equaled the composition $\hat{\boldsymbol{P}}_2\circ\hat{\boldsymbol{P}}_1^{-1}\circ
\hat{\boldsymbol{P}}_2$, but it remained to prove that the suggested nonlocal super-operator is skew-adjoint, that the bracket induced on the space of bosonic super-Hamiltonians does satisfy the Jacobi identity, and that the hierarchy flows produced by the nonlocal operator remain local.
There is a deep reason for the geometry of the $a{=}4$–SKdV to be exceptionally rich. All the three integrable $N{=}2$ supersymmetric KdV equations admit the Lax representations $L_{t_3}=[A^{(3)},L]$, see [@BonoraKrivonosSorin; @MathieuTwo; @MathieuOpen; @PopowiczLax]. For $a{=}4$, the four roots of the Lax operator $L_{a{=}4}=-({\mathcal{D}}_1{\mathcal{D}}_2+{\boldsymbol{u}})^2$, which are ${\mathcal{L}}_{1,\pm}=\pm{{\mathrm i}}({\mathcal{D}}_1{\mathcal{D}}_2+{\boldsymbol{u}})$, ${{\mathrm i}}^2=-1$, and the super-pseudodifferential operators ${\mathcal{L}}_{2,\pm}=\pm\tfrac{{{\mathrm d}}}{{{\mathrm d}}x}
+ \sum_{i>0} (\cdots)\cdot\bigl(\tfrac{{{\mathrm d}}}{{{\mathrm d}}x}\bigr)^{-i}$, generate the odd-index flows of the SKdV hierarchy via $L_{t_{2k+1}}=[({\mathcal{L}}_2^{2k+1})_{\geq0},L]$. In particular, we have $A_{a{=}4}^{(3)}=\bigl(L^{3/2}\bigr)_{\geq0}\mod({\mathcal{D}}_1{\mathcal{D}}_2+{\boldsymbol{u}})^3$. However, the *entire* $a{=}4$ hierarchy is reproduced in the Lax form via $({\mathcal{L}}_1^k{\mathcal{L}}_2)_{t_\ell}=
\bigl[\bigl({\mathcal{L}}_1^\ell{\mathcal{L}}_2\bigr)_{\geq0}, {\mathcal{L}}_1^k{\mathcal{L}}_2\bigr]$ for all $k\in{\mathbb{N}}$, c.f. [@KrivonosSorinToppan]. Hence the super-residues[^3] of the operators ${\mathcal{L}}_1^k{\mathcal{L}}_2$ are conserved. Consequently, unlike the other two, super-equation with $a{=}4$ admits twice as many constants of motion as there are for the super-equations with $a{=}-2$ or $a{=}1$. For convenience, let us recall that super-equations are homogeneous with respect to the weights $|{{\mathrm d}}/{{\mathrm d}}x|\equiv1$, $|{\boldsymbol{u}}|=1$, $|{{\mathrm d}}/{{\mathrm d}}t|=3$. Hence we conclude that, for each nonnegative integer $k$, there appears the nontrivial conserved density $\operatorname{Sres}{\mathcal{L}}_1^k{\mathcal{L}}_2$, see above, of weight $k+1$. The even weights also enter the play. Consequently, there are twice as many commuting super-flows assigned to the twice as many Hamiltonians.
The additional super-Hamiltonian $\boldsymbol{{\mathcal{H}}}^{(1)}=\tfrac{1}{2}\int{\boldsymbol{u}}^2{{\mathrm d}}\boldsymbol{\theta}{{\mathrm d}}x$ for with $a{=}4$, and the second structure , — or, equivalently, the first operator $\hat{\boldsymbol{P}}_1={{\mathrm d}}/{{\mathrm d}}x$ and the Hamiltonian $\boldsymbol{{\mathcal{H}}}^{(2)}$, or $\hat{\boldsymbol{P}}_3$ and $\boldsymbol{{\mathcal{H}}}^{(0)}=\int{\boldsymbol{u}}\,{{\mathrm d}}\boldsymbol{\theta}{{\mathrm d}}x$, see above, — generate the $N{=}2$ supersymmetric equation $$\label{Burg}
{\boldsymbol{u}}_\xi={\mathcal{D}}_1{\mathcal{D}}_2{\boldsymbol{u}}_x+4{\boldsymbol{u}}{\boldsymbol{u}}_x=
\hat{\boldsymbol{P}}_3\left(\frac{\delta}{\delta{\boldsymbol{u}}}(\boldsymbol{{\mathcal{H}}}^{(0)})\right)
= \hat{\boldsymbol{P}}_2\left(\frac{\delta}{\delta{\boldsymbol{u}}}(\boldsymbol{{\mathcal{H}}}^{(1)})\right)
= \hat{\boldsymbol{P}}_1\left(\frac{\delta}{\delta{\boldsymbol{u}}}(\boldsymbol{{\mathcal{H}}}^{(2)})\right),\quad \xi\equiv t_2.$$ Super-equation was referred to as the $N{=}2$ ‘Burgers’ equation in [@N=2Hirota; @Kiev2005] due to the recovery of ${\boldsymbol{u}}_\xi={\boldsymbol{u}}_{xx}+4{\boldsymbol{u}}{\boldsymbol{u}}_x$ on the diagonal $\theta_1=\theta_2$. On the other hand, the bosonic limit of is the tri-Hamiltonian ‘minus’ Kaup–Boussinesq system (see [@Kaup75] or [@PamukKale; @KuperIrish; @NutkuPavlov] and references therein) $$\label{BLimBurg}
u_{0;\xi}=\bigl(-u_{12}+2u_0^2\bigr)_x,\qquad
u_{12;\xi}=\bigl(u_{0;xx}+4u_0u_{12}\bigr)_x.$$ System is equivalent to the Kaup–Broer equation via an invertible substitution. In these terms, super-equation is a super-extension of the Kaup–Boussinesq system [@BrunelliDas94; @KuperSuperLongWaves; @PalitChowdhury96]. In their turn, the first three Poisson structures for with $a{=}4$ are reduced under to the respective *local* structures for , see Proposition on p. .
Our interest in the recursive production of the integrals of motion for grew after the discovery, see [@N=2Hirota], of new $n$-soliton solutions, $$\label{HirotaForm}
{\boldsymbol{u}}=\mathsf{A}(a)\cdot {\mathcal{D}}_1{\mathcal{D}}_2\log\Bigl(1+\sum_{i=1}^n \alpha_i\exp
\bigl(k_ix-k_i^3\cdot t\pm{{\mathrm i}}\,k_i\cdot\theta_1\theta_2\bigr)\Bigr),
\qquad
\mathsf{A}(a)=\left\{\begin{matrix}1,\ a{=}1,\\ \tfrac{1}{2},\ a{=}4,
\end{matrix}\right.$$ for the super-equations with $a{=}1$ or $a{=}4$ (but not $a{=}-2$ or any other $a\in{\mathbb{R}}\setminus\{1,4\}$). In formula , the wave numbers $k_i\in{\mathbb{R}}$ are arbitrary, and the phases $\alpha_i$ can be rescaled to $+1$ for non-singular $n$-soliton solutions by appropriate shifts of $n$ higher times in the SKdV hierarchy. A spontaneous decay of fast solitons and their transition into the virtual states, on the emerging background of previously invisible, slow solitons, look paradoxal for such KdV-type systems ($a{=}1$ or $a{=}4$), since they possess an infinity of the integrals of motion. The new solutions of with $a{=}1$ or $a{=}4$ are subject to the condition and therefore satisfy the bosonic limits of these $N{=}2$ super-systems. In the same way, the bosonic limit of admits multi-soliton solutions in Hirota’s form , now with the exponents $\eta_i=k_ix\pm{{\mathrm i}}k_i^2\xi\pm{{\mathrm i}}k_i\theta_1\theta_2$, see [@N=2Hirota]. This makes the role of such two-component bosonic reductions particularly important. We recall that the reduction of with $a{=}1$ yields the Kersten–Krasil’shchik equation, see [@JKKerstenEq] or [@N=2Hirota] and references therein. In this paper, we consider the bosonic limit of the $N{=}2$, $a{=}4$ SKdV equation,
\[BLim\] $$\begin{aligned}
u_{0;t}&=-u_{0;xxx}+12 u_0^2 u_{0;x}
-6\bigl(u_0u_{12}\bigr)_x,\\
u_{12;t}&=-u_{12;xxx}-6u_{12}u_{12;x}
+12 u_{0;x}u_{0;xx}+6u_0u_{0;xxx}
+12\bigl(u_0^2u_{12}\bigr)_x,\end{aligned}$$
which succeeds the Kaup–Boussinesq equation in its tri-Hamiltonian hierarchy. We construct a new Gardner deformation for it (c.f. [@PamukKale]). In general, system with $a{=}4$ admits three one-component reductions (except $u_0\not\equiv0$) and three two-component reductions, which are indicated by the edges that connect the remaining components in the diagram $$\begin{CD}
{} @. u_0 @. {} \\
@. @| @. \\
u_1 @= u_{12} @= u_2.
\end{CD}$$ System with $a{=}4$ has no three-component reductions obtained by setting to zero only one of the four fields in . We conclude this paper by presenting a Gardner deformation for the two-component boson-fermion reduction $u_0\equiv0$, $u_2\equiv0$ of the $N{=}2$, $a{=}4$–SKdV system, see on p. .
Deformation problem for $N{=}2$, $a{=}4$–SKdV equation {#SecGardner}
======================================================
In this section, we formulate the two-step algorithm for a recursive production of the bosonic super-Hamiltonians $\boldsymbol{{\mathcal{H}}}^{(k)}[{\boldsymbol{u}}]$ for the $N{=}2$ supersymmetric $a{=}4$–SKdV hierarchy. Essentially, we convert the geometric problem to an explicit computational procedure. Our scheme can be applied to other KdV-type super-systems (in particular, to with $a{=}-2$ or $a{=}1$). By definition, a classical Gardner’s deformation for an integrable evolutionary equation ${\mathcal{E}}$ is the diagram $${\mathfrak{m}}_\epsilon\colon{\mathcal{E}}(\epsilon)\to{\mathcal{E}},$$ where the equation ${\mathcal{E}}(\epsilon)$ is a parametric extension of the initial system ${\mathcal{E}}={\mathcal{E}}(0)$ and ${\mathfrak{m}}_\epsilon$ is the Miura contraction [@Gardner; @KuperIrish; @TMPh2006]. Under the assumption that ${\mathcal{E}}(\epsilon)$ be in the form of a (super-)conserved current, the Taylor coefficients $\tilde{{\boldsymbol{u}}}^{(k)}$ of the formal power series $\tilde{{\boldsymbol{u}}}=\sum_{k=0}^{+\infty}\tilde{{\boldsymbol{u}}}^{(k)}\cdot\epsilon^k$ are termwise conserved on ${\mathcal{E}}(\epsilon)$ and hence on ${\mathcal{E}}$. Therefore, the contraction ${\mathfrak{m}}_\epsilon$ yields the recurrence relations, ordered by the powers of $\epsilon$, between these densities $\tilde{{\boldsymbol{u}}}^{(k)}$, while the equality ${\mathcal{E}}(0)={\mathcal{E}}$ specifies its initial condition.
\[ExKdVe\] The contraction
\[DefKdV\] $$\begin{aligned}
{\mathfrak{m}}_\epsilon&=\smash{\bigl\{u_{12}={\tilde{u}}_{12}
\pm\epsilon{\tilde{u}}_{12;x}-\epsilon^2{\tilde{u}}_{12}^2\bigr\}}\label{KdVeKdV}\\
\intertext{maps solutions ${\tilde{u}}_{12}(x,t;\epsilon)$ of
the extended equation~${\mathcal{E}}(\epsilon)$,}
{\tilde{u}}_{12;t}&+\bigl({\tilde{u}}_{12;xx}+3{\tilde{u}}_{12}^2
-2\epsilon^2\cdot{\tilde{u}}_{12}^3\bigr)_x=0,\label{KdVe}\end{aligned}$$
to solutions $u_{12}(x,t)$ of the Korteweg–de Vries equation . Plugging the series ${\tilde{u}}_{12}=\sum_{k=0}^{+\infty}u_{12}^{(k)}\cdot\epsilon^k$ in ${\mathfrak{m}}_\epsilon$ for ${\tilde{u}}_{12}$, we obtain the chain of equations ordered by the powers of $\epsilon$, $$u_{12}=\sum_{k=0}^{+\infty}{\tilde{u}}_{12}^{(k)}\cdot\epsilon^k
\pm {\tilde{u}}_{12;x}^{(k)}\cdot\epsilon^{k+1}
-\sum_{\substack{i+j=k\\ i,j\geq0}} {\tilde{u}}_{12}^{(i)}{\tilde{u}}_{12}^{(j)}\cdot
\epsilon^{k+2}.$$ Let us fix the plus sign in by reversing $\epsilon\to-\epsilon$ if necessary. Equating the coefficients of $\epsilon^k$, we obtain the relations $$u={\tilde{u}}_{12}^{(0)},\qquad
0={\tilde{u}}_{12}^{(1)}+{\tilde{u}}_{12;x}^{(0)},\qquad
0={\tilde{u}}_{12}^{(k)}+{\tilde{u}}_{12;x}^{(k-1)}-\sum_{\substack{i+j=k-2\\ i,j\geq0}}
{\tilde{u}}_{12}^{(i)}{\tilde{u}}_{12}^{(j)},\quad k\geq2.$$ Hence, from the initial condition ${\tilde{u}}_{12}^{(0)}= u_{12}$, we recursively generate the densities $$\begin{aligned}
{\tilde{u}}_{12}^{(1)} &= - u_{12;x},\qquad
{\tilde{u}}_{12}^{(2)} = u_{12;xx} - u_{12}^2,\qquad
{\tilde{u}}_{12}^{(3)} = - u_{12;xxx} + 4u_{12;x}u_{12},\\
{\tilde{u}}_{12}^{(4)} &= u_{12;4x} - 6u_{12;xx}u_{12} - 5u_{12;x}^2 + 2u_{12}^3,\\
{\tilde{u}}_{12}^{(5)} &= - u_{12;5x} + 8u_{12;xxx}u_{12} + 18u_{12;xx}u_{12;x} - 16u_{12;x}u_{12}^2,\\
{\tilde{u}}_{12}^{(6)} &= u_{12;6x} - 10u_{12;4x}u_{12} - 28u_{12;xxx}u_{12;x} - 19u_{12;xx}^2 + 30u_{12;xx}u_{12}^2 + 50u_{12;x}^2u_{12} - 5u_{12}^4,\\
{\tilde{u}}_{12}^{(7)} &= - u_{12;7x} + 12u_{12;5x}u_{12} + 40u_{12;4x}u_{12;x} + 68u_{12;xxx}u_{12;xx} - 48u_{12;xxx}u_{12}^2\\
{}&{}\qquad{} - 216u_{12;xx}u_{12;x}u_{12} - 60u_{12;x}^3 + 64u_{12;x}u_{12}^3,\quad \text{\textit{etc}}.\end{aligned}$$ The conservation ${\tilde{u}}_{12;t}=\tfrac{{{\mathrm d}}}{{{\mathrm d}}x}\bigl(\cdot\bigr)$ implies that each coefficient $u_{12}^{(k)}$ is conserved on .
The densities $u_{12}^{(2k)}=c(k)\cdot u_{12}^k+\dots$, $c(k)=\text{const}$, determine the Hamiltonians ${\mathcal{H}}_{12}^{(k)}=\int h_{12}^{(k)}[u_{12}]\,{{\mathrm d}}x$ of the renowned KdV hierarchy. Let us show that all of them are nontrivial. Consider the zero-order part $\breve{u}_{12}^{\text{KdV}}$ such that ${\tilde{u}}_{12}\bigl([u_{12}],\epsilon\bigr)=\breve{u}_{12}^{\text{KdV}}(u_{12},\epsilon)+\dots$, where the dots denote summands containing derivatives of $u_{12}$. Taking the zero-order component of , we conclude that the generating function $\breve{u}_{12}^{\text{KdV}}$ satisfies the algebraic recurrence relation $u_{12}=\breve{u}_{12}^{\text{KdV}}-\epsilon^2\bigl(\breve{u}_{12}^{\text{KdV}}\bigr)^2$. We choose the root by the initial condition $\breve{u}_{12}^{\text{KdV}}{\bigr|}_{\epsilon=0}=u_{12}$, which yields $$\label{GenFnKdV}
\breve{u}_{12}^{\text{KdV}}=\Bigl( 1-\sqrt{1-4\epsilon^2 u_{12}}
\bigr)\bigr/(2\epsilon^2).$$ Moreover, the Taylor coefficients $\breve{u}_{12}^{(k)}(u_{12})$ in $\breve{u}_{12}^{\text{KdV}}=\sum_{k=0}^{+\infty}\breve{u}_{12}^{(k)}\cdot
\epsilon^{2k}$ equal $c(k)\cdot u_{12}^{k+1}$, where $c(k)$ are positive and grow with $k$. This is readily seen by induction over $k$ with the base $\breve{u}_{12}^{(0)}=u_{12}$. Expanding both sides of the equality $u_{12}=\breve{u}_{12}^{\text{KdV}}-\epsilon^2\cdot\bigl(\breve{u}_{12}^{\text{KdV}}\bigr)^2$ in $\epsilon^2$, we notice that $$\breve{u}_{12}^{(k)}=\sum_{\substack{i+j=k-1,\\ i,j\geq0}}
\breve{u}_{12}^{(i)}\cdot\breve{u}_{12}^{(j)}
=\sum_{i+j=k-1} c(i)c(j)\cdot u_{12}^{k+1}.$$ Therefore, the next coefficient, $c(k)=\sum_{i+j=k-1}c(i)\cdot c(j)$, is the sum over $i,j\geq0$ of products of positive numbers, whence $c(k+1)>c(k)>0$. This proves the claim.
Let us list the densities $h_{\text{KdV}}^{(k)}\sim u_{12}^{(2k)}\mod\operatorname{im}{{\mathrm d}}/{{\mathrm d}}x$ of the first seven Hamiltonians for . These will be correlated in section \[SecHam\] with the lowest seven Hamiltonians for , see [@MathieuNew] and below. We have $$\begin{aligned}
h_{\text{KdV}}^{(1)} &= u_{12}^2, \qquad
h_{\text{KdV}}^{(2)} = 2u_{12}^3 - u_{12;x}^2 + 2u_{12}^3 + u_{12;xx}, \qquad
h_{\text{KdV}}^{(3)} = 5u_{12}^4 + 5u_{12;xx}u_{12}^2 + u_{12;xx}^2, \\
h_{\text{KdV}}^{(4)} &= - 14u_{12}^5 + 70u_{12}^2u_{12;x}^2 + 14u_{12}u_{12;xxx}u_{12;x} + u_{12;xxx}^2, \\
h_{\text{KdV}}^{(5)} &= 42u_{12}^6 - 420u_{12}^3u_{12;x}^2 + 9u_{12}^2u_{12;6x} + 126u_{12}^2u_{12;xx}^2 + u_{12;4x}^2 - 7u_{12;xx}^3 - 35u_{12;x}^4, \\
h_{\text{KdV}}^{(6)} &= 1056u_{12}^7 - 18480u_{12}^4u_{12;x}^2 + 7392u_{12}^3u_{12;xx}^2 + 55u_{12}^2u_{12;8x} - 1584u_{12}^2u_{12;xxx}^2\\
{}&{}\qquad{} + 66u_{12}u_{12;4x}^2 + 3520u_{12}u_{12;xx}^3
- 6160u_{12}u_{12;x}^4 - 8u_{12;5x}^2 + 3696u_{12;xx}^2u_{12;x}^2, \\
h_{\text{KdV}}^{(7)} &= 15444u_{12}^8 - 432432u_{12}^5u_{12;x}^2 + 4004u_{12}^4u_{12;6x} + 216216u_{12}^4u_{12;xx}^2 + 2145u_{12}^3u_{12;8x} \\
{}&{}\qquad - 45760u_{12}^3u_{12;xxx}^2
+ 3861u_{12}^2u_{12;4x}^2 + 133848u_{12}^2u_{12;xx}^3 - 360360u_{12}^2u_{12;x}^4 \\
{}&{}\qquad{}- 936u_{12}u_{12;5x}^2 + 36u_{12;6x}^2 + 6552u_{12;4x}^2u_{12;xx}
+ 72072u_{12;xxx}^2u_{12;x}^2 - 28314u_{12;xx}^4.\end{aligned}$$ At the same time, the densities $u_{12}^{(2k+1)}=\tfrac{d}{dx}\bigl(\cdot\bigr)\sim0$ are trivial. Indeed, for $\omega_0\mathrel{{:}{=}}\sum_{k=0}^{+\infty}u_{12}^{(2k)}\cdot\epsilon^{2k}$ and $\omega_1\mathrel{{:}{=}}\sum_{k=0}^{+\infty}u_{12}^{(2k+1)}\cdot\epsilon^{2k}$ such that ${\tilde{u}}=\omega_0+\epsilon\cdot\omega_1$, we equate the odd powers of $\epsilon$ in and obtain $\omega_1=\tfrac{1}{2\epsilon^2}\tfrac{d}{dx}
\log\bigl(1-2\epsilon^2\omega_0\bigr)$.
In what follows, using the deformation of , we fix the coefficients of differential monomials in $u_{12}$ within a bigger deformation problem (see section \[SecBurg\]) for the two-component system .
We split the Gardner deformation problem for the $N{=}2$ supersymmetric hierarchy of with $a{=}4$ in two main and several auxiliary steps. First, we note that Miura’s contraction ${\mathfrak{m}}_\epsilon\colon{\mathcal{E}}(\epsilon)\to{\mathcal{E}}$, which encodes the recurrence relation between the conserved densities, is common for all equations of the hierarchy. Indeed, the densities (and hence any differential relations between them) are shared by all the equations. Therefore, we pass to the deformation problem for the $N{=}2$ super-Burgers equation . This makes the first simplification of the Gardner deformation problem for the $N{=}2$, $a{=}4$ super-KdV hierarchy. Second, let ${{\boldsymbol{h}}}^{(k)}$ be an $N{=}2$ super-conserved density for an evolutionary super-equation ${\mathcal{E}}$, meaning that its velocity w.r.t. a time $\tau$, $\tfrac{{{\mathrm d}}}{{{\mathrm d}}\tau}{{{\boldsymbol{h}}}}^{(k)}={\mathcal{D}}_1(\dots)+{\mathcal{D}}_2(\dots)$, is a total divergence on ${\mathcal{E}}$. By definition of ${\mathcal{D}}_i$, see , the $\theta_1\theta_2$-component $h_{12}^{(k)}$ of such ${{\boldsymbol{h}}}^{(k)}=h_0^{(k)}+\theta_1\cdot h_1^{(k)}+\theta_2\cdot h_2^{(k)}+
\theta_1\theta_2\cdot h_{12}^{(k)}$ is conserved in the classical sense, $\tfrac{{{\mathrm d}}}{{{\mathrm d}}\tau}{h}_{12}^{(k)}=\tfrac{{{\mathrm d}}}{{{\mathrm d}}x}(\dots)$ on ${\mathcal{E}}$. Let us consider the correlation between the conservation laws for the full $N{=}2$ super-system ${\mathcal{E}}$ and for its reductions that are obtained by setting certain component(s) of ${\boldsymbol{u}}$ to zero. In what follows, we study the bosonic reduction . Other reductions of the super-equation are discussed in section \[SecHam\], see on p. .
We suppose that the bosonic limit $\lim_B{\mathcal{E}}$ of the super-equation ${\mathcal{E}}$ exists, which is the case for and . By the above, each conserved super-density ${{\boldsymbol{h}}}^{(k)}[{\boldsymbol{u}}]$ determines the conserved density $h_{12}^{(k)}[u_0,u_{12}]$, which may become trivial. As in [@BonoraKrivonosSorin], we assume that the super-system ${\mathcal{E}}$ does not admit any conserved super-densities that vanish under the reduction . Then, for such $h_{12}^{(k)}$ that originates from ${{\boldsymbol{h}}}^{(k)}$ by construction, the equivalence class $\{{{\boldsymbol{h}}}^{(k)}\mod\operatorname{im}{\mathcal{D}}_i\}$ is uniquely determined by $$\int h_{12}^{(k)}[u_0,u_{12}]\,{{\mathrm d}}x=\int{{\boldsymbol{h}}}^{(k)}[{\boldsymbol{u}}]{\bigr|}_{u_1=u_2=0}{{\mathrm d}}\boldsymbol{\theta}{{\mathrm d}}x,\qquad\text{here $N{=}2$ and
${{\mathrm d}}\boldsymbol{\theta}={{\mathrm d}}\theta_1{{\mathrm d}}\theta_2$.}$$ Berezin’s definition of a super-integration, $\int{{\mathrm d}}\theta_i=0$ and $\int\theta_i\,{{\mathrm d}}\theta_i=1$, implies that the problem of recursive generation of the $N{=}2$ super-Hamiltonians $\boldsymbol{{\mathcal{H}}}^{(k)}=\int{{\boldsymbol{h}}}^{(k)}\,{{\mathrm d}}\boldsymbol{\theta}{{\mathrm d}}x$ for the SKdV hierarchy amounts to the generation of the equivalence classes $\int h_{12}^{(k)}\,{{\mathrm d}}x$ for the respective $\theta_1\theta_2$-component. We conclude that a solution of Gardner’s deformation problem for the supersymmetric system may not be subject to the supersymmetry invariance. This is a key point to further reasonings.
We stress that the equivalence class of such functions $h_{12}^{(k)}[u_0,u_{12}]$ that originate from $\boldsymbol{{\mathcal{H}}}^{(k)}$ by is, generally, much more narrow than the equivalence class $\{h_{12}^{(k)}\mod\operatorname{im}{{\mathrm d}}/{{\mathrm d}}x\}$ of all conserved densities for the bosonic limit $\lim_B{\mathcal{E}}$. Obviously, there are differential functions of the form $\tfrac{{{\mathrm d}}}{{{\mathrm d}}x}\bigl(f[u_0,u_{12}]\bigr)$ that can not be obtained[^4] as the $\theta_1\theta_2$-component of any $\bigl[{\mathcal{D}}_1(\cdot)+{\mathcal{D}}_2(\cdot)\bigr]{\bigr|}_{u_1=u_2=0}$, which is trivial in the super-sense. Therefore, let $h_{12}^{(k)}$ be *any* recursively given sequence of integrals of motion for $\lim_B{\mathcal{E}}$ (*e.g.*, suppose that they are the densities of the Hamiltonians ${\mathcal{H}}^{(k)}$ for the hierarchy of $\lim_B{\mathcal{E}}$), and let it be known that each ${\mathcal{H}}^{(k)}=\int h_{12}^{(k)}\,{{\mathrm d}}x$ does correspond to the super-analogue $\boldsymbol{{\mathcal{H}}}^{(k)}=\int{{\boldsymbol{h}}}^{(k)}\,{{\mathrm d}}\boldsymbol{\theta}{{\mathrm d}}x$. Then the reconstruction of ${{\boldsymbol{h}}}^{(k)}$ requires an intermediate step, which is the elimination of excessive, homologically trivial terms under ${{\mathrm d}}/{{\mathrm d}}x$ that preclude a given $h_{12}^{(k)}$ to be extended to the full super-density in terms of the $N{=}2$ super-field ${\boldsymbol{u}}$. This is illustrated in section \[SecHam\].
Thirdly, the gap between the two types of equivalence for the integrals of motion manifests the distinction between the deformations $\bigl(\lim_B{\mathcal{E}}\bigr)(\epsilon)$ of bosonic limits and, on the other hand, the bosonic limits $\lim_B{\mathcal{E}}(\epsilon)$ of $N{=}2$ super-deformations. The two operations, Gardner’s extension of ${\mathcal{E}}$ to ${\mathcal{E}}(\epsilon)$ and taking the bosonic limit $\lim_B{\mathcal{F}}$ of an equation ${\mathcal{F}}$, are not permutable. The resulting systems can be different. Namely, according to the classical scheme ([@Gardner], [@TMPh2006]), *each* equation in the evolutionary system $\bigl(\lim_B{\mathcal{E}}\bigr)(\epsilon)$ represents a conserved current, whence each Taylor coefficient of the respective field is conserved, see Example \[ExKdVe\]. At the same time, for $\lim_B{\mathcal{E}}(\epsilon)$, the conservation is required only for the field ${\tilde{u}}_{12}(\epsilon)$, which is the $\theta_1\theta_2$-component of the extended super-field $\tilde{{\boldsymbol{u}}}(\epsilon)$. Other equations in $\lim_B{\mathcal{E}}(\epsilon)$ can have any form.[^5] In this notation, we strengthen the problem of recursive generation of the super-Hamiltonians for the $N{=}2$ super-equation . Namely, in section \[SecBurg\] we construct true Gardner’s deformations for its two-component bosonic limit . Moreover, the known deformation for upon the component $u_{12}$ of allows to fix the coefficients of the terms that contain only $u_{12}$ or its derivatives. The solution to the Gardner deformation problem generates the recurrence relation between the nontrivial conserved densities $h_{12}^{(k)}$ which, in the meantime, depend on $u_0$ and $u_{12}$. By correlating them with the $\theta_1\theta_2$-components of the super-densities ${{\boldsymbol{h}}}^{(k)}$ that depend on ${\boldsymbol{u}}$, we derive the Hamiltonians $\boldsymbol{{\mathcal{H}}}^{(k)}$, $k\geq0$, for the $N{=}2$ supersymmetric $a{=}4$–KdV hierarchy, see section \[SecHam\].
New deformation of the Kaup–Boussinesq equation {#SecBurg}
===============================================
In this section, we construct a new Gardner’s deformation ${\mathfrak{m}}_\epsilon\colon\bigl(\lim_B{\mathcal{E}}\bigr)(\epsilon)\to\lim_B{\mathcal{E}}$ for the ‘minus’ Kaup–Boussinesq equation , which is the bosonic limit of the $N{=}2$ supersymmetric system . We will use the known deformation to fix several coefficients in the Miura contraction ${\mathfrak{m}}_\epsilon$, which ensures the difference of the new solution – from previously known deformations of , see [@PamukKale]. We prove that the new deformation is maximally nontrivial: It yields infinitely many nontrivial conserved densities, and none of the Hamiltonians is lost. In components, the $N{=}2$ super-equation reads $$\begin{aligned}
u_{0;\xi}&=\bigl(-u_{12}+2u_0^2\bigr)_x, &
u_{1;\xi}&=\bigl(u_{2;x}+4u_0u_1\bigr)_x,\\
u_{2;\xi}&=\bigl(-u_{1,x}+4u_0u_2\bigr)_x, &
u_{12;\xi} &= \bigl(u_{0;xx}+4u_0u_{12}-4u_1u_2\bigr)_x.\end{aligned}$$ Clearly, it admits the reduction ; moreover, the Kaup–Boussinesq system is the only possible limit for . Let us summarize its well-known properties [@Kaup75; @NutkuPavlov]:
\[ThTriHam\] The completely integrable Kaup–Boussinesq system inherits the local tri-Hamiltonian structure from the the two local ($\hat{P}_1$ and $\hat{P}_2$) and the nonlocal $\hat{P}_3=\hat{P}_2\circ\hat{P}_1\circ\hat{P}_2$ operators for the $N{=}2$, $a{=}4$–SKdV hierarchy under the bosonic limit $$\begin{gathered}
\binom{u_0}{u_{12}}_\xi
=\hat{A}_1^{12}\binom{\delta/\delta u_0}{\delta/\delta u_{12}}
\Bigl(\int
\bigl[2u_0^2u_{12}-\tfrac{1}{2}u_{12}^2-\tfrac{1}{2}u_{0;x}^2\bigr]\,{{\mathrm d}}x\Bigr)\\
=\hat{A}_1^{0}\binom{\delta/\delta u_0}{\delta/\delta u_{12}}
\Bigl(-\int u_0u_{12}\,{{\mathrm d}}x\Bigr)
=\hat{A}_2\binom{\delta/\delta u_0}{\delta/\delta u_{12}}
\Bigl(-\int u_{12}\,{{\mathrm d}}x\Bigr).\end{gathered}$$ The senior Hamiltonian operator $\hat{A}_2$ is $$\begin{pmatrix}
u_{0;x}+2u_0\,\tfrac{{{\mathrm d}}}{{{\mathrm d}}x} &
u_{12;x}-4u_0u_{0;x}-2u_0^2\,\tfrac{{{\mathrm d}}}{{{\mathrm d}}x}+2u_{12}\,\tfrac{{{\mathrm d}}}{{{\mathrm d}}x}
+\tfrac{1}{2}\left(\tfrac{{{\mathrm d}}}{{{\mathrm d}}x}\right)^3\vphantom{\Bigr)}\\
u_{12;x}-2u_0^2\,\tfrac{{{\mathrm d}}}{{{\mathrm d}}x}+2u_{12}\,\tfrac{{{\mathrm d}}}{{{\mathrm d}}x}
+\tfrac{1}{2}\left(\tfrac{{{\mathrm d}}}{{{\mathrm d}}x}\right)^3 &
-4u_0u_{12}\,\tfrac{{{\mathrm d}}}{{{\mathrm d}}x}-4\tfrac{{{\mathrm d}}}{{{\mathrm d}}x}\circ u_0u_{12}
-u_0\,\left(\tfrac{{{\mathrm d}}}{{{\mathrm d}}x}\right)^3
-\left(\tfrac{{{\mathrm d}}}{{{\mathrm d}}x}\right)^3\circ u_0
\end{pmatrix}.$$ The junior Hamiltonian operators $\hat{A}_1^{0}$ and $\hat{A}_1^{12}$ are obtained from $\hat{A}_2$ by the shifts of the respective fields, c.f. $$\begin{aligned}
\hat{A}_1^0&=\begin{pmatrix}
\tfrac{{{\mathrm d}}}{{{\mathrm d}}x} & -2u_{0;x}-2u_0\,\tfrac{{{\mathrm d}}}{{{\mathrm d}}x} \\
-2u_0\,\tfrac{{{\mathrm d}}}{{{\mathrm d}}x} &
-2u_{12;x}-4u_{12}\,\tfrac{{{\mathrm d}}}{{{\mathrm d}}x}-\left(\tfrac{{{\mathrm d}}}{{{\mathrm d}}x}\right)^3
\end{pmatrix}
=\frac{1}{2}\cdot{\left.\frac{{{\mathrm d}}}{{{\mathrm d}}\lambda}\right|}_{\lambda=0}
\hat{A}_2{\Bigr|}_{u_0+\lambda}\\
\intertext{and}
\hat{A}_1^{12}&=\begin{pmatrix} 0 & \tfrac{{{\mathrm d}}}{{{\mathrm d}}x} \\
\tfrac{{{\mathrm d}}}{{{\mathrm d}}x} & 0\end{pmatrix}
=\frac{1}{2}\cdot{\left.\frac{{{\mathrm d}}}{{{\mathrm d}}\mu}\right|}_{\mu=0}
\hat{A}_2{\Bigr|}_{u_{12}+\mu}.\end{aligned}$$ The three operators $\hat{A}_1^0$, $\hat{A}_1^{12}$, and $\hat{A}_2$ are Poisson compatible.
The Kaup–Boussinesq equation admits an infinite sequence of integrals of motion. We will derive them via the Gardner deformation. Unlike in [@PamukKale], from now on we always assume that is recovered under ${\tilde{u}}_0\equiv0$.
We assume that both the extension ${\mathcal{E}}(\epsilon)$ of and the contraction ${\mathfrak{m}}_\epsilon\colon{\mathcal{E}}(\epsilon)\to{\mathcal{E}}$ into are homogeneous polynomials in $\epsilon$. From now on, we denote the reduction by ${\mathcal{E}}$.
First, let us estimate the degrees in $\epsilon$ for such polynomials ${\mathcal{E}}(\epsilon)$ and ${\mathfrak{m}}_\epsilon$, by balancing the powers of $\epsilon$ in the left- and right-hand sides of with $u_0$ and $u_{12}$ replaced by the Miura contraction ${\mathfrak{m}}_\epsilon=\bigl\{u_0=u_0\bigl([{\tilde{u}}_0,{\tilde{u}}_{12}],\epsilon\bigr)$, $u_{12}=u_{12}\bigl([{\tilde{u}}_0,{\tilde{u}}_{12}],\epsilon\bigr)\bigr\}$. The time evolution in the left-hand side, which is of the form $u_\xi=\partial_{{\tilde{u}}_\xi}({\mathfrak{m}}_\epsilon)$ by the chain rule, sums the degrees in $\epsilon$: $\deg u_\xi=\deg{\mathfrak{m}}_\epsilon+\deg{\mathcal{E}}(\epsilon)$. At the same time, we notice that system is only quadratic-nonlinear. Hence its right-hand side, with ${\mathfrak{m}}_\epsilon$ substituted for $u_0$ and $u_{12}$, gives the degree $2\times\deg{\mathfrak{m}}_\epsilon$, irrespective of $\deg{\mathcal{E}}(\epsilon)$. Consequently, we obtain the balance[^6] $1:1$ for $\max\deg{\mathfrak{m}}_\epsilon:\max\deg{\mathcal{E}}(\epsilon)$. This is in contrast with the balance $1:2$ for polynomial deformations of the bosonic limit for the initial SKdV system , which is cubic-nonlinear[^7] (c.f. [@MathieuNew]).
Obviously, a lower degree polynomial extension ${\mathcal{E}}(\epsilon)$ contains fewer undetermined coefficients. This is the first profit we gain from passing to instead of . By the same argument, we conclude that ${\mathfrak{m}}_\epsilon\colon{\mathcal{E}}(\epsilon)\to{\mathcal{E}}$, viewed as the algebraic system upon these coefficients, is only *quadratic*-nonlinear w.r.t. the coefficients in ${\mathfrak{m}}_\epsilon$ (and, obviously, *linear* w.r.t. the coefficients in ${\mathcal{E}}(\epsilon)$; this is valid for any balance $\deg{\mathfrak{m}}_\epsilon:\deg{\mathcal{E}}(\epsilon)$). Hence the size of this overdetermined algebraic system is further decreased. Second, we use the unique admissible homogeneity weights for the Kaup–Boussinesq system , $$|u_0|=1,\quad |u_{12}|=2,\quad |{{\mathrm d}}/{{\mathrm d}}\xi|=2;$$ here $|{{\mathrm d}}/{{\mathrm d}}x|\equiv 1$ is the normalization. The Miura contraction ${\mathfrak{m}}_\epsilon=\bigl\{u_0={\tilde{u}}_0+\epsilon\cdot(\dots)$, $u_{12}={\tilde{u}}_{12}+\epsilon\cdot(\dots)\bigr\}$, which we assume regular at the origin, implies that $|{\tilde{u}}_0|=1$ and $|{\tilde{u}}_{12}|=2$ as well. We let $|\epsilon|=-1$ be the difference of weights for every two successive Hamiltonians for the $N{=}2$, $a{=}4$–SKdV hierarchy, see [@MathieuNew] and below. In this setup, all functional coefficients of the powers $\epsilon^k$ both in ${\mathcal{E}}(\epsilon)$ and ${\mathfrak{m}}_\epsilon$ are homogeneous differential polynomials in $u_0$, $u_{12}$, and their derivatives w.r.t. $x$. It is again important that the time $\xi$ of weight $|{{\mathrm d}}/{{\mathrm d}}\xi|=2$ in precedes the time $t$ with $|{{\mathrm d}}/{{\mathrm d}}t|=3$ in the hierarchy of , where $|\theta_i|=-\tfrac{1}{2}$ and $|{\boldsymbol{u}}|=1$. As before, we have further decreased the number of undetermined coefficients. The polynomial ansatz for Gardner’s deformation of is generated by the procedure[^8] `GenSSPoly`, which is a new possibility in the the analytic software [@SsTools]. We thus obtain the determining system ${\mathfrak{m}}_\epsilon\colon{\mathcal{E}}(\epsilon)\to{\mathcal{E}}$. Using <span style="font-variant:small-caps;">SsTools</span>, we split it to the overdetermined system of algebraic equations, which are linear w.r.t. ${\mathcal{E}}(\epsilon)$ and quadratic-nonlinear w.r.t. ${\mathfrak{m}}_\epsilon$. Moreover, we claim that this system is *triangular*. Indeed, it is ordered by the powers of $\epsilon$, since the determining system is identically satisfied at zeroth order and because equations at lower orders of $\epsilon$ involve only the coefficients of its lower powers from ${\mathfrak{m}}_\epsilon$ and ${\mathcal{E}}(\epsilon)$.
Thirdly, we use the deformation of the Korteweg–de Vries equation [@Gardner]. We recall that
- Miura’s contraction ${\mathfrak{m}}_\epsilon$ is common for all two-component systems in the bosonic limit, see , of the $N{=}2$, $a{=}4$–SKdV hierarchy;
- for any $a$, the bosonic limit of , see and , incorporates the Korteweg–de Vries equation .
Using , we fix those coefficients in ${\mathfrak{m}}_\epsilon$ which depend only on $u_{12}$ and its derivatives, but not on $u_{0}$ or its derivatives. Apparently, we discard the knowledge of such coefficients in the extension of the bosonic limit , since for us now it is not the object to be deformed. But the minimization of the algebraic system, which we have achieved by passing to , is so significant that this temporary loss in inessential. Furthermore, the above reasoning shows that the recovery of the coefficients in the extension ${\mathcal{E}}(\epsilon)$ amounts to solution of linear equations, while finding the coefficients in ${\mathfrak{m}}_\epsilon$ would cost us the necessity to solve nonlinear algebraic systems. We managed to fix some of those constants for granted.
We finally remark that the normalization of at least one coefficient in the deformation problem cancels the reduntant dilation of the parameter $\epsilon$, which, otherwise, would remain until the end. This is our fourth simplification.[^9]
We let the degrees $\deg{\mathfrak{m}}_\epsilon=\deg{\mathcal{E}}(\epsilon)$ be equal to four (c.f. [@MathieuNew]). Under this assumption, the two-component homogeneous polynomial extension ${\mathcal{E}}(\epsilon)$ of system contains $160$ undetermined coefficients. At the same time, the two components of the Miura contraction ${\mathfrak{m}}_\epsilon$ depend on $94$ coefficients. However, we decrease this number by nine, setting the coefficient of ${\tilde{u}}_{12;x}$ equal to ${+}1$ and, similarly, to ${-}1$ for ${\tilde{u}}_{12}^2$ (see , where the $\pm$ sign is absorbed by $\epsilon\mapsto {-}\epsilon$). Likewise, we set equal to zero the seven coefficients of ${\tilde{u}}_{12;xx}$, ${\tilde{u}}_{12}{\tilde{u}}_{12;x}$, ${\tilde{u}}_{12;xxx}$, ${\tilde{u}}_{12}^3$, ${\tilde{u}}_{12;x}^2$, ${\tilde{u}}_{12}{\tilde{u}}_{12;xx}$, and ${\tilde{u}}_{12;xxxx}$ in ${\mathfrak{m}}_\epsilon$.
The resulting algebraic system with the shortened list of unknowns and with the auxiliary list of nine substitutions is handled by <span style="font-variant:small-caps;">SsTools</span> and then solved by using <span style="font-variant:small-caps;">Crack</span> [@WolfCrack].
Under the above assumptions, the Gardner deformation problem for the Kaup–Boussinesq equation has a unique real solution of degree $4$. The Miura contraction ${\mathfrak{m}}_\epsilon$ is given by
\[BLimMiura\] $$\begin{aligned}
u_{0} &= {\tilde{u}}_{0} + \epsilon {\tilde{u}}_{0;x} - 2\epsilon^2{\tilde{u}}_{12}{\tilde{u}}_{0},
\label{BLimMiurau0}\\
u_{12} &= {\tilde{u}}_{12} + \epsilon\bigl({\tilde{u}}_{12;x} - 2{\tilde{u}}_{0}{\tilde{u}}_{0;x}\bigr)
+ \epsilon^2\bigl(4{\tilde{u}}_{12}{\tilde{u}}_{0}^2 -{\tilde{u}}_{12}^2 - {\tilde{u}}_{0;x}^2\bigr)
+ 4\epsilon^3 {\tilde{u}}_{12}{\tilde{u}}_{0}{\tilde{u}}_{0;x}
-4\epsilon^4 {\tilde{u}}_{12}^2{\tilde{u}}_{0}^2.\label{BLimMiurau12}\end{aligned}$$
The extension ${\mathcal{E}}(\epsilon)$ of is
\[BLimBurgE\] $$\begin{aligned}
{\tilde{u}}_{0;\xi} &= -{\tilde{u}}_{12;x}+4u_0{\tilde{u}}_{0;x}
+2\epsilon\bigl({\tilde{u}}_0{\tilde{u}}_{0;x}\bigr)_x
-4\epsilon^2\bigl({\tilde{u}}_0^2u_{12}\bigr)_x,
\\
{\tilde{u}}_{12;\xi}&={\tilde{u}}_{0;xxx}+4\bigl({\tilde{u}}_0{\tilde{u}}_{12}\bigr)_x
-2\epsilon\bigl({\tilde{u}}_0{\tilde{u}}_{12;x}\bigr)_x
-4\epsilon^2\bigl({\tilde{u}}_0{\tilde{u}}_{12}^2\bigr)_x.
$$
System preserves the first Hamiltonian operator $\hat{A}_1^\epsilon=\left(\begin{smallmatrix} 0 & {{\mathrm d}}/{{\mathrm d}}x\\ {{\mathrm d}}/{{\mathrm d}}x & 0\end{smallmatrix}\right)$ from $\hat{A}_1^{12}$ for .
The Miura contraction ${\mathfrak{m}}_\epsilon$ is shared by all equations in the Kaup–Boussinesq hierarchy. Solving the linear algebraic system, we find the extension $\bigl(\lim_B{\mathcal{E}}_{\text{SKdV}}^{a{=}4}\bigr)(\epsilon)$ for the bosonic limit of with $a{=}4$:
\[AppR\] $$\begin{aligned}
{\tilde{u}}_{0;t} &= - {\tilde{u}}_{0;xxx} - 6\bigl({\tilde{u}}_{0}{\tilde{u}}_{12}\bigr)_x
+ 12{\tilde{u}}_{0}^2{\tilde{u}}_{0;x}
+ 12\epsilon \bigl({\tilde{u}}_{0}^2{\tilde{u}}_{0;x} \bigr)_x
+ 6\epsilon^2 \bigl( {\tilde{u}}_{0}{\tilde{u}}_{12}^2 - 4{\tilde{u}}_{12}{\tilde{u}}_{0}^3 +
{\tilde{u}}_{0}{\tilde{u}}_{0;x}^2) \bigr)_x \notag \\
{}&{}\qquad{} + \epsilon^3 \bigl( (-24){\tilde{u}}_{12}{\tilde{u}}_{0}^2{\tilde{u}}_{0;x}\bigr)_x
+ \epsilon^4 \bigl( 24{\tilde{u}}_{12}^2{\tilde{u}}_{0}^3 \bigr)_x,\\
{\tilde{u}}_{12;t} &= - {\tilde{u}}_{12;xxx} - 6{\tilde{u}}_{12}{\tilde{u}}_{12;x}
+ 12\bigl({\tilde{u}}_{0}^2{\tilde{u}}_{12}\bigr)_x + 6{\tilde{u}}_{0}{\tilde{u}}_{0;xxx}
+ 12{\tilde{u}}_{0;xx}{\tilde{u}}_{0;x} \notag\\
{}&{}\qquad{}
+ 6\epsilon \bigl( {\tilde{u}}_{0;xx}{\tilde{u}}_{0;x} - 2{\tilde{u}}_{0}^2{\tilde{u}}_{12;x} \bigr)_x \notag\\
{}&{}\qquad{} + 2\epsilon^2 \bigl( {\tilde{u}}_{12}^3 - 18{\tilde{u}}_{12}^2{\tilde{u}}_{0}^2
- 6{\tilde{u}}_{12}{\tilde{u}}_{0}{\tilde{u}}_{0;xx} - 3{\tilde{u}}_{12}{\tilde{u}}_{0;x}^2
- 6{\tilde{u}}_{0}{\tilde{u}}_{12;x}{\tilde{u}}_{0;x}\bigr)_x\notag\\
{}&{}\qquad{}
+ 24\epsilon^3 \bigl( {\tilde{u}}_{12}{\tilde{u}}_{0}^3{\tilde{u}}_{12;x} \bigr)_x
+ 24\epsilon^4 \bigl( {\tilde{u}}_{12}^3{\tilde{u}}_{0}^2 \bigr)_x. \end{aligned}$$
Now we expand the fields ${\tilde{u}}_0(\epsilon)=\sum_{k=0}^{+\infty}{\tilde{u}}_0^{(k)}\cdot\epsilon^k$ and ${\tilde{u}}_{12}(\epsilon)=\sum_{k=0}^{+\infty}{\tilde{u}}_{12}^{(k)}\cdot\epsilon^k$, and plug the formal power series for ${\tilde{u}}_0$ and ${\tilde{u}}_{12}$ in ${\mathfrak{m}}_\epsilon$. Hence we start from ${\tilde{u}}_0^{(0)}=u_0$ and ${\tilde{u}}_{12}^{(0)}=u_{12}$, which is standard, and proceed with the recurrence relations between the conserved densities $u_0^{(k)}$ and $u_{12}^{(k)}$, $$\begin{gathered}
{\tilde{u}}_0^{(1)} = - u_{0;x}, \quad
{\tilde{u}}_0^{(n)} = - \tfrac{{{\mathrm d}}}{{{\mathrm d}}x}{\tilde{u}}_0^{({n-1})} + \sum\limits_{j+k=n-2}2{\tilde{u}}_{12}^{(k)}{\tilde{u}}_0^{(j)},\quad \forall n\geq2; \\{\tilde{u}}_{12}^{(1)} = 2u_{0}u_{0;x} - u_{12;x},\quad
{\tilde{u}}_{12}^{(2)} = u_{12}^2+u_{12;xx}-4u_{12}u_0^2-3u_{0;x}^2-4u_0u_{0;xx},$$$$\begin{gathered}
{\tilde{u}}_{12}^{(3)} = \sum\limits_{j+k=2}2{\tilde{u}}_0^{(j)}\tfrac{{{\mathrm d}}}{{{\mathrm d}}x}{\tilde{u}}_0^{(k)}
-\tfrac{{{\mathrm d}}}{{{\mathrm d}}x}{\tilde{u}}_{12}^{({2})}
+\sum\limits_{j+k=1}\Bigl({\tilde{u}}_{12}^{(j)}{\tilde{u}}_{12}^{(k)} + (\tfrac{{{\mathrm d}}}{{{\mathrm d}}x}{\tilde{u}}_0^{(j)})(\tfrac{{{\mathrm d}}}{{{\mathrm d}}x}{\tilde{u}}_0^{(k)})\Bigr)\\
- \sum\limits_{j+k+l=1} 4{\tilde{u}}_{12}^{(j)}{\tilde{u}}_0^{(k)}{\tilde{u}}_0^{(l)} - 4u_{12}u_{0}u_{0;x},\\
{\tilde{u}}_{12}^{(n)} = - \tfrac{{{\mathrm d}}}{{{\mathrm d}}x}{\tilde{u}}_{12}^{({n-1})}
+\sum\limits_{j+k=n-1}2{\tilde{u}}_0^{(j)}\tfrac{{{\mathrm d}}}{{{\mathrm d}}x}{\tilde{u}}_0^{(k)} + \sum\limits_{j+k=n-2}({\tilde{u}}_{12}^{(j)}{\tilde{u}}_{12}^{(k)} + (\tfrac{{{\mathrm d}}}{{{\mathrm d}}x}({\tilde{u}}_0^{(j)})
\tfrac{{{\mathrm d}}}{{{\mathrm d}}x}({\tilde{u}}_0^{(k)}))\\
-\sum\limits_{j+k+l=n-2} 4{\tilde{u}}_{12}^{(j)}{\tilde{u}}_0^{(k)}{\tilde{u}}_0^{(l)}
-\sum\limits_{j+k+l=n-3}4{\tilde{u}}_{12}^{(j)}{\tilde{u}}_0^{(k)}\tfrac{{{\mathrm d}}}{{{\mathrm d}}x}{\tilde{u}}_0^{(l)}\\
+ \sum\limits_{j+k+l+m=n-4}4{\tilde{u}}_{12}^{(j)}{\tilde{u}}_{12}^{(k)}{\tilde{u}}_0^{(l)}{\tilde{u}}_0^{(m)},\qquad \forall n\geq4. $$
\[ExBLim7dens\] Following this recurrence, let us generate the eight lowest weight nontrivial conserved densities, which start the tower of Hamiltonians for the Kaup–Boussinesq hierarchy. We begin with ${\tilde{u}}_0^{(0)}=u_0$ and ${\tilde{u}}_{12}^{(0)}=u_{12}$. Next, we obtain the densities $${\tilde{u}}_0^{(2)} =u_{0;xx} + 2u_{0}u_{12},\qquad {\tilde{u}}_{12}^{(2)}=-4u_{0;xx}u_{0}-3u_{0;x}^2+u_{12;xx}-4u_{0}^2u_{12}+u_{12}^2,$$which contribute to the tri-Hamiltonian representation of , see Proposition \[ThTriHam\]. Now we proceed with $$\begin{aligned}
{\tilde{u}}_0^{(4)}&= u_{0;4x} - 12u_{0;xx}u_{0}^2 + 6u_{0;xx}u_{12} - 18u_{0;x}^2u_{0} + 10u_{0;x}u_{12;x} + 6u_{12;xx}u_{0} - 8u_{0}^3u_{12} + 6u_{0}u_{12}^2,
\\
{\tilde{u}}_{12}^{(4)}&= - 8u_{0;4x}u_{0} - 20u_{0;xxx}u_{0;x} - 13u_{0;xx}^2 + 32u_{0;xx}u_{0}^3 - 48u_{0;xx}u_{0}u_{12} + 72u_{0;x}^2u_{0}^2 - 38u_{0;x}^2u_{12} -{}\\
{}&{}- 80u_{0;x}u_{12;x}u_{0} + u_{12;4x} - 24u_{12;xx}u_{0}^2 + 6u_{12;xx}u_{12} + 5u_{12;x}^2 + 16u_{0}^4u_{12} - 24u_{0}^2u_{12}^2 + 2u_{12}^3,
\\
{\tilde{u}}_0^{(6)}&= u_{0;6x} - 40u_{0;4x}u_{0}^2 + 10u_{0;4x}u_{12} - 200u_{0;xxx}u_{0;x}u_{0} + 28u_{0;xxx}u_{12;x} - 130u_{0;xx}^2u_{0} - {} \\
{}&{}- 198u_{0;xx}u_{0;x}^2 + 38u_{0;xx}u_{12;xx} + 80u_{0;xx}u_{0}^4 - 240u_{0;xx}u_{0}^2u_{12} + 30u_{0;xx}u_{12}^2 + 240u_{0;x}^2u_{0}^3 - {} \\
{}&{} - 380u_{0;x}^2u_{0}u_{12} + 28u_{0;x}u_{12;xxx} - 400u_{0;x}u_{12;x}u_{0}^2 + 100u_{0;x}u_{12;x}u_{12} + 10u_{12;4x}u_{0} - {} \\
{}&{}- 80u_{12;xx}u_{0}^3 + 60u_{12;xx}u_{0}u_{12} + 50u_{12;x}^2u_{0} + 32u_{0}^5u_{12} - 80u_{0}^3u_{12}^2 + 20u_{0}u_{12}^3,
\\
$$$$\begin{aligned}
{\tilde{u}}_{12}^{(6)}&= - 12u_{0;6x}u_{0} - 42u_{0;5x}u_{0;x} - 80u_{0;4x}u_{0;xx} + 160u_{0;4x}u_{0}^3 - 120u_{0;4x}u_{0}u_{12} - 49u_{0;xxx}^2 + {}\\
{}&{}+ 1200u_{0;xxx}u_{0;x}u_{0}^2 - 312u_{0;xxx}u_{0;x}u_{12} - 336u_{0;xxx}u_{12;x}u_{0} + 780u_{0;xx}^2u_{0}^2 - 206u_{0;xx}^2u_{12} + {}\\
{}&{} + 2376u_{0;xx}u_{0;x}^2u_{0} - 716u_{0;xx}u_{0;x}u_{12;x} - 456u_{0;xx}u_{12;xx}u_{0} - 192u_{0;xx}u_{0}^5 + 960u_{0;xx}u_{0}^3u_{12} -{}\\
{}&{}- 360u_{0;xx}u_{0}u_{12}^2 + 297u_{0;x}^4 - 366u_{0;x}^2u_{12;xx} - 720u_{0;x}^2u_{0}^4 + 2280u_{0;x}^2u_{0}^2u_{12} - 290u_{0;x}^2u_{12}^2 -{}\\
{}&{}- 336u_{0;x}u_{12;xxx}u_{0} + 1600u_{0;x}u_{12;x}u_{0}^3 - 1200u_{0;x}u_{12;x}u_{0}u_{12} + u_{12;6x} - 60u_{12;4x}u_{0}^2 +{}\\
{}&{}+ 10u_{12;4x}u_{12} + 28u_{12;xxx}u_{12;x} + 19u_{12;xx}^2 + 240u_{12;xx}u_{0}^4 - 360u_{12;xx}u_{0}^2u_{12} + 30u_{12;xx}u_{12}^2-{}\\
{}&{} - 300u_{12;x}^2u_{0}^2 + 50u_{12;x}^2u_{12} - 64u_{0}^6u_{12} + 240u_{0}^4u_{12}^2 - 120u_{0}^2u_{12}^3 + 5u_{12}^4,\qquad\text{\textit{etc}}.\end{aligned}$$ We will use these formulas in the next section, where, as an illustration, we re-derive the seven super-Hamiltonians of [@MathieuNew].
In the above notation, the following statements hold
- The conserved densities ${\tilde{u}}_0^{(2k)}$ and ${\tilde{u}}_{12}^{(2k)}$ of weights $2k+1$ and $2k+2$, respectively, are nontrivial for all integers $k\geq0$.
- Consider the zero-order components $\breve{u}_0(u_0,u_{12},\epsilon)$ and $\breve{u}_{12}(u_0,u_{12},\epsilon)$ of the series ${\tilde{u}}_0\bigl([u_0,u_{12}],\epsilon\bigr)$ and ${\tilde{u}}_{12}\bigl([u_0,u_{12}],\epsilon\bigr)$ with differential-polynomial coefficients. Then these generating functions are given by the formulas
$$\begin{aligned}
\left(\breve{u}_0(u_0,u_{12},\epsilon^2)\right)^2&=\frac{1}{8\epsilon^2}\cdot
\left[4\epsilon^2(u_0^2+u_{12})-1+
\sqrt{1+8\epsilon^2(u_0^2-u_{12})+16\epsilon^4(u_0^2+u_{12})^2}\right],
\label{u0zero}\\
\breve{u}_{12}(u_0,u_{12},\epsilon^2)&=
\frac{1}{2\epsilon^2}\cdot\left[
1-\sqrt{\tfrac{1}{2}-2\epsilon^2(u_{12}+u_0^2)
+\tfrac{1}{2}\sqrt{1+8\epsilon^2(u_0^2-u_{12})+16\epsilon^4(u_0^2+u_{12})^2
}}\right].\label{u12zero}\end{aligned}$$
- The generating functions for the odd-index conserved densities ${\tilde{u}}_0^{(2k+1)}$ and ${\tilde{u}}_{12}^{(2k+1)}$ are expressed via the even-index densities, see and , respectively. We claim that all the odd-index densities are trivial.
The densities ${\tilde{u}}_0^{(k)}$ and ${\tilde{u}}_{12}^{(k)}$, which are conserved for the bosonic limit of the $N{=}2$, $a{=}4$–SKdV system , retract to the conserved densities for the Korteweg–de Vries equation under $u_0\equiv0$, see Example \[ExKdVe\]. The corresponding reduction of $\breve{u}_{12}(u_0,u_{12},\epsilon)$ is the generating function . This implies that $\breve{u}_{12}=\sum_{k=0}^{+\infty} c(k)u_{12}^k\cdot\epsilon^{2k}+\dots$, whence the densities ${\tilde{u}}_{12}^{(2k)}$ are nontrivial.
Following the line of reasonings on p. , we consider the zero-order terms in Miura’s contraction , which yields
$$\begin{aligned}
u_0&=\breve{u}_0\cdot\bigl(1-2\epsilon^2\breve{u}_{12}\bigr),\label{Short0}\\
u_{12}&=\breve{u}_{12}+\epsilon^2\bigl(4\breve{u}_0^2\breve{u}_{12}-\breve{u}_{12}^2\bigr)-4\epsilon^4\breve{u}_0^2\breve{u}_{12}^2.\label{Short12}\end{aligned}$$
Therefore, $$\breve{u}_0=\frac{u_0}{1-2\epsilon^2\breve{u}_{12}}=
\sum_{k=0}^{+\infty} u_0\cdot\bigl(2\epsilon^2\breve{u}_{12}\bigr)^k.$$ Since the coefficients $c(k)$ of $u_{12}^k\cdot\epsilon^{2k}$ in $\breve{u}_{12}$ are positive, so are the coefficients of $u_0u_{12}^k\cdot\epsilon^{2k}$ in $\breve{u}_0$ for all $k\geq0$. This proves that the conserved densities ${\tilde{u}}_0^{(2k)}$ are nontrivial as well.
Second, squaring and adding it to , we obtain the equality $u_0^2+u_{12}=\breve{u}_0^2+\breve{u}_{12}-\epsilon^2\breve{u}_{12}^2$. In agreement with $\breve{u}_0{\bigr|}_{\epsilon=0}=u_0$ and $\breve{u}_{12}{\bigr|}_{\epsilon=0}=u_{12}$, we choose the root $\breve{u}_{12}=\bigl[1-\sqrt{1-4\epsilon^2\cdot
\bigl(u_{12}+u_0^2-\breve{u}_0^2\bigr)}\bigr]/(2\epsilon^2)$ of this quadratic equation. Hence yields the bi-quadratic equation upon $\breve{u}_0$, $$1-4\epsilon^2\bigl(u_{12}+u_0^2-\breve{u}_0^2\bigr)=u_0^2\bigr/\breve{u}_0^2.$$ As above, the proper choice of its root gives , whence we return to $\breve{u}_{12}$ and finally obtain .
Finally, let us substitute the expansions ${\tilde{u}}_0=
\upsilon_0(\epsilon^2)+\epsilon\cdot\upsilon_1(\epsilon^2)$ and ${\tilde{u}}_{12}=\omega_0(\epsilon^2)+\epsilon\cdot\omega_1(\epsilon^2)$ in for ${\tilde{u}}_0$ and ${\tilde{u}}_{12}$, see Example \[ExKdVe\]. By balancing the odd powers of $\epsilon$ in , it is then easy to deduce the equality $$\label{u0oddTriv}
\upsilon_1\equiv\sum_{k=0}^{+\infty} {\tilde{u}}_0^{(2k+1)}\cdot\epsilon^{2k}=
\frac{1}{4\epsilon^2}\cdot\frac{{{\mathrm d}}}{{{\mathrm d}}x}\log\bigl(
1-4\epsilon^2\cdot\upsilon_0\bigr),\qquad
\text{where }
\upsilon_0\equiv\sum_{\ell=0}^{+\infty}{\tilde{u}}_0^{(2\ell)}\cdot\epsilon^{2\ell}.$$ The balance of odd powers of $\epsilon$ in yields the algebraic equation upon $\omega_1$, whence, in agreement with the initial condition $\omega_1(0)={\tilde{u}}_{12}^{(1)}$, we choose its root $$\begin{gathered}
\omega_1=
\Bigl[1-2 \epsilon^2 \omega_0
+4 \epsilon^2 \upsilon_0^2
+4 \epsilon^4 \bigl(\upsilon_1^2
-2 \omega_0 \upsilon_0^2
+ \upsilon_0 \upsilon_{1;x}
+ \upsilon_1 \upsilon_{0;x}\bigr)
-8 \epsilon^6 \upsilon_1^2 \omega_0\\
-\Bigl(1
+4 \epsilon^2 \bigl(2\upsilon_0^2
- \omega_0\bigr)
+4 \epsilon^4 \bigl(\omega_0^2
+2 \upsilon_0 \upsilon_{1;x}
-8 \omega_0 \upsilon_0^2
+2 \upsilon_1 \upsilon_{0;x}
+2 \upsilon_1^2
+4 \upsilon_0^4\bigr)\\
+16\epsilon^6\bigl(
2 \omega_0^2 \upsilon_0^2
-2 \upsilon_1^2 \omega_0
- \omega_0 \upsilon_0 \upsilon_{1;x}
- \omega_0 \upsilon_1 \upsilon_{0;x}
-2 \upsilon_0^2 \upsilon_1 \upsilon_{0;x}
+2 \upsilon_1 \upsilon_0 \omega_{0;x}
+2 \upsilon_1^2 \upsilon_0^2
-4 \omega_0 \upsilon_0^4
+2 \upsilon_0^3 \upsilon_{1;x}\bigr)\\
+16 \epsilon^8 \bigl(\upsilon_1^4
+2 \omega_0^2 \upsilon_1^2
+4 \omega_0^2 \upsilon_0^4
-2 \upsilon_1^2 \upsilon_0 \upsilon_{1;x}
-4 \omega_0 \upsilon_0^3 \upsilon_{1;x}
+8 \upsilon_1^2 \omega_0 \upsilon_0^2
+2 \upsilon_1^3 \upsilon_{0;x}\\
+ \upsilon_0^2 \upsilon_{1;x}^2
+ \upsilon_1^2 \upsilon_{0;x}^2
+4 \omega_0 \upsilon_0^2 \upsilon_1 \upsilon_{0;x}
-2 \upsilon_0 \upsilon_{1;x} \upsilon_1 \upsilon_{0;x}\bigr)\\
+64 \epsilon^{10} \bigl(\upsilon_0 \upsilon_{1;x} \upsilon_1^2 \omega_0
-2 \omega_0^2 \upsilon_0^2 \upsilon_1^2
- \upsilon_1^3 \upsilon_{0;x} \omega_0
- \upsilon_1^4 \omega_0\bigr)
+64 \epsilon^{12} \upsilon_1^4 \omega_0^2
\Bigr)^{1/2}\Bigr]\bigr/(16 \epsilon^6 \upsilon_1 \upsilon_0).
\label{u12oddTriv}\end{gathered}$$ We claim that, using the balance of the even powers of $\epsilon$ in , the representation $\sum_{k=0}^{+\infty}{\tilde{u}}_{12}^{(2k+1)}\cdot\epsilon^{2k}\in
\operatorname{im}\tfrac{{{\mathrm d}}}{{{\mathrm d}}x}$ can be deduced, whence ${\tilde{u}}_{12}^{(2k+1)}\sim0$.
Super-Hamiltonians for $N{=}2$, $a{=}4$–SKdV hierarchy {#SecHam}
======================================================
In this section, we assign the bosonic super-Hamiltonians $\boldsymbol{{\mathcal{H}}}^{(k)}=\int\boldsymbol{h}^{(k)}[{\boldsymbol{u}}]\,{{\mathrm d}}\boldsymbol\theta{{\mathrm d}}x$ of with $a{=}4$ to the Hamiltonians $H^{(k)}=\int h_{12}^{(k)}[u_0,u_{12}]\,{{\mathrm d}}x$ of its bosonic limit . Also, we establish the no-go result on the super-field, $N{=}2$ supersymmetry invariant deformations of $a{=}4$–SKdV that retract to under the respective reduction in the super-field . At the same time, we initiate the study of Gardner’s deformations for reductions of other than , and here we find the deformations of two-component fermion-boson limit in it. However, we observe that the new solutions can not be merged with the deformation for the bosonic limit of .
From the previous section, we know the procedure for recursive production of the Hamiltonians $H^{(k)}=\int h^{(k)}\,{{\mathrm d}}x$ for the bosonic limit of the $N{=}2$, $a{=}4$–SKdV equation, here $h^{(2k)}={\tilde{u}}_0^{(2k)}$ and $h^{(2k+1)}={\tilde{u}}_{12}^{(2k)}$. In section \[SecGardner\], we explained why the reconstruction of the densities $\boldsymbol{h}^{(k)}$ for the bosonic super-Hamiltonians $\boldsymbol{{\mathcal{H}}}^{(k)}$ from $h^{(k)}\bigl[u_0,u_{12}\bigr]$ requires an intermediate step. Namely, it amounts to the proper choice of the representatives $h_{12}^{(k)}$ within the equivalence class $\bigl\{h^{(k)}\mod\operatorname{im}\tfrac{{{\mathrm d}}}{{{\mathrm d}}x}\bigr\}$ such that $h_{12}^{(k)}$ can be realized under as the $\theta_1\theta_2$-component of the super-density $\boldsymbol{h}^{(k)}$. This allows to restore the dependence on the components $u_1$ and $u_2$ of and to recover the supersymmetry invariance. The former means that each $\boldsymbol{h}^{(k)}$ is conserved on and the latter implies that $\boldsymbol{h}^{(k)}$ becomes a differential function in ${\boldsymbol{u}}$.
The correlation between *unknown* bosonic super-differential polynomials $\boldsymbol{h}^{(k)}[{\boldsymbol{u}}]$ and the densities $h^{(k)}\bigl[u_0,u_{12}\bigr]$, which are produced by the recurrence relation, is established as follows. First, we generate the homogeneous super-differential polynomial ansatz for the bosonic $\boldsymbol{h}^{(k)}$ using `GenSSPoly`, see note \[FootGenSSPoly\] on p. . Second, we split the super-field ${\boldsymbol{u}}$ using the right-hand side of and obtain the $\theta_1\theta_2$-component $h_{12}^{(k)}\bigl[u_0,u_1,u_2,u_{12}\bigr]$ of the differential function $\boldsymbol{h}^{(k)}[{\boldsymbol{u}}]$. This is done by the procedure[^10] `ToCoo`, which now is also available in <span style="font-variant:small-caps;">SsTools</span> [@SsTools]. Thirdly, we set to zero the components $u_1$ and $u_2$ of the super-field ${\boldsymbol{u}}$. This gives the ansatz $h_{12}^{(k)}\bigl[u_0,u_{12}\bigr]$ for the representative of the conserved density in the vast equivalence class. By the above, the gap between $h_{12}^{(k)}$ and the known $h^{(k)}$ amounts to $\tfrac{{{\mathrm d}}}{{{\mathrm d}}x}\bigl(f^{(k)}\bigr)$, where $f^{(k)}\bigl[u_0,u_{12}\bigr]$ is a homogeneous differential polynomial. We remark that the choice of $f$ is not unique due to the freedom in the choice of $\boldsymbol{h}^{(k)}\mod{\mathcal{D}}_1(\dots)+{\mathcal{D}}_2(\dots)$. We thus arrive at the linear algebraic equation $$\label{CutTail}
h_{12}^{(k)}-\tfrac{{{\mathrm d}}}{{{\mathrm d}}x}f^{(k)}=h^{(k)},$$ which exprimes the equality of the respective coefficients in the polynomials. The homogeneous polynomial ansatz for $f^{(k)}$ is again generated by `GenSSPoly`. Then equation is split to the algebraic system by <span style="font-variant:small-caps;">SsTools</span> and solved by <span style="font-variant:small-caps;">Crack</span> [@WolfCrack]. Hence we obtain the coefficients in $h_{12}^{(k)}$ and $f^{(k)}$. *A posteriori*, the freedom in the choice of $f^{(k)}$ is redundant, and it is convenient to set the surviving *unassigned* coefficients to zero. Indeed, they originate from the choice of a representative from the equivalence class for the super-density $\boldsymbol{h}^{(k)}[{\boldsymbol{u}}]$. This concludes the algorithm for the recursive production of homogeneous bosonic $N{=}2$ supersymmetry-invariant super-Hamiltonians $\boldsymbol{{\mathcal{H}}}^{(k)}$ for the $N{=}2$, $a{=}4$–SKdV hierarchy.
Let us reproduce the first seven super-Hamiltonians for , which were found in [@MathieuNew]. In contrast with Example \[ExBLim7dens\], we now list the *properly chosen* representatives $h^{(k)}_{12}\bigl[u_0,u_{12}\bigr]$ for the equivalence classes of conserved densities ${\tilde{u}}_{0}^{(2k)}$ and ${\tilde{u}}_{12}^{(2k)}$, here $k\leq3$. Then we expose the conserved super-densities $\boldsymbol{h}^{(k)}$ such that the respective expressions $h_{12}^{(k)}$ are obtained from the $\theta_1\theta_2$-components $\int\boldsymbol{h}^{(k)}\,{{\mathrm d}}\boldsymbol{\theta}$ by the reduction .
\[Hamsa\] $$\begin{aligned}
h_{12}^{(0)}&=u_0\sim{\tilde{u}}_0^{(0)},\qquad
\boldsymbol{h}^{(0)}=-{\mathcal{D}}_1{\mathcal{D}}_2({\boldsymbol{u}})\sim 0,\\
h_{12}^{(1)}&=u_{12}\sim{\tilde{u}}_{12}^{(0)},\qquad
\boldsymbol{h}^{(1)}={\boldsymbol{u}},\\
h_{12}^{(2)}&=-2u_{12}u_{0} \sim{\tilde{u}}_0^{(2)},\qquad
\boldsymbol{h}^{(2)}={\boldsymbol{u}}^2,\\
h_{12}^{(3)}&=\tfrac{3}{4}u_{12}^2 - 3u_{12}u_{0}^2 +\tfrac{3}{4} u_{0;x}^2 \sim {\tilde{u}}_{12}^{(2)},\qquad
\boldsymbol{h}^{(3)}={\boldsymbol{u}}^3 - \tfrac{3}{4}{\boldsymbol{u}}{\mathcal{D}}_1{\mathcal{D}}_2({\boldsymbol{u}}),\\
h_{12}^{(4)}&= 3u_{12}^2u_{0} - 4u_{12}u_{0}^3 - \tfrac{3}{2}u_{0}^2u_{0;xx} - u_{12;x}u_{0;x} \sim{\tilde{u}}_0^{(4)},\notag\\
\boldsymbol{h}^{(4)}&={\boldsymbol{u}}^4 -\tfrac{1}{2}{\boldsymbol{u}}{\boldsymbol{u}}_{xx} - \tfrac{3}{2}{\boldsymbol{u}}^2{\mathcal{D}}_1{\mathcal{D}}_2({\boldsymbol{u}}),\\
h_{12}^{(5)}&= -\tfrac{5}{4}u_{12}^3 + \tfrac{15}{2}u_{12}^2u_{0}^2 - 5u_{12}u_{0}^4 + 5u_{12}u_{0}u_{0;xx} + \tfrac{15}{8}u_{12}u_{0;x}^2
+ \tfrac{15}{2}u_{0}^2u_{0;x}^2 + \tfrac{5}{16}u_{12;x}^2 {} + \notag\\
{}&{}+ \tfrac{5}{16}u_{0;xx}^2 \sim {\tilde{u}}_{12}^{(4)},\qquad
\boldsymbol{h}^{(5)}= {\boldsymbol{u}}^5 - \tfrac{15}{16}{\boldsymbol{u}}^2 {\boldsymbol{u}}_{xx} + \tfrac{5}{8} ({\mathcal{D}}_1{\mathcal{D}}_2 {\boldsymbol{u}})^2 {\boldsymbol{u}}- \tfrac{5}{2}{\boldsymbol{u}}^3{\mathcal{D}}_1{\mathcal{D}}_2 {\boldsymbol{u}},\\
h_{12}^{(6)}&= -\tfrac{15}{4}u_{12}^3u_{0} + 15u_{12}^2u_{0}^3
- \tfrac{15}{8}u_{12}^2u_{0;xx} - 6u_{12}u_{0}^5 - \tfrac{75}{4}u_{12}u_{0}u_{0;x}^2 - \tfrac{3}{8}u_{12}u_{0;xxxx} + {} \notag\\
{} & {}+ 5u_{0}^3u_{12;xx} + 15u_{0}^3u_{0;x}^2 + \tfrac{15}{8}u_{0}u_{12;x}^2 + \tfrac{15}{8}u_{0}u_{0;xx}^2 \sim {\tilde{u}}_0^{(6)},\notag$$$$\begin{aligned}
\boldsymbol{h}^{(6)}&={\boldsymbol{u}}^6 - \tfrac{15}{8}{\boldsymbol{u}}^3 {\boldsymbol{u}}_{xx}
+ \tfrac{3}{16}{\boldsymbol{u}}{\boldsymbol{u}}_{4x} + \tfrac{15}{8}({\mathcal{D}}_1{\mathcal{D}}_2{\boldsymbol{u}})^2 - \tfrac{15}{4}{\boldsymbol{u}}^4{\mathcal{D}}_1{\mathcal{D}}_2{\boldsymbol{u}}+ \tfrac{15}{8}{\boldsymbol{u}}_{xx}{\mathcal{D}}_1{\mathcal{D}}_2{\boldsymbol{u}}- {} \notag\\
{}& {}+ \tfrac{5}{8}{\mathcal{D}}_1{\mathcal{D}}_2({\boldsymbol{u}}){\mathcal{D}}_1({\boldsymbol{u}}){\mathcal{D}}_1({\boldsymbol{u}}_x),\\
h_{12}^{(7)}&= -\tfrac{21}{8}u_{0;4x}u_{0}u_{12} + \tfrac{7}{64}u_{0;xxx}^2 + \tfrac{105}{16}u_{0;xx}^2u_{0}^2 + \tfrac{35}{32}u_{0;xx}^2u_{12}
- \tfrac{105}{8}u_{0;xx}u_{0}u_{12}^2 - \tfrac{105}{64}u_{0;4x}^4 - {}\notag\\
{}&{}- \tfrac{35}{16}u_{0;x}^2u_{12;xx} + \tfrac{105}{4}u_{0;x}^2u_{0}^4
- \tfrac{525}{8}u_{0;x}^2u_{0}^2u_{12} - \tfrac{175}{32}u_{0;x}^2u_{12}^2
+ \tfrac{7}{64}u_{12;xx}^2 + \tfrac{35}{4}u_{12;xx}u_{0}^4 + {} \notag\\
{}&{}+ \tfrac{105}{16}u_{12;x}^2u_{0}^2 - \tfrac{35}{32}u_{12;x}^2u_{12}
- 7u_{0}^6u_{12} + \tfrac{105}{4}u_{0}^4u_{12}^2 - \tfrac{105}{8}u_{0}^2u_{12}^3 + \tfrac{35}{64}u_{12}^4 \sim {\tilde{u}}_{12}^{(6)},\notag\\
\boldsymbol{h}^{(7)}&= {\boldsymbol{u}}^7 - \tfrac{105}{32}{\boldsymbol{u}}^3 {\boldsymbol{u}}_{xx}
+ \tfrac{7}{32}{\boldsymbol{u}}^2 {\boldsymbol{u}}_{4x} - \tfrac{35}{64}{\boldsymbol{u}}({\mathcal{D}}_1{\mathcal{D}}_2{\boldsymbol{u}})^3
+ \tfrac{35}{8}{\boldsymbol{u}}^3({\mathcal{D}}_1{\mathcal{D}}_2{\boldsymbol{u}})^2
- \tfrac{35}{64}({\mathcal{D}}_1{\mathcal{D}}_2{\boldsymbol{u}})^2 {\boldsymbol{u}}_{xx} - {} \notag\\
{}& {} - \tfrac{21}{4}{\boldsymbol{u}}^5{\mathcal{D}}_1{\mathcal{D}}_2{\boldsymbol{u}}+ \tfrac{105}{16}{\boldsymbol{u}}^2{\boldsymbol{u}}_{xx} {\mathcal{D}}_1{\mathcal{D}}_2{\boldsymbol{u}}+ \tfrac{315}{64}{\boldsymbol{u}}{\boldsymbol{u}}_{x}^2{\mathcal{D}}_1{\mathcal{D}}_2{\boldsymbol{u}}+ \tfrac{35}{16}{\boldsymbol{u}}({\mathcal{D}}_1{\mathcal{D}}_2{\boldsymbol{u}})({\mathcal{D}}_1{\boldsymbol{u}})({\mathcal{D}}_1{\boldsymbol{u}}_x) - {} \notag\\
{}& {} - \tfrac{7}{64}{\boldsymbol{u}}_{4x}{\mathcal{D}}_1{\mathcal{D}}_2{\boldsymbol{u}}- \tfrac{7}{8}{\boldsymbol{u}}({\mathcal{D}}_1 {\boldsymbol{u}}_{xx} )({\mathcal{D}}_1 {\boldsymbol{u}}_x).\end{aligned}$$
Of course, our super-densities $\boldsymbol{h}^{(k)}$ are equivalent to those in [@MathieuNew] up to trivial terms ${\mathcal{D}}_1(\dots)+{\mathcal{D}}_2(\dots)$.
Until now, we have not yet reported any attempt of construction of Gardner’s *super-field* deformation for , which means that the ansatz for ${\mathfrak{m}}_\epsilon$ and ${\mathcal{E}}(\epsilon)$ is written in super-functions of ${\boldsymbol{u}}$ (c.f. [@MathieuNew]). This would yield the super-Hamiltonians $\boldsymbol{{\mathcal{H}}}^{(k)}$ at once, and the intermediate deformation of a reduction for would not be necessary. At the same time, the knowledge of Gardner’s deformations for the reductions allows to inherit a part of the coefficients in the super-field ansatz by fixing them in the component expansions (*e.g.*, see , , and ).
Unfortunately, this cut-through does not work for the $N{=}2$, $a{=}4$–SKdV equation.
\[N2NoGo\] Under the assumptions that $N{=}2$ supersymmetry-invariant Gardner’s deformation ${\mathfrak{m}}_\epsilon\colon{\mathcal{E}}(\epsilon)\to{\mathcal{E}}$ of with $a{=}4$ be regular at $\epsilon=0$, be scaling-homogeneous, and retract to under the reduction $u_0=0$, $u_1=u_2=0$ in the super-field , there is no such deformation.
This rigidity statement, although under a principally different set of initial hypotheses, is contained in [@MathieuNew]. In particular, there it was supposed that $\deg{\mathfrak{m}}_\epsilon=\deg{\mathcal{E}}(\epsilon)=2$, which turns to be on the obstruction threshold, see below. We reveal the general nature of this ‘no go’ result.
Suppose there is the super-field Miura contraction ${\mathfrak{m}}_\epsilon$, $$\begin{gathered}
{\boldsymbol{u}}= \tilde{{\boldsymbol{u}}}
+\epsilon\bigl(p_3\tilde{u}^2 - p_1{\mathcal{D}}_1{\mathcal{D}}_2\tilde{{\boldsymbol{u}}} + p_2\tilde{{\boldsymbol{u}}}_x\bigr)
+\epsilon^2\Bigl(p_{15}\tilde{{\boldsymbol{u}}}^3 + p_{13}\tilde{{\boldsymbol{u}}}\tilde{{\boldsymbol{u}}}_x
+ p_{10}{\mathcal{D}}_2(\tilde{{\boldsymbol{u}}}){\mathcal{D}}_1(\tilde{{\boldsymbol{u}}}) \\
- p_{12}{\mathcal{D}}_1{\mathcal{D}}_2(\tilde{{\boldsymbol{u}}})\tilde{{\boldsymbol{u}}}
- p_{11}{\mathcal{D}}_1{\mathcal{D}}_2(\tilde{{\boldsymbol{u}}}_x) + p_{14}\tilde{{\boldsymbol{u}}}_{xx}\Bigr) + \cdots.\end{gathered}$$ To recover the deformation upon $u_{12}$ in ${\boldsymbol{u}}$, we split ${\mathfrak{m}}_\epsilon$ in components and fix the coefficients of $\epsilon{\tilde{u}}_{12;x}$ and $\epsilon^2{\tilde{u}}_{12}^2$, see . By this argument, the expansion of $\tilde{{\boldsymbol{u}}}_x$ yields $p_2=1$, while the equality $-p_{12}{\mathcal{D}}_1{\mathcal{D}}_2(\tilde{{\boldsymbol{u}}})\tilde{{\boldsymbol{u}}} + p_{10}{\mathcal{D}}_2(\tilde{{\boldsymbol{u}}}){\mathcal{D}}_1(\tilde{{\boldsymbol{u}}})
=(p_{12}-p_{10})\theta_1\theta_2u_{12}^2+\dots$ implies that $p_{12}=p_{10}-1$. Next, we generate the homogeneous ansatz for ${\mathcal{E}}(\epsilon)$, which contains $\tilde{{\boldsymbol{u}}}_t=\dots+\epsilon^2\cdot\tfrac{{{\mathrm d}}}{{{\mathrm d}}x}\bigl(q_{17}({\mathcal{D}}_2{\boldsymbol{u}})({\mathcal{D}}_1{\boldsymbol{u}}){\boldsymbol{u}}+\dots\bigr)+\dots$ in the right-hand side (the coefficient $q_{17}$ will appear in the obstruction). We stress that now both ${\mathfrak{m}}_\epsilon$ and ${\mathcal{E}}(\epsilon)$ can be formal power series in $\epsilon$ without any finite-degree polynomial truncation.
Now we split the determining equation ${\mathfrak{m}}_\epsilon\colon{\mathcal{E}}(\epsilon)\to{\mathcal{E}}$ to the sequence of super-differential polynomial equalities ordered by the powers of $\epsilon$. By the regularity assumption, the coefficients of higher powers of $\epsilon$ never contribute to the equations that arise at its lower degrees. Consequently, every contradiction obtained at a finite order in the algebraic system is universal and precludes the existence of a solution. (Of course, we assume that the contradiction is not created artificially by an excessively low order polynomial truncation of the expansions in $\epsilon$.)
This is the case for the $N{=}2$, $a{=}4$–SKdV. Using <span style="font-variant:small-caps;">Crack</span> [@WolfCrack], we solve all but two algebraic equations in the quadratic approximation. The remaining system is $$q_{17} = -p_{10}, \qquad p_{10} + q_{17} + 1=0.$$ This contradiction concludes the proof.
In Theorem \[N2NoGo\] for with $a{=}4$, we state the non-existence of the Gardner deformation in a class of differential super-polynomials in ${\boldsymbol{u}}$, that is, of $N{=}2$ supersymmetry-invariant solutions that incorporate . Still, we do *not* claim the non-existence of local regular Gardner’s deformations for the four-component system in the class of differential functions of $u_0$, $u_1$, $u_2$, and $u_{12}$.
Consequently, it is worthy to deform the reductions of other than . Clearly, if there is a deformation for the entire system, then such partial solutions contribute to it by fixing the parts of the coefficients.
Let us consider the reduction $u_0=0$, $u_2=0$ in with $a{=}4$. This is the two-component boson-fermion system $$\label{FB}
u_{1;t}=-u_{1;xxx}-3\bigl(u_1u_{12}\bigr)_x,\qquad
u_{12;t}=-u_{12;xxx}-6u_{12}u_{12;x}+3u_1u_{1;xx}.$$ Notice that system is *quadratic*-nonlinear in both fields, whence the balance $\deg{\mathfrak{m}}_\epsilon:\deg{\mathcal{E}}(\epsilon)$ for its polynomial Gardner’s deformations remains $1:1$.
We found a unique Gardner’s deformation of degree $\leq4$ for : the Miura contraction ${\mathfrak{m}}_\epsilon$ is cubic in $\epsilon$,
\[DefFB\] $$\begin{aligned}
u_1&={\tilde{u}}_1,\qquad u_{12}={\tilde{u}}_{12}
-\tfrac{1}{9} \epsilon^3 {\tilde{u}}_{1}{\tilde{u}}_{1;xx},\label{FBeFB}\\
\intertext{and the extension~${\mathcal{E}}(\epsilon)$ is given by the formulas}
{\tilde{u}}_{1;t} &= - {\tilde{u}}_{1;xxx} - 3\bigl({\tilde{u}}_1{\tilde{u}}_{12}\bigr)_x,\notag\\
{\tilde{u}}_{12;t} &= - {\tilde{u}}_{12;xxx} - 6{\tilde{u}}_{12}{\tilde{u}}_{12;x} + 3{\tilde{u}}_1{\tilde{u}}_{1;xx} + {}\notag \\
{}&{}\qquad{}+\tfrac{1}{3}\epsilon^3\Bigl(
u_1u_{1;xx}u_{12}-3u_1u_{1;x}u_{12;x}+u_{1;x}u_{1;xxx}
\Bigr)_x.\label{FBe}\end{aligned}$$
However, we observe, first, that the contraction is not recovered[^11] by under $u_1\equiv0$. Hence the deformation and its mirror copy under $u_1\leftrightarrow-u_2$ can not be merged with and to become parts of the deformation for .
Second, we recall that the fields $u_1$ and $u_2$ are, seemingly, the only local fermionic conserved densities for with $a{=}4$. Consequently, either the velocities ${\tilde{u}}_{1;t}$ and ${\tilde{u}}_{2;t}$ in Gardner’s extensions ${\mathcal{E}}(\epsilon)$ of are not expressed in the form of conserved currents (although this is indeed so at $\epsilon=0$) or the components $u_i=u_i\bigl(\bigl[{\tilde{u}}_0,{\tilde{u}}_1,{\tilde{u}}_2,{\tilde{u}}_{12}\bigr],\epsilon\bigr)$ of the Miura contractions ${\mathfrak{m}}_\epsilon$ are the identity mappings $u_i={\tilde{u}}_i$, here $i=1,2$, whence either the Taylor coefficients ${\tilde{u}}_i^{(k)}$ of ${\tilde{u}}_i$ are not termwise conserved on or there appear no recurrence relations at all. This will be the object of another paper.
Conclusion {#conclusion .unnumbered}
==========
We obtained the no-go statement for regular, scaling-homogeneous polynomial Gardner’s deformations of the $N{=}2$, $a{=}4$–SKdV equation under the assumption that the solutions retract to the original formulas by Gardner [@Gardner]. At the same time, we found a new deformation (\[BLimMiura\]–\[BLimBurgE\]) of the Kaup–Boussinesq equation that specifies the second flow in the bosonic limit of the super-hierarchy. We emphasize that other known nontrivial deformations for the Kaup–Boussinesq equation [@PamukKale] can be used for this purpose with equal success.
We exposed the two-step procedure for recursive production of the bosonic super-Hamiltonians $\boldsymbol{{\mathcal{H}}}^{(k)}$. We formulated the entire algorithm in full detail such that, with elementary modifications, it is applicable to other supersymmetric KdV-type systems.
Acknowledgements {#acknowledgements .unnumbered}
----------------
The authors thank P. Mathieu, J. W. van de Leur, and Z. Popowicz for helpful discussions. This research is partially supported by NSERC (for V.H. and T. W.) and NWO grants B61–609 and VENI 639.031.623 (for A.V.K.). A part of this research was done while A. V. K. was visiting at CRM (Montréal) and Max Planck Institute for Mathematics (Bonn), and A. O. K. was visiting at Utrecht University; the financial support and hospitality of these institutions are gratefully acknowledged. A. V. K. thanks the Organizing committee of 9th International workshop SQS’09 for support and the organizers of 8th International conference ‘Symmetry in Nonlinear Mathematical Physics’ for warm hospitality.
[99]{} [*[Bonora L., Krivonos S., Sorin A.]{}*]{} (1996) Towards the construction of $N{=}2$ supersymmetric integrable hierarchies, [*[Nuclear Phys. B]{}*]{} [**[477]{}**]{}:3, 835–854.
[*[Bourque S., Mathieu P.]{}*]{} (2001) The Painlevé analysis for $N=2$ super Korteweg-de Vries equations, [*[J. Math. Phys.]{}*]{} [**[42]{}**]{}:8, 3517–3539.
[*[Brunelli J. C., Das A.]{}*]{} (1994) The supersymmetric two boson hierarchies, [*[Phys. Lett. B]{}*]{} [**[337]{}**]{}:3-4, 303–307; The supersymmetric two boson hierarchy, in: (Quebec, PQ, 1995), 1–10, Univ. Montréal, Publ. CRM (F. Khanna and L. Vinet, eds.), Montréal, QC, 1997.
[*[Chaichian M., Kulish P. P.]{}*]{} (1987) Superconformal algebras and their relation to integrable nonlinear systems, [*[Phys. Lett. B]{}*]{} [**[183]{}**]{}:2, 169–174.
[*[Degiovanni L., Magri F., Sciacca V.]{}*]{} (2005) On deformation of Poisson manifolds of hydrodynamic type, [*[Commun. Math. Phys.]{}*]{} [**[253]{}**]{}:1, 1–24.
[*[Golovko V. A., Krasil’shchik I. S., Verbovetskiĭ A. M.]{}*]{} (2008) Variational Poisson–Nijenhuis structures for partial differential equations, [*[Theoret. and Math. Phys.]{}*]{} [**[154]{}**]{}:2, 227–239.
[*[Karasu A., Kiselev A. V.]{}*]{} (2006) Gardner’s deformations of the Boussinesq equations, [*[J. Phys. AMath.Gen.]{}*]{} [**[39]{}**]{}: 37, 11453–11460. `arXiv:nlin.SI/0603029`
[*[Kaup D. J.]{}*]{} (1975) A higher-order water-wave equation and the method for solving it, [*[Progr. Theoret. Phys.]{}*]{} [**[54]{}**]{}:2, 396–408; [*[Broer L. J. F.]{}*]{} (1975) Approximate equations for long water waves, [*[Appl. Sci. Res.]{}*]{} [**[31]{}**]{}:5, 377–395.
[*[Kersten P., Krasil’shchik J.]{}*]{} (2002) Complete integrability of the coupled KdV-mKdV system, [*[Adv. Stud. Pure Math.]{}*]{} [**[37]{}**]{}, 151–171.
[*[Kersten P. H. M., Sorin A. S.]{}*]{} (2002) Bi-Hamiltonian structure of the $N{=}2$ supersymmetric $\alpha{=}1$ KdV hierarchy, [*[Phys. Lett. A]{}*]{} [**[300]{}**]{}:4-5, 397–406. [*[Kiselev A. V.]{}*]{} (2007) Algebraic properties of Gardner’s deformations for integrable systems, [*[Theor. Math. Phys.]{}*]{} [**[152]{}**]{}:1, 963–976. `arXiv:nlin.SI/0610072` [*[Kiselev A. V., Hussin V.]{}*]{} (2009) Hirota’s virtual multi-soliton solutions of $N{=}2$ supersymmetric KdV equations, [*[Theor. Math. Phys.]{}*]{} [**[159]{}**]{}:3, 832–840. `arXiv:nlin.SI/0810.0930` [*[Kiselev A. V., Wolf T.]{}*]{} (2006) Supersymmetric representations and integrable fermionic extensions of the Burgers and Boussinesq equations, [*[SIGMA – Symmetry, Integrability and Geometry Methods and Applications]{}*]{} [**[2]{}**]{}:30, 1–19. [*[Kiselev A. V., Wolf T.]{}*]{} (2007) Classification of integrable super-systems using the <span style="font-variant:small-caps;">SsTools</span> environment, [*[Comput. Phys. Commun.]{}*]{} [**[177]{}**]{}:3, 315–328. `arXiv:nlin.SI/0609065`
[*[Kosmann-Schwarzbach Y., Magri F.]{}*]{} (1990) Poisson–Nijenhuis structures, [*[Ann. Inst. H. Poincaré ser. A Phys. Théor.]{}*]{} [**[53]{}**]{}:1, 35–81.
[*[Krivonos S., Sorin A., Toppan F.]{}*]{} (1995) On the super-NLS equation and its relation with the $N{=}2$ super-KdV equation within the coset approach, [*[Phys. Lett. A]{}*]{} [**[206]{}**]{}:3-4, 146–152.
[*[Kupershmidt B. A.]{}*]{} (1983) Deformations of integrable systems, [*[Proc. Roy. Irish Acad.]{}*]{} [**[A83]{}**]{}:1, 45–74.
[*[Kupershmidt B. A.]{}*]{} (1986) Super long waves, [*[Mech. Res. Comm.]{}*]{} [**[13]{}**]{}:1, 47–51.
[*[Laberge C. A., Mathieu P.]{}*]{} (1988) $N{=}2$ superconformal algebra and $O(2)$ integrable fermionic extensions of the Korteweg–de Vries equation, [*[Phys. Lett. B]{}*]{} [**[215]{}**]{}, 718–722.
[*[Labelle P., Mathieu P.]{}*]{} (1991) A new supersymmetric Korteweg–de Vries equation, [*[J. Math. Phys.]{}*]{} [**[32]{}**]{}:4, 923–927.
[*[Mathieu P.]{}*]{} (1988) Supersymmetric extension of the Korteweg–de Vries equation, [*[J. Math. Phys.]{}*]{} [**[29]{}**]{}:11, 2499–2506.
[*[Mathieu P.]{}*]{} (2001) Open problems for the super KdV equations, in: (A. Coley, D. Levi, R. Milson, C. Rogers, and P. Winternitz, eds). CRM Proc. Lecture Notes [**[29]{}**]{} AMS, 325–334. [*[Miura R. M., Gardner C. S., Kruskal M. D.]{}*]{} (1968) Korteweg–de Vries equation and generalizations. II. Existence of conservation laws and constants of motion, [*[J. Math. Phys.]{}*]{} [**[9]{}**]{}:8, 1204–1209.
[*[Nutku Y., Pavlov M. V.]{}*]{} (2002) Multi-Lagrangians for integrable systems, [*[J. Math. Phys.]{}*]{} [**[43]{}**]{}:3, 1441–1459.
[*[Olver P. J.]{}*]{} (1993) , Grad. Texts in Math. [**[107]{}**]{} (2nd ed.), Springer–Verlag, NY.
[*[Palit S., Chowdhury R. A.]{}*]{} (1996) A supersymmetric dispersive water wave equation, [*[J. Phys. A Math. Gen.]{}*]{} [**[29]{}**]{}:11, 2853–2859.
[*[Popowicz Z.]{}*]{} (1993) The Lax formulation of the “new” $N{=}2$ SUSY KdV equation, [*[Phys. Lett. A]{}*]{} [**[174]{}**]{}:5-6, 411–415.
[*[Sergyeyev A.]{}*]{} (2004) A simple way of making a Hamiltonian system into a bi-Hamiltonian one, [*[Acta Appl. Math.]{}*]{} [**[83]{}**]{}:1-2, 183–197.
[*[Wilson G.]{}*]{} (1981) On two constructions of conservation laws for Lax equations, [*[Quart. J. Math. Oxford]{}*]{} Ser. (2) [**[32]{}**]{}:128, 491–512.
[*[Wolf T.]{}*]{}, Applications of <span style="font-variant:small-caps;">Crack</span> in the classification of integrable systems, [**[37]{}**]{} (2004), 283–300.
[^1]: Likewise, we will extend Gardner’s deformation of to the deformation of the two-component bosonic limit for with $a{=}4$. Hence we reproduce the conservation laws for and, again, extend them to the bosonic super- Hamiltonians of the full system .
[^2]: The nonzero entries of the $(4\times4)$-matrix representation $\hat{P}_1$ for the Hamiltonian super-operator $\hat{\boldsymbol{P}}_1^{a{=}4}$ are $\bigl(\hat{P}_1\bigr)_{0,12}=\bigl(\hat{P}_1\bigr)_{2,1}
=\bigl(\hat{P}_1\bigr)_{12,0}=-\bigl(\hat{P}_1\bigr)_{1,2}={{\mathrm d}}/{{\mathrm d}}x$.
[^3]: We recall that the $N{=}2$ super-residue $\operatorname{Sres}M$ of a super-pseudodifferential operator $M$ is the coefficient of ${\mathcal{D}}_1{\mathcal{D}}_2\circ\bigl(\tfrac{{{\mathrm d}}}{{{\mathrm d}}x}\bigr)^{-1}$ in $M$ .
[^4]: Under the assumption of weight homogeneity, the freedom in the choice of such $f[u_0,u_{12}]$ is descreased, but the gap still remains.
[^5]: Still, the four components of the original $N{=}2$ supersymmetric equations within the hierarchy of *are* written in the form of conserved currents. A helpful counter-example, Gardner’s extension of the $N=1$ super-KdV equation, is discussed in [@MathieuNew; @MathieuN=1].
[^6]: This estimate is rough and can be improved by operating separately with the components of ${\mathfrak{m}}_\epsilon$ and ${\mathcal{E}}(\epsilon)$ since, in particular, the Kaup–Boussinesq system is *linear* in $u_{12}$.
[^7]: Reductions other than can produce quadratic-nonlinear subsystems of the cubic-nonlinear system , *e.g.*, if one sets $u_0=0$ and $u_2=0$, see on p. .
[^8]: \[FootGenSSPoly\]The call is `GenSSPoly(N,wglist,cname,mode)`, where
- `N` is the number of Grassmann variables $\theta_1$, $\ldots$, $\theta_N$;
- `wglist` is the list of lists `{afwlist, abwlist, wgt}`, each containing the list `afwlist` of weights for the fermionic super-fields and the list `abwlist` of weights for the bosonic super-fields; here `wgt` is the weight of the polynomial to be constructed;
- `cname` is the prefix for the names of arising undetermined coefficients (*e.g.*, `p` produces $p_1,p_2,\dots$);
- `mode` is the list of flags, which can be `fonly`, whence only fermionic polynomials are generated, or `bonly`, which yields the bosonic output.
[^9]: There is one more possibility to reduce the size of the algebraic system: this can be achieved by a thorough balance of the *differential orders* of ${\mathfrak{m}}_\epsilon$ and ${\mathcal{E}}(\epsilon)$.
[^10]: The call is `ToCoo(N,nf,nb,ex)`, where
- `N` is the number of Grassmann variables $\theta_1,\ldots,\theta_N$;
- `nf` is the number of fermionic super-fields `f(1)`,$\ldots$,`f(nf)`;
- `nb` is the number of bosonic super-fields `b(1)`,$\ldots$,`b(nb)`;
- `ex` is the super-field expression to be split in components.
For $N{=}2$, we have `f(i)=f(i,0,0)+b(i,1,0)*th(1)+b(i,0,1)*th(2)+f(i,1,1)*th(1)*th(2)`, `b(i)=b(i,0,0)+f(i,1,0)*th(1)+f(i,0,1)*th(2)+b(i,1,1)*th(1)*th(2)` as the splitting convention. The reduction is achieved by setting `b(i,0,1)`, `b(i,1,0)`, `f(j,0,1)`, and `f(j,1,0)` to zero for all $i\in[1,\mathtt{nb}]$ and $j\in[1,\mathtt{nf}]$.
[^11]: Surprisingly, the quadratic approximation in the deformation problem for is very restrictive and leads to a unique solution – for . Relaxing this constraint and thus permitting the coefficient of $\epsilon^2{\tilde{u}}_{12}^2$ in ${\mathfrak{m}}_\epsilon$ be arbitrary, we obtain two other real and two pairs of complex conjugate solutions for the deformations problem. They constitute the real and the complex orbit, respectively, under the action of the discrete symmetry $u_0\mapsto {-}u_0$, $\xi\mapsto {-}\xi$ of .
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Die übliche Interpretation der physikalischen Zustände in Quantenfeldtheorien vom Typ der Quantenelektrodynamik bereitet Schwierigkeiten aufgrund des Auftretens masseloser Teilchen, wie dem Photon. In Stoßprozessen geladener Teilchen werden unendlich viele dieser masselosen Teilchen erzeugt, die zu einer enormen Vielfalt von Zuständen führen zwischen denen theoretisch keine Interferenzen möglich sind (sie gehören zu verschiedenen Superauswahlsektoren); experimentell lassen sich diese Zustände jedoch nicht unterscheiden. Dieser scheinbare Widerspruch löst sich auf wenn man berücksichtigt, dass reale Experimente stets in zukunftsgerichteten Gebieten der Raumzeit (Lichtkegeln) durchgeführt werden. Versäumte Messungen in der Vergangenheit lassen sich aufgrund der Richtung des Zeitpfeils nicht nachholen. Die Theorie kann sich daher auf die Beschreibung und Interpretation solcher Lichtkegeldaten beschränken. In diesem Artikel wird erläutert, wie dieser Gedanke im allgemeinen Rahmen der Quantenfeldtheorie mathematisiert wird und zu einem konsistenten physikalischen Bild führt. Zustände gleicher Ladung aber inkohärenter masseloser Anteile lassen sich bei Einschränkung auf die Observablen in einem gegebenen Lichtkegel zu Ladungsklassen zusammenfassen; jede dieser Klassen vereinigt eine Vielzahl von Superauswahlsektoren. Es zeigt sich, dass zu jeder Ladungsklasse eine konjugierte Klasse existiert, die Zustände der entgegengesetzten Ladung (Antimaterie) beschreibt. Alle Zustände in einer Klasse und der entsprechenden konjugierten Klasse genügen der gleichen (Bose oder Fermi) Statistik. Obwohl auf Lichtkegeln nur die Halbgruppe zukunftsgerichteter Zeittranslationen wirkt, lässt sich sowohl die relativistische Kovarianz als auch die energetische Stabilität (Positivität der Energie) der Zustände jeder Ladungsklasse etablieren. Der Formalismus ist somit geeignet, die Eigenschaften der Zustände in Quantenfeldtheorien mit langreichweitigen Kräften in physikalisch sinnvoller Weise zu beschreiben.'
author:
- |
Detlev Buchholz\
Institut für Theoretische Physik und Courant Forschungszentrum\
“Strukturen höherer Ordnung in der Mathematik”, Universität Göttingen,\
37077 Göttingen, Deutschland
title: |
**Masselose Teilchen und Zeitpfeil\
in der relativistischen Quantenfeldtheorie**
---
Einleitung
==========
Das Verständnis der Struktur des physikalischen Zustandsraumes in Quantenfeldtheorien mit langreichweitigen Kräften, prominentestes Beispiel ist die Quantenelektrodynamik, ist ein sehr altes Problem. Seine unterschiedlichen Aspekte wurden in zahlreichen Arbeiten untersucht, siehe zum Beispiel die entsprechenden Kapitel in den Monographien [@Ha; @St; @Str] und dort angegebene Referenzen. Allen diesen Untersuchungen ist gemeinsam, dass die physikalische Raumzeit als unendlich ausgedehnter Minkowskiraum modelliert wird, in dem die quantisierte Materie und die von masselosen Teilchen erzeugte Strahlung beschrieben wird (die Gravitation bleibt außer Acht). Geht man von diesem idealisierten Bild aus, so ergibt sich als Konsequenz, dass man beliebig langwellige Strahlung bzw. masselose Teilchen beliebig kleiner Energie behandeln muss, obwohl diese in realistischen Experimenten nicht nachgewiesen werden können. Dies führt sowohl bei der Konstruktion der Theorie als auch bei ihrer Interpretation zu Schwierigkeiten, die zuweilen als Infrarotkatastrophe bezeichnet werden. Man hat zwar gelernt, wie man diese Schwierigkeiten im Rahmen der Theorie bei einigen Fragestellungen umgehen kann, etwa bei der Berechnung von “inklusiven” Wirkungsquerschnitten von Stoßprozessen. Dort summiert man über unendlich viele unbeobachtbare [niederenergetische]{} masselose Teilchen, die bei der Wechselwirkung geladener Teilchen unvermeidlich erzeugt werden. Doch ist diese Methode, die aufgrund der Überidealisierung der experimentellen Situation eingehandelten Schwierigkeiten durch *ad hoc* Rezepte wieder zu beseitigen, konzeptionell nicht befriedigend. Zudem bleiben grundlegende Fragen, etwa nach der Sektorstruktur des physikalischen Zustandsraumes, nach den Wurzeln der Teilchenstatistik oder der Existenz von Antimaterie bei diesem Zugang unbeantwortet. Bei Abwesenheit von langreichweitigen Kräften konnten diese Fragen dagegen vor langem umfassend beantwortet werden [@Ha].
Ein ganz neuer Zugang zu diesem Problemkreis, bei dem die raumzeitlichen Einschränkungen bei realen Experimenten von vornherein im Rahmen der Theorie berücksichtigt werden, wurde in einer kürzlich erschienenen Arbeit vorgeschlagen [@BuRo]. Die gewonnen Einsichten sollen hier erläutert werden ohne auf technische Details einzugehen. Vorausgesetzt werden beim Leser lediglich Kenntnisse von Grundbegriffen der Quantentheorie und der speziellen Relativitätstheorie.
Grundlagen
==========
Es sei zunächst daran erinnert, dass die Ensembles eines physikalischen Systems durch normierte Vektoren (Zustandsvektoren) $\Phi$ eines Hilbertraumes $\cH$ beschrieben werden, die Messgrößen (Observablen) wirken auf diesem Raum durch lineare hermitesche Operatoren $A = A^*$. Operatoren lassen sich addieren, per Komposition multiplizieren und mit komplexen Zahlen skalieren, sie erzeugen durch diese Operationen eine Algebra $\cA$. Die Theorie macht Vorhersagen über Mittelwerte der Messgrößen in vorgegebenen Ensembles, deren Schwankungsquadrate . Sie werden beschrieben durch Erwartungswertfunktionale $\varphi$ auf der Observablenalgebra $\cA$. Diese Funktionale erhält man aus den Hilbertraumzuständen $\Phi$ gemäß der fundamentalen Formel $$\varphi(A) = \langle \Phi, A \Phi \rangle \, ,
\quad A \in \cA \, ,$$ wobei, wie üblich, die eckige Klammer das Skalarprodukt zwischen Hilbertraumvektoren beschreibt. Es ist hier wichtig, dass man umgekehrt aus jedem solchen (linearen und positiven) Erwartungswertfunktional – wir sprechen im Folgenden kurz von Zuständen – einen Hilbertraum und eine konkrete Darstellung der Observablen auf diesem Raum rekonstruieren kann, so dass obige Relation gilt (sogenannte GNS–Konstruktion [@Ha]). Dies ist insbesondere dann von Bedeutung, wenn man Ensembles behandeln möchte, die nicht von vornherein durch Vektoren im vorgegebenen Hilbertraum beschrieben werden können (zum Beispiel ladungstragende Zustände wenn man von neutralen Zuständen ausgeht). Sie lassen sich durch geeignete Grenzwerte der zur Verfügung stehenden Funktionale darstellen. Man kann so den gegebenen Hilbertraum “verlassen” und zu einer konsistenten Beschreibung der gewünschten Ensembles unter Beibehaltung der Observablen kommen. Genauer gesagt, man erhält aus den Limesfunktionalen mittels der GNS–Konstruktion eine neue Darstellung der gegebenen Observablenalgebra auf einem Hilbertraum. In ihm gelten nach wie vor die ursprünglichen algebraischen Relationen zwischen den Observablen (man denke an Vertauschungsrelationen, Feldgleichungen ).
In der relativistischen Quantenfeldtheorie benutzt man zweckmäßigerweise das Heisenbergbild. Das heißt, die Symmetrietransformationen im zugrundeliegenden Minkowskiraum $\cM$, wie die Verschiebung des Zeitpunktes und Ortes von Messungen oder Drehungen der Messgeräte und der Übergang von einem Inertialsystem zu einem anderen wirken auf die entsprechende Observablen in $\cA$. Für jede derartige Poincaré–Transformation $\lambda$ wird diese Wirkung beschrieben durch eine lineare, multiplikative und symmetrische Abbildung $\alpha_\lambda$ von $\cA$ auf sich (Automorphismus). Man hat ebenfalls Informationen über die raumzeitlichen Lokalisierungseigenschaften der Observablen. Messungen in einem gegebenen Raumzeitgebiet $\cR \subset \cM$ werden beschrieben durch Observablen in entsprechenden Unteralgebren $\cA(\cR) \subset \cA$. Da diese Messungen auch in jedem größeren Gebiet ausgeführt werden können, gilt aus Konsistenzgründen $\cA(\cR_1) \subset \cA(\cR_2)$ falls $\cR_1 \subset \cR_2$. Weiterhin ist die Wirkung der Symmetrietransformationen $\lambda$ in $\cM$ konsistent mit der Zuordnung von Observablen zu Gebieten, d.h. Observablen in $\cA(\cR)$ werden durch die Wirkung von $\alpha_\lambda$ in die Algebra $\cA(\lambda
\cR)$ des transformierten Gebietes $\lambda \cR$ überführt, in Formeln $\alpha_\lambda \, \cA(\cR) = \cA(\lambda \cR)$.
Das Einsteinsche Kausalitätsprinzip, also die Aussage, dass sich physikalische Effekte nicht mit Überlichtgeschwindigkeit ausbreiten (dies gilt auch für die Quanteneffekte von Messungen), lässt sich in diesem Rahmen ebenfalls in sehr einfacher Weise formulieren. Dazu sei daran erinnert, dass Messungen, die sich gegenseitig nicht beeinflussen, durch Observablen beschrieben werden, deren Produkt kommutativ ist. Gemäß Kausalitätsprinzip müssen daher Paare von Observablen in raumartig getrennten (nicht durch kausale Weltlinien verbindbaren) Gebieten $\cR_1, \cR_2$ miteinander kommutieren, d.h. $A_1 A_2 = A_2 A_1$ für alle $A_1 \in \cA(\cR_1)$. $A_2 \in \cA(\cR_2)$; man bezeichnet dies als Lokalitätsbedingung. Jede physikalische Theorie, die den Grundprinzipien der [Quantentheorie]{} und der speziellen Relativitätstheorie genügt, erfüllt diese Bedingungen. Es ist bemerkenswert, dass der so abgesteckte allgemeine Rahmen vollständig ausreicht, um die hier interessierenden konzeptionellen Fragen zu diskutieren.
Geht man von der Vorstellung aus, dass der zugrunde gelegte Hilbertraum $\cH$ die Zustände sämtlicher elementarer Systeme einer Theorie beschreibt (z.B. alle Zustände endlicher Energie) so stellt sich die Frage nach dessen Sektorstruktur. Dabei trägt man der Tatsache Rechnung, dass in der Quantenfeldtheorie das Superpositionsprinzip nur eingeschränkt gilt: Zustandsvektoren unterschiedlicher Gesamtladung lassen sich zwar auf dem Papier addieren, ihre relative Phase spielt jedoch experimentell keine Rolle, sie sind nicht kohärenzfähig. Die Familien kohärenzfähiger Zuständen in $\cH$ nennt man (Superauswahl) Sektoren. Sie kann man dadurch charakterisieren, dass man von jedem gegebenen Zustand in einem Sektor zu jedem anderen Zustand im gleichen Sektor durch die Quanteneffekte von Messungen gelangen kann, die Observablen operieren transitiv auf den Sektoren. Verschiedene Sektoren lassen sich durch klassische Observablen unterscheiden, die man als Grenzwerte von geeigneten Folgen von Observablen in $\cA$ erhält. Sie vertauschen mit allen Elementen von $\cA$ und haben in jedem Sektor einen scharfen (nicht statistisch fluktuierenden) Wert, die entsprechende Superauswahlladung des Sektors.
Prominente Beispiele solcher Superauswahlladungen sind die elektrische Ladung und die sogenannte Univalenz, mit der man bosonische und fermionische Sektoren unterscheiden kann. Es zeigt sich jedoch, dass es in Quantenfeldtheorien mit langreichweitigen Kräften, die durch masselose Teilchen vermittelt werden, eine ungeheure Vielzahl von weiteren Sektoren gibt. Sie unterscheiden sich durch facettenreiche Infrarotwolken unendlich vieler masseloser Teilchen, die bei Stoßprozessen geladener Teilchen erzeugt werden. Die theoretische Beschreibung und Klassifikation aller dieser Infrarotsektoren ist mathematisch ein hoffnungsloses Problem. Zum Glück spielt sie experimentell keine Rolle. Denn es ist praktisch unmöglich unendliche Konfigurationen masseloser Teilchen mit beliebig kleiner Energie scharf zu unterscheiden. Bei Anwendungen der Theorie trägt man dieser Tatsache meist dadurch Rechnung, dass man *ad hoc* spezielle Sektoren auswählt und in Erwartungswerten über alle unbeobachtbaren masselosen Teilchen summiert. Dadurch kann man zwar eine Reihe von Schwierigkeiten umgehen, die Methode ist jedoch aus zwei Gründen konzeptionell unbefriedigend. Zum einen erfordert die Unterscheidung zwischen hoch– und niederenergetischen Teilchen die Wahl eines Inertialsystems, man bricht damit die relativistische Invarianz der Theorie. Zum anderen zerstören die Quanteneffekte der lokalisierten Observablen diese Unterscheidung wegen der Unschärferelation. Jedes Ensemble niederenergetischer Teilchen enthält nach einer lokalen Messung auch Anteile beliebig hoher Energie. Diese Tatsache erschwert die Diskussion von Konsequenzen des Einsteinschen Kausalitätsprinzips bei diesem Zugang.
Die kürzliche Lösung dieses Problems beruht auf der Einsicht, dass der Zeitpfeil bereits bei der Interpretation der mikroskopischen Theorie berücksichtigt werden muss. Um Missverständnissen vorzubeugen: der Zeitpfeil soll hier nicht erklärt werden, er wird lediglich als empirische in den theoretischen Rahmen eingebaut. Es ist ein Faktum, dass es unmöglich ist, versäumte Messungen in der Vergangenheit heute oder in der Zukunft nachzuholen. Man kann also nicht ausschließen, dass gewisse Daten, die man in der Vergangenheit hätte bestimmen können, unwiederbringlich verloren sind. Anders gesagt, die Theorie kann sich darauf beschränken, die Ergebnisse von Messungen in Raumzeitgebieten zu beschreiben und zu erklären, die in realistischen Experimenten (zumindest prinzipiell) zugänglich sind. Dies sind vorwärts gerichtete Lichtkegel $V$ mit Spitze in einem beliebig gewählten Raumzeitpunkt $a$ in der Vergangenheit; der Rand des Kegels wird von den von der Spitze ausgehenden Lichtstrahlen gebildet.
Wo die Spitze des Kegels liegt spielt praktisch keine Rolle. Man kann z.B. den Geburtstag und Geburtsort von Aristoteles wählen, der das Wort Physik ($\phi \upsilon \sigma \iota \kappa \eta$) eingeführt hat; wir alle wissen davon und befinden uns also innerhalb des entsprechenden Kegels. Eine andere Wahl wäre z.B. der Tag und Ort, an dem die Finanzierung eines geplanten Experiments genehmigt wurde. Wichtig ist nur, dass in der Vergangenheit des gewählten Punktes $a$ keine experimentellen Daten berücksichtigt werden müssen bzw. können. Natürlich erstrecken sich reale Experimente nur über einen bestimmten Zeitraum. Doch könnten im Prinzip zukünftige Generationen ein Experiment bis in alle Ewigkeit weiterführen. Lichtkegel sind also die maximalen Raumzeitgebiete in denen Experimente möglich sind.
Im Rahmen der Theorie werden die Observablen in einem gegebenen Lichtkegel $V$ durch Elemente der entsprechenden Unteralgebra $\cA(V) \subset \cA$ beschrieben. Geht man davon aus, dass nur diese Observablen in einem Experiment zur Verfügung stehen, so kann man bei gegebenem Zustand $\varphi$ eines Systems nur die entsprechenden Erwartungswerte $\varphi(A)$, $A \in \cA(V)$ bestimmen. Man erhält also nur partielle Informationen über den globalen Zustand $\varphi$ und bezeichnet dessen Einschränkung $\varphi \rest \cA(V)$ daher als partiellen Zustand. Es stellt sich dann die Frage, welche Information man aus solchen partiellen Zuständen extrahieren kann.
Bei genauerer Analyse zeigt sich, dass in Theorien, in denen es ausschließlich massive Teilchen gibt (also in einer hypothetischen Welt), die partiellen Zustände bereits die vollständige Information über den globalen Zustand enthalten, egal wie man den Lichtkegel $V$ wählt, d.h. wann und wo man mit den Messungen begonnen hat. Dies kann man auch heuristisch verstehen wenn man bedenkt, dass die Weltlinien massiver Teilchen, die sich ja mit weniger als Lichtgeschwindigkeit bewegen, irgendwann einmal in den Lichtkegel eintreten müssen. Aus den Daten in $V$ lassen sich daher die globalen Daten mit Hilfe der Theorie rekonstruieren.
Sobald masselose Teilchen in der Theorie auftreten (“Es werde Licht”) ändert sich die Situation jedoch drastisch. Werden in der Vergangenheit des gegebenen Lichtkegels $V$ durch Wechselwirkungsprozesse masselose Teilchen erzeugt, die danach ungehindert durch die Raumzeit propagieren, so können diese $V$ nicht erreichen und dort nachgewiesen werden, denn sie bewegen sich stets mit Lichtgeschwindigkeit (Huygensches Prinzip). Aus den Zuständen $\varphi \rest \cA(V)$ lassen sich diese auslaufenden masselosen Anteile daher nicht rekonstruieren. Man erhält somit in Experimenten nur eingeschränkte Informationen über die Beiträge masseloser Teilchen in den Zuständen. Wie wir noch sehen werden, ist diese Tatsache dafür verantwortlich, dass man die unterschiedlichen Infrarotwolken in den partiellen Zuständen nicht scharf unterscheiden kann. Die Zustände in verschiedenen Infrarotsektoren stimmen bei Restriktion auf Lichtkegel überein und man kann auf diese Weise Klassen von physikalisch äquivalenten (weil ununterscheidbaren) Zuständen bilden. Von Bedeutung ist dabei, dass die masselosen Teilchen in den Zuständen nicht mehr in einen energetisch weichen und einen harten Anteil aufgeteilt werden (mit den erwähnten Problemen). Stattdessen wird der masselose Teilcheninhalt getrennt in einen marginalen (nicht notwendig niederenergetischen) Teil, der in $V$ keine Beiträge liefert und einen essentiellen Teil, der Messungen in $V$ zugänglich ist. Diese Aufspaltung ist mit der relativistischen Invarianz der Theorie und dem Einsteinschen Kausalitätsprinzip verträglich und bildet den Ausgangspunkt der Untersuchungen in [@BuRo], die wir im Folgenden erläuteren wollen.
Das Vakuum
==========
Sowohl bei der Konstruktion von Theorien als auch bei deren Interpretation spielt der Zustand des Vakuums eine ausgezeichnete Rolle. Dies liegt daran, dass man eine ganze Reihe von Eigenschaften dieses speziellen Zustandes nennen kann, ohne die Theorie konstruiert zu haben: Der Vakuumzustand ist für alle inertialen Beobachter gleich, das heißt Messergebnisse in diesem Zustand ändern sich nicht, wenn man die Positionen der Messapparate durch Poincaré–Transformationen ändert. Ferner ist das Vakuum energielos, d.h. für alle inertialen Beobachter ist es der Zustand niedrigster Energie (Grundzustand).
Da wir nur Messungen in Lichtkegeln betrachten wollen, müssen wir bei der mathematischen Umsetzung dieser Charakterisierung beachten, dass die von den Poincaré–Transformationen $\lambda$ erzeugte Poincaré–Gruppe $\cP$ keine Symmetriegruppe der Kegel ist; denn kein Lichtkegel $V$ ist invariant unter beliebigen derartigen Transformationen. Die Untergruppe $\cL$ der Drehungen und Geschwindigkeitstransformationen bei festgehaltener Spitze des Kegels (Lorentz–Gruppe) lässt $V$ zwar invariant, doch ändert jede raumzeitliche Translation die Lage von $V$ im Minkowski–Raum. Es gibt jedoch Translationen, bei denen man $V$ zumindest nicht verlässt. Dies sind zukunftsgerichtete zeitartige (oder auch lichtartige) Verschiebungen. Inertiale Beobachter in $V$ müssen wenig tun, um derartige Verschiebungen ihrer Messgeräte zu bewerkstelligen: sie müssen nur warten, der Zeitpfeil tut das Übrige. Die Familie der Poincaré–Transformationen, die $V$ auf bzw. in sich abbilden, erzeugt eine Semigruppe $\cS \subset \cP$ (“semi” deshalb, weil die Elemente von $\cS$ sich zwar komponieren lassen, es gibt jedoch kein Inverses, da sich der Zeitpfeil nicht umkehren lässt).
Nach diesen Vorbemerkungen können wir nun obige formlose Beschreibung des Vakuums in unserem Formalismus präzisieren. Ein partieller Zustand $\omega \rest \cA(V)$ wird als Vakuum in $V$ gedeutet wenn $$\omega(\alpha_\lambda(A)) = \omega(A)
\quad \mbox{für alle} \quad A \in \cA(V), \ \lambda \in \cS \, .$$ Man ergänzt diese Relation noch durch einige physikalisch sinnvolle Bedingungen an die Natur der Korrelationen zwischen Paaren von Observablen in diesem Zustand, siehe [@BuRo]. Mit Hilfe der weiter oben erwähnten GNS–Konstruktion kann man dann einen Hilbertraum $\cH$ konstruieren, auf dem die Observablen in $\cA(V)$ operieren, sowie einen Zustandsvektor $\Omega \in \cH$ angeben, so dass $\omega(A) = \langle \Omega, A \Omega \rangle$ für alle $A \in \cA(V)$. Die Vektoren in $\cH$ beschreiben die durch Quanteneffekte von Messungen erzeugten lokalen Anregungen des Vakuums. Der interessante Punkt ist nun, dass die oben erwähnten charakteristischen Eigenschaften des partiellen Vakuumzustandes bereits ausreichen, um auf $\cH$ eine Darstellung der *vollen* Poincaré Gruppe $\cP$ durch unitäre Operatoren $U(\lambda)$, $\lambda \in \cP$ zu konstruieren [@BuRo]. Für ihre Einschränkung auf $\cS$ gelten die Relationen $$U(\lambda) A U(\lambda)^{-1} = \alpha_\lambda(A) \, , \quad
U(\lambda) \Omega = \Omega \, , \quad \lambda \in \cS \, .$$ Man kann also auch im Fall der auf Lichtkegel eingeschränkten Messungen alle Symmetrietransformationen durch unitäre Operatoren beschreiben, die sich aus den entsprechenden in Weise rekonstruieren lassen.
Darüber hinaus haben die Generatoren der resultierenden Raumzeittranslationen, die wie üblich als Energie–Impulsoperatoren gedeutet werden, die erwarteten spektralen Eigenschaften (Positivität der Energie für alle inertialen Beobachter). Sie lassen sich jedoch nicht in der üblichen Weise als Observablen deuten, da in ihnen die Fluktuationen des Energieinhalts in $V$ infolge von in der Vergangenheit abgestrahlten masselosen Teilchen nur pauschal berücksichtigt werden. Die Situation ist ähnlich wie in der Quantenstatistischen Mechanik, wo die Unkenntnis der Mikrozustände thermischer Systeme in die Liouvilleoperatoren statistisch eingeht. Ein charakteristisches Energie–Impulsspektrum, wie man es zum Beispiel für die lokalen Anregungen des Vakuums in der erwartet, ist in Figur \[spektrum\] dargestellt. Eingezeichnet sind die Zustände, die zu den jeweiligen Energie–Impulswerten beitragen: Die Spitze des Kegels entspricht dem Vakuum, der Rand den Energie–Impulswerten eines einzelnen Photons. Die Spektralwerte in der Nähe des Randes gehören zu Streuzuständen von Photonen (Delbrück–Streuung). Photonen genügend hoher Energie können bei Stoßprozessen Elektron–Positron Paare erzeugen, .
Ladungsklassen
==============
Auch wenn das Vakuum für die Konstruktion und Interpretation der Theorie von Bedeutung ist, so sind doch vor allem seine lokalen Anregungen und die daraus resultierenden ladungstragenden Zustände von physikalischem Interesse. Bevor wir uns der Konstruktion und Analyse dieser Zustände zuwenden, müssen wir zunächst besser verstehen, weshalb die bei der Interpretation der Theorie im Minkowskiraum auftretenden Infrarotprobleme bei unserem Zugang verschwinden.
Gemäß der Maxwellschen Elektrodynamik erzeugen beschleunigte Punktladungen Strahlung. Im Rahmen der Quantenfeldtheorie kann man diese Strahlung durch kohärente Zustände von Photonen beschreiben. Es stellt sich bei genauerer Analyse heraus, dass für Punktladungen mit verschiedenem einlaufenden und auslaufenden Impuls $p_{\mbox{\tiny \it ein}} \neq p_{\mbox{\tiny \it aus}}$ diese Zustände mit Sicherheit unendlich viele niederenergetische Photonen enthalten (sie können nicht durch Vektorzustände im Fockraum der Photonen beschrieben werden).
Man spricht daher, wie bereits mehrfach erwähnt, von Infrarotwolken. Es zeigt sich ferner, dass diese Infrarotwolken von den präzisen Werten der asymptotischen Impulse abhängen. Es gibt daher eine ungeheure Vielfalt von solchen global unterscheidbaren (im Sinne von Superauswahlsektoren) Infrarotwolken. Doch gibt es einen interessanten Spezialfall: Stimmen die beiden asymptotischen Impulse der Punktladung exakt überein, so erhält man kohärente Zustände im Fockraum der Photonen, d.h. mit endlicher Teilchenzahl. Diese Tatsache bleibt meist unbeachtet, da es ja extrem unwahrscheinlich ist, dass in realen Prozessen ein solches Ereignis auftritt. Doch spielt gerade dieser Spezialfall bei unseren Überlegungen eine wichtige Rolle. Stellt man sich nämlich die Frage, ob sich die Photonenwolken der beiden in Fig. \[stroeme\] angedeuteten Prozesse durch Messungen in einem Lichtkegel $V$ scharf lassen, so verdeutlicht Fig. \[fock\] dass dies nicht so sein sollte.
Denn durch Messung an den Photonen in $V$ wird sich nicht mit Sicherheit feststellen lassen, welche einlaufenden Impulse zwei Punktladungen mit dem gleichen auslaufenden Impuls in ferner Vergangenheit einmal hatten; durch Wechselwirkungsprozesse in der und dabei abgestrahlte marginale Photonen wird diese in den Superauswahlsektoren enthaltene globale Information völlig verwischt. Die Messergebnisse werden daher konsistent sein mit denen, die man in Zuständen mit übereinstimmenden einlaufenden und auslaufenden Impulsen erhalten würde. Anders gesagt, globale Infrarotwolken sollten sich bei Messungen in $V$ nicht von Zuständen im Fockraum der Photonen unterscheiden lassen. Tatsächlich lässt sich dies im Rahmen der Quantenfeldtheorie beweisen [@Bu].
Während die Infrarotsektoren sich also bei Messungen in Lichtkegeln nicht unterscheiden lassen, kann der Gesamtwert der von massiven Teilchen in einem Zustand getragenen Ladungen in Lichtkegeln ermittelt werden. Dies veranschaulicht Fig. \[ladung\], in der die Weltlinien von massiven geladenen Teilchen eingezeichnet sind, die sich natürlich alle mit weniger als Lichtgeschwindigkeit bewegen.
Massive Teilchen entgegengesetzter Ladung können in Stoßprozessen vernichtet werden (wobei z.B. hochenergetische Photonen entstehen) oder aber durch solche Prozesse in Paaren entgegengesetzter Ladung erzeugt werden. Diese Ereignisse ändern die eines Zustands nicht. Man kann daher aus der Analyse des massiven Teilcheninhalts in die von diesen Teilchen getragene Gesamtladung bestimmen. Diese wichtige Einsicht lässt sich im Rahmen der Quantenfeldtheorie präzise formulieren und begründen [@Bu].
In Analogie zu den Superauswahlsektoren im Minkowskiraum $\cM$ kann man daher im Rahmen der Theorie die Zustände mit gleicher Gesamtladung und auf gegebenem Lichtkegel $V$ ununterscheidbaren Infrarotwolken in Ladungsklassen einteilen. Dabei besteht jede Ladungsklasse aus Zuständen, deren Einschränkungen auf $\cA(V)$ durch die Quanteneffekte physikalischer Operationen in $V$ ineinander überführt werden können. Wir erinnern daran, dass Operationen in der Quantentheorie durch die Wirkung von unitären Operatoren auf Zustandsvektoren beschrieben werden: Stört man die Dynamik des Systems für einige Zeit durch einen in $V$ lokalisierten Operator, so kann man den Effekt dieser Störung im Wechselwirkungsbild berechnen. Nach Abschaltung der Störung erhält man einen unitären Operator $W \in \cA(V)$, der die Änderung der Zustände aufgrund dieser Operation beschreibt. Dies führt zu folgender Definition, die wir hier in etwas vereinfachter Form wiedergeben und auch so benutzen werden. Die präzise Formulierung findet sich in [@BuRo].
***Ladungsklassen:*** Zwei reine (ungemischte) Zustände $\varphi_1$, $\varphi_2$ auf der Observablenalgebra $\cA$ gehören zur gleichen Ladungsklasse, wenn es zu jedem gegebenem Lichtkegel $V$ einen entsprechenden unitären Operator $W \in \cA(V)$ gibt, so dass $$\varphi_2 \rest \cA(V) = \varphi_1 {\raisebox{1.2pt}{$\, \scriptscriptstyle \circ \, $}}\mbox{Ad} \, W \rest
\cA(V) \, ,$$ d.h. die partiellen Zustände werden durch die adjungierte Wirkung $\mbox{Ad} \, W = W \cdot W^{-1}$ von $W$ auf $\cA(V)$ ineinander überführt. Das Symbol ${\raisebox{1.2pt}{$\, \scriptscriptstyle \circ \, $}}$ bezeichnet die Komposition von Abbildungen.
Ladungen und Morphismen
=======================
Geht man, so wie bei der Konstruktion von Quantenfeldtheorien üblich, vom Vakuumzustand aus, so kann man mit Hilfe der vorangegangenen Definition alle Zustände in der Ladungsklasse des Vakuums charakterisieren, also alle neutralen Zustände, die keine globale Ladung tragen. Es stellt sich dann die Frage, wie man im Rahmen der Theorie aus diesen neutralen Zuständen geladene Zustände erhält. Auch hier ergibt sich die Antwort aus physikalische Überlegungen.
Wir wählen im Folgenden einen festen Lichtkegel $V$. Stellt man sich vor, dass inertiale Beobachter, die in der Spitze des Kegels mit unterschiedlichen Geschwindigkeiten in verschiedene Richtungen gestartet sind, alle zur gleichen Eigenzeit den Raumzeitpunkt, den sie erreicht haben, markieren, so erhält man einen Hyperboloid in $V$, den wir als Zeitschale bezeichnen, siehe Fig.\[zeitschale\]. Die asymptotisch lichtartigen Punkte auf einer solchen Zeitschale, die man nur mit Lichtgeschwindigkeit erreichen könnte, liegen raumartig zu allen kompakten Gebieten im Inneren von $V$.
Durch lokale Operationen mit genügend großem Energieübertrag kann man nun auf einer gegebenen Zeitschale aus dem Vakuum Paare von Teilchen mit entgegengesetzter Ladung erzeugen (Vakuumpolarisation), daneben entstehen Wolken von masselosen Teilchen. Um aus solchen global neutralen Paaren einen für Beobachter in $V$ effektiv ladungstragenden zu präparieren, muss die unerwünschte entgegengesetzte Ladung auf Lichtgeschwindigkeit gebracht werden. Sie befindet sich dann am asymptotisch lichtartigen Rand der Zeitschale und verschwindet dadurch aus dem kausalen Einflussgebiet von Beobachtern im Inneren von $V$. Dieser Limes würde in der Praxis unendlich viel Energie erfordern, die von der kompensierenden Ladung weggetragen wird. Man betrachtet hier also, wie häufig in der Theoretischen Physik, eine Idealisierung, da sie es gestattet, die Eigenschaften der Ladungen in Reinkultur zu analysieren. Wie in Fig. \[ladungen\] angedeutet, kann die Erzeugung von Paaren in kompakten Gebieten erfolgen, die Erzeugung der geladenen Limeszustände erfordert dagegen unendlich ausgedehnte Gebiete, die sich zur Begrenzung der von den masselosen Teilchen abgestrahlten Energie asymptotisch weiten müssen. Sie haben daher die Form hyperbolischer Kegel $\rm K$. Es ist hier wichtig, dass die Richtung und der Öffnungswinkel dieser Kegel bei der Präparation von Zuständen mit gegebener Ladung frei gewählt werden kann. Wir bezeichnen die kausalen Abhängigkeitsgebiete (Abschlüsse) der Kegel $\rm K$ mit kalligrafischen Buchstaben $\cK$ und nennen sie kurz Hyperkegel.
Nach diesen Vorbemerkungen können wir nun beschreiben, wie man in der Theorie, ausgehend vom Vakuumzustand $\omega$, geladene Zustände erhält. Dazu gibt man sich zunächst einen Hyperkegel $\cK \subset V$ vor, in dem die Präparation erfolgen soll. Dann wählt man Folgen von in $\cK$ lokalisierten unitären Operatoren, $\{W_n \in \cA(\cK)\}_{n \in \NN}$, die die Operation der Paarerzeugung beschreiben, wobei für wachsendes $n$ die kompensierende Ladung immer weiter in asymptotisch lichtartiger Richtung verschoben wird. Bei geeigneter Wahl der Operatorfolge konvergieren die resultierenden partiellen Zuständen in der Ladungsklasse des Vakuums, $\{ \omega {\raisebox{1.2pt}{$\, \scriptscriptstyle \circ \, $}}\mbox{Ad} \, W_n \rest \cA(V) \}_{n \in \NN}$, gegen einen partiellen Zustand $\varphi \rest \cA(V)$ aus der gewünschten Ladungsklasse. Die kompensierende Ladung ist in $\varphi \rest \cA(V)$ nicht mehr nachweisbar, sie befindet sich im raumartigen Komplement der in $V$ zur Verfügung stehenden Observablen.
Ausgehend von dieser physikalisch transparenten aber umständlichen Methode kann man zu einer bequemeren, aber etwas abstrakteren Beschreibung der Ladungserzeugung übergehen. Es zeigt sich nämlich, dass auch die Folgen der von den unitären Operatoren induzierten Abbildungen $\{ \mbox{Ad} \, W_n \}_{n \in \NN}$ der Observablenalgebra $\cA(V)$ auf sich selbst konvergieren, wobei wir hier nicht auf die in [@BuRo] diskutierten mathematischen Details eingehen können. Für die Grenzwerte führen wir die Notation ein $$\sigma_\cK = \lim_{n \rightarrow \infty} \, \mbox{Ad} \, W_n
\, .$$ Der Index $\cK$ erinnert daran, dass $\sigma_\cK$ Grenzwert von Operationen im Hyperkegel $\cK$ ist. Die so definierten Abbildungen $\sigma_\cK$ sind, wie ihre Approximationen, verträglich mit der algebraischen Struktur von $\cA(V)$: sie sind linear, multiplikativ und symmetrisch; der Einfachheit halber nehmen wir ferner an, dass ihr Bildbereich ebenfalls in $\cA(V)$ liegt. Wir bezeichnen diese Abbildungen kurz als (in $\cK$ lokalisierte) Morphismen.
Die in der Monographie [@Ha] erläuterten tiefen Resultate über die Sektorstruktur von Quantenfeldtheorien massiver Teilchen basieren auf der Analyse solcher Morphismen. Auch im vorliegenden Kontext von Theorien mit masselosen Teilchen erweisen sie sich als nützliches analytisches Instrument. Denn in ihnen sind in mathematisch präziser Weise alle physikalischen Informationen enthalten, die wir in den vorangegangenen Abschnitten zusammengetragen haben. Wir fassen sie in der folgenden Liste zusammen, in der einige technische Details etwas vereinfacht dargestellt sind. (Die präzise Formulierung findet sich in [@BuRo].)
***Lokalisierte Morphismen:*** Zu gegebener Ladungsklasse und beliebigem Hyperkegel $\cK \subset V$ gibt es einen auf $\cA(V)$ definierten lokalisierten Morphismus $\sigma_\cK$ mit den folgenden Eigenschaften.
- $\omega {\raisebox{1.2pt}{$\, \scriptscriptstyle \circ \, $}}\sigma_\cK \rest \cA(V)$ stimmt mit einem partiellen Zustand aus der gegebenen Ladungsklasse überein.
- Für jedes Raumzeitgebiet $\cR \subset V$, das $\cK$ enthält, gilt $\sigma_\cK(\cA(\cR)) \subseteq
\cA(\cR)$. In den Fällen wo man Gleichheit hat, bezeichnet man die Ladungsklasse und die entsprechenden Morphismen als “einfach”.
- Für jedes Raumzeitgebiet $\cR \subset V$, das raumartig von $\cK$ getrennt ist, gilt $\sigma_\cK \rest \cA(\cR) = \iota$, worin $\iota$ die identische (triviale) Abbildung bezeichnet.
- Zu jedem Paar von Morphismen $\sigma_{\cK_1}$, $\sigma_{\cK_2}$, die mit der Ladungsklasse affiliiert sind, existieren unitäre Operatoren $W_{2 1} \in \cA(V)$, die die Morphismen verknüpfen, . Man nennt die Operatoren $W_{2 1}$ Verknüpfungsoperatoren.
Punkt (a) dieser Liste wurde bereits ausführlich erläutert. Die Punkte (b) und (c) folgen aus den Eigenschaften der approximierenden Operatoren $W_n$. Im Fall von (b) gilt $W_n \in \cA(\cK) \subset \cA(\cR)$ und somit $\mbox{Ad} \, W_n (\cA(\cR)) \subseteq \cA(\cR)$ und im Fall (c) gilt aufgrund der Vertauschbarkeit von in $\cR$ und $\cK$ lokalisierten Operatoren (Lokalitätsprinzip) $\mbox{Ad} \, W_n \rest \cA(\cR) = \iota$. Diese Eigenschaften übertragen sich auf den Limes $\sigma_\cK$. In Punkt (d) kommt die Tatsache zum Ausdruck, dass die Superauswahlsektoren der bei der Erzeugung von Ladungen unvermeidlich entstehenden Infrarotwolken bei Messungen in $V$ nicht unterschieden werden können. Die Zustände $\omega {\raisebox{1.2pt}{$\, \scriptscriptstyle \circ \, $}}\sigma_{\cK_1} \rest \cA(V)$ und $\omega {\raisebox{1.2pt}{$\, \scriptscriptstyle \circ \, $}}\sigma_{\cK_2} \rest \cA(V)$ gehören daher zur gleichen Ladungsklasse und können durch die adjungierte Wirkung unitärer Operatoren in $\cA(V)$ ineinander transformiert werden. Diese Operatoren verknüpfen auch die lokalisierten Morphismen.
Wir werden uns im Folgenden auf die Diskussion der in (b) charakterisierten einfachen Ladungsklassen bzw. Morphismen beschränken. Zu ihnen gehört die uns hier besonders interessierende elektrische Ladung und die Univalenz.
Ladungskonjugation und Statistik
================================
Mit Hilfe der lokalisierten Morphismen lässt sich die Struktur der Ladungsklassen in der Quantenfeldtheorie in systematischer Weise analysieren und klassifizieren. Dazu muss man voraussetzen, dass die Theorie alle in Hyperkegeln $\cK$ prinzipiell möglichen Messungen und Operationen beschreibt, d.h. dass sich die Algebren $\cA(\cK)$ nicht durch Hinzufügen weiterer Operatoren vergrößern lassen, ohne mit dem Lokalitätsprinzip in Konflikt zu geraten. Sie sind also in diesem Sinne maximal. Diese physikalisch motivierte Forderung an die Theorie lässt sich mathematisch in Form einer Dualitätsbedingung ausdrücken [@BuRo].
Wie bereits angedeutet, werden wir uns im Folgenden auf die Diskussion der einfachen Ladungsklassen beschränken, ohne dies jedes Mal zu erwähnen. Der erste und entscheidende Schritt bei der Analyse ist der Beweis, dass sich Ladungsklassen komponieren lassen. Anschaulich gesprochen entspricht diese Komposition der Addition des Ladungsinhaltes von Zuständen. Genauer gesagt, man kann zeigen [@BuRo], dass für gegebene Ladungsklassen $L_1$, $L_2$ von Zuständen und Lichtkegel $V$ die jeweils zu diesen Klassen assoziierten, in Hyperkegeln lokalisierten Morphismen $\sigma_{\cK_1}$, $\tau_{\cK_2}$ komponiert werden können, $\tau_{\cK_2} {\raisebox{1.2pt}{$\, \scriptscriptstyle \circ \, $}}\sigma_{\cK_1}$. Die resultierenden partiellen Zustände $\omega {\raisebox{1.2pt}{$\, \scriptscriptstyle \circ \, $}}\tau_{\cK_2} {\raisebox{1.2pt}{$\, \scriptscriptstyle \circ \, $}}\sigma_{\cK_1} \rest \cA(V)$ gehören unabhängig von der Wahl von $\cK_1$, $\cK_2$ alle zu einer bestimmten Ladungsklasse $L_3$, die sich ebenfalls durch lokalisierte Morphismen $\rho_{\cK_3}$ beschreiben lässt. Insbesondere gibt es also unitäre Operatoren $W_{3 2 1} \in \cA(V)$, die diese Morphismen verknüpfen, $\rho_{\cK_3} = \mbox{Ad} \, W_{3 2 1} {\raisebox{1.2pt}{$\, \scriptscriptstyle \circ \, $}}\tau_{\cK_2} {\raisebox{1.2pt}{$\, \scriptscriptstyle \circ \, $}}\sigma_{\cK_1}$. Dies Resultat zeigt, dass bei der Komposition von Ladungen keine neuartigen Typen von Infrarotwolken entstehen, deren Superauswahlsektoren sich in $V$ unterscheiden lassen.
Bei unserer heuristischen Diskussion der Ladungserzeugung waren wir von der Existenz von entgegengesetzten Ladungen ausgegangen, die sich gegenseitig neutralisieren. Es zeigt sich, dass auch diese empirische Tatsache bereits in der Struktur der lokalisierten Morphismen kodiert ist [@BuRo]: Zu jeder Ladungsklasse $L$ von Zuständen und gegebenem Lichtkegel $V$ mit dazu assoziierten in Hyperkegeln lokalisierten Morphismen $\sigma_\cK$ existiert eine konjugierte Ladungsklasse $\overline{L}$ mit entsprechend lokalisierten Morphismen $\overline{\sigma}_\cK$ für die gilt $\sigma_\cK {\raisebox{1.2pt}{$\, \scriptscriptstyle \circ \, $}}\overline{\sigma}_\cK =
\overline{\sigma}_\cK {\raisebox{1.2pt}{$\, \scriptscriptstyle \circ \, $}}\sigma_\cK = \iota$. Das heißt, alle Effekte der Ladungen in $L$ lassen sich durch Komposition mit bestimmten Konfigurationen der konjugierten Ladungen aus $\overline{L}$ vollständig neutralisieren und umgekehrt. Die mit $\overline{L}$ affiliierten Zustände werden daher als zu $L$ ladungskonjugierte Antimaterie interpretiert.
Eine weitere fundamentale Eigenschaft von Ladungen, nämlich die Tatsache, dass sie bei Vertauschung in Zuständen (je nach Ladungstyp) entweder der Bose– oder der Fermistatistik genügen, lässt sich ebenfalls aus den Eigenschaften der lokalisierten Morphismen herleiten [@BuRo]. Dazu betrachtet man bei gegebener Ladungsklasse $L$ und Lichtkegel $V$ zwei assoziierte, beliebig lokalisierte Morphismen $\sigma_{\cK_1}$, $\sigma_{\cK_2}$ und vergleicht die resultierenden komponierten Morphismen $\sigma_{\cK_1} {\raisebox{1.2pt}{$\, \scriptscriptstyle \circ \, $}}\sigma_{\cK_2}$ bzw. $\sigma_{\cK_2} {\raisebox{1.2pt}{$\, \scriptscriptstyle \circ \, $}}\sigma_{\cK_1}$. Wie oben erläutert, gehören diese komponierten Morphismen zur gleichen Ladungsklasse (mit verdoppelter Ladung) und sind durch unitäre Verknüpfungsoperatoren miteinander verbunden. Tatsächlich gibt es eine kanonische Wahl für diesen Verknüpfungsoperator, der in einem feldtheorischen Rahmen als gruppentheoretischer Kommutator von ladungstragenden (und daher nicht observablen) unitären Operatoren gedeutet werde könnte.
Um diesen Sachverhalt zu beleuchten, nehmen wir für einen Moment an, dass es ladungstragende, in $\cK_j$ lokalisierte unitäre Operatoren $V_j$ gibt so dass $\sigma_{\cK_j} = \mbox{Ad} \, V_j$, $j = 1,2$. Der Operator $W_{2 1} = V_2 V_1^{-1}$ wäre dann ein Verknüpfungsoperator zwischen $\sigma_{\cK_1}$ und $\sigma_{\cK_2}$ im Sinne der oben angegebenen Definition und der gruppentheoretische Kommutator $V_2 V_1 V_2^{-1} V_1^{-1}$ wäre ein Verknüpfungsoperator zwischen $\sigma_{\cK_1} {\raisebox{1.2pt}{$\, \scriptscriptstyle \circ \, $}}\sigma_{\cK_2} = \mbox{Ad} \, V_1 V_2$ und $\sigma_{\cK_2} {\raisebox{1.2pt}{$\, \scriptscriptstyle \circ \, $}}\sigma_{\cK_1} = \mbox{Ad} \, V_2 V_1$. Der interessante Punkt ist nun, dass sich dieser gruppentheoretische Kommutator auch ohne Kenntnis von ladungstragenden Operatoren berechnen lässt gemäß $$V_2 V_1 V_2^{-1} V_1^{-1} =
(V_2 V_1^{-1}) V_1 ( V_1 V_2^{-1}) V_1^{-1} =
W_{2 1} \, \mbox{Ad} \, V_1 (W_{2 1}^{-1}) =
W_{2 1} \sigma_{\cK_1}(W_{2 1}^{-1}) \, .$$ Auf der rechten Seite dieser Gleichung steht eine Größe, die in unserem nur auf Observablen aufgebauten Formalismus bestimmt werden kann.
Diese Überlegungen legen es nahe, für die gegebenen Morphismen $\sigma_{\cK_1}$, $\sigma_{\cK_2}$ und entsprechendem Verknüpfungsoperator $W_{2 1} \in \cA(V)$ den Operator $\varepsilon(\sigma_{\cK_1}, \sigma_{\cK_2}) =
W_{2 1} \, \sigma_{\cK_1}(W_{2 1}^{-1})$ einzuführen. Man stellt fest, dass $\varepsilon(\sigma_{\cK_1}, \sigma_{\cK_2})$ nicht von der Wahl von $W_{2 1}$ abhängt sowie die komponierten Morphismen $\sigma_{\cK_1} {\raisebox{1.2pt}{$\, \scriptscriptstyle \circ \, $}}\sigma_{\cK_2}$ und $\sigma_{\cK_2} {\raisebox{1.2pt}{$\, \scriptscriptstyle \circ \, $}}\sigma_{\cK_1}$ verknüpft. In diesem Sinne ist
der Operator kanonisch. Der physikalisch bedeutsame Punkt ist nun, dass für Morphismen, die in raumartig getrennten Hyperkegeln $\cK_1, \cK_2$ lokalisiert sind, siehe Fig. \[statistik\], der Operator $\varepsilon(\sigma_{\cK_1}, \sigma_{\cK_2})$ eine besonders einfache Form hat. Man kann nämlich zeigen [@BuRo], dass es zu jeder Ladungsklasse $L$ eine Zahl $\varepsilon_L \in \{\pm 1 \}$ (genannt Statistikparameter) gibt, so dass $\varepsilon(\sigma_{\cK_1}, \sigma_{\cK_2}) = \varepsilon_L 1$ für alle mit der Klasse assoziierten, in raumartig getrennten Hyperkegeln lokalisierten Morphismen $\sigma_{\cK_1},
\sigma_{\cK_2}$. Die physikalische Bedeutung dieses Ergebnisses wird klar, wenn man es durch die oben betrachteten ladungstragenden Operatoren $V_j$, $j=1,2$ reformuliert: Falls $\varepsilon_L = 1$ müssen diese Operatoren bei raumartigen Abständen miteinander kommutieren (bosonischer Fall), falls $\varepsilon_L = -1$ müssen sie antikommutieren (fermionischer Fall). Jede Ladungsklasse $L$ genügt also entweder der Bose oder der Fermi Statistik. Andere Möglichkeiten gibt es für die hier betrachteten einfachen Ladungen nicht. Man kann ferner zeigen [@BuRo], dass die Statistik der konjugierten Klasse $\overline{L}$ stets mit der von $L$ übereinstimmt, $\varepsilon_{\overline{L}} = \varepsilon_L^{ }$.
Es sei betont, dass diese Ergebnisse ganz wesentlich von der Tatsache abhängen, dass die physikalische Raumzeit vierdimensional ist. In hypothetischen Welten niedrigerer Dimension würde die gleiche Analyse ergeben, dass Ladungsklassen auch einer anderen (z.B. anyonischen) Statistik genügen können. Man gelangt somit, wie im massiven Fall [@Ha], auch in Theorien mit masselosen Teilchen zu einem vertieften Verständnis der Wurzeln der Teilchenstatistik.
Kovarianz und Spektrum
======================
Ein weiterer physikalisch wichtiger Punkt ist die Frage, wie man im vorliegend Rahmen die relativistische Invarianz der Ladungsklassen und die energetischen Eigenschaften der entsprechenden Zustände beschreibt. Wie bereits erwähnt, wirkt auf einen gegebenen Lichtkegel $V$ in natürlicher Weise nur die Semigruppe $\cS \subset \cP$ der Lorentztransformationen und zukunftsgerichteten Translationen, die $V$ auf bzw. in sich abbilden. Wir müssen daher zunächst diskutieren, wie man in physikalisch sinnvoller Weise den Transport von lokalisierten Morphismen $\sigma$ . (Da wir im folgenden einen festen Morphismus betrachten spielt dessen Lokalisationsgebiet keine Rolle.)
Zur Lösung dieses Problems gehen wir von einigen physikalisch motivierten Annahmen aus. Für eine gegebene Transformation $\lambda \in \cS$ sollte es im Prinzip möglich sein, die durch $\sigma$ beschriebene Ladungserzeugung in $V$ in exakt der gleichen Weise auf dem transformierten Lichtkegel zu bewerkstelligen (Wiederholbarkeit des Experiments). Es sollte also Morphismen ${}^\lambda \sigma$ auf $\cA(V)$ geben für die gilt ${}^\lambda \sigma {\raisebox{1.2pt}{$\, \scriptscriptstyle \circ \, $}}\alpha_\lambda
= \alpha_\lambda {\raisebox{1.2pt}{$\, \scriptscriptstyle \circ \, $}}\sigma$, wobei daran erinnert sei, dass . Der Ladungsinhalt ändert sich beim Transport nicht, die Morphismen ${}^\lambda \sigma$ gehören somit alle zur gleichen Ladungsklasse und es gibt unitäre Verknüpfungsoperatoren $W_\lambda \in \cA(V)$, so dass $\mbox{Ad} \, W_\lambda {\raisebox{1.2pt}{$\, \scriptscriptstyle \circ \, $}}{}^\lambda \sigma = \sigma$, $\lambda \in \cS$. Die relativistische Invarianz der Theorie kommt dann dadurch zum Ausdruck, dass die durch $W_\lambda$ beschriebenen physikalischen Operationen auf allen transportierten Lichtkegeln $\mu V$, $\mu \in \cS$ in analoger Weise operieren, d.h. $\alpha_\mu(W_\lambda)$ ist auch Verknüpfungsoperator zwischen den transportierten Morphismen ${}^{\mu \lambda} \sigma$ und ${}^{\mu} \sigma$ für $\lambda, \mu \in \cS$. (Hier benutzt man die Tatsache dass $\cS$ eine Semigruppe ist, dass also das dem sukzessiven Transport entsprechende Produkt erklärt ist.) Wir fassen diese physikalisch erwarteten Eigenschaften in folgender Definition zusammen.
***Kovariante Morphismen:*** Ein Morphismus $\sigma$ auf $\cA(V)$ heißt kovariant, wenn es eine dazu assoziierte Familie von transportierten Morphismen ${}^\lambda \sigma$, $\lambda \in \cS$ gibt mit den oben angegebenen Eigenschaften.
Es ist bemerkenswert, dass diese zwar physikalisch gut begründete aber implizite Charakterisierung der kovarianten Morphismen ausreicht, um die gewohnte Formulierung der relativistischen Kovarianz zu etablieren [@BuRo]. Ausgehend von der durch das Vakuum fixierten unitären Darstellung $U$ der vollen Poincaré Gruppe $\cP$ kann man nämlich zeigen, dass die gemäß $U_\sigma(\lambda) = U(\lambda) W_\lambda$, $\lambda \in \cS$ definierten unitären Operatoren die Transformation der Observablen in der Ladungsklasse von $\sigma$ implementieren, $$\mbox{Ad} \, U_\sigma(\lambda) {\raisebox{1.2pt}{$\, \scriptscriptstyle \circ \, $}}\sigma = \sigma {\raisebox{1.2pt}{$\, \scriptscriptstyle \circ \, $}}\alpha_\lambda \, , \quad \lambda \in \cS \, .$$ Darüber hinaus existiert stets eine Wahl der Verknüpfungsoperatoren $W_\lambda$, so dass sich die Operatoren $U_\sigma(\lambda)$, $\lambda \in \cS$ eindeutig zu einer unitären Darstellung von $\cP$ (bzw. der Überlagerungsgruppe von $\cP$) fortsetzen lassen. Beobachter in $V$ können somit in eindeutiger Weise über den Energieinhalt und Drehimpuls (Spin) von Zuständen in der gegebenen Ladungsklasse sprechen. Diese ergeben sich aus den Generatoren der Darstellung $U_\sigma$, die jedoch nicht direkt als Observablen werden können. Observablen im üblichen Sinne sind die unitären (und damit im Sinne der Spektraltheorie normalen) Verknüpfungsoperatoren $W_\lambda = U(\lambda)^{-1} U_\sigma(\lambda) \in \cA(V)$, $\lambda \in \cS$ aus denen man z.B. Informationen über den Energieunterschied zwischen Zuständen in der Ladungsklasse des Vakuums und der geladenen Klasse erhält. Mit der getroffenen Wahl des Vakuums als Referenzzustand und der sich daraus eindeutig ergebenden Darstellung $U$ von $\cP$ kann man diese Informationen dann mithilfe der Darstellungen $U_\sigma$ in gewohnter Weise beschreiben.
Ein weiteres Resultat, das die physikalische Konsistenz unseres theoretischen Zugangs untermauert, betrifft die Form des Energie–Impuls Spektrums in den Ladungsklassen. Durch eine detaillierte Analyse des sich bei der Komposition von konjugierten Ladungsklassen ergebenden Spektrums lässt sich zeigen [@BuRo], dass die Generatoren der raumzeitlichen Verschiebungen $U_\sigma \rest \RR^4$ in allen Ladungsklassen die relativistische Spektrumsbedingung erfüllen (Positivität der Energie in allen Inertialsystemen). Diese energetische Stabilität der Ladungsklassen ist also eine Konsequenz der Struktur der Observablen und des Vakuums und muss nicht extra postuliert werden.
Fazit
=====
Die vorangegangene Diskussion hat verdeutlicht, dass die notorischen Infrarotprobleme bei der Interpretation von Quantenfeldtheorien mit langreichweitigen Kräften auf einer Überidealisierung beruhen: Nimmt man im Rahmen der Theorie wie gewöhnlich an, dass der masselose Teilcheninhalt der Zustände durch Messungen im gesamten Minkowskiraum bestimmt werden kann, so ist man gezwungen, eine Vielzahl unendlicher Infrarotwolken masseloser Teilchen in Betracht zu ziehen. Berücksichtigt man jedoch die Existenz des Zeitpfeils, so ist offensichtlich, dass sich Messungen bestenfalls in zukunftsgerichteten Lichtkegeln durchführen lassen. Versäumte Messungen in der Vergangenheit bedeuten im Allgemeinen einen unwiederbringlichen Verlust an Informationen, und die Theorie kann sich daher auf die Beschreibung und Interpretation von Lichtkegeldaten beschränken.
Wie wir gesehen haben, führt diese Beschränkung zu einer sowohl mit den Symmetrien des Minkowskiraums als auch dessen kausaler Struktur verträglichen Infrarotregularisierung. Das aus der Einsteinschen Kausalitätsforderung ableitbare Huygenschen Prinzips bewirkt die geometrische Aufteilung des globalen masselosen Teilcheninhalts in einen in Lichtkegeln experimentell zugänglichen essentiellen und einen in der Vergangenheit abgestrahlten und daher für alle Zeiten verlorenen marginalen Anteil. Angesichts dieser Fakten wird im Rahmen der Theorie verständlich, weshalb sich globale Infrarotwolken nicht scharf unterscheiden lassen.
Auch die Frage, welche physikalischen Aspekte einer Theorie in Lichtkegeln bestimmt werden können, konnte beantwortet werden. Dies sind zum einen die Eigenschaften massiver Teilchen, insbesondere deren Ladungen und Statistik, sowie die Existenz und Eigenschaften von Antimaterie. Darüber hinaus lassen sich aus den auf Lichtkegel eingeschränkten partiellen Zuständen auch deren energetische und rotatorischen Eigenschaften bestimmen. Materie kann also auch in diesem neuen theoretischen Zugang detailliert beschrieben werden.
Der in Lichtkegeln nachweisbare masselose Teilcheninhalt lässt sich durch Zustände mit endlicher Teilchenzahl (also im Fockraum) beschreiben. Es sei betont, dass dies auch masselose Teilchen einschließt, die den Lichtkegel aus fernen Galaxien erreichen. Diese Teilchen wurden zwar vor Äonen erzeugt, doch kommen sie nicht aus der Vergangenheit des Beobachters, sondern aus zum Beobachter zuvor raumartig getrennt liegenden Gebieten. Stellt man sich nun vor, dass durch diese Teilchen Informationen über das Ergebnis von Experimenten unserer Kollegen übermittelt werden, so kann man sich die Frage stellen, ob diese Informationen im Widerspruch zu den in hiesigen Lichtkegeln gewonnenen physikalischen Einsichten stehen könnten. Die Antwort ist beruhigend: Auch die extragalaktischen Kollegen können ihre Experimente bestenfalls in Lichtkegeln durchführen und man kann zeigen [@BuRo], dass die Vorhersagen der Theorie nicht von der Wahl eines Lichtkegels abhängen. Somit ist die innere Konsistenz unserer zwar ungewohnten aber dafür problemfreien Interpretation der relativistischen Quantenfeldtheorie gesichert.
Danksagung {#danksagung .unnumbered}
==========
Mein Dank gilt dem Evangelischen Studienwerk, besonders Herrn Prof. Dr. Eberhard Müller, für die langjährige fruchtbare Zusammenarbeit und stete Förderung des Promotionsschwerpunktes “Wechselwirkung”.
[99]{}
D. Buchholz, The physical state space of quantum electrodynamics.\
Commun. Math. Phys. [**85**]{} (1982) 49
D. Buchholz und J. E. Roberts, New light on infrared problems: Sectors, statistics, symmetries and spectrum.\
e–print arXiv:1304.2794
R. Haag, *Local Quantum Physics: Fields: Particles, Algebras*.\
Berlin, Springer (1992)
O. Steinmann, *Perturbative Quantum Electrodynamics and Axiomatic Field Theory*.\
Berlin Springer (2000)
F. Strocchi, *An Introduction to Non-Perturbative Foundations of Quantum Field Theory*.\
Oxford University Press (2013)
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'For a small sessile or pendant droplet it is generally assumed that gravity does not play any role once the Bond number is small. This is even assumed for evaporating binary sessile or pendant droplets, in which convective flows can be driven due to selective evaporation of one component and the resulting concentration and thus surface tension differences at the air-liquid interface. However, recent studies have shown that in such droplets gravity indeed can play a role and that natural convection can be the dominant driving mechanism for the flow inside evaporating binary droplets (Edwards et al., Phys. Rev. Lett. **121**, 184501 (2018); Li et al., Phys. Rev. Lett. **122**, 114501 (2019)). In this study, we derive and validate a quasi-stationary model for the flow inside evaporating binary sessile and pendant droplets, which successfully allows to predict the prevalence and the intriguing interaction of Rayleigh and/or Marangoni convection on the basis of a phase diagram for the flow field expressed in terms of the Rayleigh and Marangoni numbers.'
author:
- 'Christian Diddens, Yaxing Li, Detlef Lohse'
bibliography:
- 'refs.bib'
title: Competing Marangoni and Rayleigh convection in evaporating binary droplets
---
Introduction {#sec:intro:intro}
============
Evaporating droplets frequently occur in nature and applications, be it a rain droplet evaporating on a leaf, a droplet on a hot surface in spray cooling, a droplet of insecticides sprayed on a leaf or an inkjet-printed ink droplet on paper. Many of such droplets are multicomponent, i.e. consisting of a mixture of liquids. From the physical point of view, an evaporating multicomponent droplet in a host gas is paradigmatic for combined multi-phase and multi-component flow including a phase transition. Scientifically, this process encompasses the various fields of fluid mechanics, thermodynamics and also aspects from the field of chemistry. The evaporation dynamics is also relevant for the deposit left behind the evaporation of a particle-laden droplet. Here, pioneering work was done by @Deegan1997a around 20 years ago, when they identified the coffee-stain effect – i.e. the phenomenon of finding a typical ring structure of deposited particles after the evaporation of a coffee droplet – and successfully explained it by the combination of a non-uniform evaporation rate along the droplet interface and a pinned contact line. In applications, one usually wants to prevent such coffee-stain effect, e.g. for obtaining a homogeneous deposition pattern in inkjet printing [@Park2006; @Kuang2014; @Hoath2016; @Sefiane2014]. For reviews on evaporating pure droplets we refer to @Cazabat2010 and @Erbil2012.
Preventing the coffee-stain effect can be achieved by altering the flow inside the droplet during the drying process by inducing gradients in the acting forces. Focussing on the interfacial forces first, a tangential gradient of the surface tension along the liquid-gas interface leads to the well-known Marangoni effect, i.e. a tangential traction that drives the liquid towards positions of higher surface tension [@Scriven1960; @Pearson1958]. By that, the entire flow in the droplet can be altered from the typical outwards flow towards the contact line to a recirculating flow driven by a persistent Marangoni effect [@Hu2006]. For the case of a pure droplet, the necessary gradient in surface tension can be generated by thermal effects, e.g. either self-induced by latent heat of evaporation or externally imposed by heating or cooling the substrate [@Girard2006; @Sodtke2008; @Dunn2009; @Tam2009].
The other mechanism to induce Marangoni flow is known as solutal Marangoni effect, which is usually much stronger. For solutal Marangoni flow, the droplet must consist of more than one component, e.g. a solvent and one or more surfactants [@Still2012; @Marin2016; @Kwiecinski2019] or a solvent and possibly multiple co-solvents [@Sefiane2003a; @Christy2011; @Tan2016a; @Li2018] or dissolved salts [@Soulie2015; @Marin2019]. For a recent perspective review on droplets consisting of more than one component, we refer to @Lohse2020. The difference in the volatilities of the individual constituents leads to preferential evaporation of one or the other component and thereby compositional gradients are induced. Since the surface tension is a function of the composition and due to the non-uniform evaporation profile, a surface tension gradient along the liquid-gas interface can build up and result in a similar Marangoni circulation as in the thermally-driven case. The nature of the resulting flow can be quite different, mostly depending on whether the evaporation process leads to an overall decreasing or increasing surface tension, i.e. whether the more volatile component has a higher or lower surface tension than the less volatile component.
In a binary droplet consisting e.g. of water and glycerol, with water being more volatile and having the higher surface tension, the overall surface tension decreases during the preferential evaporation of water and the resulting Marangoni flow is usually regular, axisymmetric and directed towards the position of the lowest evaporation rate of water, i.e. towards the contact line for contact angles above $\SI{90}{\degree}$ and towards the apex for contact angles below $\SI{90}{\degree}$ [@Diddens2017a; @Diddens2017c].
On the contrary, e.g. in case of a binary droplet consisting of water and ethanol, where the overall surface tension increases due to the predominant evaporation of ethanol, the typical Marangoni effect is way more violent and chaotic [@Christy2011; @Bennacer2014]. Here, in particular, the axial symmetry of the droplet is usually broken, leading to a complicated scenario of initially chaotic flow driven by the solutal Marangoni effect and followed by either thermal Marangoni flow or the typical coffee-stain flow, when almost only water is left [@Diddens2017]. Remarkably, the presence of a strong Marangoni effect can also have a significant influence on the shape and wetting behaviour of droplets [@Tsoumpas2015; @Karpitschka2017]. Finally, the evaporation of mixture droplets can show a variety of additional intriguing phenomena, e.g. multiple phase changes and microdroplet nucleation in ternary droplets like ouzo [@Tan2016a; @Tan2017], and phase segregation in binary droplets [@Li2018] or rather homogeneous deposition patterns by an interplay of Marangoni flow, surfactants and polymers [@Kim2016].
As highlighted above, besides the gradient in the surface tension, i.e. in the interfacial forces, also gradients in the mass density, i.e. in the bulk force due to gravity, can influence the flow by natural convection. Similar to the surface tension, the mass density is a function of the temperature and, in the case of mixtures, of the composition, so that thermally and solutally driven natural convection can be realized in evaporating droplets. Flow driven by natural convection is one of the most important fields of fluid mechanics, as e.g. in Rayleigh-Bénard systems, however, these are usually investigated at large spatial dimensions. For small droplets, on the other hand, one would naively expect that the flow in case of thermal or solutal gradients is predominantly driven by the Marangoni effect, since a small droplet is associated with a small Bond number and hence surface tension effects would dominate over gravity. As a consequence, most studies on droplet evaporation focus on the Marangoni effect, but disregard the presence of natural convection by this argument.
Recent studies, however, showed that even for small droplets with small Bond numbers, the internal flow can be decisively determined by natural convection and not by Marangoni flow [@Edwards2018; @Li2019]. This has been even found at the later stages of water-ethanol droplets, which initially show a very intense chaotic Marangoni flow [@Edwards2018]. Obviously, these findings give rise to the following question: Under what circumstances which kind of flow pattern can be found in an evaporating binary droplet, i.e. when is the flow dominated by the Marangoni effect and when by natural convection?
In this manuscript, we answer this questions by carefully investigating both kinds of driving forces and their mutual interaction. The corresponding effects can be quantified by non-dimensional numbers, namely the Marangoni number for flow due to surface tension gradients and the Rayleigh (or Archimedes/Grashof) number for the natural convection. By considering quasi-stationary instants during the drying process, these numbers successfully allow to predict the flow inside the droplets on the basis of phase diagrams in the $\mbox{\it Ra}$-$\mbox{\it Ma}$ parameter space. We also validated these phase diagrams with full simulations and corresponding experiments.
The paper is organised as follows: We will first present the complete set of dynamical equations describing the evaporation of a binary mixture droplet. In section \[sec:motiv:motiv\], these equations are solved to discuss an illustrative example case. We will then introduce the quasi-stationary approximation in section \[sec:simpmodel:simpmodel\] and discuss the phase diagrams obtained by this model in section \[sec:phasediag:phasediag\]. The paper ends with a conclusion and a comparison with experimental data in the appendix.
Governing Equations {#sec:detmodel:detmodel}
===================
![Schematic of the model. The problem is considered to be axisymmetric and isothermal. The flow and the advection-diffusion equation for the composition in the droplet is solved with consideration of gravity and the composition-dependence of the liquid mass density, dynamic viscosity and diffusivity. The transport in the gas phase is assumed to be diffusion-limited. At the interface, Raoult’s law is used to enforce the vapour-liquid equilibrium, mass transfer due to evaporation is considered and the Marangoni shear stress due to a composition-dependent surface tension is taken into account. []{data-label="fig:detmodel:scheme"}](fig01){width="99.00000%"}
The evaporation of a mixture droplet is a multi-phase and free interface problem with multi-component dynamics in both the liquid and gas phase. For a binary droplet, the liquid is constituted by two components, $\alpha=\text{A},\,\text{B}$, whereas the gas phase is in general a ternary gas mixture of the ambient medium, e.g. air, and the vapours of the components $\text{A}$ and $\text{B}$. When the droplet evaporates at a temperature $T$ far below the boiling point and in absence of forced or strong natural convection, the diffusive dynamics in the gas phase can be approximated by the quasi-stationary Laplace equation [@Deegan2000a; @Hu2002a; @Popov2005; @Diddens2017a] $$\nabla^2c_{\alpha}=0
\label{eq:detmodel:laplgas}$$ for the vapour concentrations $c_{\alpha}$, i.e. the partial mass densities. The corresponding boundary conditions are given by the vapour-liquid equilibrium according to Raoult’s law at the liquid-gas interface and the ambient vapour concentration at infinity, i.e. $$\begin{aligned}
c_{\alpha}&=c_{\alpha}^{\text{eq}}=c_{\alpha}^{\text{pure}}\gamma_{\alpha}x_{\alpha} &\text{at the liquid-gas interface and} \label{eq:detmodel:raoults} \\
c_{\alpha}&=c_{\alpha}^\infty &\text{far away from the droplet}\,,
\label{eq:detmodel:laplgasbc}\end{aligned}$$ where $c_{\alpha}^{\text{pure}}$ is the saturation vapour concentration in case of the pure liquid $\alpha$ and $x_{\alpha}$ is the liquid mole fraction. The activity coefficients $\gamma_{\alpha}$ account for thermodynamic non-idealities and are functions of the composition, i.e. of $x_{\alpha}$. Neglecting the small contribution of the Stefan flow at temperatures sufficiently below the boiling point, the evaporation rates $j_{\alpha}$ are given by the diffusive fluxes at the liquid-gas interface, i.e. $$j_{\alpha}=-D_{\alpha}^{\text{vap}}\partial_nc_{\alpha} \,.
\label{eq:detmodel:evaprate}$$ While the dynamics in the gas phase can be considered in the diffusive and quasi-stationary limit, convection can be dominant in the liquid phase, which can be attributed to the typical diffusion coefficients, namely $D_{\alpha}^{\text{vap}}\sim \SI{e-5}{\meter^2\per\second}$ in the gas phase and $D\sim\SI{e-9}{\meter^2\per\second}$ in the liquid phase. Therefore, the liquid phase has to be described by the full convection-diffusion equation for the liquid mass fraction $y_{\alpha}$, which is expressed for the component A only due to the identity $y_{\text{A}}+y_{\text{B}}=1$, i.e. $$\rho\left(\partial_t y_{\text{A}} + \boldsymbol{u}\bcdot\bnabla y_{\text{A}} \right) = \bnabla\bcdot\left(\rho D \bnabla y_{\text{A}}\right) \,.
\label{eq:detmodel:massfracconvA}$$ The liquid density $\rho$ and the mutual diffusivity $D$ are in general functions of the composition, i.e. of $y_{\text{A}}$. The mass transfer rates $j_{\alpha}$ due to evaporation induce a change in composition at the liquid-gas interface. Using the mass transfer expression $j_{\alpha}=\rho y_{\alpha}(\boldsymbol{u}_\alpha-\boldsymbol{u}_{\text{I}})\bcdot\boldsymbol{n}$ with $\boldsymbol{u}_\alpha$ denoting the velocity of component $\alpha$ and $\boldsymbol{u}_{\text{I}}$ and $\boldsymbol{n}$ denoting the interface velocity and normal, respectively, this compositional change can be expressed by a Robin boundary condition for Eq. in the frame co-moving with the interface in normal direction, namely $$-\rho D\bnabla{}y_{\text{A}}\bcdot \boldsymbol{n}= y_{\text{B}}j_{\text{A}}- y_{\text{A}}j_{\text{B}}= (1-y_{\text{A}})j_{\text{A}}-y_{\text{A}}j_{\text{B}}\,.
\label{eq:detmodel:massfluxbc}$$ Finally, the flow in the droplet is given by the Navier-Stokes equations $$\begin{aligned}
\rho\left(\partial_t\boldsymbol{u}+\boldsymbol{u}\bcdot \bnabla\boldsymbol{u}\right)&=-\bnabla p + \bnabla\bcdot\left( \mu \left(\bnabla \boldsymbol{u} + (\bnabla \boldsymbol{u})^\text{t} \right) \right) + \rho g \boldsymbol{e}_z \label{eq:detmodel:nseq} \\
\partial_t \rho + \bnabla\bcdot(\rho\boldsymbol{u})&=0\,.
\label{eq:detmodel:contieq}\end{aligned}$$ Here, we have chosen the $z$-axis to point towards the apex of the droplet, i.e. a sessile droplet and a pendant droplet can be realized by negative and positive values for $g$, respectively. Note that the viscosity $\mu$ and the mass density $\rho$ are in general functions of the composition $y_{\text{A}}$. A dependency on the temperature is disregarded in the following due to the fact that thermal effects at lower temperatures are usually considerably inferior to the impact of solutal gradients. For the contact line dynamics, we are focussing here on a pinned contact line, i.e. evaporation in the constant radius mode (*CR-mode*, @Picknett1977 [@Stauber2014]). To resolve the incompatibility of a no-slip boundary condition at the substrate and the evaporative mass loss at the contact line, we impose a Navier-slip boundary condition with a small slip-length in the nanometre scale instead. This effectively resembles a no-slip boundary condition in the main part of the droplet-substrate interface, but still allows for a consistent mass transfer at the contact line. The free liquid-gas interface is subject to the kinematic boundary condition considering the mass transfer, i.e. $$\left(\boldsymbol{u}-\boldsymbol{u}_{\text{I}}\right)\bcdot\boldsymbol{n}=\frac{1}{\rho}(j_{\text{A}}+j_{\text{B}})
\label{eq:detmodel:kinematicbc}$$ and furthermore to the Laplace pressure in normal direction $$-p + \mu \boldsymbol{n}\bcdot\left(\bnabla \boldsymbol{u} + (\bnabla \boldsymbol{u})^\text{t} \right)\cdot \boldsymbol{n} = \sigma\kappa\,,
\label{eq:detmodel:laplpress}$$ where the traction in the gas phase has been neglected due to the viscosity ratio. Here, $\sigma$ is the local surface tension, $\kappa$ the curvature of the interface and $p$ denotes the pressure difference with respect to the ambient gas pressure. Finally, also the Marangoni shear stress in tangential direction has to be considered: $$\mu \boldsymbol{n}\bcdot\left(\bnabla \boldsymbol{u} + (\bnabla \boldsymbol{u})^\text{t} \right) \cdot \boldsymbol{t} = \bnabla_{t} \sigma \bcdot \boldsymbol{t} \,.
\label{eq:detmodel:marashear}$$ Here, $\bnabla_{t}=(\mathbf{1}-\boldsymbol{n}\boldsymbol{n})\bcdot\bnabla$ is the surface gradient operator. A sketch of the model is depicted in [figure]{} \[fig:detmodel:scheme\].
Numerical solution of the dynamical equations for an instructive example {#sec:motiv:motiv}
========================================================================
In order to solve the given set of equations numerically, we have generalised the sharp-interface arbitrary Lagrangian-Eulerian finite element method described in @Diddens2017c by considering the gravitational force, and also validated it by a more general reimplementation of the same model with the finite element package <span style="font-variant:small-caps;">oomph-lib</span>[^1] [@Heil2006], which allows for interface deformations and considers the general continuity equation . The latter method has been successfully validated against various experiments [@Li2018; @Li2019; @Gauthier2019; @YanshenLi2019].
In [figure]{} \[fig:motiv:example\], a simulation of a sessile glycerol-water droplet (initially 5 wt.% glycerol) with an initial volume of and an initial contact angle of evaporating at a constant temperature of and a relative humidity of is shown. The contact line remains pinned during drying and glycerol (liquid B) is assumed to be non-volatile due to its low volatility compared to water (liquid A), i.e. $c_{\text{B}}=c_{\text{B}}^\infty=0$ and $j_{\text{B}}=0$. For more details about these kind of simulations we refer to @Diddens2017 [@Diddens2017c], where however we did not consider of the influence of gravity.
![Simulation of a glycerol-water droplet revealing rich flow patterns during the evaporation process. The water vapour mass fraction is shown in the gas phase, whereas the glycerol mass fraction (left) and the velocity magnitude (right) are shown inside the droplet. (a) Initially, both Rayleigh and Marangoni convection support the flow from the apex to the contact line. (b) Although the contact angle is still above $\theta>\SI{90}{\degree}$, a Marangoni-induced counter-rotating vortex (black) emerges close to the interface, whereas the bulk flow is driven by natural convection (white). (c) Due to the increased evaporation rate at the contact line for $\theta<\SI{90}{\degree}$, the Marangoni-driven vortex grows in size until (d) the vortex driven by natural convection disappears. See supplementary movie 1 for the entire simulation.[]{data-label="fig:motiv:example"}](fig02){width="100.00000%"}
Initially, in [figure]{} \[fig:motiv:example\](a), one can see a single vortex in the entire droplet, directed from the apex towards the contact line. This vortex is generated for two reasons, namely Marangoni convection and natural convection (Rayleigh convection). Due to the enhanced evaporation rate of water at the apex at the high contact angle, the water content is predominantly reduced at the top of the droplet, resulting in a lower surface tension compared to the region near the contact line. This drives a Marangoni flow towards the contact line. Since glycerol is more dense than water, the glycerol-rich outer shell of the droplet also sinks down due to natural convection, which also results in a flow from the apex to the contact line due to the spherical geometry. Hence, both mechanisms support recirculating flow in the same direction.
Remarkably, in [figure]{} \[fig:motiv:example\](b), the situation changes. The contact angle is still above , i.e. still having the highest water evaporation rate at the top of the droplet. According to the afore-mentioned discussion, one would still naively expect the same kind of single vortex flow. However, the simulation clearly shows two vortices, one in the bulk driven by natural convection (white) and another one close to the interface, which is driven by Marangoni flow in the opposite direction (black). The reason why the Marangoni flow is reversed, i.e. why there is more water at the top of the droplet although the evaporation rate of water is still dominant at the apex, is the fact that there is enhanced water replenishment by diffusion at the apex, which compensates for the rather small difference in the evaporation rates at the top and near the contact line. This can be seen by the rather steep concentration gradient in normal direction at the apex compared to the region near the contact line. The reason of the steep concentration gradient in normal direction close at the apex is the upward directed convective water replenishment from the bulk, which is governed by the internal vortex driven by natural convection. This means that sufficiently strong natural convection in the bulk can reverse the Marangoni flow at the interface, although one would not anticipate this by just considering the profile of the evaporation rate at this contact angle.
Upon further evaporation, in [figure]{} \[fig:motiv:example\](c), the contact angle falls below , resulting in a higher water evaporation rate near the contact line. Hence, less water is present at the contact line as compared to the apex due two facts, namely the effect of preferential evaporation at the contact line and the lower replenishing diffusive flux of water from the bulk at the contact line. Thereby, the Marangoni flow gets enhanced compared to the situation in [figure]{} \[fig:motiv:example\](b) and the relative size of the Marangoni-induced vortex at the interface grows at the expense of the counter-rotating bulk vortex by natural convection.
Finally, in [figure]{} \[fig:motiv:example\](d), the contact angle becomes rather small so that the Marangoni flow at the interface is even stronger due to the enhanced non-uniformity of the evaporation rate. Furthermore, the influence of natural convection also diminishes rather quickly, i.e. with cubic power in terms of the length scale according to the Rayleigh number (see later for its definition), due to the reduced volume of the droplet. This results in the depicted situation, i.e. that the flow direction within the entire droplet is completely determined by the Marangoni effect.
In a nutshell, one can infer from the direct numerical simulation results in [figure]{} \[fig:motiv:example\] that there can be multiple flow scenarios during the drying of a single binary droplet, driven by an interplay of natural (i.e. Rayleigh) convection and Marangoni convection. One also clearly sees that, for a particle laden droplet, the coffee-stain effect would not occur as there is no noticeable flow towards the contact line (which for pure evaporating droplets transports the suspended particles to the rim of the pinned droplet) as compared to the strongly recirculating flow due to Marangoni flow and gravity.
Quasi-stationary approximation of the dynamical equations {#sec:simpmodel:simpmodel}
=========================================================
After discussing some possible flow scenarios by considering a representative numerical example in the previous section, we will now focus on a simplification of the full model described in section \[sec:detmodel:detmodel\]. We generalise again from the particular case of a water-glycerol droplet to the general case of a binary droplet, where both liquids and are allowed to evaporate. The goal is to find the simplest model possible that allows to predict the expected flow scenario in the droplet by a minimum number of non-dimensional quantities.
Evaporation Numbers {#sec:simpmodel:evnos}
-------------------
As shown in the example simulation, the liquid recirculates multiple times during the evaporation process due to the fast flow in the droplet. Hence, the typical liquid velocity $\boldsymbol{u}$ is much larger than the normal interface movement velocity $\boldsymbol{u}_{\text{I}}$. Moreover, this leads to a rather well-mixed droplet, i.e. with typical compositional deviations of about a few percent in terms of mass fractions. These observations allow for several simplifications of the model. First of all, the liquid composition is expanded into two terms, i.e. $y_{\text{A}}(\boldsymbol{x},t)=y_{\text{A},0}+y$, namely the spatially averaged composition $y_{\text{A},0}$, which slowly evolves over the entire drying time, and the small local composition deviations $y(\boldsymbol{x},t)$. Since the composition-dependent liquid properties are usually rather smooth functions of the composition, this separation can be transferred to a first order Taylor expansion of the liquid properties, i.e. $$\begin{aligned}
\rho&=\rho_0+y\:\partial_{y_\text{A}}\rho\,, &\sigma&=\sigma_0+y\:\partial_{y_\text{A}}\sigma\,, \nonumber \\
\mu&=\mu_0+y\:\partial_{y_\text{A}}\mu\,, &D&=D_{0}+y\:\partial_{y_\text{A}}D\,, \label{eq:simpmodel:compoexpand}\\
c_{\text{A}}^{\text{eq}}&=c_{\text{A},0}^{\text{eq}}+y\:\partial_{y_\text{A}}c_{\text{A}}^{\text{eq}} \,, &c_{\text{B}}^{\text{eq}}&=c_{\text{B},0}^{\text{eq}}+y\:\partial_{y_\text{A}}c_{\text{B}}^{\text{eq}} \,. \nonumber
\label{eq:simpmodel:compoexpandbla}\end{aligned}$$ Since the averaged composition $y_{\text{A},0}$ evolves slowly, this expansion can be done at any specific time of interest during the evaporation process. In particular, this means that the coefficients of the Taylor expansions can be treated as constants during some time close to the considered instant. This allows us to introduce the following non-dimensionalized scales $$\boldsymbol{x}=V^{1/3}\tilde{\boldsymbol{x}}\,,\qquad t=\frac{V^{2/3}}{D_{0}}\tilde{t}\,,\qquad \boldsymbol{u}=\frac{D_{0}}{V^{1/3}}\tilde{\boldsymbol{u}} \,,
\label{eq:simpmodel:scales}$$ where the spatial scale is chosen in that way, that the nondimensionalized droplet volume $\tilde{V}$ becomes unity.
In a next step, the vapour fields are decomposed in a similar manner as , namely in a normalized contribution $\tilde{c}_0$ which is one at the interface and zero at infinity and a contribution $\tilde{c}_\Delta$ accounting for the effect of local composition variations on the vapour concentration via Raoult’s law to the first order, i.e. $$c_{\alpha}=\left(c_{\alpha,0}^{\text{eq}}-c_{\alpha}^\infty\right)\tilde{c}_0 + c_{\alpha}^\infty + (\partial_{y_\text{A}}c_{\alpha}^{\text{eq}})\tilde{c}_\Delta \,.
\label{eq:simpmodel:vapsep}$$ The Laplace equation splits into two Laplace equations, i.e. $\tilde\nabla^2 \tilde{c}_0=0$ and $\tilde\nabla^2 \tilde{c}_\Delta=0$ and the boundary conditions are transformed to $$\begin{aligned}
\tilde{c}_0&=1\text{ and }\tilde{c}_\Delta=y &\text{at the liquid-gas interface and} \label{eq:simpmodel:raoults} \\
\tilde{c}_0&=\tilde{c}_\Delta=0 &\text{far away from the droplet}\,.
\label{eq:simpmodel:laplgasbc}\end{aligned}$$ Thereby, the evaporation rates separate in the same way, i.e. $$\begin{aligned}
j_{\alpha}=\frac{D_{\alpha}^{\text{vap}}}{V^{1/3}}\left[\left(c_{\alpha,0}^{\text{eq}}-c_{\alpha}^\infty\right)\tilde{j}_0 + (\partial_{y_\text{A}}c_{\alpha}^{\text{eq}})\tilde{j}_y \right] \,,
\label{eq:simpmodel:evaprate}\end{aligned}$$ where $\tilde{j}_0=-\tilde\partial_n\tilde{c}_0$ only depends on the shape of the droplet, i.e. resembles the normalized evaporation profile of a homogeneous droplet, and $\tilde{j}_y=-\tilde\partial_n\tilde{c}_\Delta$ is a linear functional of $y$, i.e. the Dirichlet-to-Neumann map, accounting for deviations in the evaporation rate due to a varying interfacial composition via the composition-dependent vapour-liquid equilibrium, i.e. Raoult’s law.
When dropping terms of quadratic order in $y$, the convection-diffusion equation within the droplet becomes $$\partial_{\tilde t} y_{\text{A},0} + \partial_{\tilde t} y + \tilde{\boldsymbol{u}}\bcdot\tilde\bnabla y = \tilde\nabla^2 y
\label{eq:simpmodel:massfracconvA}$$ and the corresponding interface boundary condition reads $$-\tilde\bnabla{}y\bcdot \boldsymbol{n}= \mbox{\it Ev}_y\tilde{j}_0 +\mbox{\it Ev}_{\text{vap}}\tilde{j}_y - \mbox{\it Ev}_{\text{tot}} y \tilde{j}_0
\label{eq:simpmodel:massfluxbc}$$ with the non-dimensional evaporation numbers $$\begin{aligned}
\mbox{\it Ev}_y=&\frac{1}{\rho_0D_{0}}\Big[(1-y_{\text{A},0})D_{\text{A}}^{\text{vap}}(c_{\text{A},0}^{\text{eq}}-c_{\text{A}}^\infty)-y_{\text{A},0}D_{\text{B}}^{\text{vap}}(c_{\text{B},0}^{\text{eq}}-c_{\text{B}}^\infty)\Big] \label{eq:simpmodel:evzero} \\
\mbox{\it Ev}_{\text{vap}}=&\frac{1}{\rho_0D_{0}}\Big[(1-y_{\text{A},0})D_{\text{A}}^{\text{vap}}\partial_{y_\text{A}}c_{\text{A}}^{\text{eq}} -y_{\text{A},0}D_{\text{B}}^{\text{vap}}\partial_{y_\text{A}}c_{\text{B}}^{\text{eq}}\Big] \label{eq:simpmodel:evy} \\
\mbox{\it Ev}_{\text{tot}}=&\frac{1}{\rho_0D_{0}}\Big[D_{\text{A}}^{\text{vap}}(c_{\text{A},0}^{\text{eq}}-c_{\text{A}}^\infty) +D_{\text{B}}^{\text{vap}}(c_{\text{B},0}^{\text{eq}}-c_{\text{B}}^\infty)\Big]\,.
\label{eq:simpmodel:evtot}\end{aligned}$$ The number $\mbox{\it Ev}_y$ quantifies the intensity of the concentration gradient induced in the liquid by preferential evaporation of one of the components, i.e. it compares the differences of the two diffusive vapour transports in the gas phase with the mutual diffusion in the binary liquid. Since the resulting composition gradient along the interface and in the bulk is the driving mechanism for Marangoni flow and natural convection, this number will become important to quantify these processes later on. Note that dependent on the volatilities of the components and their mass fractions in the liquid, $\mbox{\it Ev}_y$ may be positive or negative.
$\mbox{\it Ev}_{\text{vap}}$ is an estimate for the influence of local variations in the liquid concentration on the preferential evaporation, i.e. the linear feedback due to the quasi-stationary diffusion in the gas phase. If the composition is rather uniform in the droplet, which is usually by fast recirculating convection, the term $\mbox{\it Ev}_{\text{vap}}\tilde{j}_y$ provides only a minor contribution in , meaning that the profile of the evaporation rates is similar to the one of a pure droplet. Since $\partial_{y_\text{A}}c_{\text{A}}^{\text{eq}}>0$ and $\partial_{y_\text{A}}c_{\text{B}}^{\text{eq}}<0$, i.e. the vapour concentration of A increases and B decreases for an increasing fraction of A in the liquid, $\mbox{\it Ev}_{\text{vap}}$ is always positive. Large numbers of $\mbox{\it Ev}_{\text{vap}}$ can actually arise towards the end of the drying of a glycerol-water droplet, as discussed later on in section \[sec:validation:validation\].
Finally, $\mbox{\it Ev}_{\text{tot}}$ is a measure for the total evaporation speed, i.e. for the typical interface speed $\tilde{\boldsymbol{u}}_{\text{I}}$ and the volume evolution. Note that the total evaporation speed and volume evolution is a measure for the flow towards a pinned contact line, i.e. the flow leading to the coffee-stain effect. If none of the components condensates, i.e. both either evaporate or are non-volatile, the modulus of $\mbox{\it Ev}_y$ is smaller than $\mbox{\it Ev}_{\text{tot}}$. Nevertheless, since the deviation from the average composition is small, i.e. $y\ll 1$, and the Marangoni convection and/or natural convection are sufficiently large, the contribution of the latter to the flow can still be dominant compared to $\mbox{\it Ev}_{\text{tot}} y \tilde{j}_0$ in .
Nondimensionalized flow {#sec:simpmodel:nseq}
-----------------------
For the Navier-Stokes equations, we employ the established Boussinesq approximation, which is valid as long as $y\partial_{y_\text{A}}\rho$ is small compared to $\rho_0$ [@Gray1976]. Due to the usually small composition gradients, this assumption is valid here. Therefore, except for the body force term $\rho g \boldsymbol{e}_z$, only the zeroth order terms proportional to $\rho_0$ are kept, whereas $y\partial_{y_\text{A}}\rho$-terms are disregarded. With the same argument, terms proportional to $y\partial_{y_\text{A}}\mu$ and $y\partial_{y_\text{A}}\sigma$ can be disregarded, whenever there is a dominant term proportional to $\mu_0$ and $\sigma_0$, respectively. Following this argument, the Navier-Stokes equations can be written as $$\begin{aligned}
\mbox{\it Sc}^{-1}\left(\partial_{\tilde t}\tilde{\boldsymbol{u}}+\tilde{\boldsymbol{u}}\bcdot \tilde\bnabla\tilde{\boldsymbol{u}}\right)&=-\tilde\bnabla \tilde p + \tilde\bnabla\bcdot \left(\tilde\bnabla \tilde{\boldsymbol{u}} + (\tilde\bnabla \tilde{\boldsymbol{u}})^\text{t} \right) + \mbox{\it Ra}^* y \boldsymbol{e}_z \label{eq:simpmodel:nseq} \\
\tilde\bnabla\bcdot\tilde{\boldsymbol{u}}&=0\,.
\label{eq:simpmodel:contieq}\end{aligned}$$ Here, the shifted nondimensionalized pressure, the Schmidt number and the incomplete Rayleigh number read $$\tilde p=\frac{V^{2/3}}{D_{0}\mu_0}\left(p-\rho g z\right) \,, \qquad \mbox{\it Sc}=\frac{\mu_0}{D_{0}\rho_0}\,, \qquad \mbox{\it Ra}^*=\frac{V g\partial_{y_\text{A}}\rho}{D_{0}\mu_0}\,.
\label{eq:simpmodel:reynra}$$ The Schmidt number for liquids is usually $\mbox{\it Sc}> \num{e3}$ which suggests that the lhs of can be disregarded. However, since the chosen velocity and time scale in does not necessarily coincide with the actual present scales, this argument is only valid for small Reynolds numbers. In small droplets with rather low volatilities and regular Marangoni flow, however, this assumption is surely met, e.g. $\mbox{\it Re} < 0.05$ in the case of the simulation in [figure]{} \[fig:motiv:example\]. The incomplete Rayleigh number $\mbox{\it Ra}^*$ deviates from the conventional definition of the Rayleigh number just by the lack of an estimate for the composition difference, i.e. a term like $\Delta y_{}$. The dynamic boundary conditions at the interface, and , read in the Boussinesq approximation $$-\tilde p + \boldsymbol{n}\bcdot\left(\tilde\bnabla \tilde{\boldsymbol{u}} + (\tilde\bnabla \tilde{\boldsymbol{u}})^\text{t} \right)\cdot \boldsymbol{n} = \frac{1}{\mbox{\it Ca}^*}\left(\tilde\kappa+\mbox{\it Bo}\: \tilde z\right)
\label{eq:simpmodel:laplpress}$$ $$\boldsymbol{n}\bcdot\left(\tilde\bnabla \tilde{\boldsymbol{u}} + (\tilde\bnabla \tilde{\boldsymbol{u}})^\text{t} \right) \cdot \boldsymbol{t} = \mbox{\it Ma}^* \tilde\bnabla_{t} y \bcdot \boldsymbol{t} \,.
\label{eq:simpmodel:marashear}$$ Here, the non-dimensional number $\mbox{\it Ca}^*$, the Bond number and the incomplete Marangoni number read $$\mbox{\it Ca}^*=\frac{D_{0}\mu_0}{V^{1/3}\sigma_0}\,, \qquad \mbox{\it Bo}=\frac{\rho_0 g V^{2/3}}{\sigma_0}\,, \qquad \mbox{\it Ma}^*=\frac{V^{1/3}\partial_{y_\text{A}}\sigma}{D_{0}\mu_0}\,.
\label{eq:simpmodel:mastar}$$ Note that the definition of $\mbox{\it Ca}^*$ does not coincide with the capillary number, i.e. it does not consider the actually present typical velocity scale, i.e. the intensity of the capillary shape relaxations during evaporation cannot be inferred from $\mbox{\it Ca}^*$. However, both the real capillary number $\mbox{\it Ca}=\mu_0U/\sigma_0$ and $\mbox{\it Ca}^*$ are small in the systems considered here ($\mbox{\it Ca}<\num{1e-6}$ and $\mbox{\it Ca}^*<\num{1e-7}$ in the simulation depicted in [figure]{} \[fig:motiv:example\]). Similar to $\mbox{\it Ra}^*$, the incomplete Marangoni number $\mbox{\it Ma}^*$ lacks in an estimate for the composition difference, i.e. $\Delta y$, as compared to the conventional definition. Finally, the kinematic boundary condition becomes $$\left(\tilde{\boldsymbol{u}}-\tilde{\boldsymbol{u}_{\text{I}}}\right)\bcdot\boldsymbol{n}=\mbox{\it Ev}_{\text{tot}} \tilde{j}_0 + \mbox{\it Ev}_{\text{tot,vap}} \tilde{j}_y \,,
\label{eq:simpmodel:kinematicbc}$$ where $$\mbox{\it Ev}_{\text{tot,vap}}=\frac{1}{\rho_0D_{0}}\Big[D_{\text{A}}^{\text{vap}}\partial_{y_\text{A}}c_{\text{A}}^{\text{eq}} +D_{\text{B}}^{\text{vap}}\partial_{y_\text{A}}c_{\text{B}}^{\text{eq}}\Big]
\label{eq:simpmodel:evtoty}$$ is the analogue of $\mbox{\it Ev}_{\text{vap}}$ for the total evaporation rate, i.e. the effect of a change in the saturation pressure due to a locally deviating composition on the total evaporation rate.
Estimation of the outwards flow {#sec:simpmodel:coffstain}
-------------------------------
![Bulk flow for small capillary number $\mbox{\it Ca}$ and Bond number $\mbox{\it Bo}$ in absence of Marangoni flow and natural convection for a total evaporation number $\mbox{\it Ev}_{\text{tot}}=1$ and a contact angle of (a) $\theta=\SI{60}{\degree}$ and (b) $\theta=\SI{120}{\degree}$. The evaporation rate $\mbox{\it Ev}_{\text{tot}}\tilde{j}_0$ causes a volume loss, which prescribes the normal interface movement $\tilde{\boldsymbol{u}}_{\text{I}}\bcdot\boldsymbol{n}$ (depicted on the left sides). The bulk flow (right sides) is governed by Stokes flow with the normal boundary condition $\tilde{\boldsymbol{u}}\bcdot\boldsymbol{n}=\tilde{\boldsymbol{u}_{\text{I}}}\bcdot\boldsymbol{n}+\mbox{\it Ev}_{\text{tot}}\tilde{j}_0$. Apparently, the typical purely capillary-driven velocity is of order $\mbox{\it Ev}_{\text{tot}}$. []{data-label="fig:simpmodel:coffstain"}](fig03){width="100.00000%"}
Before focusing on natural convection and Marangoni flow in the droplet, it is beneficial to obtain an estimate for the velocity in the droplet in absence of these mechanisms, i.e. $\mbox{\it Ma}^*=\mbox{\it Ra}^*=0$. This case, exemplified e.g. by a pure isothermal droplet, combined with a pinned contact line represents the purely capillary-driven outward flow, which causes the coffee-stain effect.
For small droplets, the capillary number $\mbox{\it Ca}$ is small, so that the surface tension forces according to lead to an intense relaxing traction, whenever the droplet deviates from the equilibrium shape. Since the Bond number $\mbox{\it Bo}$ is small as well for small droplets, the hydrostatic term in can be neglected, leading to a spherical cap with a homogeneous curvature $\tilde\kappa$ as equilibrium shape. Hence, the shape evolution and thereby the interface velocity $\tilde{\boldsymbol{u}_{\text{I}}}$ is solely given by the evaporation rate and the contact line kinetics, which is assumed to be pinned here. Since the term $\mbox{\it Ev}_{\text{tot,vap}} \tilde{j}_y$ in is proportional to $y$, it can be disregarded with respect to $\mbox{\it Ev}_{\text{tot}} \tilde{j}_0$ in accordance with the Boussinesq approximation. As a consequence, one ends up with a linear Stokes flow problem, where the entire bulk velocity is given by the instantaneous shape relaxation, which is proportional to the rate of evaporation, i.e. to $\mbox{\it Ev}_{\text{tot}}$.
By integrating the evaporation rate $\mbox{\it Ev}_{\text{tot}} \tilde{j}_0$ one obtains the volume loss and thereby one can reconstruct the normal velocity of the interface $\tilde{\boldsymbol{u}}_{\text{I}}\bcdot\boldsymbol{n}$. The flow in the bulk $\tilde{\boldsymbol{u}}$ is subsequently given by solving the Stokes flow with the normal boundary condition $\tilde{\boldsymbol{u}}\bcdot\boldsymbol{n}=\tilde{\boldsymbol{u}_{\text{I}}}\bcdot\boldsymbol{n}+\mbox{\it Ev}_{\text{tot}}\tilde{j}_0$. In [figure]{} \[fig:simpmodel:coffstain\], representative solutions for the bulk flow $\tilde{\boldsymbol{u}}$ with unity evaporation number, i.e. $\mbox{\it Ev}_{\text{tot}}=1$, are depicted. It is apparent that the typical bulk velocity is of the order unity, i.e. $\|\tilde{\boldsymbol{u}}\|\sim 1$. Since the flow is proportional to $\mbox{\it Ev}_{\text{tot}}$, the typical velocity corresponding to an arbitrary evaporation number $\mbox{\it Ev}_{\text{tot}}$ is hence $\|\tilde{\boldsymbol{u}}\|\sim\mbox{\it Ev}_{\text{tot}}$. This holds also for the typical interface movement, i.e. $\|\tilde{\boldsymbol{u}}_{\text{I}}\|\sim\mbox{\it Ev}_{\text{tot}}$.
Quasi-stationary limit
----------------------
Knowing the fact that the capillary flow due to the volume loss is on the order of $\mbox{\it Ev}_{\text{tot}}$, we now focus on the contributions to the flow by Marangoni forces and natural convection. In a first step, one can consider the case where $\mbox{\it Ev}_{\text{tot}}=0$, i.e. no total mass transfer and hence a constant volume and shape of the droplet. This scenario can be realized by tuning the ambient humidities of A and B so that the evaporative mass loss of component A is balanced by the condensation of component B. In this case $\mbox{\it Ev}_{\text{tot}}=0$ and $\mbox{\it Ev}_y>0$ holds. Again, due to the small capillary number and the small Bond number, one can assume a spherical cap shape with volume $\tilde{V}=1$ and contact angle $\theta$, which are both constant now. Furthermore, there is no interface movement, $\tilde{\boldsymbol{u}}_{\text{I}}\bcdot\boldsymbol{n}=0$, and no total mass transfer, $\tilde{\boldsymbol{u}}\bcdot\boldsymbol{n}=0$.
By averaging over the droplet volume $\tilde{V}=1$, defining the integrated evaporation rate $$\tilde J=\int \tilde{j}_0 \mathrm{d}\tilde A
\label{eq:simpmodel:intevap}$$ and considering only the zeroth order term in the boundary condition in accordance with the Boussinesq approximation, one can separate the average composition $y_{\text{A},0}$ and the deviation $y$ as follows: $$\begin{aligned}
\partial_{\tilde t} y_{\text{A},0} &= -\mbox{\it Ev}_y \tilde J_0 \label{eq:simpmodel:avgevo} \\
\partial_{\tilde t} y + \tilde{\boldsymbol{u}}\bcdot\tilde\bnabla y &= \tilde\nabla^2 y +\mbox{\it Ev}_y \tilde J_0 \,.
\label{eq:simpmodel:advdiffuwithcorrection}\end{aligned}$$ Here, the term $\mbox{\it Ev}_y \tilde J_0$ assures that $y_{\text{A},0}$ is indeed the average composition and that the average of $y$ remains zero, i.e. the term compensates for the imposed composition gradient at the liquid-gas interface. As already stated in section \[sec:simpmodel:evnos\], this splitting holds only for limited time, since a variation in $y_{\text{A},0}$ leads to a change in the liquid properties which where used for the nondimensionalization. Usually, however, the coupled dynamics of flow $\tilde{\boldsymbol{u}}$ and compositional differences $y$ due to Marangoni and natural convection is considerable faster than $\mbox{\it Ev}_y \tilde J_0$, This was already apparent from the simulations depicted in [figure]{} \[fig:motiv:example\] and it will be validated later on in section \[sec:validation:validation\]. Furthermore, this observation allows to focus on stationary solutions. Finally, upon introducing $$\begin{aligned}
\xi =\frac{y}{\mbox{\it Ev}_y}\,,
\label{eq:simpmodel:tildey}\end{aligned}$$ one ends up at the following set of coupled equations: $$\begin{aligned}
\tilde{\boldsymbol{u}}\bcdot\tilde\bnabla \xi &= \tilde\nabla^2 \xi +\tilde J_0 \label{eq:simpmodel:cdessnd} \\
-\tilde\bnabla \tilde p + \tilde\bnabla\bcdot \left(\tilde\bnabla \tilde{\boldsymbol{u}} + (\tilde\bnabla \tilde{\boldsymbol{u}})^\text{t} \right) + \mbox{\it Ra}\, \xi \boldsymbol{e}_z&=0 \label{eq:simpmodel:nseqStokes} \\
\tilde\bnabla\bcdot\tilde{\boldsymbol{u}}&=0
\label{eq:simpmodel:divfree}\end{aligned}$$ subject to the following boundary conditions $$\begin{aligned}
-\tilde\bnabla{}\xi\bcdot \boldsymbol{n}&= \tilde{j}_0 \label{eq:simpmodel:simpMl}\\
\tilde{\boldsymbol{u}}\bcdot\boldsymbol{n}&=0\label{eq:simpmodel:nooutfluxss}\\
\boldsymbol{n}\bcdot\left(\tilde\bnabla \tilde{\boldsymbol{u}} + (\tilde\bnabla \tilde{\boldsymbol{u}})^\text{t} \right) \cdot \boldsymbol{t} &= \mbox{\it Ma}\: \tilde\bnabla_{t} \xi \bcdot \boldsymbol{t}
\label{eq:simpmodel:marashearss}\end{aligned}$$ at the liquid-gas interface and $$\begin{aligned}
\tilde\bnabla{}\xi\bcdot \boldsymbol{n}&= 0 \label{eq:simpmodel:nopene}\\
\boldsymbol{u}&=0
\label{eq:simpmodel:noslipss}\end{aligned}$$ at the liquid-substrate interface. Note that the simplified kinematic boundary condition is now compatible with the no-slip boundary condition at the contact line, i.e. a slip length is not required. Besides the contact angle $\theta$, only two parameters enter the system, namely the Marangoni number and the Rayleigh number, which read $$\mbox{\it Ma}=\mbox{\it Ma}^*\mbox{\it Ev}_y=\frac{V^{1/3}\partial_{y_\text{A}}\sigma}{D_{0}\mu_0}\mbox{\it Ev}_y\,\,,\qquad \mbox{\it Ra}=\mbox{\it Ra}^*\mbox{\it Ev}_y=\frac{V g\partial_{y_\text{A}}\rho}{D_{0}\mu_0}\mbox{\it Ev}_y\,.
\label{eq:simpmodel:MaNoRaNo}$$ Note that the characteristic numbers for both mechanisms are proportional to the induced composition gradient due to mass transfer, i.e. $\mbox{\it Ev}_y$. Of course, in particular the tangential gradient along the interface is also strongly dependent on the contact angle $\theta$, since this determines the profile of the evaporation rate $\tilde{j}_0$.
These equations are not only valid for the specific assumed case of $\mbox{\it Ev}_{\text{tot}}=0$, but also when the combination of Marangoni and Rayleigh flow $\tilde{\boldsymbol{u}}$ predicted by this model is considerably faster than the capillary flow, i.e. a flow situation with outwards flow leading to the coffee-stain effect. According to the estimations in section \[sec:simpmodel:coffstain\], this is the case if $\tilde{\boldsymbol{u}}\gg \mbox{\it Ev}_{\text{tot}}$.
Phase diagram for combined Marangoni and Rayleigh convection {#sec:phasediag:phasediag}
============================================================
Procedure
---------
Unfortunately, the analytical treatment of the model equations - is hampered by the geometry, which demands rather complicated toroidal coordinates, and by the very strong nonlinear coupling of $\tilde{\boldsymbol{u}}$ and $\xi$ due to the advection term. Therefore we investigate the system by numerical means. Our analysis is limited to axisymmetric solutions and we only consider the case $\mbox{\it Ev}_{\text{vap}}=0$, i.e. neglecting the feedback of the altered gas composition due to the liquid-vapour equilibrium on the local evaporation rate. Finally, we will focus on the case $\mbox{\it Ma}\geq 0$, for which evaporation leads to an overall reduction of the surface tension, as in the case of the water-glycerol depicted in [figure]{} \[fig:motiv:example\]. This results in a regular flow, i.e. no chaotic behaviour can be found, at least not for moderate flow conditions. In the case of negative Marangoni numbers, chaotic flow patterns cannot be excluded due to the Marangoni instability [@Christy2011; @Bennacer2014; @Machrafi2010]. Of course, this spatio-temporal evolving type of flow cannot be captured within the assumption of a quasi-stationary process. One can, on the other hand, test the linear stability of the quasi-stationary solutions in the case of negative Marangoni numbers to find the transition to chaotic flow, but since also the axial symmetry is usually broken in case of negative Marangoni numbers, one also would have to generalize the entire solution procedure from axisymmetric cylindrical coordinates to the full three-dimensional problem, as done by @Diddens2017.
In order to find solutions of the system, we employed a finite element method on an axisymmetric mesh with triangular elements. We used linear basis functions for $\xi$ and $\tilde p$ and quadratic basis functions for $\tilde{\boldsymbol{u}}$, i.e. typical Taylor-Hood elements. The equations have been implemented in both <span style="font-variant:small-caps;">FEniCS</span>[^2] [@LoggMardalEtAl2012a] and <span style="font-variant:small-caps;">oomph-lib</span> for mutual validation. The condition of zero velocity in normal direction, i.e. Eq. , has been implemented by Lagrange multipliers. For an enhanced stability in the Newton method during the solution process, it has been found beneficial to replace $\tilde J_0$ in by a Lagrange multiplier which ensures that the average of $\xi$ is zero. This removes the null space with respect to a constant shift in $\xi$ and a corresponding adjustment of the pressure $\tilde p$.
Due to the nonlinear advection term, it is in general possible that multiple solutions exist for a given parameter combination $(\theta,\mbox{\it Ma},\mbox{\it Ra})$. For the parameter ranges considered in the following, however, we are confident that we found the generic solutions due to the following strategy: For every considered contact angle $\theta$, we performed adiabatic scans along $\mbox{\it Ra}$ in increasing and decreasing direction for fixed $\mbox{\it Ma}$ and vice versa. During that, no hysteresis, i.e. bistable regions, have been found. Furthermore, by tracing these parameter paths with continuation, we have not detected any unstable branches. This has been furthermore validated by investigating the eigenvalues with a shift-inverted Arnoldi method[^3]. Finally, for each parameter combination, we performed temporal integrations of the unsteady model equations starting from a homogeneous state $\xi=0$. Since these runs converged to the same solutions as obtained by the steady parameter scans, we are sure that all solutions discussed in the following are indeed generic and stable. Note, however, that this is in general not true outside the considered parameter ranges.
Phase diagrams
--------------
![Qualitative flow types as function of the Marangoni number $\mbox{\it Ma}$ and the Rayleigh number $\mbox{\it Ra}$ for a small contact angle $\theta=60^\circ$. For sessile droplets ($\mbox{\it Ra}>0$) with large Marangoni numbers and small Rayleigh numbers, Marangoni flow dominates in the droplet and results in a circulating flow from the contact line along the free interface towards the apex (*Ma dominant*, black streamlines). On the contrary, if $\mbox{\it Ma}$ is small and $\mbox{\it Ra}$ is large, gravity-driven flow dominates with a flow direction from the apex along the free interface to the contact line (*Ra dominant*, white streamlines). While the other mechanism, i.e. natural convection in the regime *Ma dominant* and Marangoni flow in the regime *Ra dominant*, can still quantitatively influence the flow, only a single vortex can be found which is driven by the dominant mechanism. In between these regions, however, there is a regime where the bulk flow is driven by natural convection whereas the flow close to the interface is dominated by the Marangoni effect (*Ma vs. Ra*). Here, two counter-rotating vortices can be seen. For pendant droplets ($\mbox{\it Ra}<0$), both mechanisms driving the flow in the same direction (*Ma & Ra same dir.*). Hence, one cannot identify the main driving mechanism from the direction of the flow, so that the streamlines are coloured grey. If both mechanisms are sufficiently strong, however, the bulk flow due to natural convection can become so intense that the composition gradient along the interface changes direction and a flow reversal due to the Marangoni effect can arise in the vicinity of the interface (*Ra reverses Ma*).[]{data-label="fig:phasediag:MaRaLowCa"}](fig04){width="100.00000%"}
![Qualitative flow types as function of the Marangoni number $\mbox{\it Ma}$ and the Rayleigh number $\mbox{\it Ra}$ for a high contact angle $\theta=120^\circ$. Since the direction of the Marangoni flow is reversed in comparison to the case $\theta<90^\circ$ (cf. [figure]{} \[fig:phasediag:MaRaLowCa\]), the diagram qualitatively flips upside down. Now, for sessile droplets ($\mbox{\it Ra}>0$) both mechanisms act in the same direction (*Ma & Ra same dir.*) and for sufficiently intense driving, the natural convection in the bulk can reverse the composition gradient at the interface, leading to a Marangoni-driven reversal close to the interface (*Ra reverses Ma*). For pendant droplets ($\mbox{\it Ra}<0$), either Marangoni flow or natural convection dominates (*Ma dominant*/*Ra dominant*), or the bulk flow is driven by natural convection, whereas the interfacial flow is governed by Marangoni flow (*Ma vs. Ra*). []{data-label="fig:phasediag:MaRaHighCa"}](fig05){width="100.00000%"}
The phase diagrams for small and large contact angles, i.e. for $\theta=\SI{60}{\degree}$ and $\theta=\SI{120}{\degree}$, are depicted in figures [figure]{} \[fig:phasediag:MaRaLowCa\] and [figure]{} \[fig:phasediag:MaRaHighCa\], respectively. Here we have assumed, as in the case of the glycerol-water droplet, that the blue liquid A (e.g. water) is more volatile, less dense but associated with a higher surface tension than the red liquid B (e.g. glycerol). This means by definition that $\mbox{\it Ma}>0$ holds and that a sessile droplet is described by $\mbox{\it Ra}>0$ whereas a pendant droplet is given by $\mbox{\it Ra}<0$ .
Depending on the Marangoni number $\mbox{\it Ma}$, the Rayleigh number $\mbox{\it Ra}$ and the contact angle $\theta$, different qualitative flow scenarios can be found. For high Marangoni numbers and small Rayleigh numbers, the Marangoni flow dominates (*Ma dominant*) and vice versa (*Ra dominant*). In between, however, for sessile droplets with a contact angle below $\SI{90}{\degree}$ and for pendant droplets with a contact angle above $\SI{90}{\degree}$, there is a region where the Marangoni effect determines the flow direction at the interface, whereas the bulk flow is driven by natural convection (*Ma vs. Ra*). In the opposite cases, i.e. for pendant droplets with $\theta<\SI{90}{\degree}$ and sessile droplet with $\theta>\SI{90}{\degree}$, both mechanisms drive a flow in the same direction, so that one cannot directly distinguish between the two mechanisms driving the flow (*Ma & Ra same dir.*). In the limit of very strong driving of both mechanisms, however, natural convection can become so intense, that the surface tension gradient is reversed, leading to a Marangoni-induced reversal of the flow at the interface (*Ra reverses Ma*). This effect can be explained by the distortion of the internal composition field due to natural convection. For pendant droplets with $\theta<\SI{90}{\degree}$, the composition gradient in the bulk in normal direction is much more pronounced near the contact line as opposed to the apex. As a consequence, the diffusive replenishment of the blue liquid at the interface is enhanced near the contact line so that in fact more blue liquid, i.e. the one with higher surface tension, can be found near the contact line instead of at the apex – despite of its higher volatility and the pronounced evaporation rate at the contact line. The resulting Marangoni flow is therefore reversed as anticipated by considering the profile of the evaporation rate alone. The same explanation holds for the case $\theta>\SI{90}{\degree}$, except that one finds a more pronounced normal composition gradient near the apex as compared to the region close to the contact line, and the situation is reversed.
All transitions between the afore-mentioned regimes are continuous. The drawn phase boundaries are defined by the emergence or disappearance of a second vortex. There is no bifurcation and/or hysteresis present at the boundaries of the regimes. In supplementary movie 2, a path through the parameter space is traversed and the corresponding stationary solution is shown, which illustrates the behaviour of the flow upon crossing the phase region boundaries, i.e. how the stationary solution gradually changes between single and two-vortex solutions.
Finally, we also investigate the contact angle dependence of the phase diagrams by showing the corresponding regions for $\theta=\SI{40}{\degree},\,\SI{60}{\degree}$ and $\SI{80}{\degree}$ in [figure]{} \[fig:phasediag:CaDep\](a) and for $\theta=\SI{100}{\degree},\,\SI{120}{\degree}$ and $\SI{140}{\degree}$ in [figure]{} \[fig:phasediag:CaDep\](b). Obviously, the phase boundaries are shifted, but qualitative differences in the phase diagrams cannot be found.
![Influence of the contact angle $\theta$ on the boundaries of the phase diagram for (a) $\theta<\SI{90}{\degree}$ and (b) $\theta>\SI{90}{\degree}$.[]{data-label="fig:phasediag:CaDep"}](fig06){width="100.00000%"}
Note again that we have assumed in the phase diagrams that the more volatile liquid (blue) is less dense in the insets in the phase diagrams. In the other case, the droplets depicted in the insets are required to be mirrored vertically, as the Rayleigh number is then negative for sessile droplets and positive for pendant droplets. Furthermore, it is noteworthy that these diagrams are for pinned droplets (CR-mode) and droplets with a constant contact angle (CA-mode), as only stationary solutions are considered anyhow. As long as the dominant velocity contribution is given by recirculating flow due to Marangoni and/or natural convection, any capillary flow due to shape relaxations can be disregarded. Finally, the diagrams can also be used for condensation instead of evaporation, as long as the $\mbox{\it Ma}>0$ holds. For condensation of component A, $\mbox{\it Ev}_y<0$ holds, so that $\mbox{\it Ma}>0$ is true if component A has the lower surface tension, i.e. $\partial_{y_\text{A}}\sigma<0$. We therefore anticipate that the diagrams can predict the flow when ethanol condenses on a pure water droplet, whereas it would fail to predict the flow when water condenses on a pure glycerol droplet ($\mbox{\it Ma}<0$). In fact, the latter case has been investigated experimentally, showing indeed chaotic cellular flow structures [@Shin2016].
Validation of the quasi-stationary approximation against the full numerical simulation {#sec:validation:validation}
======================================================================================
Since there were a number of assumptions made in the simplification of the problem, it is necessary to validate the predicted flow by comparing it with results of the full numerical simulation, i.e. with the full set of equations as described in section \[sec:detmodel:detmodel\]. We focus on the representative simulation depicted in [figure]{} \[fig:motiv:example\]. At each instant in time, we have extracted the spatially averaged water mass fraction $y_{\text{A},0}$, the volume $V$ to determine the spatial scale $\sqrt[3]{V}$ and the contact angle $\theta$ from the simulation. From $y_{\text{A},0}$, we obtain $\rho_0$ and $\partial_{y_\text{A}}\rho$, $\sigma_0$ and $\partial_{y_\text{A}}\sigma$, as well as $\mu_0$, $D_{0}$ and $c_{\text{A},0}^{\text{eq}}$ from the composition-dependent properties of binary glycerol-water mixtures. This allows to calculate the normalized evaporation-induced composition gradient $\mbox{\it Ev}_y$ and the characteristic numbers $\mbox{\it Ra}$ and $\mbox{\it Ma}$. On the basis of these numbers and the contact angle, we solve the simplified quasi-stationary model and re-dimensionalise the resulting velocity and composition field as well as the evaporation rate using the scales . This procedure allows to compare the full unsteady evolution of the droplet with the corresponding predictions at each instant by the simplified quasi-stationary model.
![Comparison of the full simulation (left) from [figure]{} \[fig:motiv:example\] and the corresponding result predicted by the quasi-stationary model (right) at different times. The colour-code inside shows the glycerol concentration, whereas the streamlines indicate the velocity field. In the gas phase, the water vapour and the corresponding evaporation rate is depicted. (a) Initially, the full simulation has not yet attained the quasi-stationary limit, so that the intensity of the composition deviation is overpredicted in the quasi-stationary model. In (b-d), the quasi-stationary model predicts the result of the full simulation up to a deviation that can be barely seen by eye. See supplementary movie 3 for the comparison between full simulation and quasi-stationary model over the entire simulation time. []{data-label="fig:validation:numcompare"}](fig07){width="100.00000%"}
The results are depicted for several instants in [figure]{} \[fig:validation:numcompare\], where the full simulation is shown on the left and the corresponding prediction of the quasi-stationary model is depicted on the right. Initially, i.e. in [figure]{} \[fig:validation:numcompare\](a), the full simulation has not attained the quasi-stationary limit. Hence, the quasi-stationary model slightly overpredicts the composition variations, i.e. it shows more glycerol (red) near the interface and more water (blue) in the bulk. Therefore, the flow field also slightly differs, i.e. the transient full simulation shows a single vortex, whereas the quasi-stationary model predicts the presence of a small counter-rotating vortex near the apex. Furthermore, a very gentle deviation in the spherical cap shape due to the gravitational effect in the full simulation can be seen at the apex as well (regime *Ra reverses Ma*). At later time steps, i.e. in [figure]{} \[fig:validation:numcompare\](b-d), however, the flow and the composition predicted by the quasi-stationary model match the results of the full simulation almost perfectly, be it in terms of the composition field, the flow pattern, the shape or the evaporation rate. This result substantiates the fact that the capillary outwards flow, which has been disregarded in the quasi-stationary model, can indeed be neglected as long as there is a prominent recirculating flow due to Marangoni and/or Rayleigh convection.
![Comparison of characteristic quantities of the full simulation of [figure]{} \[fig:motiv:example\] and the corresponding results predicted by the quasi-stationary model. (a) The three input parameters for the quasi-stationary model $\mbox{\it Ra}$, $\mbox{\it Ma}$ and $\theta$ extracted from the full simulation. (b) Comparison of the rms velocity. (c) Comparison of the maximum and minimum glycerol content inside the droplet. (d) Evaporation numbers extracted from the full simulation. []{data-label="fig:validation:CompareQuants"}](fig08){width="100.00000%"}
To assess the quality of the quasi-stationary model in more detail, we have extracted some characteristic quantities of both simulations, i.e. from the full simulation of [figure]{} \[fig:motiv:example\] and the corresponding quasi-stationary limit at each instant. In [figure]{} \[fig:validation:CompareQuants\](a), the time evolution of the three key parameters, namely the Rayleigh number $\mbox{\it Ra}$, the Marangoni number $\mbox{\it Ma}$ and the contact angle $\theta$ is shown. These numbers were used as input for the quasi-stationary model. The rms of the velocity inside the droplet is depicted in [figure]{} \[fig:validation:CompareQuants\](b). Again one can see an initial disagreement due to the fact that the full simulation has not yet attained its quasi-stationary limit. After that, i.e. after about $\SI{50}{\second}$ to $\SI{100}{\second}$, the rms-velocity is well predicted until it shows again a disagreement towards the end of the drying time. The reason for the overpredicted velocity in the quasi-stationary model can be found in [figure]{} \[fig:validation:CompareQuants\](c), where the minimum and maximum glycerol concentration in the droplet according to both simulations are plotted against time. While it shows good agreement in the main part of the drying, the quasi-stationary model shows an enhanced maximum glycerol concentration towards the end of the drying, i.e. when almost only glycerol is left in the droplet. In fact, the glycerol concentration predicted by the quasi-stationary model even exceeds the physically realistic threshold of $\SI{100}{\percent}$. Obviously, this overprediction of the composition differences explains the elevated prediction of the rms velocity. The reason of the overpredicted composition difference can finally be seen in [figure]{} \[fig:validation:CompareQuants\](d), where the evaporation numbers are depicted. When the droplet almost entirely consists of glycerol, the evaporation number $\mbox{\it Ev}_{\text{vap}}$, quantifying the reduction of the water vapour pressure for vanishing water at the interface on the evaporation dynamics (cf. Eq. ), becomes very large ($\mbox{\it Ev}_{\text{vap}}\to 18$ at $t=\SI{1000}{\second}$). This effect is not considered in the quasi-stationary model, since it assumes the averaged composition $y_{\text{A},0}$ to predict the vapour-liquid equilibrium, not the local composition at the interface. Thereby, the amount of water vapour is strongly overestimated which results in a high evaporation rate and thereby in an unrealistically high composition difference. Obviously, the quasi-stationary model loses validity when $\mbox{\it Ev}_{\text{vap}}$ becomes too large, meaning that the dependence of the vapour-liquid equilibrium on the local interface composition cannot be neglected any more. For a more detailed model, this effect can easily be incorporated into the quasi-stationary model, but it would introduce a fourth parameter besides $\mbox{\it Ma}$, $\mbox{\it Ra}$ and $\theta$ into the set of equations, which is beyond the scope of this article.
Conclusion {#sec:conclusion:conclusion}
==========
During the evaporation of a binary droplet, multiple flow scenarios can be found, which is a result of an interplay of differences in the volatilities, mass densities and surface tensions of the two constituents. The difference in the volatilities induces compositional gradients in the bulk and also, due to the in general non-homogeneous evaporation rate, along the interface. Due to the composition-dependent mass density and surface tension, natural convection and Marangoni flow can set in, leading to a recirculating flow in the droplet that is usually much faster than the typical capillary outwards flow towards the contact line, which can be seen in pure droplets and leads to the coffee-stain effect in particle-laden droplets.
Based on justified assumptions, we simplified the full model equations to a quasi-stationary model that only requires three parameters, namely the contact angle, the Rayleigh and the Marangoni number. Both, the Rayleigh and Marangoni number, linearly scale with a non-dimensional evaporation number, $\mbox{\it Ev}_y$, which is a measure for the induced composition gradient by the preferential evaporation of one or the other component. By numerically solving for stationary solutions of the simplified model, we have explored the phase space in terms of these three quantities. The obtained phase diagrams allow for the prediction of the flow types in sessile and pendant binary droplets, with contact angles below and above .
We found in total five different flow patterns: If one of both mechanisms, i.e. either natural convection or Marangoni flow, gets sufficiently strong as compared to the other one, it can dominate and control the flow direction in the entire droplet. This scenario can usually be seen in the case when the corresponding number, namely the Rayleigh or Marangoni number, respectively, is much larger than the other. In these cases, a single vortex can be seen in the droplet. If both mechanisms drive the flow into a different direction and are comparably strong in terms of their non-dimensional numbers, one can find two vortices, one in the bulk driven by natural convection, and a counter-rotating vortex at the interface due to Marangoni flow. The fourth flow type is the case, when both mechanisms act in the same direction, so that one cannot distinguish the particular cause of the driving and only a single vortex is present. Remarkably, however, in particular regimes in the phase space, the Marangoni flow can be reversed due to the natural convection in the bulk, leading to the fifth solution, where again two vortices can be found. In this situation, the bulk flow driven by natural convection deforms the internal composition field so that the diffusion dynamics in the liquid are altered, which eventually reverses the composition gradient at the interface and hence the Marangoni flow.
To use the phase diagrams presented in this article, several requirements have to be fulfilled: First of all, the influence of thermal effects must be negligible compared to the solutal ones. The two liquids must be miscible and the droplet must not be too large, so that the capillary number and the Bond number are small in order to guarantee a spherical cap shape during the evaporation. Furthermore, the Reynolds number must be small and the spatial variations in the composition must be small enough in order to allow for a first order Taylor expansion of the composition-dependent liquid properties according to . Also the requirements for the Bousinessq approximation must hold. The Marangoni number, as defined in , must be positive, i.e. evaporation leads to an overall decrease of the surface tension, so that Marangoni-unstable chaotic flow can be excluded and the recirculating flow must be sufficiently faster than the movement of the interface. Finally, the influence of a change in the local composition on the vapour-liquid equilibrium may not be too strong, as it has been discussed on the basis on the evaporation number $\mbox{\it Ev}_{\text{vap}}$ describing the feedback of local composition changes on the evaporation rate in section \[sec:validation:validation\].
If all these requirements are fulfilled and the composition-dependence of the required physical properties are known, the phase diagrams of this article allow for a prediction of the qualitative flow pattern in an evaporating binary droplet, probably with exception of a short initial transient phase.
The method described in this article can be directly transferred to thermally driven Marangoni flow and natural convection in a pure droplet. Instead of a convection-diffusion equation for one component, one would have to consider the convection-diffusion equation for the temperature field. The boundary conditions will be different, e.g. a Dirichlet boundary condition of constant temperature at a highly conducting substrate and non-dimensional evaporative cooling instead of the number $\mbox{\it Ev}_y$, but the methodological principle can remain the same. Also a generalization to negative Marangoni numbers could be interesting, but it would require the consideration of the problem in three dimensions. This would allow to predict axial symmetry breaking and also bifurcations into chaotic Marangoni flow regimes by performing a linear stability analysis of the quasi-stationary solutions.
This work is part of an Industrial Partnership Programme (IPP) of the Netherlands Organization for Scientific Research (NWO). This research programme is co-financed by Canon Production Printing Holding B.V., University of Twente and Eindhoven University of Technology. DL gratefully acknowledges support by his ERC-Advanced Grant DDD (project number 740479).
Declaration of Interests {#declaration-of-interests .unnumbered}
========================
The authors report no conflict of interest.
Appendix A. Comparison with experiments and relation to the Grashof number {#sec:appendix:appendix .unnumbered}
==========================================================================
Since even the detailed full model is subject to some assumptions, e.g. the diffusion-limited vapour transport and the disregard of thermal effects, we also performed experiments on various sessile and pendant binary droplets with different volumes and contact angles. The details of the experimental setup are described by @Li2019. Here, we are more interested in a qualitative agreement, i.e. whether the flow direction is dominated by natural convection or not.
Simultaneously, we address the Grashof number (also known as Archimedes number) in the following. This number, defined as $$\mbox{\it Gr}=gh^3\rho_0(\rho_\text{A,pure}-\rho_\text{B,pure})/\mu^2\,,
\label{eq:appendix:grass}$$ with $h$ being the height of the droplet and $\mu$ the averaged viscosity of both liquids, was used in our previous publication [@Li2019] as an indicator whether the flow in the droplet is dominated by natural convection ($\mbox{\it Gr}\gg 1$) or not ($\mbox{\it Gr}\ll 1$). Compared to the non-dimensional numbers presented in this manuscript, i.e. the evaporation number $\mbox{\it Ev}_y$ and the Rayleigh number $\mbox{\it Ra}$, this number is independent of the current droplet composition. Instead, it just takes the pure densities of both fluids and the averaged density and viscosity into account. This means that the Grashof number $\mbox{\it Gr}$ is easily accessible, whereas the non-dimensional numbers used throughout this manuscript require the knowledge of the instantaneous average composition and the full composition-dependence of all properties, which is not always possible in an experimental setup.
Therefore, we are interested to substantiate the argumentation by @Li2019 that the much simpler Grashof number $\mbox{\it Gr}$ can be used as indicator whether to expect natural convection ($\mbox{\it Gr}\gg 1$) or not ($\mbox{\it Gr}\ll 1$). To investigate the validity, we replace the Rayleigh number by the Grashof number in the following. If one assumes that $\partial_{y_\text{A}}\rho$ is independent of the composition, i.e. a linear dependence of the mass density on the mass fractions, one can obtain the Grashof number via the relation $$\mbox{\it Gr}=\frac{3}{\pi}\, \frac{1-\cos \theta}{2+\cos \theta}\,\frac{\mbox{\it Ra}}{\mbox{\it Ev}_y\mbox{\it Sc}}\,,
\label{eq:appendix:grashof}$$ where the factor depending on the contact angle $\theta$ is a consequence of the different characteristic length scales, i.e. $\sqrt[3]{V}$ for $\mbox{\it Ra}$ and $h$ for $\mbox{\it Gr}$. While the Schmidt number $\mbox{\it Sc}=\mu/(\rho D)$ for liquids is typically $\mbox{\it Sc}\sim\mathcal{O}(10^4-10^5)$, for moderately volatile liquids like e.g. water at typical ambient conditions, $\mbox{\it Ev}_y\sim\mathcal{O}(10^{-1}-10^0)$ holds. In order to obtain diagrams independent of these quantities, we set the factor $\mbox{\it Ev}_y\mbox{\it Sc}=1000$ in for the determination of the boundaries in the phase diagrams.
![Same as [figure]{} \[fig:phasediag:MaRaLowCa\] and [figure]{} \[fig:phasediag:MaRaHighCa\], but expressed in terms of $\mbox{\it Gr}$ instead of $\mbox{\it Ra}$. Obviously, the onset of gravity-driven flow, even in presence of rather strong Marangoni driving happens close to $\mbox{\it Gr}=1$ (indicated by the grey line). Furthermore, experimental data of @Li2019 is also indicated.[]{data-label="fig:appendix:ArMont"}](fig09){width="100.00000%"}
The phase diagrams rescaled to the Grashof number via this way are depicted in [figure]{} \[fig:appendix:ArMont\]. One can infer from these diagrams that even in competition with a strong Marangoni effect, the onset of gravity-driven bulk flow happens approximately at a Grashof number of $\mbox{\it Gr}\sim\mathcal{O}(1)$ for a contact angle of $\theta=\SI{70}{\degree}$ in (a) and $\theta=\SI{100}{\degree}$ in (b). Furthermore, the experimental results of @Li2019 are indicated as dots. The 1[,]{}2-propanediol-water droplets with a contact angle of $\theta=\SI{70}{\degree}$ discussed in @Li2019 clearly show the effect of natural convection for an apex height of $h=\SI{800}{\micro\meter}$, whereas is was not visible for $h=\SI{410}{\micro\meter}$. This clearly coincides with the prediction of the phase diagram in [figure]{} \[fig:appendix:ArMont\](a). The experiments on glycerol-water droplets with $\theta=\SI{100}{\degree}$, as discussed in the supplementary information of @Li2019, reveal an absence of observable natural convection for $h=\SI{154}{\micro\meter}$, whereas the presence of natural convection was found at heights $h\geq\SI{320}{\micro\meter}$, with increasing velocity for elevated heights. Also this can be inferred from the $\mbox{\it Ma}$-$\mbox{\it Gr}$-diagram depicted in [figure]{} \[fig:appendix:ArMont\](b). Thus we conclude that the Grashof number $\mbox{\it Gr}$ is indeed an indicator for the presence or absence of decisive natural convection in a binary droplet. The $\mbox{\it Ma}$-$\mbox{\it Ra}$-diagrams presented in this manuscript, however, provide a much more detailed prediction of the possible flow scenarios.
[^1]: [oomph-lib.maths.man.ac.uk](oomph-lib.maths.man.ac.uk)
[^2]: <https://fenicsproject.org/>
[^3]: using *Spectra* <https://spectralib.org>
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'The *canonical degree* of a curve $C$ on a surface $X$ is $K_X\cdot C$. Our main result, Theorem \[thm:MAIN\], is that on a surface of general type there are only finitely many curves with negative self–intersection and sufficiently large canonical degree. Our proof strongly relies on results by Miyaoka. We extend our result both to surfaces not of general type and to non–negative curves, and give applications, e.g. to finiteness of negative curves on a general blow–up of $\mathbb P^ 2$ at $n\geq 10$ general points (a result related to *Nagata’s Conjecture*). We finally discuss a conjecture by Vojta concerning the asymptotic behaviour of the ratio between the canonical degree and the geometric genus of a curve varying on a surface. The results in this paper go in the direction of understanding the *bounded negativity* problem.'
address:
- 'Università degli Studi di Roma Tor Vergata, Via della Ricerca Scientifica 1, 00133, Rome, Italy.'
- 'Laboratoire de Mathématiques et Applications de lÕuniversité de Poitiers, Téléport 2 – BP 30179, Boulevard Marie et Pierre Curie, 86962 Futuroscope Chasseneuil Cedex, France.'
author:
- 'Ciro Ciliberto, Xavier Roulleau'
title: On finiteness of curves with high canonical degree on a surface
---
Introduction.
=============
Let $C$ be a projective curve on a smooth projective complex surface $X$. By *curve* we mean an irreducible, reduced 1–dimensional scheme. We denote by $g=g(C)$ its geometric genus and by $p=p_{a}(C)$ its arithmetic genus, i.e. $C^{2}+K\cdot C=2p_{a}-2$, where $K=K_{X}$ is a canonical divisor of $X$. We set $\delta=\delta(C)=p-g$. We call a curve *negative* if $C^{2}<0$. The *canonical degree* of $C$ is $k_C=K\cdot C$, often simply denoted by $k$. If $g(C)\neq 1$, we set $$\beta_C=\frac{k_C}{g(C)-1},$$ often simply denoted by $\beta$. For a surface $X$ we set $$a_X=3c_{2}(X)-K_{X}^{2},$$ often simply denoted by $a$. If the Kodaira dimension $\kappa=\kappa(X)$ is non–negative, one has $a\geq 0$.
The main result of this paper concerns negative curves with high canonical degree:
\[thm:MAIN\] (A) Let $C$ be a negative curve not isomorphic to $\mathbb{P}^{1}$ on a surface $X$ with $\kappa\geq0$. If $a=0$, there are no rational curves on $X$, i.e. either $a\not = 0$ or $g>0$. Moreover $$\label{eq:1}
k_C\leq3(g-1)+\frac{3}{4}a+\frac{1}{4}\sqrt{9a^{2}+24a(g-1)}.$$ Furthermore, if $g>1$ then $$\label{eq:3}
\beta_C \leq3+\frac{3}{4}a+\frac{1}{4}\sqrt{9a^{2}+24a}\leq4+\frac{3}{2}a.$$ If, in addition, $\beta=3+\epsilon>3$, then $$\label{eq:4}
g\leq1+\frac{3a(\epsilon+1)}{2\epsilon^{2}}.$$ (B) Suppose $\kappa(X)=2$. Then for any $\epsilon>0$ there are at most finitely many negative curves $C$ on $X$ such that $k_C\geq(3+\epsilon)(g-1)>0$.
From (B) it follows that:
Let $X$ be a surface of general type. There is a function $B(\epsilon), $ defined for $\epsilon \in ]0,\infty[$ such that for all negative curves $C$ we have $$k_C\leq(3+\epsilon)(g-1)+B(\epsilon) \,\,\, \text{and}\,\,\, -C^{2}\leq(1+\epsilon)(g-1)+B(\epsilon).$$
On some smooth quaternionic Shimura surfaces $X$ there are infinitely many totally geodesic curves (see [@Chinburg]). Such a curve $C$ is also called a *Shimura curve* and satisfies $k_C=4(g-1)$. We thus obtain the following corollary, which was one of the main result of [@BR]:
On a Shimura surface, there exist finitely many (may be none) negative Shimura curves.
The proof of Theorem \[thm:MAIN\], contained in §\[sec:proof\], strongly relies on a result by Miyaoka’s (see [@Miyaoka Cor 1.4] stated as Theorem \[thm:(Miyaoka)\] below). In particular, the inequality is very similar to [@Miyaoka formula (3)], which has a slightly lower growth in $g$, but applies only to minimal surfaces.
In §\[sec:nagata\] we make an extension of Theorem \[thm:MAIN\] which works also in the case $\kappa=-\infty$, and we prove a finiteness result for negative curves on a general blow–up of $\mathbb P^ 2$ at $n\geq 10$ general points. This is a bounded negativity result which is reminiscent of the famous *Nagata’s Conjecture*, predicting that there is no negative curve on such a surface except for $(-1)$–rational curves.
In §\[sec:pos\], using again Miyaoka’s result, we prove a boundedness theorem for non–negative curves of high canonical degree. In §\[sec:voita\] we discuss a conjecture by Vojta concerning the asymptotic behaviour of $k_C/(g-1)$ when $C$ varies among all curves on a surface. We introduce an invariant related to Vojta’s conjecture and we prove a bound for it.
The results in this paper go in the direction of understanding *bounded negativity* (see [@BR]). The *Bounded Negativity Conjecture* (BNC) predicts that on a surface of general type over $\mathbb{C}$ the self–intersection of negative curves is bounded below. Nagata’s conjecture, which we mentioned above, is also a sort of bounded negativity assertion. As a general reference on both bounded negativity and Nagata’s conjecture, see [@Harbourne]. Also Vojta’s conjecture is related to bounded negativity, as we discuss in §\[sec:voita\]. As a consequence of Theorem \[thm:MAIN\], we have the following information on negative curves:
Suppose BNC fails, so that there exists a sequence $(C_{n})_{n \in \mathbb N}$ of negative curves of genus $g_{n}$ with $\lim g_{n}=\infty$ and $\lim C_{n}^{2}=-\infty$. Then $$\limsup_n\frac{K\cdot C_{n}}{g_{n}-1}\leq3.$$
In conclusion, the authors would like to thank B. Harbourne and J. Roé for useful exchanges of ideas about the application to Nagata’s Conjecture in §\[sec:nagata\].
The proof of the main theorem {#sec:proof}
=============================
Our proof relies on the following result by Miyaoka’s (see [@Miyaoka Cor 1.4]):
\[thm:(Miyaoka)\] Let $C$ be curve on a surface $X$ with $\kappa\geq 0$. Then for all $\alpha\in[0,1]$, we have: $${\alpha^{2}}(C^{2}+3k_C-6g+6)-4\alpha(k_C-3g+3)+2a\geq0.\label{eq:Miyaoka 1}$$ Suppose $C$ is not isomorphic to $\mathbb{P}^{1}$ and $k_C>3(g-1)$. Then $$2(k_C-3g+3)^{2}-a(C^{2}+3k_C-6g+6)\leq0.\label{eq:Miyaoka 1bis}$$ Suppose in addition $K^{2}>0$. Then $$(\frac{c_{2}}{K^{2}}-1)k_C^{2}+(4(g-1)+a)k_C-2(g-1)(3(g-1)+a)\geq(\frac{c_{2}}{K^{2}}-\frac{1}{3})[k_C^{2}-C^{2}K^{2}]\geq0.\label{eq:Miyaoka 2}$$
Inequality is [@Miyaoka Thm 1.3, (i)] and is [@Miyaoka Thm 1.3, (ii)]. As for this is [@Miyaoka Thm 1.3, (iii)], which is stated there under the assumption that $X$ is minimal of general type and $C\not\simeq\mathbb{P}^{1}$. However Miyaoka’s argument works more generally under the weaker assumption $K^{2}>0$.
We are now ready for the:
Let us prove (A). Let $C$ be a negative curve on $X$ not isomorphic to $\mathbb{P}^{1}$. Then $-aC^{2}\ge 0$, with equality if and only if $a=0$. If $k_C\leq3(g-1)$, there is nothing to prove. Let us suppose $k=k_C>3(g-1)$ and set $\mathfrak g=g-1$. By , one has $$\label{eq:p}
P(k):=2(k-3\mathfrak g)^{2}-a(3k-6\mathfrak g ) \leq 0,$$ with strict inequality if $a>0$.
If $a=0$ and $g=0$, the polynomial $P$ is positive, thus this cannot occur. In the remaining cases, $k_C$ is less than or equal to the largest root of $P$, whence we get .
Suppose $g>1$ and let $\epsilon>0$ be such that $k_C=(3+\epsilon)(g-1)$. By we obtain $$\label{eq:big}
\epsilon(g-1)\leq\frac{3}{4}a+\frac{1}{4}\sqrt{9a^{2}+24a(g-1)},$$ hence $$\epsilon\leq\frac{3a}{4(g-1)}+\frac{1}{4(g-1)}\sqrt{9a^{2}+24a(g-1)}\leq\frac{3}{4}a+\frac{1}{4}\sqrt{9a^{2}+24a}\leq\frac{3}{2}a+1.$$ which yields . On the other hand, reads $$4\epsilon(g-1)-3a\leq\sqrt{9a^{2}+24a(g-1)},$$ and by squaring one gets , finishing the proof of (A).
Next we prove (B). Let $\beta_0>3$. By and , negative curves with $\beta>\beta_0$ have bounded genus $g$, therefore by also $k_C$ is bounded, hence the arithmentic genus $p$ is bounded.
Suppose $K$ is big. By [@Lazarsfeld Cor. 2.2.7] there exist $m\in\mathbb{N}^{*}$, an ample divisor $A$ and an effective divisor $Z$ such that $$mK\equiv A+Z.$$ Since $Z$ is effective, the set of integers $Z\cdot C$, when $C$ varies among negative curves, is bounded from below, therefore the degree $A\cdot C=(mK-Z) \cdot C$ of these curves with respect to the ample divisor $A$ is bounded. Hence, by results of Chow–Grothendieck [@Grothendieck], [@Mumford Lecture 15], one has only finitely many components of the Hilbert scheme containing points corresponding to such curves. Since they are negative, these components contain only one curve, proving the assertion.
Surfaces not of general type {#sec:nagata}
============================
We want to deduce from Theorem \[thm:MAIN\] a result valid for any smooth surface.
\[prop:nongen\]Let $Y$ be any smooth projective surface. Let $\eta\in {{\rm Pic}}(Y)$ be such that $\vert K_{Y}+\eta\vert$ is big and $\vert 2\eta \vert$ contains a base point free pencil. Let $\beta_{0}>3$. Then there are finitely many negative curves $D$ on $Y$ such that $$\label{eq:nongen}
k_D \geq\beta_{0}(g-1)+\frac{\beta_{0}-2}{2}D\cdot \eta.$$
Under the hypotheses there is a smooth curve $B\equiv 2\eta$ intersecting all negative curves of $X$ only at smooth points with intersection multiplicity 1. Let us make a double cover $f:X\to Y$ branched along $B$. Then for all negative curve $D$ of $X$, $C=f^ *(D)$ is irreducible, negative and $g(C)=2g(D)-1+\eta\cdot D$ by Hurwitz formula. Since $f_*(K_X)=K_Y\oplus (K_{Y}+\eta)$, then $\kappa(X)=2$ and we finish by applying (B) of Theorem \[thm:MAIN\] to $X$ and to $C=f^{*}(D)$.
As an application, we take $Y_n$ to be the plane blow-up at $n$ general points. Then ${\rm Pic}(Y_n)\cong \mathbb Z^ {n+1}$ generated by the classes of the pull–back $L$ of a line and of minus the exceptional divisors $E_1,\ldots, E_{n}$ over the blown–up points. We write $D=(d,m_{1},\dots,m_{n})$ to denote the class of a curve with components $d,m_{1},\dots,m_{n}$ with respect to this basis. We may use exponential notation to denote repeated $m_i$’s. Thus $-K=(3, 1^ {n})$.
\[prop:nagata\] Fix $\beta_0>3$. There are finitely many irreducible curves of class $D=(d,m_{1},\dots,m_{n})$ on $Y_n$ such that $$\label{eq:nagata}
\frac {D^ 2}d \leq \frac {2-\beta_0} {\beta_0} \big (1+ \frac M d\big )\,\,\, \text {where} \,\,\,\,\, M=\sum_{i=1}^ n m_i.$$
We apply Theorem \[prop:nongen\], by taking $\eta=4L$. Indeed $-K+\eta=(1,-1^ {n})$ is big.
Recall that $$\epsilon_n=\inf \{\frac d M, \,\,\, \text {for all effective} \,\,\, D=(d,m_{1},\dots,m_{n}),\,\,\, \text {such that}\,\,\, M >0 \}$$ is the *Seshadri constant* of $Y_n$. Nagata’s Conjecture (see [@Nagata]) is equivalent to say that $\epsilon _n=1/\sqrt n$ if $n\ge 10$ (see [@HarbourneRoe]).
Proposition \[prop:nagata\] can be seen as weak form of Nagata’s Conjecture. Indeeed, let us look at the *homogenous case* $D=(d,m^ n)$ with $n\geq 10$. Nagata’s Conjecture predicts that, if the $n$ blown–up points are in very general position, there is no irreducible such curve with $D^ 2<0$ (see [@CM; @Nagata]), i.e., with $d<\sqrt n m$. The conclusion of Proposition \[prop:nagata\] is not absence of curves, but finiteness of their set, under a stronger assumption than Nagata’s. Let us look at the difference between the two assumptions. In the $(m,d)$–plane applies to pairs $(m,d)$ in the first quarter below the hyperbola with equation $$\label{eq:hyper}
\beta_0d^ 2+d(\beta_0-2)-n\beta_0m^ 2+(\beta_0-2)nm= 0$$ drawn in black in Figure \[fig:1\]. One of its asymptotes (drawn in blue) is parallel to the *Nagata line* $d=\sqrt n m$ (drawn in green).
[Picture.png]{}
Since for all effective divisor $D=(d,m_1,\ldots,m_n)$ one has $d/M\ge \epsilon_n$, one has approximations $d/M\ge e_n$ to Nagata’s conjecture for any lower approximation $e_n$ of $\epsilon_n$. The best known in general is the one in [@HarbourneRoe] $$\label{eq:hr}
\epsilon _n \geq e_n=\sqrt {\frac 1n \big (1 - \frac 1 {f(n)}\big )}$$ where $f(n)$ is, for most $n$, an explicitly given quadratic function of $n$ (see ). For $n=10$ in the homogeneous case the best result is $e_{10}=228/721$ (see [@Petra]).
The hyperbola meets the line $d=e_n m$, therefore Proposition \[prop:nagata\] gives some information in an unlimited region where the above approximations to Nagata do not work.
Proposition \[prop:nagata\] implies that there are finitely many irreducible curves of class $D=(d,m_{1},\dots,m_{n})$ on $Y_n$ such that $$\label{eq:sesh}
\frac {D^ 2}d \leq \frac {2-\beta_0} {\beta_0} \big (1+ \frac 1 \epsilon_n \big),$$ where $\epsilon_n$ can be replaced by $e_n$ in Harbourne–Roé’s approximation . This result is not surprising. Indeed, J. Roé pointed out to us an easy argument which shows that there is no irreducible curve of class $D=(d,m_{1},\dots,m_{n})$ on $Y_n$ such that $$\frac {D^ 2} d< - \frac 1 {n\epsilon_n}$$ which is better than , and the difference $$\frac {\beta_0-2} {\beta_0} \big (1+ \frac 1 \epsilon_n \big)- \frac 1 {n\epsilon_n}$$ tends to $\frac {\beta_0-2} {\beta_0}\sim \frac 1 3$ for $n\to \infty$.
A boundedness result for non–negative curves {#sec:pos}
============================================
With the usual notation, for a curve $C$ on the surface $X$ with $C^ 2\neq 0$, we set $x_C:=\frac{\delta_C}{C^{2}}$, with the usual convention that the index $C$ can be dropped if there is no ambiguity.
\[thm:MAIN2\] Consider real numbers $x_{0}>\frac{1}{2}$ and $\beta_{0}>3$. Let $C$ be a curve on $X$, with $\kappa(X)\geq 0$, satisfying the following conditions:\
(1) $C^{2}>0$, $k_C=\beta(g-1)$ with $\beta>\beta_{0}$ and $g>1$;\
(2) $x_C>x_{0}$.\
Then $$g\leq a\frac{(\beta-2)}{(\beta-3)^{2}}\frac{3x_{0}-1}{2x_{0}-1}+1\label{eq:positif 1},$$ $$k_C\leq a\frac{(\beta-2)}{\beta(\beta-3)^{2}}\frac{3x_{0}-1}{2x_{0}-1},\label{eq:positif 2}$$ $$k_C\leq 2(g-1) +a\frac {(\beta-2)^ 2}{(\beta-3)^ 2}\cdot \frac{3x_{0}-1}{2x_{0}-1}.\label{eq:positif 3}$$ If $\kappa(X)=2$, then the Hilbert scheme of curves on $X$ satisfying (1) and (2) has finitely many irreducible components.
One has $$k_C-2(g-1)=2\delta-C^ 2=(\beta-2)(g-1)>0.
$$ Hence by , we have $$P(\alpha):=\alpha^{2}(3\delta-C^{2})+\alpha\frac{2(\beta-3)}{\beta-2}(C^{2}-2\delta)+a\geq0\,\label{eq:Positive}$$ for $\alpha \in [0,1]$. Since the coefficient of the leading term of $P$ is positive, the minimum of $P(\alpha)$ is attained for $$\alpha_{0}=\frac{(\beta-3)(2\delta-C^{2})}{(\beta-2)(3\delta-C^{2})}.$$ Since $\beta>3$, we have $\alpha_{0}\in ]0,1[$, and, by we have $$P(\alpha_0)=-\frac{(\beta-3)^{2}(2\delta-C^{2})^{2}}{(\beta-2)^{2}(3\delta-C^{2})}+a\geq0$$ Thus $$\frac{a}{\mu}\geq\frac{(2\delta-C^{2})^{2}}{(3\delta-C^{2})}\,\,\, \text {where}\,\,\, \mu=\frac{(\beta-3)^{2}}{(\beta-2)^{2}},$$ hence $$\frac{a}{\mu}\cdot \frac{3\delta-C^{2}}{2\delta-C^{2}}\geq2\delta-C^{2}.$$ We have $$\frac{3\delta-C^{2}}{2\delta-C^{2}}=\frac{3x-1}{2x-1}<\frac{3x_{0}-1}{2x_{0}-1}$$ because $\frac{3x-1}{2x-1}$ is decreasing for $x>x_{0}>\frac{1}{2}$, hence $$(\beta-2)(g-1)=k_C-2(g-1)=2\delta-C^{2}\leq\frac{a}{\mu}\cdot \frac{3x_{0}-1}{2x_{0}-1},$$ which implies , and . Moreover both $g$ and $k_C$ are bounded from above and, if $\kappa(X)=2$, we conclude with the same argument at the end of the proof of Theorem \[thm:MAIN\].
Let be $\beta_{0}>3$ and let $(C_{n})_{n\in \mathbb N}$ be a sequence of curves on $X$ with $\kappa=2$ such that $k_{C_{n}}>\beta_{0}(g(C_{n})-1)$, $C_{n}^{2}>0$ and $\lim g(C_{n})=\infty$. Then $$\label{eq:limsup}
\lim_n \frac {\delta_{C_n} } {C_{n}^{2}} = \frac{1}{2},$$ moreover $\lim_n \frac {g(C_n)} {\delta_{C_{n}}}=\lim_n \frac{K\cdot C_n}{\delta_n}=0.$
Let $C$ be a curve with $k_C=(3+\epsilon)(g-1)$, $\epsilon >0$. Since $(1+\epsilon)(g-1)=2\delta-C^2$, we get $\frac{\delta}{C^2}-\frac{1}{2}=(1+\epsilon)\frac{g-1}{2C^2}\geq 0$. Therefore $\lim\inf_n \frac{\delta_{n}}{C_{n}^{2}} \geq \frac{1}{2}$. On the other hand, by Theorem \[thm:MAIN2\] we obtain $\lim\sup_n \frac{\delta_{n}}{C_{n}^{2}} \leq \frac{1}{2}$. The remaining limits are readily computed.
For Shimura curves on Shimura surfaces, we have $K\cdot C=4(g-1)$ and, if there is one, there are infinitely many of them, with the genus going to infinity.
On a conjecture by Vojta {#sec:voita}
========================
The results in §\[sec:pos\] are reminiscent of the following conjecture (see [@Autissier]), which predicts that curves of bounded geometric genus on a surface of general type form a bounded family:
\[conj:1\] Let $X$ be a smooth projective surface. There exist constants $A,\, B$ such that for any curve $C$ we have $$k_C\leq A(g-1)+B.$$
If this conjecture is satisfied for $X$ with $\kappa(X)=2$, then $X$ contains finitely many curves of genus $0$ or $1$. This is known to hold for minimal surfaces with big cotangent bundle (see [@Bogomolov; @Deschamps]).
A stronger version of Conjecture \[conj:1\] is the following conjecture by Vojta (see again [@Autissier]):
\[conj:2\] For any real number $\epsilon>0$, we can take $A=4+\epsilon$ in Conjecture \[conj:1\] (and $B=B(\epsilon)$ a function of $\epsilon$).
An even stronger, more recent version, predicts that $A=2+\epsilon$ (see [@Autissier]).
If $C$ is a smooth curve on $X$, then $k_C=2(g-1)-C^2$, therefore if BNC holds, then Vojta’s conjecture holds for smooth curves with $A=2$. This suggests a close relationship between Vojta’s conjecture and BNC.
Miyaoka proves in [@Miyaoka] that Conjecture \[conj:1\] also holds if $K^{2}>c_{2}$ and he gives explicit values for $A$ and $B$, but they are far away from the ones predicted by Conjecture \[conj:2\]. Moreover Miyaoka proves that $k_C\leq 3(g-1)$ for (smooth) compact ball quotient surfaces on which the equality is attained by an infinite number of curves, i.e., Shimura curves, if they exists.
In [@Autissier] one proves that for surfaces whose universal cover is the bi–disk, one has $$k_C\leq4(g-1).$$ This is sharp since for Shimura curves on Shimura surfaces, one has $k_C=4(g-1)$.
For $X$ a surface, we define $$\Lambda_{X}=\sup_{(C_{n})_{n\in\mathbb{N}}} \big \{ \limsup_{n}\frac{K\cdot C_{n}}{g_{n}-1}\big \}$$ where $(C_{n})_{n\in\mathbb{N}}$ varies among all sequences of curves $C_{n}$ in $X$ of genus $g_{n}=g(C_{n})>1$ with $\lim_{n}g_{n}=\infty$. Conjecture \[conj:1\] says that $\Lambda_X<\infty$.
If $X$ has trivial canonical bundle, then $\Lambda_X=0$. Apart form this case, and the aforementioned cases studied in [@Autissier; @Miyaoka], nothing is known about $\Lambda_X$. The following result gives us a piece of information:
\[thm:voita\] Let $X$ be a surface of general type. Let $L$ be a very ample divisor on $X$, and let $\gamma$ be the arithmetic genus of curves in $\vert L\vert$. Then $$\Lambda_X\geq \frac {K\cdot L}{L^ 2+\gamma-1}>0.$$
Look at the surface $X$ embedded in $\mathbb P^ r$, with $r\geq 3$, via $L$. Then take a general projection $\pi: X\to \mathbb P^ 2$. Consider a general rational curve of degree $n$ in $\mathbb P^ 2$ and let $C_n$ be its pull–back via $\pi$. Then $C_n\in \vert nL\vert$.
By Hurwitz formula, the ramification divisor $R$ of $\pi$ is such that $R\equiv K+3L$. So Hurwitz formula again, implies that the geometric genus $g_n$ of $C_n$ satisfies $$2g_n-2=nL\cdot K+(3n-2)L^ 2.$$ Therefore $$\frac {K\cdot C_n}{g_n-1}= \frac {L\cdot K}{L^ 2+\gamma-1- \frac {L^ 2}n}$$ and this proves the assertion.
Suppose that, in the setting of Theorem \[thm:voita\], one has $K=mL$, with $m>0$. Then $$\Lambda_X\geq \frac {2m}{m+3}.$$ So there are sequences $(X_n)_{n\in \mathbb N}$ of surfaces, e.g., complete intersections of increasing degree in projective space, with $m\to \infty$, and therefore $\Lambda_{X_n}\to 2$ (from below).
P. Autissier, A. Chambert-Loir, C. Gasbarri, On the canonical degrees of curves in varieties of general type, Geom. Funct. Anal. 22 (2012), no. 5, 1051–1061.
Th. Bauer, B. Harbourne, A. L. Knutsen, A. Küronya, S. Müller-Stach, X. Roulleau, T. Szemberg, Negative curves on algebraic surfaces, Duke Math. J. 162 (2013), 1877–1894.
F.A. Bogomolov, Families of curves on a surface of general type, Doklady Akademii Nauk SSSR, (5)236 (1977), 1041–1044.
T. Chinburg, M. Stover, Negative curves on surfaces and applications to hermitian locally symmetric spaces, Math–ArXiv 1105.1154.
C. Ciliberto, R. Miranda, Degenerations of planar linear systems, J. Reine Angew. Math., 501 (1998), 191–220.
M. Deschamps, Courbes de genre géométrique borné sur une surface de type général \[d’après F. A. Bogomolov\]. (French) Séminaire Bourbaki, 30e année (1977/78), Exp. No. 519, pp. 233–247, Lecture Notes in Math., 710, Springer, Berlin, 1979.
A. Grothendieck, Techniques de construction et théorèmes d’existence en géométrie algébrique IV, les schémas de Hilbert, Séminaire Bourbaki, Vol. 6, Exp. No. 221, 249–276, Soc. Math. France, Paris, 1995.
B. Harbourne, Global aspects of the geometry of surfaces. Ann. Univ. Paedagog. Crac. Stud. Math. 9 (2010), 5–41.
B. Harbourne, J. Roé, Discrete behavior of Seshadri constants on surfaces. Journal of Pure and Applied Algebra, 212 (2008), 616–627.
R. Lazarsfeld, Positivity in Algebraic geometry I. Ergebnisse der Mathematik und ihrer Grenzgebiete. 3. Folge / A Series of Modern Surveys in Mathematics, Springer 2004.
Y. Miyaoka, The orbibundle Miyaoka-Yau-Sakai inequality and an effective Bogomolov-McQuillan theorem, Publ. Res. Inst. Math. Sci. 44 (2008), no. 2, 403–417.
D. Mumford, Lectures on curves on an algebraic surface. With a section by G. M. Bergman. Annals of Mathematics Studies, No. 59 Princeton University Press, Princeton, N.J. 1966
M. Nagata, On rational surfaces, II. Mem. Coll. Sci. Univ. Kyoto, Ser. A, Math. 33 (1960), 271–293.
I. Petrakiev, Homogeneous Interpolation and Some Continued Fractions, Trans. Amer. Math. Soc. Ser. B 1 (2014), 23–44.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We outline the theory of sets with distributive operations: multishelves and multispindles, with examples provided by semi-lattices, lattices and skew lattices. For every such a structure we define multi-term distributive homology and show some of its properties. The main result is a complete formula for the homology of a finite distributive lattice. We also indicate the answer for unital spindles and conjecture the general formula for semi-lattices and some skew lattices. Then we propose a generalization of a lattice as a set with a number of idempotent operations satisfying the absorption law.'
address:
- |
Department of Mathematics, George Washington University\
Washington, DC 20052\
and Gdańsk University, Poland
- |
Department of Mathematics, Columbia University\
New York, NY 10027
author:
- 'Józef H. Przytycki'
- 'Krzysztof K. Putyra'
title: Homology of distributive lattices
---
[^1]
[^2]
Introduction {#chpt:intro}
============
Distributive structures {#chpt:monoids}
=======================
Multi-term chain complexes {#chpt:complex}
==========================
Decompositions of a chain complex {#chpt:decomp}
=================================
Computation of multi-term homology {#chpt:hom}
==================================
Odds and ends {#chpt:ends}
=============
> 0
[^1]: JHP was partially supported by the NSA-AMS 091111 grant, by the Polish Scientific Grant: Nr. N-N201387034, and by the GWU REF grant.
[^2]: KP was supported by the NSF grant DMS-1005750 in summer 2011.\
${}^1$ Paweł Waszkiewicz (1973-2011) was a faculty member of Theoretical Computer Science at Jagiellonian University in Krakow. He obtained PhD at the University of Birmingham, UK, in 2002 in the theory of domains and formal languages. Although his career has been ceased in a tragic way nine years later, he has already published 21 papers. The second author is indebted to him for being introduced to the fascinating world of categories, posets and domains.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Synthetic medical data which preserves privacy while maintaining utility can be used as an alternative to real medical data, which has privacy costs and resource constraints associated with it. At present, most models focus on generating cross-sectional health data which is not necessarily representative of real data. In reality, medical data is longitudinal in nature, with a single patient having multiple health events, non-uniformly distributed throughout their lifetime. These events are influenced by patient covariates such as comorbidities, age group, gender etc. as well as external temporal effects (e.g. flu season). While there exist seminal methods to model time series data, it becomes increasingly challenging to extend these methods to medical event time series data. Due to the complexity of the real data, in which each patient visit is an event, we transform the data by using summary statistics to characterize the events for a fixed set of time intervals, to facilitate analysis and interpretability. We then train a generative adversarial network to generate synthetic data. We demonstrate this approach by generating human sleep patterns, from a publicly available dataset. We empirically evaluate the generated data and show close univariate resemblance between synthetic and real data. However, we also demonstrate how stratification by covariates is required to gain a deeper understanding of synthetic data quality.'
author:
- |
Saloni Dash\
BITS Pilani, Goa Campus\
Goa, India\
`[email protected]` Ritik Dutta\
IIT Gandhinagar\
Gandhinagar, India\
`[email protected]` Isabelle Guyon\
UPSud/INRIA U. Paris-Saclay\
Paris-Saclay, France\
`[email protected]` Adrien Pavao\
UPSud/INRIA U. Paris-Saclay\
Paris-Saclay, France\
`[email protected]` Andrew Yale\
Rensselaer Polytechnic Institute\
Troy, New York\
`[email protected]`\
Kristin P. Bennett\
Rensselaer Polytechnic Institute\
Troy, New York\
`[email protected]`
bibliography:
- 'ref.bib'
title: Synthetic Event Time Series Health Data Generation
---
Introduction
============
Advancements in health informatics have largely been aided by the use of massive amounts of medical data. Researchers build computational models with this data, which facilitate improved healthcare provisions for the general public. Examples include drug discovery, personalized medicine and medical education. However, most of the Electronic Health Records (EHR) accumulated by health organizations are not easily accessible to the public due to privacy concerns. This stifles healthcare research. EHR datasets are rarely released, making research unreproducible. Research is biased towards the few publicly available de-identified medical datasets which address only limited aspects of health care, for example the MIMIC-III (Medical Information Mart for Intensive Care) dataset. [@johnson2016mimic]. Generating synthetic health data which maintains the utility of the data as well as preserves the privacy of the patients is a potential solution. Machine Learning models like Bayesian Networks [@avino2018generating], Generative Adversarial Networks [@medgan; @yale2019privacy] (medGAN, HealthGAN) etc. have been used successfully to generate health data. Nonetheless, the data generated by these methods is not representative of real medical records, as they contain only one record per patient. In real life, patient data consists of stream of in-patient and out-patient visits and other treatment events through time. Due to it’s temporal nature, longitudinal data is conducive for causal analysis studies and in overall offers greater utility when compared to cross-sectional data. In addition, covariates such as age, gender, and race are critical for retrospective observational studies. Therefore, building models for generating longitudinal synthetic data with such covariates could greatly facilitate healthcare research.
Our ultimate goal is to build an end-to-end system that generates longitudinal synthetic health data which captures temporal relations of patient records along with their covariates, preserves utility as well as privacy, and is computationally efficient, scalable and free of cost. This system should ideally not require human domain knowledge and should provide quantitative assessments of privacy, resemblance, and utility instead of just relying on theoretical guarantees. There exist few methods that generate time series medical data. Synthea [@synthea] uses modules informed by clinicians and real-world statistics to simulate patient records from birth to present day. It guarantees privacy as the records are not generated using real patients and claims to preserve utility by using health care practitioners and real statistics to build rules that synthesize this data. But the time series of events that it produces do not necessarily closely model actual patient trajectories especially for patients with multiple comorbidities. Methods such as the Recurrent Conditional GAN (RCGAN) [@esteban2017real] successfully synthesize time series of continuously-captured real-valued vitals (e.g. heart rate, respiratory rate, mean arterial pressure etc.) But how this model can be extended to generate events that happen at irregular intervals with covariates requires further research.
Method
======
We demonstrate our approach by generating sleep patterns using a publicly available health dataset. We use as inspiration a sleep study that examined habitual sleep times with respect to covariates including gender, age, and day of the week [@basner2007american]. The event data comes from American Time Use Survey (ATUS), a federally administered, continuous survey on time use in the United States sponsored by the Bureau of Labor Statistics and conducted by the U.S. Census Bureau [@atus] The survey measures how people divide their time among life’s activities in a nationally representative sample. There are many different types of events per person. However, we choose to restrict our data to only the sleep activities of the people for primarily two reasons. Firstly, sleeping patterns are intuitive to understand, analyse and are extremely sensitive to covariates such as age or day of the week. Secondly, both short and long habitual sleep times are closely associated with increased mortality risks [@hammond1964some], diabetes [@gottlieb2005association] and hypertension [@gottlieb2006association] which makes the sleep activity data highly suitable for causal and preemptive studies, as explained previously.
Our strategy is to transform event data consisting of one record per event, into cross-sectional data consisting of one record per subject. For each person, the day starts at 4:00 am on the same day and ends at the hour they wake up the next day. ATUS also includes covariates for each patient and event. To ensure uniformity, we limit the day to 10:00 am the next day, which gives us a 30 hour day for each person. We choose 10:00 am because 96% of the people are awake by that time. Consequently, we divide the 30 hour day into 30 events of one hour each, and compute a summary statistic of sleep in that hour, i.e. the number of minutes spent sleeping in that hour. We then treat these events as features of the dataset and append them to the covariates of age, sex, day of the week and month of the year. In this manner, we transform the data to cross-sectional data consisting of 34 (including covariates) features per patient (a matrix) while retaining it’s temporal properties.
Once the data is transformed to matrix form any method can be used to generate the data. For our preliminary work, we used HealthGAN[@yale2019privacy], a Wasserstein Generative Adversarial Network, to generate our synthetic dataset. Yale et. al proposed HealthGAN and provided an empirical assessment of different data generators and quantified utility and privacy losses. Out of the several baseline methods including Gaussian Multivariate [@duda2012pattern] and Parzen Windows [@parzen1962estimation], their proposed model, the HealthGAN, outperformed all of them in terms of privacy preservation as well as maintaining data utility. We use the HealthGAN directly to generate our synthetic dataset. HealthGAN is well suited for this task becuase it is designed to perform well on binary features. The 30 average sleep features are almost binary since they are 0 when a person awake and 60 if a person sleeps for the entire hour. We use the data transformation method detailed in the HealthGAN paper [@yale2019privacy] to process our data. Other generative methods [@xiao2017joint; @esteban2017real] are left for future work.
Experimental Results
====================
We first determine whether the generator is able to capture the average sleep patterns of the population. Figure 2 shows that the synthetic average sleep per hour closely resembles the real data in the average case. An individual is awake by 10:00 am and asleep by 12:00 am. The generator is also able to learn covariate statistics reasonably well. As shown in Figure 3, the covariate probabilities of the real and synthetic data lie closely along the diagonal. For instance, 49% of the days are week days (derived from the day of the week covariate) in the real data and 48% of the days are week days in the synthetic data. Similar probabilities are derived from the age, sex and month covariates and are compared in the probability plot in Figure 3.
\[fig:meanplot\]
\[fig:dimprob\]
In order to qualitatively assess the utility of the generated data, we examine the average sleep time by age and day of the week as analysed by Basner et. al [@basner2007american] in their sleep study. Such analysis with respect to covariates is very typical of health care studies.
[.5]{}
[.5]{}
We observe certain distinct trends when we examine the data stratified with respect to the covariates age group and day of the week. In Figure 4(a) in the real data, the average sleep time for weekends is different than week days across all age groups. On an average, teenagers and young adults sleep (age group 15-24) significantly more than another age group, especially on weekends (Sat-Sun). Adults between 35-54 years in general require less sleep than other age groups, and with older age groups, the average sleep time increases proportionally. In Figure 4(b), the synthetic data successfully captures these trends but underestimates the average sleep time for teens in general. This may be due to irregularities and high variances in the sleep patterns of young people.
To investigate this further, we plot the median, first and third quartiles of average sleep per hour of people aged between 15-24. Figure 5 shows the quantile plot of the average sleep times for the age group 15-24 years. It can be observed from the plot that there is high variance in the waking and sleeping times in the real data which is not captured by the synthetic data. Thus we can see that for health data, it is not sufficient to simply report average case behavior. It is essential to investigate behavior by covariates to make sure that patterns with respect to critical groups are well modeled.
Conclusion and Future Work
==========================
The potential of generating synthetic health data which respects privacy and maintains utility is groundbreaking. A significant amount of research has been conducted for generating cross-sectional data, however the problem of generating event based time series health data, which is illustrative of real medical data has largely been unexplored. We stress the importance and utility of generating such data with a special emphasis on the covariates; an exceedingly significant aspect of medical time series data.
We provide an innovative methodology of transforming the longitudinal data to cross-sectional data without sacrificing the temporal properties, and then using state of the art generators to synthesize the transformed data. The benefit of this approach is that it can be readily used with any synthetic data generation method and facilitates effective evaluation of the quality of synthetic data. We test our model by synthesizing sleep patterns of Americans from a publicly available dataset and empirically show close univariate and covariate resemblances between the real and synthetic data. We also evaluate the utility of the generated data by comparing sleep trends inspired from a sleep study, where we conclude that while the generator is able to capture significant general trends but faces challenges on subgroups with irregular sleep patterns. We underscore the importance of empirically evaluating the quality of synthetic medical data with respect to critical covariates. We leave application of these methods to actual EHR data with many event types and using alternative data generators for future work.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We consider a class $K$ of structures, e.g. trees with $\omega+1$ levels, metric spaces and mainly, classes of Abelian groups like the one mentioned in the title and the class of reduced separable (Abelian) $p$-groups. We say $M\in K$ is universal for $K$ if any member $N$ of $K$ of cardinality not bigger than the cardinality of $M$ can be embedded into $M$. This is a natural, often raised, problem. We try to draw consequences of cardinal arithmetic to non–existence of universal members for such natural classes.'
author:
- |
Saharon Shelah\
Institute of Mathematics\
The Hebrew University\
Jerusalem, Israel\
and Rutgers University\
Department of Mathematics\
New Brunswick, NJ USA
bibliography:
- 'lista.bib'
- 'listb.bib'
- 'listx.bib'
date: September 1995
title: 'Non-existence of Universals for Classes like reduced torsion free Abelian groups under embeddings which are not necessarily pure'
---
\#1\#2[[ \[proposition\][\#2]{}]{}]{}
[^1]
Table of Contents {#table-of-contents .unnumbered}
=================
§0 - INTRODUCTION
§1 - THEIR PROTOTYPE IS ${\frak K}^{tr}_{\langle\lambda_n:n< \omega
\rangle}$, NOT ${\frak K}^{tr}_\lambda$!
§2 - ON STRUCTURES LIKE $(\prod\limits_n
\lambda_n,E_m)_{m<\omega}$, $\eta E_m\nu =:\eta(m) = \nu(m)$
§3 - REDUCED TORSION FREE GROUPS; NON-EXISTENCE OF UNIVERSALS
§4 - BELOW THE CONTINUUM THERE MAY BE UNIVERSAL STRUCTURES
§5 - BACK TO ${\frak K}^{rs(p)}$, REAL NON-EXISTENCE RESULTS
§6 - IMPLICATIONS BETWEEN THE EXISTENCE OF UNIVERSALS
§7 - NON-EXISTENCE OF UNIVERSALS FOR TREES WITH SMALL DENSITY
§8 - UNIVERSALS IN SINGULAR CARDINALS
§9 - METRIC SPACES AND IMPLICATIONS
§10 - ON MODULES
§11 - OPEN PROBLEMS
REFERENCES
Introduction
============
[**Context.**]{} In this paper, model theoretic notions (like superstable, elementary classes) appear in the introduction but not in the paper itself (so the reader does not need to know them). Only naive set theory and basic facts on Abelian groups (all in [@Fu]) are necessary for understanding the paper. The basic definitions are reviewed at the end of the introduction. On the history of the problem of the existence of universal members, see Kojman, Shelah [@KjSh:409]; for more direct predecessors see Kojman, Shelah [@KjSh:447], [@KjSh:455] and [@Sh:456], but we do not rely on them. For other advances see [@Sh:457], [@Sh:500] and Džamonja, Shelah [@DjSh:614]. Lately [@Sh:622] continue this paper.
A class ${\frak K}$ is a class of structures with an embeddability notion. If not said otherwise, an embedding, is a one to one function preserving atomic relations and their negations. If ${\frak K}$ is a class and $\lambda$ is a cardinal, then ${\frak K}_\lambda$ stands for the collection of all members of ${\frak K}$ of cardinality $\lambda$.\
We similarly define ${\frak K}_{\leq\lambda}$.
A member $M$ of ${\frak K}_\lambda$ is universal, if every $N\in {\frak
K}_{\le \lambda}$, embeds into $M$. An example is $M=:\bigoplus\limits_\lambda
{\Bbb Q}$, which is universal in ${\frak K}_\lambda$ if ${\frak K}$ is the class of all torsion-free Abelian groups, under usual embeddings.
We give some motivation to the present paper by a short review of the above references. The general thesis in these papers, as well as the present one is:
\[0.1\] General Abelian groups and trees with $\omega+1$ levels behave in universality theorems like stable non-superstable theories.
The simplest example of such a class is the class ${\frak K}^{tr} =:$ trees $T$ with $(\omega+1)$-levels, i.e. $T\subseteq {}^{\omega\ge}\alpha$ for some $\alpha$, with the relations $\eta E^0_n\nu =: \eta\restriction n=\nu
\restriction n\ \&\ \lg(\eta)\geq n$. For ${\frak K}^{tr}$ we know that $\mu^+<\lambda={{\rm cf}}(\lambda)
<\mu^{\aleph_0}$ implies there is no universal for ${\frak K}^{tr}_\lambda$ (by [@KjSh:447]). Classes as ${\frak K}^{rtf}$ (defined in the title), or ${\frak K}^{rs(p)}$ (reduced separable Abelian $p$-groups) are similar (though they are not elementary classes) when we consider pure embeddings (by [@KjSh:455]). But it is not less natural to consider usual embeddings (remembering they, the (Abelian) groups under consideration, are reduced). The problem is that the invariant has been defined using divisibility, and so under non-pure embedding those seemed to be erased.
Then in [@Sh:456] the non-existence of universals is proved restricting ourselves to $\lambda>2^{\aleph_0}$ and $(< \lambda)$-stable groups (see there). These restrictions hurt the generality of the theorem; because of the first requirement we lose some cardinals. The second requirement changes the class to one which is not established among Abelian group theorists (though to me it looks natural).
Our aim was to eliminate those requirements, or show that they are necessary. Note that the present paper is mainly concerned essentially with results in ZFC, but they have roots in “difficulties" in extending independence results thus providing a case for the
\[0.2\] Even if you do not like independence results you better look at them, as you will not even consider your desirable ZFC results when they are camouflaged by the litany of many independence results you can prove things.
Of course, independence has interest [*per se*]{}; still for a given problem in general a solution in ZFC is for me preferable on an independence result. But if it gives a method of forcing (so relevant to a series of problems) the independence result is preferable (of course, I assume there are no other major differences; the depth of the proof would be of first importance to me).
As occurs often in my papers lately, quotations of [**pcf**]{} theory appear.
This paper is also a case of
\[0.3\] Assumption of cases of not GCH at singular (more generally $pp\lambda>
\lambda^+$) are “good", “helpful" assumptions; i.e. traditionally uses of GCH proliferate mainly not from conviction but as you can prove many theorems assuming $2^{\aleph_0}=\aleph_1$ but very few from $2^{\aleph_0}>\aleph_1$, but assuming $2^{\beth_\omega}>\beth^+_\omega$ is helpful in proving.
Unfortunately, most results are only almost in ZFC as they use extremely weak assumptions from [**pcf**]{}, assumptions whose independence is not know. So practically it is not tempting to try to remove them as they may be true, and it is unreasonable to try to prove independence results before independence results on [**pcf**]{} will advance.
In §1 we give an explanation of the earlier difficulties: the problem (of the existence of universals for ${\frak K}^{rs(p)}$) is not like looking for ${\frak K}^{tr}$ (trees with $\omega+1$ levels) but for ${\frak K}^{tr}_{\langle\lambda_\alpha:\alpha<
\omega\rangle}$ where
($\oplus$)
: $\lambda^{\aleph_0}_n<\lambda_{n+1}<\mu$, $\lambda_n$ are regular and $\mu^+<\lambda=\lambda_\omega={{\rm cf}}(\lambda)<\mu^{\aleph_0}$ and ${\frak K}^{tr}_{\langle\lambda_n:n<\omega\rangle}$ is $$\{T:T\mbox{ a tree with $\omega+1$ levels, in level $n< \omega$ there are
$\lambda_n$ elements}\}.$$
We also consider ${\frak K}^{tr}_{\langle\lambda_\alpha:\alpha\leq\omega
\rangle}$, which is defined similarly but the level $\omega$ of $T$ is required to have $\lambda_\omega$ elements.
For ${\frak K}^{rs(p)}$ this is proved fully, for ${\frak K}^{rtf}$ this is proved for the natural examples.
In §2 we define two such basic examples: one is ${\frak K}^{tr}_{\langle
\lambda_\alpha:\alpha \le \omega \rangle}$, and the second is ${\frak K}^{fc}_{\langle \lambda_\alpha:\alpha\leq\omega\rangle}$. The first is a tree with $\omega+1$ levels; in the second we have slightly less restrictions. We have $\omega$ kinds of elements and a function from the $\omega$-th-kind to the $n$th kind. We can interpret a tree $T$ as a member of the second example: $P^T_\alpha = \{x:x
\mbox{ is of level }\alpha\}$ and $$F_n(x) = y \quad\Leftrightarrow \quad x \in P^T_\omega\ \&\ y \in
P^T_n\ \&\ y <_T x.$$ For the second we recapture the non-existence theorems.
But this is not one of the classes we considered originally.
In §3 we return to ${\frak K}^{rtf}$ (reduced torsion free Abelian groups) and prove the non-existence of universal ones in $\lambda$ if $2^{\aleph_0}
< \mu^+<\lambda={{\rm cf}}(\lambda)<\mu^{\aleph_0}$ and an additional very weak set theoretic assumption (the consistency of its failure is not known).
Note that (it will be proved in [@Sh:622]):
($\otimes$)
: if $\lambda<2^{\aleph_0}$ then ${\frak K}^{rtf}_\lambda$ has no universal members.
Note: if $\lambda=\lambda^{\aleph_0}$ then ${\frak K}^{tr}_\lambda$ has universal member also ${\frak K}^{rs(p)}_\lambda$ (see [@Fu]) but not ${\frak K}^{rtf}_\lambda$ (see [@Sh:e Ch IV, VI]).
We have noted above that for ${\frak K}^{rtf}_\lambda$ requiring $\lambda\geq 2^{\aleph_0}$ is reasonable: we can prove (i.e. in ZFC) that there is no universal member. What about ${\frak K}^{rs(p)}_\lambda$? By §1 we should look at ${\frak K}^{tr}_{\langle\lambda_i:i\le\omega\rangle}$, $\lambda_\omega=\lambda<2^{\aleph_0}$, $\lambda_n<\aleph_0$.
In §4 we prove the consistency of the existence of universals for ${\frak K}^{tr}_{\langle\lambda_i:i \le\omega\rangle}$ when $\lambda_n\leq \omega$, $\lambda_\omega=\lambda< 2^{\aleph_0}$ but of cardinality $\lambda^+$; this is not the original problem but it seems to be a reasonable variant, and more seriously, it shoots down the hope to use the present methods of proving non-existence of universals. Anyhow this is ${\frak K}^{tr}_{\langle \lambda_i:i \le\omega\rangle}$ not ${\frak K}^{rs(p)}_{\lambda_\omega}$, so we proceed to reduce this problem to the previous one under a mild variant of MA. The intentions are to deal with “there is universal of cardinality $\lambda$" in Džamonja Shelah [@DjSh:614].
The reader should remember that the consistency of e.g.
> [*$2^{\aleph_0}>\lambda>\aleph_0$ and there is no $M$ such that $M\in
> {\frak K}^{rs(p)}$ is of cardinality $<2^{\aleph_0}$ and universal for ${\frak K}^{rs(p)}_\lambda$* ]{}
is much easier to obtain, even in a wider context (just add many Cohen reals).
As in §4 the problem for ${\frak K}^{rs(p)}_\lambda$ was reasonably resolved for $\lambda<2^{\aleph_0}$ (and for $\lambda = \lambda^{\aleph_0}$, see [@KjSh:455]), we now, in §5 turn to $\lambda>2^{\aleph_0}$ (and $\mu,\lambda_n$) as in $(\oplus)$ above. As in an earlier proof we use $\langle C_\delta: \delta\in S\rangle$ guessing clubs for $\lambda$ (see references or later here), so $C_\delta$ is a subset of $\delta$ (so the invariant depends on the representation of $G$ but this disappears when we divide by suitable ideal on $\lambda$). What we do is: rather than trying to code a subset of $C_\delta$ (for $\bar G=\langle G_i:i<\lambda\rangle$ a representation or filtration of the structure $G$ as the union of an increasing continuous sequence of structures of smaller cardinality) by an element of $G$, we do it, say, by some set $\bar x=\langle x_t:t\in
{{\rm Dom}}(I)\rangle$, $I$ an ideal on ${{\rm Dom}}(I)$ (really by $\bar x/I$). At first glance if ${{\rm Dom}}(I)$ is infinite we cannot list [*a priori*]{} all possible such sequences for a candidate $H$ for being a universal member, as their number is $\ge\lambda^{\aleph_0}=\mu^{\aleph_0}$. But we can find a family $${{\cal}F}\subseteq\{\langle x_t:t\in A\rangle:\ A\subseteq{{\rm Dom}}(I),\ A\notin
I,\ x_t\in\lambda\}$$ of cardinality $<\mu^{\aleph_0}$ such that for any $\bar{x}=\langle x_t:t\in
{{\rm Dom}}(I)\rangle$, for some $\bar y\in {{\cal}F}$ we have $\bar y=\bar
x\restriction {{\rm Dom}}(\bar y)$.
As in §3 there is such ${{\cal}F}$ except when some set theoretic statement related to [**pcf**]{} holds. This statement is extremely strong, also in the sense that we do not know to prove its consistency at present. But again, it seems unreasonable to try to prove its consistency before the [**pcf**]{} problem was dealt with. Of course, we may try to improve the combinatorics to avoid the use of this statement, but are naturally discouraged by the possibility that the [**pcf**]{} statement can be proved in ZFC; thus we would retroactively get the non-existence of universals in ZFC.
In §6, under weak [**pcf**]{} assumptions, we prove: if there is a universal member in ${\frak K}^{fc}_\lambda$ then there is one in ${\frak K}^{rs(p)}_\lambda$; so making the connection between the combinatorial structures and the algebraic ones closer.
In §7 we give other weak [**pcf**]{} assumptions which suffice to prove non-existence of universals in ${\frak K}^x_{\langle\lambda_\alpha:\alpha
\le\omega\rangle}$ (with $x$ one of the “legal” values): $\max{{\rm pcf}}\{\lambda_n:n<\omega\}=\lambda$ and ${{\cal}P}(
\{\lambda_n:n<\omega\})/J_{<\lambda}\{\lambda_n:n<\omega\}$ is an infinite Boolean Algebra (and $(\oplus)$ holds, of course).
In [@KjSh:409], for singular $\lambda$ results on non-existence of universals (there on orders) can be gotten from these weak [**pcf**]{} assumptions.
In §8 we get parallel results from, in general, more complicated assumptions.
In §9 we turn to a closely related class: the class of metric spaces with (one to one) continuous embeddings, similar results hold for it. We also phrase a natural criterion for deducing the non-existence of universals from one class to another.
In §10 we deal with modules and in §11 we discuss the open problems of various degrees of seriousness.
The sections are written in the order the research was done.
\[0.4\] Note that we deal with trees with $\omega+1$ levels rather than, say, with $\kappa+1$, and related situations, as those cases are quite popular. But inherently the proofs of §1-§3, §5-§9 work for $\kappa+1$ as well (in fact, [**pcf**]{} theory is stronger).
For a structure $M$, $\|M\|$ is its cardinality.
For a model, i.e. a structure, $M$ of cardinality $\lambda$, where $\lambda$ is regular uncountable, we say that $\bar M$ is a representation (or filtration) of $M$ if $\bar M=\langle M_i:i<\lambda\rangle$ is an increasing continuous sequence of submodels of cardinality $<\lambda$ with union $M$.
For a set $A$, we let $[A]^\kappa = \{B:B \subseteq A \mbox{ and }
|B|=\kappa\}$.
For a set $C$ of ordinals, $${{\rm acc}}(C)=\{\alpha\in C: \alpha=\sup(\alpha \cap C)\}, \mbox{(set of
accumulation points)}$$ $${{\rm nacc}}(C)=C\setminus {{\rm acc}}(C) \ (=\mbox{ the set of non-accumulation points}).$$ We usually use $\eta$, $\nu$, $\rho$ for sequences of ordinals; let $\eta\vartriangleleft\nu$ means $\eta$ is an initial segment of $\nu$.
Let ${{\rm cov}}(\lambda, \mu, \theta, \sigma)= \min\{|{{\cal}P}|: {{\cal}P}\subseteq
[\lambda]^{<\mu}$, and for every $A\in [\lambda]^{<\theta}$ for some $\alpha<
\sigma$ and $B_i\in {{\cal}P}$ for $i< \alpha$ we have $A\subseteq
\bigcup\limits_{i< \alpha} B_i\}$. Remember that for an ordinal $\alpha$, e.g. a natural number, $\alpha=\{\beta:\beta<\alpha\}$.
${\frak K}^{rs(p)}$ is the class of (Abelian) groups which are $p$-groups (i.e. $(\forall x\in G)(\exists n)[p^nx = 0]$) reduced (i.e. have no divisible non-zero subgroups) and separable (i.e. every cyclic pure subgroup is a direct summand). See [@Fu].
For $G\in{\frak K}^{rs(p)}$ define a norm $\|x\|=\inf\{2^{-n}:
p^n \mbox{ divides } x\}$. Now every $G\in {\frak K}^{rs(p)}$ has a basic subgroup $B=\bigoplus\limits_{\scriptstyle n<\omega\atop\scriptstyle
i<\lambda_n} {\Bbb Z} x^n_i$, where $x^n_i$ has order $p^{n+1}$, and every $x\in G$ can be represented as $\sum\limits_{\scriptstyle n<\omega\atop
\scriptstyle i<\lambda_n} a^n_ix^n_i$, where for each $n$, $w_n(x)=\{i<
\lambda_n:a^n_ix^n_i\ne 0\}$ is finite.
${\frak K}^{rtf}$ is the class of Abelian groups which are reduced and torsion free (i.e. $G \models nx = 0$, $n>0$$\Rightarrow\qquad x = 0$).
For a group $G$ and $A\subseteq G$ let $\langle A\rangle_G$ be the subgroup of $G$ generated by $A$, we may omit the subscript $G$ if clear from the context.
Group will mean an Abelian group, even if not stated explicitly.
Let $H\subseteq_{pr} G$ means $H$ is a pure subgroup of $G$.
Let $nG=\{nx: x\in G\}$ and let $G[n]=\{x\in G: nx=0\}$.
${\frak K}$ will denote a class of structures with the same vocabulary, with a notion of embeddability, equivalently a notion $\leq_{{\frak K}}$ of submodel.
Their prototype is ${\frak K}^{tr}_{\langle \lambda_n:n<\omega
\rangle}$ not ${\frak K}^{tr}$!
==============================================================
If we look for universal member in ${\frak K}^{rs(p)}_\lambda$, thesis \[0.1\] suggests to us to think it is basically ${\frak K}^{tr}_\lambda$ (trees with $\omega+1$ levels, i.e. ${\frak K}^{tr}_{\lambda}$ is our prototype), a way followed in [@KjSh:455], [@Sh:456]. But, as explained in the introduction, this does not give answer for the case of usual embedding for the family of all such groups. Here we show that for this case the thesis should be corrected. More concretely, the choice of the prototype means the choice of what we expect is the division of the possible classes. That is for a family of classes a choice of a prototype assert that we believe that they all behave in the same way.
We show that looking for a universal member $G$ in ${\frak K}^{rs(p)}_\lambda$ is like looking for it among the $G$’s with density $\le\mu$ ($\lambda,\mu$, as usual, as in $(\oplus)$ from §0). For ${\frak K}^{rtf}_\lambda$ we get weaker results which still cover the examples usually constructed, so showing that the restrictions in [@KjSh:455] (to pure embeddings) and [@Sh:456] (to $(<\lambda)$-stable groups) were natural.
\[1.1\] Assume that $\mu=\sum\limits_{n<\omega}\lambda_n=\lim\sup\limits_n\lambda_n$, $\mu\le\lambda\le\mu^{\aleph_0}$, and $G$ is a reduced separable $p$-group such that $$|G|=\lambda\quad\mbox{ and }\quad\lambda_n(G)=:\dim((p^n G)[p]/
(p^{n+1}G)[p])\le\mu$$ (this is a vector space over ${\Bbb Z}/p {\Bbb Z}$, hence the dimension is well defined).\
[*Then*]{} there is a reduced separable $p$-group $H$ such that $|H|=\lambda$, $H$ extends $G$ and $(p^nH)[p]/(p^{n+1}H)[p]$ is a group of dimension $\lambda_n$ (so if $\lambda_n\geq \aleph_0$, this means cardinality $\lambda_n$).
\[1.1A\] So for $H$ the invariants from [@KjSh:455] are trivial.
(See Fuchs [@Fu]). We can find $z^n_i$ (for $n<\omega$, $i<\lambda_n(G)\le\mu$) such that:
(a)
: $z^n_i$ has order $p^n$,
(b)
: $B=\sum\limits_{n,i}\langle z^n_i \rangle_G$ is a direct sum,
(c)
: $B$ is dense in $G$ in the topology induced by the norm $$\|x\|=\min\{2^{-n}:p^n \mbox{ divides } x \mbox{ in } G\}.$$
For each $n<\omega$ and $i<\lambda_n(G)$ ($\le\mu$) choose $\eta^n_i\in
\prod\limits_{m<\omega}\lambda_m$, pairwise distinct such that for $(n^1,i^1)
\neq (n^2,i^2)$ for some $n(*)$ we have: $$\lambda_n \ge \lambda_{n(*)}\qquad \Rightarrow\qquad \eta^{n^1}_{i^1}(n)
\neq \eta^{n^2}_{i^2}(n).$$ Let $H$ be generated by $G$, $x^m_i$ ($i<\lambda_m$, $m<\omega$), $y^{n,k}_i$ ($i<\lambda_n$, $n<\omega$, $n\le k<\omega)$ freely except for:
($\alpha$)
: the equations of $G$,
($\beta$)
: $y^{n,n}_i = z^n_i$,
($\gamma$)
: $py^{n,k+1}_i - y^{n,k}_i = x^k_{\eta^n_i(k)}$,
($\delta$)
: $p^{n+1}x^n_i = 0$,
($\varepsilon$)
: $p^{k+1}y^{n,k}_i = 0$.
Now check. $\square_{\ref{1.1}}$
\[1.2\]
1. ${{\bf t}}$ denotes a sequence $\langle t_i:i<\omega\rangle$, $t_i$ a natural number $>1$.
2. For a group $G$ we define $$G^{[{{\bf t}}]}=\{x\in G:\bigwedge_{j<\omega}[x\in (\prod_{i<j} t_i)
G]\}.$$
3. We can define a semi-norm $\|-\|_{{{\bf t}}}$ on $G$ $$\|x\|_{{{\bf t}}}=\min\{2^{-i}:x\in (\prod_{j<i} t_j)G\}$$ and so the semi-metric $$d_{{{\bf t}}}(x,y)=\|x-y\|_{{{\bf t}}}.$$
\[1.2A\] So, if $\|-\|_{{{\bf t}}}$ is a norm, $G$ has a completion under $\|-\|_{{{\bf t}}}$, which we call $\|-\|_{{{\bf t}}}$-completion; if ${{\bf t}}=\langle i!:i<\omega
\rangle$ we refer to $\|-\|_{{{\bf t}}}$ as $\Bbb Z$-adic norm, and this induces $\Bbb Z$-adic topology, so we can speak of $\Bbb Z$-adic completion.
\[1.3\] Suppose that
($\otimes_0$)
: $\mu=\sum\limits_n\lambda_n$ and $\mu\le\lambda\le
\mu^{\aleph_0}$ for simplicity, $2<2\cdot\lambda_n\le\lambda_{n+1}$ (maybe $\lambda_n$ is finite!),
($\otimes_1$)
: $G$ is a torsion free group, $|G|=\lambda$, and $G^{[{{\bf t}}]}=\{0\}$,
($\otimes_2$)
: $G_0\subseteq G$, $G_0$ is free and $G_0$ is ${{\bf t}}$-dense in $G$ (i.e. in the topology induced by the metric $d_{{{\bf t}}}$), where ${{\bf t}}$ is a sequence of primes.
[*Then*]{} there is a torsion free group $H$, $G\subseteq H$, $H^{[{{\bf t}}]}
=\{0\}$, $|H|=\lambda$ and, under $d_{{{\bf t}}}$, $H$ has density $\mu$.
Let $\{x_i:i<\lambda\}$ be a basis of $G_0$. Let $\eta_i\in
\prod\limits_{n<\omega} \lambda_n$ for $i<\mu$ be distinct such that $\eta_i(n+1)\geq \lambda_n$ and $$i\ne j\qquad \Rightarrow\qquad (\exists m)(\forall n)[m \le n \quad
\Rightarrow\quad \eta_i(n) \ne \eta_j(n)].$$ Let $H$ be generated by $$G,\ \ x^m_i \mbox{ (for $i<\lambda_m$, $m<\omega$), }\ y^n_i \mbox{
(for $i<\mu$, $n<\omega$)}$$ freely except for
(a)
: the equations of $G$,
(b)
: $y^0_i = x_i$,
(c)
: $t_n\, y^{n+1}_i + y^n_i = x^n_{\eta_i(n)}$.
[**Fact A**]{} $H$ extends $G$ and is torsion free.
\[Why? As $H$ can be embedded into the divisible hull of $G$.\]
[**Fact B**]{} $H^{[{{\bf t}}]}= \{0\}$.
Let $K$ be a countable pure subgroup of $H$ such that $K^{[{{\bf t}}]}\ne
\{0\}$. Now without loss of generality $K$ is generated by
(i)
: $K_1\subseteq G\cap\mbox{ [the $d_{{{\bf t}}}$--closure of $\langle x_i:
i\in I\rangle_G]$]}$, where $I$ is a countable infinite subset of $\lambda$ and $K_1\supseteq\langle x_i:i\in I\rangle_G$,
(ii)
: $y^m_i$, $x^n_j$ for $i\in I$, $m<\omega$ and $(n,j)\in
J$, where $J\subseteq \omega\times \lambda$ is countable and $$i\in I,\ n<\omega\qquad\Rightarrow\qquad (n,\eta_i(n))\in J.$$
Moreover, the equations holding among those elements are deducible from the equations of the form
(a)$^-$
: equations of $K_1$,
(b)$^-$
: $y^0_i=x_i$ for $i \in I$,
(c)$^-$
: $t_n\,y^{n+1}_i+y^n_i=x^n_{\eta_i(n)}$ for $i\in I,n<\omega$.
We can find $\langle k_i:i<\omega\rangle$ such that $$[n\ge k_i\ \&\ n\ge k_j\ \&\ i \ne j\qquad \Rightarrow\qquad \eta_i(n)\ne
\eta_j(n)].$$ Let $y \in K\setminus\{0\}$. Then for some $j$, $y\notin (\prod\limits_{i<j}
t_i)G$, so for some finite $I_0\subseteq I$ and finite $J_0\subseteq J$ and $$y^* \in\langle\{x_i:i\in I_0\}\cup\{x^n_\alpha:(n,\alpha)\in J_0\}
\rangle_K$$ we have $y-y^*\in(\prod\limits_{i<j} t_i) G$. Without loss of generality $J_0
\cap\{(n,\eta_i(n)):i\in I,\ n\ge k_i\}=\emptyset$. Now there is a homomorphism $\varphi$ from $K$ into the divisible hull $K^{**}$ of $$K^* = \langle\{x_i:i\in I_0\}\cup\{x^n_j:(n,j)\in J_0\}\rangle_G$$ such that ${{\rm Rang}}(\varphi)/K^*$ is finite. This is enough.
[**Fact C**]{} $H_0=:\langle x^n_i:n<\omega,i<\lambda_n
\rangle_H$ is dense in $H$ by $d_{{{\bf t}}}$.
Straight as each $x_i$ is in the $d_{{{\bf t}}}$-closure of $H_0$ inside $H$.
Noting then that we can increase the dimension easily, we are done. $\square_{\ref{1.3}}$
On structures like $(\prod\limits_n \lambda_n,E_m)_{m<\omega}$, $\eta E_m \nu =: \eta(m)=\nu(m)$
================================================================================================
\[2.1\] We discuss the existence of universal members in cardinality $\lambda$, $\mu^+<\lambda<\mu^{\aleph_0}$, for certain classes of groups. The claims in §1 indicate that the problem is similar not to the problem of the existence of a universal member in ${\frak K}^{tr}_\lambda$ (the class of trees with $\lambda$ nodes, $\omega+1$ levels) but to the one where the first $\omega$ levels, are each with $<\mu$ elements. We look more carefully and see that some variants are quite different.
The major concepts and Lemma (\[2.3\]) are similar to those of §3, but easier. Since detailed proofs are given in §3, here we give somewhat shorter proofs.
\[2.2\] For a sequence $\bar\lambda=\langle\lambda_i:i\le\delta\rangle$ of cardinals we define:
(A)
: ${\frak K}^{tr}_{\bar \lambda}=\{T:\,T$ is a tree with $\delta
+1$ levels (i.e. a partial order such that
for $x\in T$, ${{\rm lev}}_T(x)=:{{\rm otp}}(\{y:y<x\})$ is an ordinal $\le\delta$) such
that:${{\rm lev}}_i(T)=:\{x\in T:{{\rm lev}}_T(x)=i\}$ has cardinality $\le\lambda_i\}$,
(B)
: ${\frak K}^{fc}_{\bar\lambda}=\{M:\,M=(|M|,P_i,F_i)_{i\le\delta}$, $|M|$ is the disjoint union of
$\langle P_i:i\le\delta\rangle$, $F_i$ is a function from $P_\delta$ to $P_i$, $\|P_i\|\le\lambda_i$,
$F_\delta$ is the identity (so can be omitted)$\}$,
(C)
: If $[i\le\delta\quad \Rightarrow\quad \lambda_i=\lambda]$ then we write $\lambda$, $\delta+1$ instead of $\langle\lambda_i:i\le\delta\rangle$.
\[2.2A\] Embeddings for ${\frak K}^{tr}_{\bar\lambda}$, ${\frak K}^{fc}_{\bar\lambda}$ are defined naturally: for ${\frak K}^{tr}_{\bar\lambda}$ embeddings preserve $x<y$, $\neg x<y$, ${{\rm lev}}_T(x)=\alpha$; for ${\frak K}^{fc}_{\bar\lambda}$ embeddings are defined just as for models.
If $\delta^1=\delta^2=\delta$ and $[i<\delta\quad\Rightarrow\quad\lambda^1_i
\le\lambda^2_i]$ and $M^\ell\in{\frak K}^{fc}_{\bar\lambda^\ell}$, (or $T^\ell
\in{\frak K}^{tr}_{\bar\lambda^\ell}$) for $\ell=1,2$, then an embedding of $M^1$ into $M^2$ ($T^1$ into $T^2$) is defined naturally.
\[2.3\] Assume $\bar\lambda=\langle\lambda_i:i\le\delta\rangle$ and $\theta$, $\chi$ satisfy (for some $\bar C$):
(a)
: $\lambda_\delta$, $\theta$ are regular, $\bar C=\langle C_\alpha:
\alpha\in S\rangle$, $S\subseteq\lambda=:\lambda_\delta$, $C_\alpha\subseteq
\alpha$, for every club $E$ of $\lambda$ for some $\alpha$ we have $C_\alpha
\subseteq E$, $\lambda_\delta<\chi\le |C_\alpha|^\theta$ and ${{\rm otp}}(C_\alpha)
\ge\theta$,
(b)
: $\lambda_i\le\lambda_\delta$,
(c)
: there are $\theta$ pairwise disjoint sets $A\subseteq\delta$ such that $\prod\limits_{i\in A}\lambda_i\ge\lambda_\delta$.
[*Then*]{}
($\alpha$)
: there is no universal member in ${\frak K}^{fc}_{\bar
\lambda}$;moreover
($\beta$)
: if $M_\alpha\in {\frak K}^{fc}_{\bar\lambda}$ or even $M_\alpha\in {\frak K}^{fc}_{\lambda_\delta}$ for $\alpha<\alpha^*<\chi$ [*then*]{} some $M\in {\frak K}^{fc}_{\bar\lambda}$ cannot be embedded into any $M_\alpha$.
\[2.3A\] Note that clause $(\beta)$ is relevant to our discussion in §1: the non-universality is preserved even if we increase the density and, also, it is witnessed even by non-embeddability in many models.
Let $\langle A_\varepsilon:\varepsilon<\theta\rangle$ be as in clause (c) and let $\eta^\varepsilon_\alpha\in\prod\limits_{i\in A_\varepsilon}
\lambda_i$ for $\alpha<\lambda_\delta$ be pairwise distinct. We fix $M_\alpha
\in {\frak K}^{fc}_{\lambda_\delta}$ for $\alpha<\alpha^*<\chi$.
For $M\in {\frak K}^{fc}_{\bar\lambda}$, let $\bar M=(|M|,P^M_i,
F^M_i)_{i\le\delta}$ and let $\langle M_\alpha: \alpha<
\lambda_\delta\rangle$ be a representation (=filtration) of $M$; for $\alpha\in S$, $x\in P^M_\delta$, let
$$\begin{array}{rl}
{{\rm inv}}(x,C_\alpha;\bar M)=\big\{\beta\in C_\alpha:&\mbox{for some }\varepsilon
<\theta\mbox{ and } y\in M_{\min(C_\alpha\setminus (\beta+1))}\\
&\mbox{we have }\ \bigwedge\limits_{i\in A_\varepsilon} F^M_i(x)=F^M_i(y)\\
&\mbox{\underbar{but} there is no such } y\in M_\beta\big\}.
\end{array}$$ $${{\rm Inv}}(C_\alpha,\bar M)=\{{{\rm inv}}(x,C_\alpha,\bar M):x\in P^M_\delta\}.$$ $${{\rm INv}}(\bar M,\bar C)=\langle{{\rm Inv}}(C_\alpha,\bar M):\alpha\in S\rangle.$$ $${{\rm INV}}(\bar M,\bar C)={{\rm INv}}(\bar M,\bar C)/{{\rm id}}^a(\bar C).$$ Recall that $${{\rm id}}^a(\bar C)=\{T\subseteq\lambda:\mbox{ for some club $E$ of
$\lambda$ for no $\alpha\in T$ is $C_\alpha\subseteq E$}\}.$$ The rest should be clear (for more details see proofs in §3), noticing
\[2.3B\]
1. ${{\rm INV}}(\bar M,\bar C)$ is well defined, i.e. if $\bar M^1$, $\bar M^2$ are representations (=filtrations) of $M$ then ${{\rm INV}}(\bar M^1,\bar
C)={{\rm INV}}(\bar M^2,\bar C)$.
2. ${{\rm Inv}}(C_\alpha,\bar M)$ has cardinality $\le\lambda$.
3. ${{\rm inv}}(x,C_\alpha;\bar M)$ is a subset of $C_\alpha$ of cardinality $\le \theta$.
$\square_{\ref{2.3}}$
\[2.4\] If $\mu=\sum\limits_{n<\omega}\lambda_n$ and $\lambda^{\aleph_0}_n<
\lambda_{n+1}$ and $\mu^+<\lambda_\omega={{\rm cf}}(\lambda_\omega)<\mu^{\aleph_0}$, [*then*]{} in ${\frak K}^{fc}_{\langle\lambda_\alpha:\alpha\le\omega\rangle}$ there is no universal member and even in ${\frak K}^{fc}_{\langle
\lambda_\omega:\alpha\le\omega\rangle}$ we cannot find a member universal for it.
Should be clear or see the proof in §3. $\square_{\ref{2.4}}$
Reduced torsion free groups: Non-existence of universals
========================================================
We try to choose torsion free reduced groups and define invariants so that in an extension to another such group $H$ something survives. To this end it is natural to stretch “reduced" near to its limit.
\[3.1\]
1. ${\frak K}^{tf}$ is the class of torsion free (abelian) groups.
2. ${\frak K}^{rtf}=\{G\in {\frak K}^{tf}:{\Bbb Q}$ is not embeddable into $G$ (i.e. $G$ is reduced)$\}$.
3. ${{\bf P}}^*$ denotes the set of primes.
4. For $x\in G$, ${{\bf P}}(x,G)=:\{p\in{{\bf P}}^*: \bigwedge\limits_n x\in
p^n G\}$.
5. ${\frak K}^x_\lambda=\{G\in{\frak K}^x:\|G\|=\lambda\}$.
6. If $H\in {\frak K}^{rtf}_\lambda$, we say $\bar H$ is a representation or filtration of $H$ if $\bar H=\langle
H_\alpha:\alpha<\lambda\rangle$ is increasing continuous and $H=\bigcup\limits_{\alpha<\lambda} H_\alpha$, $H\in {\frak K}^{rtf}$ and each $H_\alpha$ has cardinality $<\lambda$.
\[3.2\]
1. If $G\in {\frak K}^{rtf}$, $x\in G\setminus\{0\}$, $Q\cup{{\bf P}}(x,G)
\subsetneqq{{\bf P}}^*$, $G^+$ is the group generated by $G,y,y_{p,\ell}$ ($\ell
<\omega$, $p\in Q$) freely, except for the equations of $G$ and $$y_{p,0}=y,\quad py_{p,\ell+1}=y_{p,\ell}\quad \mbox{ and }\quad
y_{p,\ell}=z\mbox{ when } z\in G,p^\ell z=x$$ [*then*]{} $G^+\in {\frak K}^{rtf}$, $G\subseteq_{pr}G^+$ (pure extension).
2. If $G_i\in {\frak K}^{rtf}$ ($i<\alpha$) is $\subseteq_{pr}$-increasing [*then*]{} $G_i\subseteq_{pr}\bigcup\limits_{j<\alpha}G_j\in{\frak K}^{rtf}$ for every $i<\alpha$.
The proof of the following lemma introduces a method quite central to this paper.
\[3.3\] Assume that
$(*)^1_\lambda$
: $2^{\aleph_0}+\mu^+<\lambda={{\rm cf}}(\lambda)<
\mu^{\aleph_0}$,
$(*)^2_\lambda$
: for every $\chi<\lambda$, there is $S\subseteq
[\chi]^{\le\aleph_0}$, such that:
(i)
: $|S|<\lambda$,
(ii)
: if $D$ is a non-principal ultrafilter on $\omega$ and $f:D
\longrightarrow\chi$ [*then*]{} for some $a\in S$ we have $$\bigcap \{X\in D:f(X)\in a\}\notin D.$$
[*Then*]{}
($\alpha$)
: in ${\frak K}^{rtf}_\lambda$ there is no universal member (under usual embeddings (i.e. not necessarily pure)),
($\beta$)
: moreover, $G_i\in {\frak
K}^{rtf}_\lambda$, for $i<i^*<\mu^{\aleph_0}$ $G\in
{\frak K}^{rtf}_\lambda$ not embeddable into any one of $G_i$.
Before we prove \[3.3\] we consider the assumptions of \[3.3\] in \[3.4\], \[3.5\].
\[3.4\]
1. In \[3.3\] we can replace $(*)^1_\lambda$ by
$(**)^1_\lambda$ (i)
: $2^{\aleph_0}<\mu<\lambda={{\rm cf}}(\lambda)<
\mu^{\aleph_0}$,
(ii)
: there is $\bar C=\langle C_\delta:\delta\in S^*\rangle$ such that $S^*$ is a stationary subset of $\lambda$, each $C_\delta$ is a subset of $\delta$ with ${{\rm otp}}(C_\delta)$ divisible by $\mu$, $C_\delta$ closed in $\sup(C_\delta)$ (which normally $\delta$, but not necessarily so) and $$(\forall\alpha)[\alpha\in {{\rm nacc}}(C_\delta)\quad \Rightarrow\quad {{\rm cf}}(\alpha)
>2^{\aleph_0}]$$ (where ${{\rm nacc}}$ stands for “non-accumulation points”), and such that $\bar C$ guesses clubs of $\lambda$ (i.e. for every club $E$ of $\lambda$, for some $\delta\in S^*$ we have $C_\delta\subseteq E$) and $[\delta\in S^*\quad \Rightarrow\quad {{\rm cf}}(\delta)=\aleph_0]$.
2. In $(*)^1_\lambda$ and in $(*)^2_\lambda$, without loss of generality $(\forall\theta<\mu)[\theta^{\aleph_0}<\mu]$ and ${{\rm cf}}(\mu)=\aleph_0$.
1) This is what we actually use in the proof (see below).
2\) Replace $\mu$ by $\mu'=\min\{\mu_1:\mu^{\aleph_0}_1\ge\mu$ (equivalently $\mu^{\aleph_0}_1=\mu^{\aleph_0}$)$\}$. $\square_{\ref{3.4}}$
Compare to, say, [@KjSh:447], [@KjSh:455]; the new assumption is $(*)^2_\lambda$, note that it is a very weak assumption, in fact it might be that it is always true.
\[3.5\] Assume that $2^{\aleph_0}<\mu<\lambda<\mu^{\aleph_0}$ and $(\forall \theta<
\mu)[\theta^{\aleph_0}<\mu]$ (see \[3.4\](2)). Then each of the following is a sufficient condition to $(*)^2_\lambda$:
($\alpha$)
: $\lambda<\mu^{+\omega_1}$,
($\beta$)
: if ${\frak a}\subseteq{{\rm Reg}}\cap\lambda\setminus\mu$ and $|{\frak a}|\le 2^{\aleph_0}$ then we can find $h:{\frak a}\longrightarrow
\omega$ such that: $$\lambda>\sup\{\max{{\rm pcf}}({\frak b}):{\frak b}\subseteq {\frak a}\mbox{
countable, and $h\restriction {\frak b}$ constant}\}.$$
Clause $(\alpha)$ implies Clause $(\beta)$: just use any one-to-one function $h:{{\rm Reg}}\cap\lambda\setminus\mu\longrightarrow\omega$.
Clause $(\beta)$ implies (by [@Sh:410 §6] + [@Sh:430 §2]) that for $\chi<\lambda$ there is $S\subseteq [\chi]^{\aleph_0}$, $|S|<\lambda$ such that for every $Y\subseteq\chi$, $|Y|=2^{\aleph_0}$, we can find $Y_n$ such that $Y=\bigcup\limits_{n<\omega} Y_n$ and $[Y_n]^{\aleph_0}
\subseteq S$. (Remember: $\mu>2^{\aleph_0}$.) Without loss of generality (as $2^{\aleph_0} < \mu < \lambda$):
$(*)$
: $S$ is downward closed.
So if $D$ is a non-principal ultrafilter on $\omega$ and $f:D\longrightarrow
\chi$ then letting $Y={{\rm Rang}}(f)$ we can find $\langle Y_n:n<\omega\rangle$ as above. Let $h:D\longrightarrow\omega$ be defined by $h(A)=\min\{n:f(A)\in
Y_n\}$. So $$X\subseteq D\ \ \&\ \ |X|\le\aleph_0\ \ \&\ \ h\restriction X
\mbox{ constant }\Rightarrow\ f''(X)\in S\quad\mbox{(remember
$(*)$)}.$$ Now for each $n$, for some countable $X_n\subseteq D$ (possibly finite or even empty) we have: $$h \restriction X_n\ \mbox{ is constantly } n,$$ $$\ell <\omega \ \&\ (\exists A\in D)(h(A)=n\ \&\ \ell\notin A)
\Rightarrow (\exists B\in X_n)(\ell\notin B).$$ Let $A_n=:\bigcap\{A:A\in X_n\}=\bigcap\{A:A\in D\mbox{ and } h(X)=n\}$. If the desired conclusion fails, then $\bigwedge\limits_{n<\omega}A_n\in D$. So $$(\forall A)[A\in D\quad \Leftrightarrow\quad\bigvee_{n<\omega} A\supseteq
A_n].$$ So $D$ is generated by $\{A_n:n<\omega\}$ but then $D$ cannot be a non-principal ultrafilter. $\square_{\ref{3.5}}$
The case when $D$ is a principal ultrafilter is trivial.
of Lemma \[3.3\] Let $\bar C=\langle C_\delta:\delta\in S^*\rangle$ be as in $(**)^1_{\bar \lambda}$ (ii) from \[3.4\] (for \[3.4\](1) its existence is obvious, for \[3.3\] - use [@Sh:e VI, old III 7.8]). Let us suppose that $\bar A=\langle A_\delta:\delta\in S^*\rangle$, $A_\delta
\subseteq{{\rm nacc}}(C_\delta)$ has order type $\omega$ ($A_\delta$ like this will be chosen later) and let $\eta_\delta$ enumerate $A_\delta$ increasingly. Let $G_0$ be freely generated by $\{x_i:i<\lambda\}$.
Let $R$ be $$\begin{array}{rl}
\big\{\bar a: &\bar a=\langle a_n:n<\omega\rangle\mbox{ is a sequence of
pairwise disjoint subsets of } {{\bf P}}^*,\\
&\mbox{with union }{{\bf P}}^* \mbox{ for simplicity, such that}\\
&\mbox{for infinitely many }n,\ a_n\ne\emptyset\big\}.
\end{array}$$ Let $G$ be a group generated by $$G_0 \cup \{y^{\alpha,n}_{\bar a},z^{\alpha,n}_{\bar a,p}:\ \alpha<\lambda,
\ \bar a\in R,\ n<\omega,\ p \mbox{ prime}\}$$ freely except for:
(a)
: the equations of $G_0$,
(b)
: $pz^{\alpha,n+1}_{\bar a,p}=z^{\alpha,n}_{\bar a,p}$ when $p\in a_n$, $\alpha<\lambda$,
(c)
: $z^{\delta,0}_{\bar a,p}=y^{\delta,n}_{\bar a}-
x_{\eta_\delta(n)}$ when $p\in a_n$ and $\delta\in S^*$.
Now $G\in {\frak K}^{rtf}_\lambda$ by inspection.
Before continuing the proof of \[3.3\] we present a definition and some facts.
\[3.7\] For a representation $\bar H$ of $H\in {\frak K}^{rtf}_\lambda$, and $x\in H$, $\delta\in S^*$ let
1. ${{\rm inv}}(x,C_\delta;\bar H)=:\{\alpha\in C_\delta:$ for some $Q\subseteq
{{\bf P}}^*$, there is $y\in H_{\min[C_\delta\setminus(\alpha+1)]}$ such that $Q\subseteq{{\bf P}}(x-y,H)$ but for no $y\in H_\alpha$ is $Q\subseteq{{\bf P}}(x-y,H)\}$
(so ${{\rm inv}}(x,C_\delta;\bar H)$ is a subset of $C_\delta$ of cardinality $\le
2^{\aleph_0}$).
2. ${{\rm Inv}}^0(C_\delta,\bar H)=:\{{{\rm inv}}(x,C_\delta;\bar H):x\in
\bigcup\limits_i H_i\}$.
3. ${{\rm Inv}}^1(C_\delta,\bar H)=:\{a:a\subseteq C_\delta$ countable and for some $x\in H$, $a\subseteq{{\rm inv}}(x,C_\delta;\bar H)\}$.
4. ${{\rm INv}}^\ell(\bar H,\bar C)=:{{\rm Inv}}^\ell(H,\bar H,\bar C)=:\langle
{{\rm Inv}}^\ell(C_\delta;\bar H):\delta\in S^*\rangle$ for $\ell\in\{0,1\}$.
5. ${{\rm INV}}^\ell(H,\bar C)=:{{\rm INv}}^\ell(H,\bar H,\bar C)/{{\rm id}}^a(\bar C)$, where $${{\rm id}}^a(\bar C)=:\{T\subseteq\lambda:\mbox{ for some club $E$ of
$\lambda$ for no $\delta\in T$ is }C_\delta\subseteq E\}.$$
6. If $\ell$ is omitted, $\ell = 0$ is understood.
\[3.8\]
1. ${{\rm INV}}^\ell(H,\bar C)$ is well defined.
2. The $\delta$-th component of ${{\rm INv}}^\ell(\bar H,\bar C)$ is a family of $\le\lambda$ subsets of $C_\delta$ each of cardinality $\le 2^{\aleph_0}$ and if $\ell=1$ each member is countable and the family is closed under subsets.
3. [*If*]{} $G_i\in{\frak K}^{rtf}_\lambda$ for $i<i^*$, $i^*<
\mu^{\aleph_0}$, $\bar G^i=\langle\bar G_{i,\alpha}:\alpha<\lambda\rangle$ is a representation of $G_i$,
[*then*]{} we can find $A_\delta\subseteq{{\rm nacc}}(C_\delta)$ of order type $\omega$ such that: $i<i^*$, $\delta\in S^*\qquad \Rightarrow$for no $a$ in the $\delta$-th component of ${{\rm INv}}^\ell(G_i,\bar G^i,\bar C)$ do we have $|a \cap A_\delta|\ge\aleph_0$.
Straightforward. (For (3) note ${{\rm otp}}(C_\delta)\ge\mu$, so there are $\mu^{\aleph_0}>\lambda$ pairwise almost disjoint subsets of $C_\delta$ each of cardinality $\aleph_0$ and every $A\in{{\rm Inv}}(C_\delta,\bar G^i)$ disqualifies at most $2^{\aleph_0}$ of them.) $\square_{\ref{3.8}}$
\[3.9\] Let $G$ be as constructed above for $\langle A_\delta:\delta\in
S^*\rangle,A_\delta\subseteq{{\rm nacc}}(C_\delta)$, ${{\rm otp}}(A_\delta)=\omega$ (where $\langle A_\delta:\delta\in S^*\rangle$ are chosen as in \[3.8\](3) for the sequence $\langle G_i:i<i^* \rangle$ given for proving \[3.3\], see $(\beta)$ there).
Assume $G \subseteq H\in {\frak K}^{rtf}_\lambda$ and $\bar H$ is a filtration of $H$. [*Then*]{} $$\begin{array}{rr}
B=:\big\{\delta:A_\delta\mbox{ has infinite intersection with some}&\ \\
a\in{{\rm Inv}}(C_\delta,\bar H)\big\}&=\ \lambda\ \mod\ {{\rm id}}^a(\bar C).
\end{array}$$
We assume otherwise and derive a contradiction. Let for $\alpha
<\lambda$, $S_\alpha\subseteq [\alpha]^{\le \aleph_0}$, $|S_\alpha|<\lambda$ be as guaranteed by $(*)^2_\lambda$.
Let $\chi>2^\lambda$, ${\frak A}_\alpha\prec (H(\chi),\in,<^*_\chi)$ for $\alpha<\lambda$ increasing continuous, $\|{\frak A}_\alpha\|<\lambda$, $\langle {\frak A}_\beta:\beta\le\alpha\rangle\in {\frak A}_{\alpha+1}$, ${\frak A}_\alpha\cap\lambda$ an ordinal and: $$\langle S_\alpha:\alpha<\lambda\rangle,\ G,\ H,\ \bar C,\ \langle
A_\delta:\delta\in S^* \rangle,\ \bar H,\ \langle x_i, y^\delta_{\bar a},
z^{\delta,n}_{\bar a,p}:\;i,\delta,\bar a,n,p \rangle$$ all belong to ${\frak A}_0$ and $2^{\aleph_0}+1\subseteq {\frak A}_0$. Then $E=\{\delta<\lambda:{\frak A}_\delta \cap\lambda=\delta\}$ is a club of $\lambda$. Choose $\delta\in S^* \cap E\setminus B$ such that $C_\delta
\subseteq E$. (Why can we? As to ${{\rm id}}^a(\bar C)$ belong all non stationary subsets of $\lambda$, in particular $\lambda\setminus E$, and $\lambda\setminus S^*$ and $B$, but $\lambda\notin {{\rm id}}^a(\bar
C)$.) Remember that $\eta_\delta$ enumerates $A_\delta$ (in the increasing order). For each $\alpha\in A_\delta$ (so $\alpha\in E$ hence ${\frak A}_\alpha \cap \lambda=\alpha$ but $\bar H\in {\frak
A}_\alpha$ hence $H\cap {\frak A}_\alpha= H_\alpha$) and $Q\subseteq{{\bf P}}^*$ choose, if possible, $y_{\alpha,Q}\in H_\alpha$ such that: $$Q\subseteq{{\bf P}}(x_\alpha-y_{\alpha,Q},H).$$ Let $I_\alpha=:\{Q\subseteq{{\bf P}}^*:y_{\alpha,Q}$ well defined$\}$. Note (see \[3.4\] $(**)^1_\lambda$ and remember $\eta_\delta(n)\in A_\delta\subseteq
{{\rm nacc}}(C_\delta)$) that ${{\rm cf}}(\alpha)>2^{\aleph_0}$ (by (ii) of \[3.4\] $(**)^1_\lambda$) and hence for some $\beta_\alpha<\alpha$, $$\{ y_{\alpha,Q}:Q\in I_\alpha\}\subseteq H_{\beta_\alpha}.$$ Now:
$\otimes_1$
: $I_\alpha$ is downward closed family of subsets of ${{\bf P}}^*$, ${{\bf P}}^*\notin I_\alpha$ for $\alpha
\in A_\delta$.
\[Why? See the definition for the first phrase and note also that $H$ is reduced for the second phrase.\]
$\otimes_2$
: $I_\alpha$ is closed under unions of two members (hence is an ideal on ${{\bf P}}^*$).
\[Why? If $Q_1,Q_2\in I_\alpha$ then (as $x_\alpha\in G\subseteq H$ witnesses this): $$\begin{array}{rl}
({{\cal}H}(\chi),\in,<^*_\chi)\models &(\exists x)(x\in H\ \&\
Q_1\subseteq{{\bf P}}(x- y_{\alpha,Q_1},H)\ \&\\
&Q_2\subseteq{{\bf P}}(x-y_{\alpha,Q_2},H)).
\end{array}$$ All the parameters are in ${\frak A}_\alpha$ so there is $y\in
{\frak A}_\alpha\cap H$ such that $$Q_1\subseteq{{\bf P}}(y-y_{\alpha,Q_1},H)\quad\mbox{ and }\quad Q_2\subseteq
{{\bf P}}(y-y_{\alpha,Q_2},H).$$ By algebraic manipulations, $$Q_1\subseteq {{\bf P}}(x_\alpha-y_{\alpha,Q_1},H),\ Q_1\subseteq{{\bf P}}(y-y_{\alpha,
Q_1},H)\quad\Rightarrow\quad Q_1\subseteq{{\bf P}}(x_\alpha-y,H);$$ similarly for $Q_2$. So $Q_1\cup Q_2\subseteq{{\bf P}}(x_\alpha-y,H)$ and hence $Q_1\cup Q_2\in I_\alpha$.\]
$\otimes_3$
: If $\bar Q=\langle Q_n:n\in\Gamma\rangle$ are pairwise disjoint subsets of ${{\bf P}}^*$, for some infinite $\Gamma\subseteq\omega$, then for some $n\in\Gamma$ we have $Q_n\in I_{\eta_\delta(n)}$.
\[Why? Otherwise let $a_n$ be $Q_n$ if $n\in \Gamma$, and $\emptyset$ if $n\in \omega\setminus \Gamma$, and let $\bar a=\langle a_n: n<
\omega\rangle$. Now $n\in\Gamma\quad\Rightarrow\quad\eta_\delta(n)\in
{{\rm inv}}(y^{\delta, 0}_{\bar a},C_\delta;\bar H)$ and hence $$A_\delta\cap{{\rm inv}}(y^{\delta, 0}_{\bar a},C_\delta;\bar
H)\supseteq\{\eta_\delta(n):n\in \Gamma\},$$ which is infinite, contradicting the choice of $A_\delta$.\]
$\otimes_4$
: for all but finitely many $n$ the Boolean algebra ${{\cal}P}({{\bf P}}^*)/I_{\eta_\delta(n)}$ is finite.
\[Why? If not, then by $\otimes_1$ second phrase, for each $n$ there are infinitely many non-principal ultrafilters $D$ on ${{\bf P}}^*$ disjoint to $I_{\eta_\delta(n)}$, so for $n<\omega$ we can find an ultrafilter $D_n$ on ${{\bf P}}^*$ disjoint to $I_{\eta_\delta(n)}$, distinct from $D_m$ for $m<n$. Thus we can find $\Gamma\in [\omega]^{\aleph_0}$ and $Q_n\in D_n$ for $n\in\Gamma$ such that $\langle Q_n:n\in\Gamma\rangle$ are pairwise disjoint (as $Q_n\in D_n$ clearly $|Q_n|=\aleph_0$). Why? Look: if $B_n
\in D_0\setminus D_1$ for $n\in\omega$ then $$(\exists^\infty n)(B_n \in D_n)\quad\mbox{ or }\quad(\exists^\infty n)
({{\bf P}}^*\setminus B_n \in D_n),$$ etc. Let $Q_n=\emptyset$ for $n\in\omega\setminus\Gamma$, now $\bar Q=
\langle Q_n:n<\omega\rangle$ contradicts $\otimes_3$.\]
$\otimes_5$
: If the conclusion (of \[3.9\]) fails, then for no $\alpha\in A_\delta$ is ${{\cal}P}({{\bf P}}^*)/I_\alpha$ finite.
\[Why? If not, choose such an $\alpha$ and $Q^*\subseteq{{\bf P}}^*$, $Q^*
\notin I_\alpha$ such that $I=I_\alpha\restriction Q^*$ is a maximal ideal on $Q^*$. So $D=:{{\cal}P}(Q^*)\setminus I$ is a non-principal ultrafilter. Remember $\beta=\beta_\alpha<\alpha$ is such that $\{y_{\alpha,Q}:Q\in I_\alpha\}\subseteq H_\beta$. Now, $H_\beta\in
{\frak A}_{\beta+1}$, $|H_\beta|<\lambda$. Hence $(*)^2_\lambda$ from \[3.3\] (note that it does not matter whether we consider an ordinal $\chi<\lambda$ or a cardinal $\chi<\lambda$, or any other set of cardinality $< \lambda$) implies that there is $S_{H_\beta}\in
{\frak A}_{\beta+1}$, $S_{H_\beta}\subseteq [H_\beta]^{\le \aleph_0}$, $|S_{H_\beta}|<\lambda$ as there. Now it does not matter if we deal with functions from an ultrafilter on $\omega$ an ultrafilter on $Q^*$. We define $f:D\longrightarrow H_\beta$ as follows: for $U\in D$ we let $f(U)=y_{\alpha,Q^* \setminus U}$. (Note: $Q^*\setminus U\in I_\alpha$, hence $y_{\alpha,Q^* \setminus U}$ is well defined.) So, by the choice of $S_{H_\beta}$ (see (ii) of $(*)^2_\lambda$), for some countable $f'
\subseteq f$, $f'\in {\frak A}_{\beta+1}$ and $\bigcap\{U:U\in{{\rm Dom}}(f')\}
\notin D$ (reflect for a minute). Let ${{\rm Dom}}(f')=\{U_0,U_1,\ldots\}$. Then $\bigcup\limits_{n<\omega}(Q^*\setminus U_n)\notin I_\alpha$. But as in the proof of $\otimes_2$, as $$\langle y_\alpha,(Q^* \setminus U_n):n<\omega\rangle\in
{\frak A}_{\beta+1}\subseteq {\frak A}_\alpha,$$ we have $\bigcup\limits_{n<\omega}(Q^*\setminus U_n)\in I_\alpha$, an easy contradiction.\]
Now $\otimes_4$, $\otimes_5$ give a contradiction. $\square_{\ref{3.3}}$
\[3.10\] We can deal similarly with $R$-modules, $|R|<\mu$ $R$ has infinitely many prime ideals $I$. Also the treatment of ${\frak K}^{rs(p)}_\lambda$ is similar to the one for modules over rings with one prime.
Note: if we replace “reduced" by $$x\in G \setminus\{0\}\quad \Rightarrow\quad (\exists p\in{{\bf P}}^*)(x\notin
pG)$$ then here we could have defined $${{\bf P}}(x,H)=:\{p\in {{\bf P}}^*:x\in pH\}$$ and the proof would go through with no difference (e.g. choose a fixed partition $\langle {{\bf P}}^*_n: n< \omega\rangle$ of ${{\bf P}}^*$ to infinite sets, and let ${{\bf P}}'(x, H)=\{n: x\in pH\mbox{ for every }p\in
{{\bf P}}^*_n\}$). Now the groups are less divisible.
\[3.11\] We can get that the groups are slender, in fact, the construction gives it.
Below the continuum there may be universal structures
=====================================================
Both in [@Sh:456] (where we deal with universality for $(<\lambda)$-stable (Abelian) groups, like ${\frak K}^{rs(p)}_\lambda$) and in §3, we restrict ourselves to $\lambda>2^{\aleph_0}$, a restriction which does not appear in [@KjSh:447], [@KjSh:455]. Is this restriction necessary? In this section we shall show that at least to some extent, it is.
We first show under MA that for $\lambda<2^{\aleph_0}$, any $G\in
{\frak K}^{rs(p)}_\lambda$ can be embedded into a “nice" one; our aim is to reduce the consistency of “there is a universal in ${\frak K}^{rs(p)}_\lambda$" to “there is a universal in ${\frak K}^{tr}_{\langle\aleph_0:n<\omega\rangle\char 94\langle \lambda
\rangle}$". Then we proceed to prove the consistency of the latter. Actually a weak form of MA suffices.
\[4.2\]
1. $G\in {\frak K}^{rs(p)}_\lambda$ is [*tree-like*]{} if:
(a)
: we can find a basic subgroup $B=
\bigoplus\limits_{\scriptstyle i<\lambda_n\atop\scriptstyle n<\omega}
{\Bbb Z} x^n_i$, where $$\lambda_n=\lambda_n(G)=:\dim\left((p^nG)[p]/p^{n+1}(G)[p]\right)$$ (see Fuchs [@Fu]) such that: ${\Bbb Z} x^n_i \cong {\Bbb Z}/p^{n+1}
{\Bbb Z}$ and
$\otimes_0$
: every $x\in G$ has the form $$\sum\limits_{n,i}\{a^n_i p^{n-k} x^n_i:n\in [k,\omega)\mbox{ and }
i<\lambda\}$$ where $a^n_i\in{\Bbb Z}$ and $$n<\omega\quad \Rightarrow\quad w_n[x]=:\{i:a^n_i\, p^{n-k}x^n_i\neq 0\}
\mbox{ is finite}$$
(this applies to any $G\in {\frak K}^{rs(p)}_\lambda$ we considered so far; we write $w_n[x]=w_n[x,\bar Y]$ when $\bar Y=\langle x^n_i:n,i\rangle$). Moreover
(b)
: $\bar Y=\langle x^n_i:n,i\rangle$ is tree-like inside $G$, which means that we can find $F_n:\lambda_{n+1}\longrightarrow\lambda_n$ such that letting $\bar F=\langle F_n:n<\omega\rangle$, $G$ is generated by some subset of $\Gamma(G,\bar Y,\bar F)$ where: $$\hspace{-0.5cm}\begin{array}{rl}
\Gamma(G,\bar Y,\bar F)=\big\{x:&\mbox{for some }\eta\in\prod\limits_{n
<\omega}\lambda_n, \mbox{ for each } n<\omega \mbox{ we have}\\
&F_n(\eta(n+1))=\eta(n)\mbox{ and }x=\sum\limits_{n\ge k}
p^{n-k}x^n_{\eta(n)}\big\}.
\end{array}$$
2. $G\in {\frak K}^{rs(p)}_\lambda$ is [*semi-tree-like*]{} if above we replace (b) by
(b)$'$
: we can find a set $\Gamma\subseteq\{\eta:\eta$ is a partial function from $\omega$ to $\sup\limits_{n<\omega} \lambda_n$ with $\eta(n)< \lambda_n\}$ such that:
($\alpha$)
: $\eta_1\in\Gamma,\ \eta_2\in\Gamma,\ \eta_1(n)=\eta_2(n)
\quad \Rightarrow\quad\eta_1\restriction n=\eta_2\restriction n$,
($\beta$)
: for $\eta\in\Gamma$ and $n\in{{\rm Dom}}(\eta)$, there is $$y_{\eta,n}=\sum \{p^{m-n}x^m_{\eta(m)}:m \in {{\rm Dom}}(\eta)\mbox{ and }
m \ge n\}\in G,$$
($\gamma$)
: $G$ is generated by $$\{x^n_i:n<\omega,i<\lambda_n\}\cup\{y_{\eta,n}:\eta\in\Gamma,n\in
{{\rm Dom}}(\eta)\}.$$
3. $G\in {\frak K}^{rs(p)}_\lambda$ is [*almost tree-like*]{} if in (b)$'$ we add
($\delta$)
: for some $A\subseteq\omega$ for every $\eta\in\Gamma$, ${{\rm Dom}}(\eta)=A$.
\[4.3\]
1. Suppose $G\in {\frak K}^{rs(p)}_\lambda$ is almost tree-like, as witnessed by $A\subseteq\omega$, $\lambda_n$ (for $n<\omega$), $x^n_i$ (for $n\in A$, $i<\lambda_n$), and if $n_0<n_2$ are successive members of $A$, $n_0<n<n_2$ then $\lambda_n\ge\lambda_{n_0}$ or just $$\lambda_n\ge|\{\eta(n_0):\eta\in\Gamma\}|.$$ [*Then*]{} $G$ is tree-like (possibly with other witnesses).
2. If in \[4.2\](3) we just demand $\eta\in\Gamma\quad\Rightarrow
\quad\bigvee\limits_{n<\omega}{{\rm Dom}}(\eta)\setminus n=A\setminus n$; then changing the $\eta$’s and the $y_{\eta,n}$’s we can regain the “almost tree-like".
1\) For every successive members $n_0<n_2$ of $A$ for $$\alpha\in S_{n_0}=:\{\alpha:(\exists\eta)[\eta\in\Gamma\ \&\ \eta(n_0)
=\alpha]\},$$ choose ordinals $\gamma(n_0,\alpha,\ell)$ for $\ell\in (n_0,n_2)$ such that $$\gamma(n_0,\alpha_1,\ell)=\gamma(n_0,\alpha_2,\ell)\quad\Rightarrow
\quad\alpha_1=\alpha_2.$$ We change the basis by replacing for $\alpha\in S_{n_0}$, $\{x^n_\alpha\}\cup
\{x^\ell_{\gamma(n_0,\alpha,\ell)}:\ell\in (n_0,n_2)\}$ (note: $n_0<n_2$ but possibly $n_0+1=n_2$), by: $$\begin{array}{rl}
\biggl\{ x^{n_0}_\alpha + px^{n_0+1}_{\gamma(n_0,\alpha,n_0+1)},
&x^{n_0+1}_{\gamma(n_0,\alpha,n_0+1)} +
px^{n_0+2}_{\gamma(n_0,\alpha,n_0+2)},\ldots, \\
&x^{n_2-2}_{\gamma(n_0,\alpha,n_2-2)} +
px^{n_2-1}_{\gamma(n_0,\alpha,n_2-1)},
x^{n_2-1}_{\gamma(n_0,\gamma,n_2-1)} \biggr\}.
\end{array}$$
2\) For $\eta\in \Gamma$ let $n(\eta)=\min\{ n: n\in A\cap{{\rm Dom}}(\eta)$ and ${{\rm Dom}}(\eta)\setminus n=A\setminus n\}$, and let $\Gamma_n=\{\eta\in \Gamma: n(\eta)=n\}$ for $n\in A$. We choose by induction on $n< \omega$ the objects $\nu_\eta$ for $\eta\in \Gamma_n$ and $\rho^n_\alpha$ for $\alpha< \lambda_n$ such that: $\nu_\eta$ is a function with domain $A$, $\nu_\eta\restriction (A\setminus
n(\eta))=\eta\restriction (A\setminus n(\eta))$ and $\nu_\eta\restriction (A\cap n(\eta))= \rho^n_{\eta(n)}$, $\nu_\eta(n)< \lambda_n$ and $\rho^n_\alpha$ is a function with domain $A\cap n$, $\rho^n_\alpha(\ell)< \lambda_\ell$ and $\rho^n_\alpha
\restriction (A\cap \ell) = \rho^\ell_{\rho^n_\alpha(\ell)}$ for $\ell\in A\cap n$. There are no problems and $\{\nu_\eta: \eta\in
\Gamma_n\}$ is as required. $\square_{\ref{4.3}}$
\[4.1\] Let $\lambda<2^{\aleph_0}$. Any $G\in {\frak K}^{rs(p)}_\lambda$ can be embedded into some $G'\in {\frak K}^{rs(p)}_\lambda$ with countable density which is tree-like.
By \[4.3\] it suffices to get $G'$ “almost tree-like" and $A\subseteq\omega$ which satisfies \[4.3\](1). The ability to make $A$ thin helps in proving Fact E below. By \[1.1\] without loss of generality $G$ has a base (i.e. a dense subgroup of the form) $B=\bigoplus\limits_{\scriptstyle n<\omega\atop\scriptstyle i<\lambda_n}
{\Bbb Z} x^n_i$, where ${\Bbb Z} x^n_i\cong{\Bbb Z}/p^{n+1}{\Bbb Z}$ and $\lambda_n=\aleph_0$ (in fact $\lambda_n$ can be $g(n)$ if $g\in
{}^\omega\omega$ is not bounded (by algebraic manipulations), this will be useful if we consider the forcing from [@Sh:326 §2]).
Let $B^+$ be the extension of $B$ by $y^{n,k}_i$ ($k<\omega$, $n<\omega$, $i<\lambda_n$) generated freely except for $py^{n,k+1}_i=y^{n,k}_i$ (for $k<\omega$), $y^{n,\ell}_i=p^{n-\ell}x^n_i$ for $\ell\le n$, $n<\omega$, $i<\lambda_n$. So $B^+$ is a divisible $p$-group, let $G^+ =:
B^+\bigoplus\limits_B
G$. Let $\{z^0_\alpha:\alpha<\lambda\}\subseteq G[p]$ be a basis of $G[p]$ over $\{p^n x^n_i:n,i<\omega\}$ (as a vector space over ${\Bbb Z}/p{\Bbb
Z}$ i.e. the two sets are disjoint, their union is a basis); remember $G[p]=\{x\in G:px=0\}$. So we can find $z^k_\alpha\in G$ (for $\alpha<
\lambda$, $k<\omega$ and $k\ne 0$) such that $$pz^{k+1}_\alpha-z^k_\alpha=\sum_{i\in w(\alpha,k)} a^{k,\alpha}_i
x^k_i,$$ where $w(\alpha,k)\subseteq\omega$ is finite (reflect on the Abelian group theory).
We define a forcing notion $P$ as follows: a condition $p \in P$ consists of (in brackets are explanations of intentions):
(a)
: $m<\omega$, $M\subseteq m$,
\[$M$ is intended as $A\cap\{0,\ldots,m-1\}$\]
(b)
: a finite $u\subseteq m\times\omega$ and $h:u\longrightarrow
\omega$ such that $h(n,i)\ge n$,
\[our extensions will not be pure, but still we want that the group produced will be reduced, now we add some $y^{n,k}_i$’s and $h$ tells us how many\]
(c)
: a subgroup $K$ of $B^+$: $$K=\langle y^{n,k}_i:(n,i)\in u,k<h(n,i)\rangle_{B^+},$$
(d)
: a finite $w\subseteq\lambda$,
\[$w$ is the set of $\alpha<\lambda$ on which we give information\]
(e)
: $g:w\rightarrow m + 1$,
\[$g(\alpha)$ is in what level $m'\le m$ we “start to think" about $\alpha$\]
(f)
: $\bar\eta=\langle\eta_\alpha:\alpha\in w\rangle$ (see (i)),
\[of course, $\eta_\alpha$ is the intended $\eta_\alpha$ restricted to $m$ and the set of all $\eta_\alpha$ forms the intended $\Gamma$\]
(g)
: a finite $v\subseteq m\times\omega$,
\[this approximates the set of indices of the new basis\]
(h)
: $\bar t=\{t_{n,i}:(n,i)\in v\}$ (see (j)),
\[approximates the new basis\]
(i)
: $\eta_\alpha\in {}^M\omega$, $\bigwedge\limits_{\alpha\in w}
\bigwedge\limits_{n\in M} (n,\eta_\alpha(n))\in v$,
\[toward guaranteeing clause $(\delta)$ of \[4.2\](3) (see \[4.3\](2))\]
(j)
: $t_{n,i}\in K$ and ${\Bbb Z} t_{n,i} \cong {\Bbb Z}/p^n
{\Bbb Z}$,
(k)
: $K=\bigoplus\limits_{(n,i)\in v} ({\Bbb Z} t_{n,i})$,
\[so $K$ is an approximation to the new basic subgroup\]
(l)
: if $\alpha\in w$, $g(\alpha)\le\ell\le m$ and $\ell\in M$ then $$z^\ell_\alpha-\sum\{t^{n-\ell}_{n,\eta_\alpha(n)}:\ell\le n\in
{{\rm Dom}}(\eta_\alpha)\}\in p^{m-\ell}(K+G),$$
\[this is a step toward guaranteeing that the full difference (when ${{\rm Dom}}(\eta_\alpha)$ is possibly infinite) will be in the closure of $\bigoplus\limits_{\scriptstyle n\in [i,\omega)\atop\scriptstyle
i<\omega} {\Bbb Z} x^n_i$\].
We define the order by:
$p \le q$ if and only if
$(\alpha)$
: $m^p\le m^q$, $M^q \cap m^p = M^p$,
$(\beta)$
: $u^p\subseteq u^q$, $h^p\subseteq h^q$,
$(\gamma)$
: $K^p\subseteq_{pr} K^q$,
$(\delta)$
: $w^p\subseteq w^q$,
$(\varepsilon)$
: $g^p\subseteq g^q$,
$(\zeta)$
: $\eta^p_\alpha\trianglelefteq\eta^q_\alpha$, (i.e. $\eta^p_\alpha$ is an initial segment of $\eta^q_\alpha$)
$(\eta)$
: $v^p\subseteq v^q$,
$(\theta)$
: $t^p_{n,i}=t^q_{n,i}$ for $(n,i)\in v^p$.
[**A Fact**]{} $(P,\le)$ is a partial order.
[*Proof of the Fact:*]{} Trivial.
[**B Fact**]{} $P$ satisfies the c.c.c. (even is $\sigma$-centered).
[*Proof of the Fact:*]{} It suffices to observe the following.
Suppose that
$(*)$(i)
: $p,q \in P$,
(ii)
: $M^p=M^q$, $m^p=m^q$, $h^p=h^q$, $u^p=u^q$, $K^p=K^q$, $v^p=v^q$, $t^p_{n,i}=t^q_{n,i}$,
(iii)
: $\langle\eta^p_\alpha:\alpha\in w^p\cap w^q\rangle =
\langle\eta^q_\alpha:\alpha\in w^p\cap w^q\rangle$,
(iv)
: $g^p\restriction (w^p \cap w^q)=g^q \restriction(w^p
\cap w^q)$.
Then the conditions $p,q$ are compatible (in fact have an upper bound with the same common parts): take the common values (in (ii)) or the union (for (iii)).
[**C Fact**]{} For each $\alpha<\lambda$ the set ${{\cal}I}_\alpha=:\{p\in P:\alpha\in w^p\}$ is dense (and open).
[*Proof of the Fact:*]{} For $p\in P$ let $q$ be like $p$ except that: $$w^q=w^p\cup\{\alpha\}\quad\mbox{ and }\quad g^q(\beta)=\left\{
\begin{array}{lll}
g^p(\beta) &\mbox{if}& \beta\in w^p \\
m^p &\mbox{if}& \beta=\alpha,\ \beta\notin w^p.
\end{array}\right.$$
[**D Fact**]{} For $n<\omega$, $i<\omega$ the following set is a dense subset of $P$: $${{\cal}J}^*_{(n,i)}=\{p\in P:x^n_i\in K^p\ \&\ (\forall n<m^p)(\{n\}\times
\omega)\cap u^p \mbox{ has }>m^p\mbox{ elements}\}.$$
[*Proof of the Fact:*]{} Should be clear.
[**E Fact**]{} For each $m<\omega$ the set ${{\cal}J}_m
=:\{p\in P:m^p\ge m\}$ is dense in $P$.
[*Proof of the Fact:*]{} Let $p\in P$ be given such that $m^p
<m$. Let $w^p=\{\alpha_0,\ldots,\alpha_{r-1}\}$ be without repetitions; we know that in $G$, $pz^0_{\alpha_\ell}=0$ and $\{z^0_{\alpha_\ell}:
\ell<r\}$ is independent $\mod\ B$, hence also in $K+G$ the set $\{z^0_{\alpha_\ell}:\ell<r\}$ is independent $\mod\ K$. Clearly
(A)
: $pz^{k+1}_{\alpha_\ell}=z^k_{\alpha_\ell}\mod\ K$ for $k\in [g(\alpha_\ell),m^p)$, hence
(B)
: $p^{m^p}z^{m^p}_{\alpha_\ell}=z^{g(\alpha_\ell)}_{\alpha_\ell}
\mod\ K$.
Remember
(C)
: $z^{m^p}_{\alpha_\ell}=\sum\{a^{k,\alpha_\ell}_i p^{k-m^p} x^k_i:
k \ge m^p,i\in w(\alpha_\ell,k)\}$,
and so, in particular, (from the choice of $z^0_{\alpha_\ell}$) $$p^{m^p+1}z^{m^p}_{\alpha_\ell}=0\quad\mbox{ and }\quad
p^{m^p}z^{m^p}_{\alpha_\ell}\ne 0.$$ For $\ell<r$ and $n\in [m^p,\omega)$ let $$s^n_\ell=:\sum\big\{a^{k,\alpha_\ell}_i p^{k-m^p} x^k_i:k \ge m^p
\mbox{ but }k<n\mbox{ and } i\in w(\alpha_\ell,k)\big\}.$$ But $p^{k-m^p}x^k_i = y^{k,m^p}_i$, so $$s^n_\ell=\sum\big\{a^{k,\alpha_\ell}_i y^{k,m^p}_i:k\in [m^p,n)
\mbox{ and }i\in (\alpha_\ell,k)\big\}.$$ Hence, for some $m^*>m,m^p$ we have: $\{p^m\,s^{m^*}_\ell:\ell<r\}$ is independent in $G[p]$ over $K[p]$ and therefore in $\langle x^k_i:k\in
[m^p,m^*],i<\omega\rangle$. Let $$s^*_\ell=\sum\big\{a^{k,\alpha_\ell}_i:k\in [m^p,m^*)\mbox{ and }i\in
w(\alpha_\ell,k)\}.$$ Then $\{ s^*_\ell:\ell<r\}$ is independent in $$B^+_{[m,m^*)}=\langle y^{l,m^*-1}_i:k\in [m^p,m^*)\mbox{ and }i<\omega
\rangle.$$ Let $i^*<\omega$ be such that: $w(\alpha_\ell,k)\subseteq\{0,\ldots,i^*-1\}$ for $k\in [m^p,m^*)$, $\ell=1,\ldots,r$. Let us start to define $q$: $$\begin{array}{c}
m^q=m^*,\quad M^q=M^p\cup\{m^*-1\},\quad w^q=w^p,\quad g^q=g^p,\\
u^q=u^p\cup ([m^p,m^*)\times\{0,\ldots,i^*-1\}),\\
h^q\mbox{ is } h^p\mbox{ on } u^p\mbox{ and }h^q(k,i)=m^*-1\mbox{
otherwise},\\
K^q\mbox{ is defined appropriately, let } K'=\langle x^n_i:n\in [m^p,m^*),
i<i^*\rangle.
\end{array}$$ Complete $\{s^*_\ell:\ell<r\}$ to $\{s^*_\ell:\ell<r^*\}$, a basis of $K'[p]$, and choose $\{t_{n,i}:(n,i)\in v^*\}$ such that: $[p^mt_{n,i}=0\ \
\Leftrightarrow\ \ m>n]$, and for $\ell<r$ $$p^{m^*-1-\ell}t_{m^*-1,\ell} = s^*_\ell.$$ The rest should be clear.
The generic gives a variant of the desired result: almost tree-like basis; the restriction to $M$ and $g$ but by \[4.3\] we can finish. $\square_{\ref{4.2}}$
\[MA$_\lambda$($\sigma$-centered)\] \[4.4\] For $(*)_0$ to hold it suffices that $(*)_1$ holds where
$(*)_0$
: in ${\frak K}^{rs(p)}_\lambda$, there is a universal member,
$(*)_1$
: in ${\frak K}^{tr}_{\bar\lambda}$ there is a universal member, where:
(a)
: $\lambda_n=\aleph_0$, $\lambda_\omega=\lambda$, $\ell g(\bar\lambda)
=\omega+1$
(b)
: $\lambda_\omega=\lambda$, $\lambda_n\in [n,\omega)$, $\ell g
(\bar \lambda)=\omega+1$.
\[4.4A\] Any $\langle\lambda_n:n<\omega\rangle$, $\lambda_n<\omega$ which is not bounded suffices.
For case (a) - by \[4.1\].
For case (b) - the same proof. $\square_{\ref{4.4}}$
\[4.5\] Assume $\lambda<2^{\aleph_0}$ and
(a)
: there are $A_i\subseteq\lambda$, $|A_i|=\lambda$ for $i<2^\lambda$ such that $i\ne j \Rightarrow |A_i \cap A_j| \le \aleph_0$.
Let $\bar\lambda=\langle \lambda_\alpha:\alpha\le\omega\rangle$, $\lambda_n
= \aleph_0$, $\lambda_\omega=\lambda$.
[*Then*]{} there is $P$ such that:
$(\alpha)$
: $P$ is a c.c.c. forcing notion,
$(\beta)$
: $|P|=2^\lambda$,
$(\gamma)$
: in $V^P$, there is $T\in {\frak K}^{tr}_{\bar\lambda}$ into which every $T' \in ({\frak K}^{tr}_{\bar \lambda})^V$ can be embedded.
Let $\bar T=\langle T_i:i<2^\lambda\rangle$ list the trees $T$ of cardinality $\le\lambda$ satisfying $${}^{\omega >}\omega\subseteq T \subseteq {}^{\omega \ge} \omega\quad
\mbox{ and }\quad T\cap {}^\omega\omega\mbox{ has cardinality $\lambda$, for
simplicity.}$$ Let $T_i\cap {}^\omega\omega=\{\eta^i_\alpha:\alpha\in A_i \}$.
We shall force $\rho_{\alpha,\ell}\in {}^\omega\omega$ for $\alpha<
\lambda$, $\ell<\omega$, and for each $i<2^\lambda$ a function $g_i:A_i
\longrightarrow\omega$ such that: there is an automorphism $f_i$ of $({}^{\omega>}\omega,\triangleleft)$ which induces an embedding of $T_i$ into $\left(({}^{\omega>}\omega)\cup \{\rho_{\alpha,g_i(\alpha)}:\alpha<
\lambda\},\triangleleft\right)$. We shall define $p\in P$ as an approximation.
A condition $p\in P$ consists of:
(a)
: $m<\omega$ and a finite subset $u$ of ${}^{m \ge}\omega$, closed under initial segments such that $\langle\rangle\in u$,
(b)
: a finite $w\subseteq 2^\lambda$,
(c)
: for each $i\in w$, a finite function $g_i$ from $A_i$ to $\omega$,
(d)
: for each $i\in w$, an automorphism $f_i$ of $(u,\triangleleft)$,
(e)
: a finite $v\subseteq\lambda\times\omega$,
(f)
: for $(\alpha,n)\in v$, $\rho_{\alpha,n}\in u\cap
({}^m\omega)$,
such that
(g)
: if $i\in w$ and $\alpha\in{{\rm Dom}}(g_i)$ then:
$(\alpha)$
: $(\alpha,g_i(\alpha))\in v$,
$(\beta)$
: $\eta^i_\alpha\restriction m\in u$,
$(\gamma)$
: $f_i(\eta^i_\alpha\restriction m)=\rho_{\alpha,
g_i(\alpha)}$,
(h)
: $\langle\rho_{\alpha,n}:(\alpha,n)\in v\rangle$ is with no repetition (all of length $m$),
(i)
: for $i\in w$, $\langle\eta^i_\alpha\restriction m:\alpha\in
{{\rm Dom}}(g_i)\rangle$ is with no repetition.
The order on $P$ is: $p \le q$ if and only if:
$(\alpha)$
: $u^p \subseteq u^q$, $m^p\le m^q$,
$(\beta)$
: $w^p \subseteq w^q$,
$(\gamma)$
: $f^p_i \subseteq f^q_i$ for $i\in w^p$,
$(\delta)$
: $g^p_i \subseteq g^q_i$ for $i\in w^p$,
$(\varepsilon)$
: $v^p \subseteq v^q$,
$(\zeta)$
: $\rho^p_{\alpha,n}\trianglelefteq\rho^q_{\alpha,n}$, when $(\alpha,n) \in v^p$,
$(\eta)$
: if $i\ne j\in w^p$ then for every $\alpha\in A_i\cap
A_j\setminus ({{\rm Dom}}(g^p_i)\cap {{\rm Dom}}(g^p_j))$ we have $g^q_i(\alpha)\ne
g^q_j(\alpha)$.
[**A Fact**]{} $(P,\le)$ is a partial order.
[*Proof of the Fact:*]{} Trivial.
[**B Fact**]{} For $i<2^\lambda$ the set $\{p:i\in
w^p\}$ is dense in $P$.
[*Proof of the Fact:*]{} If $p\in P$, $i\in 2^\lambda
\setminus w^p$, define $q$ like $p$ except $w^q=w^p\cup\{i\}$, ${{\rm Dom}}(g^q_i)=\emptyset$.
[**C Fact**]{} If $p\in P$, $m_1\in
(m^p,\omega)$, $\eta^*\in u^p$, $m^*<\omega$, $i\in w^p$, $\alpha\in\lambda
\setminus{{\rm Dom}}(g^p_i)$ [*then*]{} we can find $q$ such that $p\le q\in
P$, $m^q>m_1$, $\eta^* \char 94\langle m^*\rangle\in u^q$ and $\alpha\in
{{\rm Dom}}(g_i)$ and $\langle\eta^j_\beta\restriction m^q:j\in w^q$ and $\beta\in
{{\rm Dom}}(g^q_j)\rangle$ is with no repetition, more exactly $\eta^{j(1)}_{\beta_1}\setminus m^q= \eta^{j(2)}_{\beta_2}\restriction
m^q \Rightarrow \eta^{j(1)}_{\beta_1}=\eta^{j(2)}_{\beta_2}$.
[*Proof of the Fact:*]{} Let $n_0\le m^p$ be maximal such that $\eta^i_\alpha\restriction n_0 \in u^p$. Let $n_1<\omega$ be minimal such that $\eta^i_\alpha\restriction n_1\notin\{\eta^i_\beta\restriction n_1:
\beta\in{{\rm Dom}}(g^p_i)\}$ and moreover the sequence $$\langle\eta^j_\beta\restriction n_1:j\in w^p\ \&\ \beta\in{{\rm Dom}}(g^p_j)\
\ \mbox{ or }\ \ j=i\ \&\ \beta=\alpha\rangle$$ is with no repetition. Choose a natural number $m^q>m^p+1,n_0+1,n_1+2$ and let $k^*=:3+\sum\limits_{i\in w^p}|{{\rm Dom}}(g^p_i)|$. Choose $u^q\subseteq
{}^{m^q\ge}\omega$ such that:
(i)
: $u^p\subseteq u^q\subseteq {}^{m^q\ge}\omega$, $u^q$ is downward closed,
(ii)
: for every $\eta\in u^q$ such that $\ell g(\eta)<m^q$, for exactly $k^*$ numbers $k$, $\eta\char 94\langle k\rangle\in u^q\setminus
u^p$,
(iii)
: $\eta^j_\beta\restriction\ell\in u^q$ when $\ell\le m^q$ and $j\in w^p$, $\beta\in{{\rm Dom}}(g^p_j)$,
(iv)
: $\eta^i_\alpha\restriction\ell\in u^q$ for $\ell\le m^q$,
(v)
: $\eta^*\char 94\langle m^* \rangle\in u^q$.
Next choose $\rho^q_{\beta,n}$ (for pairs $(\beta,n)\in v^p)$ such that: $$\rho^p_{\beta,n}\trianglelefteq\rho^q_{\beta,n}\in u^q\cap {}^{m^q}
\omega.$$ For each $j\in w^p$ separately extend $f^p_j$ to an automorphism $f^q_j$ of $(u^q,\triangleleft)$ such that for each $\beta\in{{\rm Dom}}(g^p_j)$ we have: $$f^q_j(\eta^j_\beta\restriction m^q)=\rho^q_{\beta,g_j}(\beta).$$ This is possible, as for each $\nu\in u^p$, and $j\in w^p$, we can separately define $$f^q_j\restriction\{\nu':\nu\triangleleft\nu'\in u^q\ \mbox{ and }\ \nu'
\restriction (\ell g(\nu)+1)\notin u^p\}$$ –its range is $$\{\nu':f^p_j(\nu)\triangleleft \nu'\in u^q\ \mbox{ and }\ \nu'
\restriction (\ell g(\nu)+1)\notin u^p\}.$$ The point is: by Clause (ii) above those two sets are isomorphic and for each $\nu$ at most one $\rho^p_{\beta,n}$ is involved (see Clause (h) in the definition of $p \in P$). Next let $w^q=w^p$, $g^q_j=g^p_j$ for $j\in w
\setminus\{i\}$, $g^q_i\restriction{{\rm Dom}}(g^p_i)=g^p_i$, $g^q_i(\alpha)=
\min(\{n:(\alpha,n)\notin v^p\})$, ${{\rm Dom}}(g^q_i)={{\rm Dom}}(g^p_i)\cup\{\alpha\}$, and $\rho^q_{\alpha,g^q_i(\alpha)}=f^g_i(\eta^i_\alpha\restriction m^q)$ and $v^q=v^p\cup\{(\alpha,g^q_i(\alpha))\}$.
[**D Fact**]{} $P$ satisfies the c.c.c.
[*Proof of the Fact:*]{} Assume $p_\varepsilon\in P$ for $\varepsilon<\omega_1$. By Fact C, without loss of generality each $$\langle\eta^j_\beta\restriction m^{p_\varepsilon}:j\in w^{p_\varepsilon}
\mbox{ and }\beta\in{{\rm Dom}}(g^{p_\varepsilon}_j)\rangle$$ is with no repetition. Without loss of generality, for all $\varepsilon
<\omega_1$ $$U_\varepsilon=:\big\{\alpha<2^\lambda:\alpha\in w^{p_\varepsilon}\mbox{ or }
\bigvee_{i\in w^p}[\alpha\in{{\rm Dom}}(g_i)]\mbox{ or }\bigvee_k(k,\alpha)\in
v^{p_\varepsilon}\big\}$$ has the same number of elements and for $\varepsilon\ne\zeta<\omega_1$, there is a unique one-to-one order preserving function from $U_\varepsilon$ onto $U_\zeta$ which we call ${{\rm OP}}_{\zeta,\varepsilon}$, which also maps $p_\varepsilon$ to $p_\zeta$ (so $m^{p_\zeta}=m^{p_\varepsilon}$; $u^{p_\zeta}
=u^{p_\varepsilon}$; ${{\rm OP}}_{\zeta,\varepsilon}(w^{p_\varepsilon})=
w^{p_\zeta}$; if $i\in w^{p_\varepsilon}$, $j={{\rm OP}}_{\zeta,\varepsilon}(i)$, then $f_i\circ{{\rm OP}}_{\varepsilon,\zeta}\equiv f_j$; and [*if*]{} $\beta=
{{\rm OP}}_{\zeta,\varepsilon}(\alpha)$ and $\ell<\omega$ [*then*]{} $$(\alpha,\ell)\in v^{p_\varepsilon}\quad\Leftrightarrow\quad (\beta,\ell)
\in v^{p_\zeta}\quad\Rightarrow\quad\rho^{p_\varepsilon}_{\alpha,\ell}=
\rho^{p_\zeta}_{\beta,\ell}).$$ Also this mapping is the identity on $U_\zeta\cap U_\varepsilon$ and $\langle U_\zeta:\zeta<\omega_1\rangle$ is a $\triangle$-system.
Let $w=:w^{p_0}\cap w^{p_1}$. As $i\ne j\ \Rightarrow\ |A_i\cap A_j|\le
\aleph_0$, without loss of generality
$(*)$
: if $i\ne j\in w$ then $$U_{\varepsilon}\cap (A_i\cap A_j)\subseteq w.$$
We now start to define $q\ge p_0,p_1$. Choose $m^q$ such that $m^q\in
(m^{p_\varepsilon},\omega)$ and $$\begin{array}{ll}
m^q>\max\big\{\ell g(\eta^{i_0}_{\alpha_0}\cap\eta^{i_1}_{\alpha_1})+1:&
i_0\in w^{p_0},\ i_1 \in w^{p_1},\ {{\rm OP}}_{1,0}(i_0)=i_1,\\
\ &\alpha_0\in{{\rm Dom}}(g^{p_0}_{i_0}),\ \alpha_1\in{{\rm Dom}}(g^{p_1}_{i_1}),\\
\ &{{\rm OP}}_{1,0}(\alpha_0)=\alpha_1\big\}.
\end{array}$$ Let $u^q\subseteq {}^{m^q\ge}\omega$ be such that:
(A)
: $u^q\cap\left({}^{m^{p_0}\ge}\omega\right)=u^q\cap\left(
{}^{m^{p_1}\ge}\omega\right)=u^{p_0}=u^{p_1}$,
(B)
: for each $\nu\in u^q$, $m^{p_0}\le\ell g(\nu)<m^q$, for exactly two numbers $k<\omega$, $\nu\char 94 \langle k\rangle\in u^q$,
(C)
: $\eta^i_\alpha\restriction\ell\in u^q$ for $\ell\le m^q$ : $i\in w^{p_0}$, $\alpha\in{{\rm Dom}}(g^{p_0}_i)$ $i\in w^{p_1}$, $\alpha\in{{\rm Dom}}(g^{p_1}_i)$.
\[Possible as $\{\eta^i_\alpha\restriction m^{p_\varepsilon}:i\in
w^{p_\varepsilon},\alpha\in{{\rm Dom}}(g^{p_\varepsilon}_i)\}$ is with no repetitions (the first line of the proof).\]
Let $w^q=:w^{p_0}\cup w^{p_1}$ and $v^q=:v^{p_0}\cup v^{p_1}$ and for $i \in w^q$ $$g^q_i=\left\{\begin{array}{lll}
g^{p_0}_i &\mbox{\underbar{if}}& i\in w^{p_0}\setminus w^{p_1},\\
g^{p_1}_i &\mbox{\underbar{if}}& i\in w^{p_1}\setminus w^{p_0},\\
g^{p_0}_i \cup g^{p_1}_i &\mbox{\underbar{if}}& i\in w^{p_0}\cap w^{p_1}.
\end{array}\right.$$ Next choose $\rho^q_{\alpha,\ell}$ for $(\alpha,\ell)\in v^q$ as follows. Let $\nu_{\alpha,\ell}$ be $\rho^{p_0}_{\alpha,\ell}$ if defined, $\rho^{p_1}_{\alpha,\ell}$ if defined (no contradiction). If $(\alpha,\ell)
\in v^q$ choose $\rho^q_{\alpha,\ell}$ as any $\rho$ such that:
$\otimes_0$
: $\nu_{\alpha,\ell}\triangleleft\rho\in u^q\cap
{}^{(m^q)}\omega$.
But not all choices are O.K., as we need to be able to define $f^q_i$ for $i\in w^q$. A possible problem will arise only when $i\in w^{p_0}\cap
w^{p_1}$. Specifically we need just (remember that $\langle
\rho^{p_\varepsilon}_{\alpha,\ell}:(\alpha,\ell)\in v^{p_\varepsilon}
\rangle$ are pairwise distinct by clause (b) of the Definition of $p\in P$):
$\otimes_1$
: if $i_0\in w^{p_0},(\alpha_0,\ell)=(\alpha_0,
g_{i_0}(\alpha_0)),\alpha_0\in{{\rm Dom}}(g^{p_0}_{i_0})$, $i_1={{\rm OP}}_{1,0}(i_0)$ and $\alpha_1={{\rm OP}}_{1,0}(\alpha_0)$ and $i_0=i_1$
[*then*]{} $\ell g(\eta^{i_0}_{\alpha_0}\cap\eta^{i_1}_{\alpha_1})=\ell g
(\rho^q_{\alpha_0,\ell}\cap\rho^q_{\alpha_1,\ell})$.
We can, of course, demand $\alpha_0\neq \alpha_1$ (otherwise the conclusion of $\otimes_1$ is trivial). Our problem is expressible for each pair $(\alpha_0,\ell),(\alpha_1,\ell)$ separately as: first the problem is in defining the $\rho^q_{(\alpha,\ell)}$’s and second, if $(\alpha'_1,\ell')$, $(\alpha'_2,\ell)$ is another such pair then $\{(\alpha_1,\ell),
(\alpha_2,\ell)\}$, $\{(\alpha'_1,\ell'),(\alpha'_2,\ell')\}$ are either disjoint or equal. Now for a given pair $(\alpha_0,\ell),(\alpha_1,\ell)$ how many $i_0=i_1$ do we have? Necessarily $i_0\in w^{p_0}\cap w^{p_1}=w$. But if $i'_0\ne i''_0$ are like that then $\alpha_0\in A_{i'_0}\cap
A_{i''_0}$, contradicting $(*)$ above because $\alpha_0\neq
\alpha_1={{\rm OP}}_{1,0}(\alpha_0)$. So there is at most one candidate $i_0=i_1$, so there is no problem to satisfy $\otimes_1$. Now we can define $f^q_i$ (i$\in w^q$) as in the proof of Fact C.
The rest should be clear. $\square_{\ref{4.4}}$
\[4.6\] Suppose $V\models GCH$, $\aleph_0<\lambda<\chi$ and $\chi^\lambda=\chi$. Then for some c.c.c. forcing notion $P$ of cardinality $\chi$, not collapsing cardinals nor changing cofinalities, in $V^P$:
(i)
: $2^{\aleph_0}=2^\lambda=\chi$,
(ii)
: ${\frak K}^{tr}_\lambda$ has a universal family of cardinality $\lambda^+$,
(iii)
: ${\frak K}^{rs(p)}_\lambda$ has a universal family of cardinality $\lambda^+$.
First use a preliminary forcing $Q^0$ of Baumgartner [@B], adding $\langle A_\alpha:\alpha<\chi\rangle$, $A_\alpha\in
[\lambda]^\lambda$, $\alpha\neq\beta\quad \Rightarrow\quad |A_\alpha\cap
A_\beta|\le\aleph_0$ (we can have $2^{\aleph_0}=\aleph_1$ here, or $[\alpha\ne \beta\quad \Rightarrow\quad A_\alpha\cap A_\beta$ finite\], but not both). Next use an FS iteration $\langle P_i,\dot{Q}_i:i<\chi\times
\lambda^+\rangle$ such that each forcing from \[4.4\] appears and each forcing as in \[4.5\] appears. $\square_{\ref{4.6}}$
\[4.7\] We would like to have that there is a universal member in ${\frak K}^{rs(p)}_\lambda$; this sounds very reasonable but we did not try.
In our framework, the present result shows limitations to ZFC results which the methods applied in the previous sections can give.
Back to ${\frak K}^{rs(p)}$, real non-existence results
=======================================================
By §1 we know that if $G$ is an Abelian group with set of elements $\lambda$, $C\subseteq\lambda$, then for an element $x\in G$ the distance from $\{y:y<\alpha\}$ for $\alpha\in C$ does not code an appropriate invariant. If we have infinitely many such distance functions, e.g. have infinitely many primes, we can use more complicated invariants related to $x$ as in §3. But if we have one prime, this approach does not help.
If one element fails, can we use infinitely many? A countable subset $X$ of $G$ can code a countable subset of $C$: $$\{\alpha\in C:\mbox{ closure}(\langle X\rangle_G)\cap\alpha\nsubseteq
\sup(C\cap\alpha)\},$$ but this seems silly - we use heavily the fact that $C$ has many countable subsets (in particular $>\lambda$) and $\lambda$ has at least as many. However, what if $C$ has a small family (say of cardinality $\le\lambda$ or $<\mu^{\aleph_0}$) of countable subsets such that every subset of cardinality, say continuum, contains one? Well, we need more: we catch a countable subset for which the invariant defined above is infinite (necessarily it is at most of cardinality $2^{\aleph_0}$, and because of §4 we are not trying any more to deal with $\lambda\le 2^{\aleph_0}$). The set theory needed is expressed by $T_J$ below, and various ideals also defined below, and the result itself is \[5.7\].
Of course, we can deal with other classes like torsion free reduced groups, as they have the characteristic non-structure property of unsuperstable first order theories; but the relevant ideals will vary: the parallel to $I^0_{\bar\mu}$ for $\bigwedge\limits_n \mu_n=\mu$, $J^2_{\bar\mu}$ seems to be always O.K.
\[5.1\]
1. For $\bar\mu=\langle\mu_n:n<\omega\rangle$ let $B_{\bar\mu}$ be $$\bigoplus\{K^n_\alpha:n<\omega,\alpha<\mu_n\},\qquad K^n_\alpha=
\langle{}^* t^n_\alpha\rangle_{K^n_\alpha}\cong {\Bbb Z}/p^{n+1} {\Bbb Z}.$$ Let $B_{\bar\mu\restriction n}=\bigoplus\{K^m_\alpha:\alpha<\mu_m,m<n\}
\subseteq B_{\bar\mu}$ (they are in ${\frak K}^{rs(p)}_{\le\sum\limits_{n}
\mu_n}$). Let $\hat B$ be the $p$-torsion completion of $B$ (i.e. completion under the norm $\|x\|=\min\{2^{-n}:p^n\mbox{ divides }x\}$).
2. Let $I^1_{\bar \mu}$ be the ideal on $\hat B_{\bar \mu}$ generated by $I^0_{\bar \mu}$, where $$\begin{array}{ll}
I^0_{\bar\mu}=\big\{A\subseteq\hat B_{\bar\mu}:&\mbox{for every large
enough }n,\\
&\mbox{for no }y\in\bigoplus\{K^m_\alpha:m\le n\mbox{ and }\alpha<\mu_m\}\\
&\mbox{but }y\notin\bigoplus\{K^m_\alpha:m<n\mbox{ and }\alpha<\mu_m\}
\mbox{ we have}:\\
&\mbox{for every }m\mbox{ for some }z\in\langle A\rangle\mbox{ we have:}\\
&p^m\mbox{ divides }z-y\big\}.
\end{array}$$ (We may write $I^0_{\hat B_{\bar\mu}}$, but the ideal depends also on $\langle\bigoplus\limits_{\alpha<\mu_n} K^n_\alpha:n<\omega\rangle$ not just on $\hat B_{\bar\mu}$ itself).
3. For $X,A\subseteq\hat B_{\bar\mu}$, $$\mbox{ recall }\ \ \langle A\rangle_{\bar
B_{\bar\mu}}=\big\{\sum\limits_{n<n^*} a_ny_n:y_n
\in A,\ a_n\in{\Bbb Z}\mbox{ and } n^*\in{\Bbb N}\big\},$$ $$\mbox{ and let }\ \ c\ell_{\hat B_{\bar\mu}}(X)=\{x:(\forall
n)(\exists y\in X)(x-y\in p^n \hat B_{\bar \mu})\}.$$
4. Let $J^1_{\bar \mu}$ be the ideal which $J^{0.5}_{\bar \mu}$ generates, where $$\begin{array}{ll}
J^{0.5}_{\bar\mu}=\big\{A\subseteq\prod\limits_{n<\omega}\mu_n:
&\mbox{for some }n<\omega\mbox{ for no }m\in [n,\omega)\\
&\mbox{and }\beta<\gamma<\mu_m\mbox{ do we have}:\\
&\mbox{for every }k\in [m,\omega)\mbox{ there are }\eta,\nu\in A\mbox{
such}\\
&\mbox{that:}\ \ \eta(m)=\beta,\,\nu(m)=\gamma,\,\eta\restriction m=
\nu\restriction m\\
&\mbox{and }\eta\restriction (m,k)=\nu\restriction (m,k)\big\}.
\end{array}$$
5. $$\begin{array}{rr}
J^0_{\bar\mu}=\{A\subseteq\prod\limits_{n<\omega}\mu_n:
&\mbox{for some }n<\omega\mbox{ and } k,\mbox{ the mapping }\eta\mapsto
\eta \restriction n\\
&\mbox{is }(\le k)\mbox{-to-one }\}.
\end{array}$$
6. $J^2_{\bar\mu}$ is the ideal of nowhere dense subsets of $\prod\limits_n\mu_n$ (under the following natural topology: a neighbourhood of $\eta$ is $U_{\eta,n}=\{\nu:\nu\restriction n=\eta
\restriction n\}$ for some $n$).
7. $J^3_{\bar \mu}$ is the ideal of meagre subsets of $\prod\limits_n
\mu_n$, i.e. subsets which are included in countable union of members of $J^2_{\bar \mu}$.
\[5.2\]
1. $I^0_{\bar\mu}$, $J^0_{\bar\mu}$, $J^{0.5}_{\bar\mu}$ are $(<\aleph_1)$-based, i.e. for $I^0_{\bar \mu}$: if $A\subseteq\hat
B_{\bar\mu}$, $A\notin I^0_{\bar\mu}$ then there is a countable $A_0
\subseteq A$ such that $A_0\notin I^0_{\bar\mu}$.
2. $I^1_{\bar\mu}$, $J^0_{\bar\mu}$, $J^1_{\bar\mu}$, $J^2_{\bar\mu}$, $J^3_{\bar\mu}$ are ideals, $J^3_{\bar\mu}$ is $\aleph_1$-complete.
3. $J^0_{\bar\mu}\subseteq J^1_{\bar\mu}\subseteq J^2_{\bar\mu}
\subseteq J^3_{\bar\mu}$.
4. There is a function $g$ from $\prod\limits_{n<\omega}\mu_n$ into $\hat B_{\bar\mu}$ such that for every $X\subseteq\prod\limits_{n<\omega}
\mu_n$: $$X\notin J^1_{\bar\mu}\quad \Rightarrow\quad g''(X)\notin
I^1_{\bar\mu}.$$
E.g. 4) Let $g(\eta)=\sum\limits_{n<\omega}p^n({}^*t^n_{\eta(n)})$.
Let $X \subseteq\prod\limits_{n<\omega}\mu_n$, $X\notin J^1_{\bar\mu}$. Assume $g''(X)\in\bar I^1_{\bar\mu}$, so for some $\ell^*$ and $A_\ell
\subseteq\hat B_{\bar\mu}$, ($\ell<\ell^*$) we have $A_\ell\in I^0_{\bar\mu}$, and $g''(X)\subseteq\bigcup\limits_{\ell<\ell^*} A_\ell$, so $X=
\bigcup\limits_{\ell<\ell^*} X_\ell$, where $$X_\ell=:\{\eta\in X:g(\eta)\in A_\ell\}.$$ As $J^1_{\bar \mu}$ is an ideal, for some $\ell<\ell^*$, $X_\ell\notin
J^1_{\bar\mu}$. So by the definition of $J^1_{\bar\mu}$, for some infinite $\Gamma\subseteq\omega$ for each $m\in\Gamma$ we have $\beta_m<\gamma_m<
\mu_m$ and for every $k\in [m,\omega)$ we have $\eta_{m,k},\nu_{m,k}$, as required in the definition of $J^1_{\bar \mu}$. So $g(\eta_{m,k}),
g(\nu_{m,k}) \in A_\ell$ (for $m\in \Gamma$, $k\in (m,\omega)$). Now $${}^* t^m_{\gamma_m} - {}^* t^m_{\beta_m}=g(\eta_{m,k})-g(\nu_{m,k})
\mod\ p^k \hat B_{\bar \mu},$$ but $g(\eta_{m,k})-g(\nu_{m,k})\in\langle A_\ell\rangle_{\hat B_{\bar\mu}}$. Hence $$(\exists z\in\langle A_\ell\rangle_{\hat B_{\bar\mu}})[{}^* t^m_{\gamma_m}
-{}^* t^m_{\beta_m}=z\mod\ p^k \hat B_{\bar\mu}],$$ as this holds for each $k$, ${}^* t^m_{\gamma_m}-{}^* t^m_{\beta_m}\in
c \ell(\langle A_\ell\rangle_{\hat B_{\bar\mu}})$.
This contradicts $A_\ell
\in I^0_{\bar \mu}$. $\square_{\ref{5.2}}$
\[5.3\] Let $I\subseteq{{\cal}P}(X)$ be downward closed (and for simplicity $\{\{x\}:
x\in X\}\subseteq I$). Let $I^+={{\cal}P}(X)\setminus I$. Let $$\begin{array}{ll}
{{\bf U}}^{<\kappa}_I(\mu)=:\min\big\{|{{\cal}P}|:&{{\cal}P}\subseteq [\mu]^{<\kappa},
\mbox{ and for every } f:X\longrightarrow\mu\mbox{ for some}\\
&Y\in {{\cal}P},\mbox{ we have }\{x\in X:f(x)\in Y\}\in
I^+\big\}.
\end{array}$$ Instead of $<\kappa^+$ in the superscript of ${{\bf U}}$ we write $\kappa$. If $\kappa>|{{\rm Dom}}(I)|^+$, we omit it (since then its value does not matter).
\[5.4\]
1. If $2^{<\kappa}+|{{\rm Dom}}(I)|^{<\kappa}\le\mu$ we can find $F\subseteq$ partial functions from ${{\rm Dom}}(I)$ to $\mu$ such that:
(a)
: $|F|={{\bf U}}^{<\kappa}_I(\mu)$,
(b)
: $(\forall f:X\longrightarrow\mu)(\exists Y\in I^+)[f\restriction
Y \in F]$.
2. Such functions (as ${{\bf U}}^{<\kappa}_I(\mu)$) are investigated in [**pcf**]{} theory ([@Sh:g], [@Sh:410 §6], [@Sh:430 §2], [@Sh:513]).
3. If $I\subseteq J\subseteq {{\cal}P}(X)$, then ${{\bf U}}^{<\kappa}_I(\mu)\le
{{\bf U}}^{<\kappa}_J(\mu)$, hence by \[5.2\](3), and the above $${{\bf U}}^{<\kappa}_{J^0_{\bar\mu}}(\mu)\le {{\bf U}}^{<\kappa}_{J^1_{\bar\mu}}(\mu)
\le {{\bf U}}^{<\kappa}_{J^2_{\bar\mu}}(\mu)\le
{{\bf U}}^{<\kappa}_{J^3_{\bar\mu}}(\mu)$$ and by \[5.2\](4) we have ${{\bf U}}^{<\kappa}_{I^1_{\bar \mu}}\leq
{{\bf U}}^{<\kappa}_{J^1_{\bar \mu}}(\mu).$
4. On ${{\rm IND}}_\theta(\bar\kappa)$ (see \[5.5A\] below) see [@Sh:513].
\[5.5A\] ${{\rm IND}}'_\theta(\langle\kappa_n:n<\omega\rangle)$ means that for every model $M$ with universe $\bigcup\limits_{n<\omega}\kappa_n$ and $\le\theta$ functions, for some $\Gamma\in [\omega]^{\aleph_0}$ and $\eta\in
\prod\limits_{n<\omega}\kappa_n$ we have: $$n\in\Gamma\quad\Rightarrow\quad\eta(n)\notin c\ell_M\{\eta(\ell):\ell
\ne n\}.$$
Actually if $\theta\geq \aleph_0$, this implies that we can fix $\Gamma$, hence replacing $\langle \kappa_n: n< \omega\rangle$ by an infinite subsequence we can have $\Gamma=\omega$.
\[5.5\]
1. If $\mu_n\rightarrow (\kappa_n)^2_{2^\theta}$ and ${{\rm IND}}'_\theta(\langle
\kappa_n:n<\omega\rangle)$ [*then*]{} $\prod\limits_{n<\omega}\mu_n$ is not the union of $\le\theta$ sets from $J^1_{\bar \mu}$.
2. If $\theta=\theta^{\aleph_0}$ and $\neg{{\rm IND}}'_\theta(\langle\mu_n:
n<\omega\rangle$) then $\prod\limits_{n<\omega}\mu_n$ is the union of $\le\theta$ members of $J^1_{\bar\mu}$.
3. If $\lim\sup\limits_n \mu_n$ is $\ge 2$, then $\prod\limits_{n<\omega}
\mu_n\notin J^3_{\bar\mu}$ (so also the other ideals defined above are not trivial by \[5.2\](3), (4)).
1) Suppose $\prod\limits_{n<\omega}\mu_n$ is $\bigcup\limits_{i<\theta} X_i$, and each $X_i\in J^1_{\bar\mu}$. We define for each $i<\theta$ and $n<k<\omega$ a two-place relation $R^{n,k}_i$ on $\mu_n$:
$\beta R^{n,k}_i \gamma$ if and only if
there are $\eta,\nu\in X_i\subseteq\prod\limits_{\ell<k}\mu_\ell$ such that $$\eta\restriction [0,n)=\nu\restriction [0,n)\quad\mbox{and }\
\eta\restriction (n,k)=\nu\restriction (n,k)\quad\mbox{and }\ \eta(n)
=\beta,\ \nu(n)=\gamma.$$ Note that $R^{n,k}_i$ is symmetric and $$n<k_1<k_2\ \&\ \beta R^{n,k_2}_i \gamma\quad \Rightarrow\quad \beta
R^{n,k_1}_i \gamma.$$ As $\mu_n\rightarrow (\kappa_n)^2_{2^\theta}$, we can find $A_n\in
[\mu_n]^{\kappa_n}$ and a truth value ${{\bf t}}^{n,k}_i$ such that for all $\beta<\gamma$ from $A_n$, the truth value of $\beta R^{n,k}_i\gamma$ is ${{\bf t}}^{n,k}_i$. If for some $i$ the set $$\Gamma_i=:\{n<\omega:\mbox{ for every }k\in (n,\omega)\mbox{ we have }
{{\bf t}}^{n,k}_i=\mbox{ true}\}$$ is infinite, we get a contradiction to “$X_i\in J^1_{\bar \mu}$", so for some $n(i)<\omega$ we have $n(i)=\sup(\Gamma_i)$.
For each $n<k<\omega$ and $i<\theta$ we define a partial function $F^{n,k}_i$ from $\prod\limits_{\scriptstyle \ell<k,\atop\scriptstyle\ell
\ne n} A_\ell$ into $A_n$:
> $F(\alpha_0\ldots\alpha_{n-1},\alpha_{n+1},\ldots,\alpha_k)$ is the first $\beta\in A_n$ such that for some $\eta\in X_i$ we have $$\begin{array}{c}
> \eta\restriction [0,n)=\langle\alpha_0,\ldots,\alpha_{n-1}\rangle,\quad
> \eta(n)=\beta,\\
> \eta\restriction (n,k)=\langle\alpha_{n+1},\ldots,\alpha_{k-1}\rangle.
> \end{array}$$
So as ${{\rm IND}}'_\theta(\langle\kappa_n:n<\omega\rangle)$ there is $\eta=\langle
\beta_n:n<\omega\rangle\in\prod\limits_{n<\omega} A_n$ such that for infinitely many $n$, $\beta_n$ is not in the closure of $\{\beta_\ell:\ell
<\omega,\,\ell\ne n\}$ by the $F^{n,k}_i$’s. As $\eta\in\prod\limits_{n<
\omega} A_n\subseteq\prod\limits_{n<\omega}\mu_n=\bigcup\limits_{i<\theta}
X_i$, necessarily for some $i<\theta$, $\eta\in X_i$. Let $n\in(n(i),\omega)$ be such that $\beta_n$ is not in the closure of $\{\beta_\ell:\ell<\omega\mbox{ and }\ell\neq n\}$ and let $k>n$ be such that ${{\bf t}}^{n,k}_i=\mbox{ false}$. Now $\gamma=:
F^{n,k}_i(\beta_0,\ldots,\beta_{n-1},\beta_{n+1},\ldots,\beta_{k-1})$ is well defined $\le\beta_n$ (as $\beta_n$ exemplifies that there is such $\beta$) and is $\ne \beta_n$ (by the choice of $\langle\beta_\ell:\ell<\omega\rangle$), so by the choice of $n(i)$ (so of $n$, $k$ and earlier of ${{\bf t}}^{n, k}_i$ and of $A_n$) we get contradiction to “$\gamma<\beta_n$ are from $A_n$".
2) Let $M$ be an algebra with universe $\sum\limits_{n<\omega}
\mu_n$ and $\le\theta$ functions (say $F^n_i$ for $i<\theta$, $n<\omega$, $F^n_i$ is $n$-place) exemplifying $\neg{{\rm IND}}'_\theta(\langle\mu_n:n<\omega
\rangle)$. Let $$\Gamma=:\{\langle(k_n,i_n):n^*\le n<\omega\rangle:n^*<\omega\mbox{ and }
\bigwedge_n n<k_n<\omega\mbox{ and }i_n<\theta\}.$$ For $\rho=\langle(k_n,i_n):n^*\le n<\omega\rangle\in\Gamma$ let $$\begin{array}{rl}
A_\rho=:&\big\{\eta\in\prod\limits_{n<\omega}\mu_n:\mbox{for every }n
\in [n^*,\omega)\mbox{ we have}\\
&\qquad\eta(n)=F^{k_n-1}_{i_n}\left(\eta(0),\ldots,\eta(n-1),\eta(n+1),
\ldots,\eta(k_n)\right)\big\}.
\end{array}$$ So, by the choice of $M$, $\prod\limits_{n<\omega}\mu_n=\bigcup\limits_{\rho
\in\Gamma} A_\rho$. On the other hand, it is easy to check that $A_\rho\in
J^1_{\bar \mu}$. $\square_{\ref{5.5}}$
\[5.6\] If $\mu=\sum\limits_{n<\omega}\lambda_n$, $\lambda^{\aleph_0}_n<\lambda_{n+1}$ and $\mu<\lambda={{\rm cf}}(\lambda)<\mu^{+\omega}$\
then ${{\bf U}}^{\aleph_0}_{I^0_{\langle\lambda_n:n<\omega\rangle}}(\lambda)=
\lambda$ and even ${{\bf U}}^{\aleph_0}_{J^3_{\langle \lambda_n:n<\omega\rangle}}
(\lambda)=\lambda$.
See [@Sh:410 §6], [@Sh:430 §2], and [@Sh:513] for considerably more.
\[5.7\] Assume $\lambda>2^{\aleph_0}$ and
$(*)$(a)
: $\prod\limits_{n<\omega}\mu_n<\mu$ and $\mu^+<\lambda=
{{\rm cf}}(\lambda)<\mu^{\aleph_0}$,
(b)
: $\hat B_{\bar\mu}\notin I^0_{\bar\mu}$ and $\lim_n\sup\mu_n$ is infinite,
(c)
: ${{\bf U}}^{\aleph_0}_{I^0_{\bar\mu}}(\lambda)=\lambda$ (note $I^0_{\bar \mu}$ is not required to be an ideal).
[*Then*]{} there is no universal member in ${\frak K}^{rs(p)}_\lambda$.
Let $S\subseteq\lambda$, $\bar C=\langle C_\delta:\delta\in S\rangle$ guesses clubs of $\lambda$, chosen as in the proof of \[3.3\] (so $\alpha
\in{{\rm nacc}}(C_\delta)\ \Rightarrow\ {{\rm cf}}(\alpha)>2^{\aleph_0}$). Instead of defining the relevant invariant we prove the theorem directly, but we could define it, somewhat cumbersomely (like [@Sh:e III,§3]).
Assume $H\in {\frak K}^{rs(p)}_\lambda$ is a pretender to universality; without loss of generality with the set of elements of $H$ equal to $\lambda$.
Let $\chi=\beth_7(\lambda)^+$, ${\bar{\frak A}}=\langle {\frak A}_\alpha:
\alpha<\lambda\rangle$ be an increasing continuous sequence of elementary submodels of $({{\cal}H}(\chi),\in,<^*_\chi)$, ${\bar {\frak A}}\restriction
(\alpha+1)\in {\frak A}_{\alpha+1}$, $\|{\frak A}_\alpha\|<\lambda$, ${\frak
A}_\alpha\cap\lambda$ an ordinal, ${\frak A}=\bigcup\limits_{\alpha<\lambda}
{\frak A}_\alpha$ and $\{H,\langle\mu_n:n<\omega\rangle,\mu,\lambda\}\in
{\frak A}_0$, so $B_{\bar \mu},\hat B_{\bar \mu} \in {\frak A}_0$ (where $\bar
\mu=\langle\mu_n:n<\omega\rangle$, of course).
For each $\delta\in S$, let ${{\cal}P}_\delta=:[C_\delta]^{\aleph_0}\cap
{\frak A}$. Choose $A_\delta\subseteq C_\delta$ of order type $\omega$ almost disjoint from each $a\in {{\cal}P}_\delta$, and from $A_{\delta_1}$ for $\delta_1\in\delta\cap S$; its existence should be clear as $\lambda< \mu^{\aleph_0}$. So
$(*)_0$
: every countable $A\in {\frak A}$ is almost disjoint to $A_\delta$.
By \[5.2\](2), $I^0_{\bar\mu}$ is $(<\aleph_1)$-based so by \[5.4\](1) and the assumption (c) we have
$(*)_1$
: for every $f:\hat B_{\bar\mu}\longrightarrow\lambda$ for some countable $Y\subseteq \hat B_{\bar\mu}$, $Y\notin I^0_{\bar\mu}$, we have $f\restriction Y\in {\frak A}$
(remember $(\prod\limits_{n<\omega}\mu_n)^{\aleph_0}=\prod\limits_{n<\omega}
\mu_n$).
Let $B$ be $\bigoplus\{G^n_{\alpha,i}:n<\omega,\alpha<\lambda,\,i<
\sum\limits_{k<\omega}\mu_k\}$, where $$G^n_{\alpha,i}=\langle x^n_{\alpha,i}\rangle_{G^n_{\alpha,i}}\cong{\Bbb
Z}/p^{n+1}{\Bbb Z}.$$ So $B$, $\hat B$, $\langle(n,\alpha,i,x^n_{\alpha,i}):n<\omega,\alpha<\lambda,
i<\sum\limits_{k<\omega}\mu_k\rangle$ are well defined. Let $G$ be the subgroup of $\hat B$ generated by: $$\begin{array}{rl}
B\cup\big\{x\in\hat B: &\mbox{for some }\delta\in S,\, x\mbox{ is in the
closure of }\\
&\bigoplus\{G^n_{\alpha,i}:n<\omega,i<\mu_n,\alpha\mbox{ is the }n\mbox{th
element of } A_\delta\}\big\}.
\end{array}$$ As $\prod\limits_{n<\omega}\mu_n<\mu<\lambda$, clearly $G\in {\frak
K}^{rs(p)}_\lambda$, without loss of generality the set of elements of $G$ is $\lambda$ and let $h:G\longrightarrow H$ be an embedding. Let $$E_0=:\{\delta<\lambda:({\frak A}_\delta,h \restriction \delta,\;G
\restriction\delta)\prec({\frak A},h,G)\},$$ $$E=:\{\delta<\lambda:{{\rm otp}}(E_0\cap\delta)=\delta\}.$$ They are clubs of $\lambda$, so for some $\delta\in S$, $C_\delta\subseteq E$ (and $\delta\in E$ for simplicity). Let $\eta_\delta$ enumerate $A_\delta$ increasingly.
There is a natural embedding $g = g_\delta$ of $B_{\bar \mu}$ into $G$: $$g({}^* t^n_i) = x^n_{\eta_\delta(n),i}.$$ Let $\hat g_\delta$ be the unique extension of $g_\delta$ to an embedding of $\hat B_{\bar\mu}$ into $G$; those embeddings are pure, (in fact $g''_\delta
(\hat B_{\bar\mu})\setminus g''_\delta(B_\mu)\subseteq G\setminus G\cap
{\frak A}_\delta$). So $h\circ\hat g_\delta$ is an embedding of $\hat B_{\bar
\mu}$ into $H$, not necessarily pure but still an embedding, so the distance function can become smaller but not zero and $$h\circ\hat g_\delta(\hat B_{\bar\mu})\setminus h\circ g_\delta(B_\mu)
\subseteq H\setminus {\frak A}_\delta.$$ Remember $\hat B_{\bar\mu}\subseteq {\frak A}_0$ (as it belongs to ${\frak
A}_0$ and has cardinality $\prod\limits_{n<\omega}\mu_n<\lambda$ and $\lambda\cap {\frak A}_0$ is an ordinal). By $(*)_1$ applied to $f=h\circ\hat g$ there is a countable $Y \subseteq \hat B_{\bar \mu}$ such that $Y \notin I^0_{\bar\mu}$ and $f \restriction Y \in {\frak
A}$. But, from $f \restriction Y$ we shall below reconstruct some countable sets not almost disjoint to $A_\delta$, reconstruct meaning in ${\frak A}$, in contradiction to $(*)_0$ above.
As $Y\notin I^0_{\bar \mu}$ we can find an infinite $S^*\subseteq\omega\setminus m^*$ and for $n\in
S^*$, $z_n\in\bigoplus\limits_{\alpha<\mu_n} K^n_\alpha\setminus\{0\}$ and $y^\ell_n\in\hat B_{\bar\mu}$ (for $\ell<\omega$) such that:
$(*)_2$
: $z_n+y_{n,\ell}\in\langle Y\rangle_{\hat B_{\bar\mu}}$,and
$(*)_3$
: $y_{n,\ell}\in p^\ell\,\hat B_{\bar\mu}$.
Without loss of generality $pz_n=0\ne z_n$ hence $p\,y^\ell_n=0$. Let $$\nu_\delta(n)=:\min(C_\delta\setminus (\eta_\delta(n)+1)),\quad z^*_n=
(h\circ\hat g_\delta)(z_n)\quad\mbox{ and }\quad y^*_{n,\ell}=(h\circ\hat
g_\delta)(y_{n,\ell}).$$ Now clearly $\hat g_\delta(z_n)=g_\delta(z_n)=x^n_{\eta_\delta(n),i}\in
G\restriction\nu_\delta(n)$, hence $(h\circ\hat g_\delta)(z_n)\notin H
\restriction\eta_\delta(n)$, that is $z^*_n\notin
H\restriction\eta_\delta(n)$.
So $z^*_n\in H_{\nu_\delta(n)}\setminus H_{\eta_\delta(n)}$ belongs to the $p$-adic closure of ${{\rm Rang}}(f\restriction Y)$. As $H$, $G$, $h$ and $f\restriction Y$ belongs to ${\frak A}$, also $K$, the closure of ${{\rm Rang}}(f\restriction Y)$ in $H$ by the $p$-adic topology belongs to ${\frak A}$, and clearly $|K|\leq 2^{\aleph_0}$, hence $$A^*=\{\alpha\in C_\delta: K\cap H_{\min(C_\delta\setminus
(\alpha+1))}\setminus H_\alpha \mbox{ is not empty}\}$$ is a subset of $C_\delta$ of cardinality $\leq 2^{\aleph_0}$ which belongs to ${\frak A}$, hence $[A^*]^{\aleph_0}\subseteq {\frak A}$ but $A_\delta\subseteq A^*$ so $A_\delta\in {\frak A}$, a contradiction. $\square_{\ref{5.7}}$
Implications between the existence of universals
================================================
\[6.1\] Let $\bar n=\langle n_i:i<\omega\rangle$, $n_i\in [1,\omega)$. Remember $$J^2_{\bar n}=\{A\subseteq\prod_{i<\omega} n_i:A \mbox{ is nowhere
dense}\}.$$ Assume $\lambda\ge 2^{\aleph_0}$, $T^{\aleph_0}_{J^3_{\bar n}}(\lambda)=
\lambda$ or just $T^{\aleph_0}_{J^2_{\bar n}}(\lambda)=\lambda$ for every such $\bar n$, and $$n<\omega\quad\Rightarrow\quad\lambda_n\le\lambda_{n+1}\le\lambda_\omega
=\lambda\quad\mbox{ and}$$ $$\lambda\le\prod_{n<\omega}\lambda_n\quad\mbox{ and }\quad\bar\lambda=
\langle\lambda_i:i\le\omega\rangle.$$
1. If in ${\frak K}^{fc}_{\bar \lambda}$ there is a universal member
[*then*]{} in ${\frak K}^{rs(p)}_\lambda$ there is a universal member.
2. If in ${\frak K}^{fc}_\lambda$ there is a universal member for ${\frak K}^{fc}_{\bar \lambda}$
[*then*]{} in $${\frak K}^{rs(p)}_{\bar\lambda}=:\{G\in {\frak K}^{rs(p)}_\lambda:\lambda_n
(G)\le\lambda\}$$ there is a universal member (even for ${\frak K}^{rs(p)}_\lambda$).
($\lambda_n(G)$ were defined in \[1.1\]).
1. Similarly for “there are $M_i\in {\frak K}_{\lambda_1}$ ($i<\theta$) with $\langle M_i:i<\theta\rangle$ being universal for ${\frak K}_\lambda$”.
2. The parallel of \[1.1\] holds for ${\frak K}^{fc}_\lambda$.
3. By §5 only the case $\lambda$ singular or $\lambda=\mu^+\ \&\
{{\rm cf}}(\mu)=\aleph_0\ \& \ (\forall \alpha<
\mu)(|\alpha|^{\aleph_0}<\mu)$ is of interest for \[6.1\].
1) By \[1.1\], (2) $\Rightarrow$ (1).
More elaborately, by part (2) of \[6.1\] below there is $H\in {\frak
K}^{rs(p)}_{\bar \lambda}$ which is universal in ${\frak
K}^{rs(p)}_{\bar \lambda}$. Clearly $|G|=\lambda$ so $H\in {\frak
K}^{rs(p)}_\lambda$, hence for proving part (1) of \[6.1\] it suffices to prove that $H$ is a universal member of ${\frak
K}^{rs(p)}_\lambda$. So let $G\in {\frak K}^{rs(p)}_\lambda$, and we shall prove that it is embeddable into $H$. By \[1.1\] there is $G'$ such that $G\subseteq G'\in {\frak K}^{rs(p)}_{\bar \lambda}$. By the choice of $H$ there is an embedding $h$ of $G'$ into $H$. So $h\restriction G$ is an embedding of $G$ into $H$, as required.
2) Let $T^*$ be a universal member of ${\frak K}^{fc}_{\bar
\lambda}$ (see §2) and let $P_\alpha = P^{T^*}_\alpha$.
Let $\chi>2^\lambda$. Without loss of generality $P_n=\{n\}\times
\lambda_n$, $P_\omega=\lambda$. Let $$B_0=\bigoplus\{G^n_t:n<\omega,t\in P_n \},$$ $$B_1=\bigoplus \{G^n_t: n< \omega\mbox{ and }t\in P_n\},$$ where $G^n_t\cong {\Bbb Z}/p^{n+1}{\Bbb Z}$, $G^n_t$ is generated by $x^n_t$. Let ${\frak B}\prec ({{\cal}H}(\chi),\in,<^*_\chi)$, $\|{\frak B}\|=
\lambda$, $\lambda+1\subseteq {\frak B}$, $T^*\in {\frak B}$, hence $B_0$, $B_1\in {\frak B}$ and $\hat B_0, \hat B_1\in {\frak B}$ (the torsion completion of $B$). Let $G^* =\hat B_1\cap {\frak B}$.
Let us prove that $G^*$ is universal for ${\frak K}^{rs(p)}_{\bar
\lambda}$ (by \[1.1\] this suffices). Let $G \in {\frak K}^{rs(p)}_{\lambda}$, so by \[1.1\] without loss of generality $B_0 \subseteq G\subseteq\hat B_0$. We define $R$: $$\begin{array}{rl}
R=\big\{\eta:&\eta\in\prod\limits_{n<\omega}\lambda_n\mbox{ and for
some }
x\in G\mbox{ letting }\\
&x=\sum\{a^n_i\,p^{n-k}\,x^n_i:n<\omega,i\in w_n(x)\}\mbox{ where }\\
&w_n(x)\in [\lambda_n]^{<\aleph_0},a^n_i\,p^{n-k}\,x^n_i\ne 0\mbox{
we have }\\
&\bigwedge\limits_n\eta(n)\in w_n(x)\cup \{\ell:
\ell+|w_n(x)|\leq n\}\big\}.
\end{array}$$ Lastly let $M =:(R\cup\bigcup\limits_{n<\omega}\{n\}\times\lambda_n,\,P_n,\,
F_n)_{n<\omega}$ where $P_n=\{n\}\times\lambda_n$ and $F_n(\eta)=(n,\eta(n))$, so clearly $M\in {\frak K}^{fc}_{\bar\lambda}$. Consequently, there is an embedding $g:M\longrightarrow T^*$, so $g$ maps $\{n\}\times\lambda_n$ into $P^{T^*}_n$ and $g$ maps $R$ into $P^{T^*}_\omega$. Let $g(n,\alpha)=
(n,g_n(\alpha))$ (i.e. this defines $g_n$). Clearly $g\restriction (\cup
P^M_n)=g\restriction (\bigcup\limits_n\{n\}\times\lambda_n)$ induces an embedding $g^*$ of $B_0$ to $B_1$ (by mapping the generators into the generators).
The problem is why:
$(*)$
: if $x=\sum\{a^n_i\,p^{n-k}\,x^n_i:n<\omega,i\in w_n(x)\}\in G$
[*then*]{} $g^*(x)=\sum\{a^n_i\,p^{n-k}\,g^*(x^n_i):n<\omega,i\in w_n(x)\}\in
G^*$.
As $G^*=\hat B_1\cap {\frak B}$, and $2^{\aleph_0}+1\subseteq {\frak B}$, it is enough to prove $\langle g^{\prime\prime}(w_n(x)):n<\omega\rangle\in
{\frak B}$. Now for notational simplicity $\bigwedge\limits_n [|w_n(x)|\ge n+1]$ (we can add an element of $G^*\cap {\frak B}$ or just repeat the arguments). For each $\eta
\in\prod\limits_{n<\omega} w_n(x)$ we know that $$g(\eta)=\langle g(\eta(n)):n<\omega\rangle\in T^*\quad\mbox{ hence is in }
{\frak B}$$ (as $T^*\in {\frak B}$, $|T^*|\le\lambda$). Now by assumption there is $A\subseteq\prod\limits_{n<\omega} w_n(x)$ which is not nowhere dense such that $g
\restriction A\in {\frak B}$, hence for some $n^*$ and $\eta^* \in
\prod\limits_{\ell<n^*}w_\ell(x)$, $A$ is dense above $\eta^*$ (in $\prod\limits_{n<\omega} w_n(x)$). Hence $$\langle\{\eta(n):\eta\in A\}:n^* \le n<\omega\rangle=\langle w_n[x]:n^*\le
n<\omega\rangle,$$ but the former is in ${\frak B}$ as $A\in {\frak B}$, and from the latter the desired conclusion follows. $\square_{\ref{6.1}}$
Non-existence of universals for trees with small density
========================================================
For simplicity we deal below with the case $\delta=\omega$, but the proof works in general (as for ${\frak K}^{fr}_{\bar\lambda}$ in §2). Section 1 hinted we should look at ${\frak K}^{tr}_{\bar\lambda}$ not only for the case $\bar\lambda=\langle\lambda:\alpha\le\omega\rangle$ (i.e. ${\frak
K}^{tr}_\lambda$), but in particular for $$\bar\lambda=\langle\lambda_n:n<\omega\rangle\char 94\langle\lambda\rangle,
\qquad \lambda^{\aleph_0}_n<\lambda_{n+1}<\mu<\lambda={{\rm cf}}(\lambda)<
\mu^{\aleph_0}.$$ Here we get for this class (embeddings are required to preserve levels), results stronger than the ones we got for the classes of Abelian groups we have considered.
\[7.1\] Assume that
(a)
: $\bar\lambda=\langle\lambda_\alpha:\alpha\le\omega\rangle$, $\lambda_n<\lambda_{n+1}<\lambda_\omega$, $\lambda=\lambda_\omega$, all are regulars,
(b)
: $D$ is a filter on $\omega$ containing cobounded sets,
(c)
: ${{\rm tcf}}(\prod \lambda_n /D)=\lambda$ (indeed, we mean $=$, we could just use $\lambda\in{{\rm pcf}}_D(\{\lambda_n:n<\omega\})$),
(d)
: $(\sum\limits_{n<\omega}\lambda_n)^+<\lambda<\prod\limits_{n<
\omega}\lambda_n$.
[*Then*]{} there is no universal member in ${\frak K}^{tr}_{\bar\lambda}$.
We first notice that there is a sequence $\bar P=\langle P_\alpha:
\sum\limits_{n<\omega}\lambda_n<\alpha<\lambda\rangle$ such that:
1. $|P_\alpha|<\lambda$,
2. $a\in P_\alpha\quad\Rightarrow\quad a$ is a closed subset of $\alpha$ of order type $\leq\sum\limits_{n<\omega}\lambda_n$,
3. $a\in\bigcup\limits_{\alpha<\lambda} P_\alpha\ \&\ \beta\in{{\rm nacc}}(a)
\quad \Rightarrow\quad a\cap\beta\in P_\beta$,
4. For all club subsets $E$ of $\lambda$, there are stationarily many $\delta$ for which there is an $a\in\bigcup\limits_{\alpha<\lambda} P_\alpha$ such that $${{\rm cf}}(\delta)=\aleph_0\ \&\ a\in P_\delta\ \&\ {{\rm otp}}(a)=\sum_{n<\omega}
\lambda_n\ \&\ a\subseteq E.$$
\[Why? If $\lambda=(\sum\limits_{n<\omega}\lambda_n)^{++}$, then it is the successor of a regular, so we use [@Sh:351 §4], i.e. $$\{\alpha<\lambda:{{\rm cf}}(\alpha)\le(\sum_{n<\omega}\lambda_n)\}$$ is the union of $(\sum\limits_{n<\omega}\lambda_n)^+$ sets with squares.\
If $\lambda>(\sum\limits_{n<\omega}\lambda_n)^{++}$, then we can use [@Sh:420 §1], which guarantees that there is a stationary $S\in
I[\lambda]$.\]
We can now find a sequence $$\langle f_\alpha,g_{\alpha,a}:\alpha<\lambda,a\in P_\alpha\rangle$$ such that:
(a)
: $\bar f=\langle f_\alpha:\alpha<\lambda\rangle$ is a $<_D$-increasing cofinal sequence in $\prod\limits_{n<\omega}\lambda_n$,
(b)
: $g_{\alpha,a}\in\prod\limits_{n<\omega}\lambda_n$,
(c)
: $\bigwedge\limits_{\beta<\alpha} f_\beta<_D g_{\alpha,a}<_D
f_{\alpha+1}$,
(d)
: $\lambda_n>|a|\ \&\ \beta\in{{\rm nacc}}(a)\quad \Rightarrow\quad
g_{\beta,a\cap\beta}(n)<g_{\alpha,a}(n)$.
\[How? Choose $\bar f$ by ${{\rm tcf}}(\prod\limits_{n<\omega}\lambda_n/D)=\lambda$. Then choose $g$’s by induction, possibly throwing out some of the $f$’s; this is from [@Sh:g II, §1].\]
Let $T\in {\frak K}^{tr}_{\bar \lambda}$.
We introduce for $x\in{{\rm lev}}_\omega(T)$ and $\ell<\omega$ the notation $F^T_\ell(x) = F_\ell(x)$ to denote the unique member of ${{\rm lev}}_\ell(T)$ which is below $x$ in the tree order of $T$.
For $a\in\bigcup\limits_{\alpha<\lambda} P_\alpha$, let $a=\{
\alpha_{a,\xi}:\xi<{{\rm otp}}(a)\}$ be an increasing enumeration. We shall consider two cases. In the first one, we assume that the following statement $(*)$ holds. In this case, the proof is easier, and maybe $(*)$ always holds for some $D$, but we do not know this at present.
[[$(*)$]{}]{}
: There is a partition $\langle A_n:n < \omega \rangle$ of $\omega$ into sets not disjoint to any member of $D$.
In this case, let for $n\in\omega$, $D_n$ be the filter generated by $D$ and $A_n$. Let for $a\in\bigcup\limits_{\alpha<\lambda} P_\alpha$ with ${{\rm otp}}(a)=
\sum\limits_{n<\omega}\lambda_n$, and for $x\in{{\rm lev}}_\omega(T)$, $${{\rm inv}}(x,a,T)=:\langle\xi_n(x,a,T):n<\omega\rangle,$$ where $$\begin{array}{rl}
\xi_n(x,a,T)=:\min\big\{\xi<{{\rm otp}}(a)\!:&\mbox{ for some }m<\omega
\mbox{ we have }\\
& \langle F^T_\ell(x):
\ell<\omega\rangle<_{D_n} g_{\alpha', a'}\mbox{ where }\\
&\alpha'=\alpha_{a,\omega\xi+m}\mbox{ and } a'=
a\cap\alpha'\big\}.
\end{array}$$ Let $${{\rm INv}}(a,T)=:\{{{\rm inv}}(x,a,T):x\in T\ \&\ {{\rm lev}}_T(x)=\omega\},$$ $$\begin{array}{rl}
{{\rm INV}}(T)=:\big\{c:&\mbox{for every club } E\subseteq\lambda,\mbox{ for some }
\delta\mbox{ and } a\in P\\
&\mbox{we have }{{\rm otp}}(a)=\sum\lambda_n\ \&\ a\subseteq E\ \&\ a\in P_\delta\\
&\mbox{and for some } x\in T\mbox{ of } {{\rm lev}}_T(x)=\omega,\ c={{\rm inv}}(x,a,T)
\big\}.
\end{array}$$ (Alternatively, we could have looked at the function giving each $a$ the value ${{\rm INv}}(a,T)$, and then divide by a suitable club guessing ideal as in the proof in §3, see Definition \[3.7\].)
Clearly
[**Fact**]{}: ${{\rm INV}}(T)$ has cardinality $\le\lambda$.
The main point is the following
[**Main Fact**]{}: If ${{\bf h}}:T^1\longrightarrow T^2$ is an embedding, [*then*]{} $${{\rm INV}}(T^1)\subseteq{{\rm INV}}(T^2).$$
[*Proof of the [**Main Fact**]{} under $(*)$*]{} We define for $n
\in\omega$ $$E_n=:\big\{\delta<\lambda_n:\,\delta>\bigcup_{\ell<n}\lambda_\ell\mbox{ and
}\left(\forall x\in{{\rm lev}}_n(T^1)\right)({{\bf h}}(x)<\delta \Leftrightarrow x<\delta)
\big\}.$$ We similarly define $E_\omega$, so $E_n$ ($n\in\omega$) and $E_\omega$ are clubs (of $\lambda_n$ and $\lambda$ respectively). Now suppose $c\in{{\rm INV}}(T_1)
\setminus{{\rm INV}}(T_2)$. Without loss of generality $E_\omega$ is (also) a club of $\lambda$ which exemplifies that $c\notin{{\rm INV}}(T_2)$. For $h\in
\prod\limits_{n<\omega}\lambda_n$, let $$h^+(n)=:\min(E_n\setminus h(n)),\quad\mbox{ and }\quad\beta[h]=\min\{\beta
<\lambda:h<f_\beta\}.$$ (Note that $h<f_{\beta[h]}$, not just $h<_D f_{\beta[h]}$.) For a sequence $\langle h_i:i<i^*\rangle$ of functions from $\prod\limits_{n<\omega}
\lambda_n$, we use $\langle h_i:i<i^* \rangle^+$ for $\langle h^+_i:i<i^*
\rangle$. Now let $$E^*=:\big\{\delta<\lambda:\mbox{if }\alpha<\delta\mbox{ then }
\beta[f^+_\alpha]<\delta\mbox{ and }\delta\in{{\rm acc}}(E_\omega)\big\}.$$ Thus $E^*$ is a club of $\lambda$. Since $c\in{{\rm INV}}(T_1)$, there is $\delta<
\lambda$ and $a\in P_\delta$ such that for some $x\in{{\rm lev}}_\omega(T_1)$ we have $$a\subseteq E^*\ \&\ {{\rm otp}}(a)=\sum_{n<\omega}\lambda_n\ \&\ c={{\rm inv}}(x,a,T_1).$$ Let for $n\in\omega$, $\xi_n=:\xi_n(x,a,T_1)$, so $c=\langle\xi_n:n<\omega
\rangle$. Also let for $\xi<\sum\limits_{n<\omega}\lambda_n$, $\alpha_\xi=:
\alpha_{a,\xi}$, so $a=\langle\alpha_\xi:\xi<\sum\limits_{n<\omega}\lambda_n
\rangle$ is an increasing enumeration. Now fix an $n<\omega$ and consider ${{\bf h}}(x)$. Then we know that for some $m$
($\alpha$)
: $\langle F^{T_1}_\ell(x):\ell<\omega\rangle<_{D_n}
g_{\alpha'}$ where $\alpha'=\alpha_{\omega\xi_n+m}$and
($\beta$)
: for no $\xi<\xi_n$ is there such an $m$.
Now let us look at $F^{T_1}_\ell(x)$ and $F^{T_2}_\ell({\bold h}(x))$. They are not necessarily equal, but
($\gamma$)
: $\min(E_\ell\setminus F^{T_1}_\ell(x))=\min(E_\ell
\setminus F^{T_2}_\ell({{\bf h}}(x))$
(by the definition of $E_\ell$). Hence
($\delta$)
: $\langle F^{T_1}_\ell(x):\ell<\omega\rangle^+=\langle
F^{T_2}_\ell({{\bf h}}(x)):\ell<\omega\rangle^+$.
Now note that by the choice of $g$’s
($\varepsilon$)
: $(g_{\alpha_\varepsilon,a\cap\alpha_\varepsilon})^+
<_{D_n} g_{\alpha_{\varepsilon+1},a\cap\alpha_{\varepsilon+1}}$.
From $(\delta)$ and $(\varepsilon)$ it follows that $\xi_n({{\bf h}}(x),a,
T^2)=\xi_n(x,a,T^1)$. Hence $c\in{{\rm INV}}(T^2)$. $\square_{\mbox{Main Fact}}$
Now it clearly suffices to prove:
[**Fact A:**]{} For each $c=\langle\xi_n:n<\omega
\rangle\in {}^\omega(\sum\limits_{n<\omega}\lambda_n)$ we can find a $T\in
{\frak K}^{tr}_{\bar\lambda}$ such that $c\in{{\rm INV}}(T)$.
[*Proof of the Fact A in case $(*)$ holds*]{} For each $a\in
\bigcup\limits_{\delta<\lambda} P_\delta$ with ${{\rm otp}}(a)=\sum\limits_{n\in
\omega}\lambda_n$ we define $x_{c,a}=:\langle x_{c,a}(\ell):\ell<\omega
\rangle$ by:
> if $\ell\in A_n$, then $x_{c,a}(\ell)=\alpha_{a,\omega\xi_n+\delta}$.
Let $$T=\bigcup_{n<\omega}\prod_{\ell<n}\lambda_\ell\cup\big\{x_{c,a}:a\in
\bigcup_{\delta<\lambda} P_\delta\ \&\ {{\rm otp}}(a)=\sum_{n<\omega}\lambda_n
\big\}.$$ We order $T$ by $\triangleleft$.
It is easy to check that $T$ is as required. $\square_A$
Now we are left to deal with the case that $(*)$ does not hold. Let $${{\rm pcf}}(\{\lambda_n:n<\omega\})=\{\kappa_\alpha:\alpha\le\alpha^*\}$$ be an enumeration in increasing order so in particular $$\kappa_{\alpha^*}=\max{{\rm pcf}}(\{\lambda_n:n<\omega\}).$$ Without loss of generality $\kappa_{\alpha^*}=\lambda$ (by throwing out some elements if necessary) and $\lambda\cap{{\rm pcf}}(\{\lambda_n:n<\omega\})$ has no last element (this appears explicitly in [@Sh:g], but is also straightforward from the pcf theorem). In particular, $\alpha^*$ is a limit ordinal. Hence, without loss of generality $$D=\big\{A\subseteq\omega:\lambda>\max{{\rm pcf}}\{\lambda_n:n\in\omega\setminus A\}
\big\}.$$ Let $\langle {\frak a}_{\kappa_\alpha}:\alpha\le\alpha^*\rangle$ be a generating sequence for ${{\rm pcf}}(\{\lambda_n:n<\omega\})$, i.e. $$\max{{\rm pcf}}({\frak a}_{\kappa_\alpha})=\kappa_\alpha\quad\mbox{ and }\quad
\kappa_\alpha\notin{{\rm pcf}}(\{\lambda_n:n<\omega\}\setminus {\frak
a}_{\kappa_\alpha}).$$ (The existence of such a sequence follows from the pcf theorem). Without loss of generality, $${\frak a}_{\alpha^*} = \{ \lambda_n:n < \omega\}.$$ Now note
If ${{\rm cf}}(\alpha^*)=\aleph_0$, then $(*)$ holds.
Why? Let $\langle\alpha(n):n<\omega\rangle$ be a strictly increasing cofinal sequence in $\alpha^*$. Let $\langle B_n:n<\omega\rangle$ partition $\omega$ into infinite pairwise disjoint sets and let $$A_\ell=:\big\{k<\omega:\bigvee_{n\in B_\ell}[\lambda_k\in {\frak
a}_{\kappa_{\alpha(n)}}\setminus\bigcup_{m<n} {\frak a}_{\kappa_{\alpha(m)}}]
\big\}.$$ To check that this choice of $\langle A_\ell:\ell<\omega\rangle$ works, recall that for all $\alpha$ we know that $\alpha_{\kappa_\alpha}$ does not belong to the ideal generated by $\{{\frak a}_{\kappa_\beta}: \beta<\alpha\}$ and use the pcf calculus. $\square$
Now let us go back to the general case, assuming ${{\rm cf}}(\alpha^*)>\aleph_0$. Our problem is the possibility that $${{\cal}P}(\{\lambda_n:n<\omega\})/J_{<\lambda}[\{\lambda_n:n<\omega\}].$$ is finite. Let now $A_\alpha=:\{n:\lambda_n\in {\frak a}_\alpha\}$, and $$\begin{array}{lll}
J_\alpha &=: &\big\{A\subseteq\omega:\max{{\rm pcf}}\{\lambda_\ell:\ell\in A\}<
\kappa_\alpha\big\}\\
J'_\alpha &=: &\big\{A\subseteq\omega:\max{{\rm pcf}}\{\lambda_\ell:\ell\in A\}\cap
{\frak a}_{\kappa_\alpha}<\kappa_\alpha\big\}.
\end{array}$$ We define for $T\in {\frak K}^{tr}_{\bar\lambda}$, $x\in{{\rm lev}}_\omega(T)$, $\alpha<\alpha^*$ and $a\in\bigcup\limits_{\delta<\lambda} P_\delta$: $$\begin{array}{rl}
\xi^*_\alpha(x,a,T)=:\min\big\{\xi:&\bigvee_m [\langle F^T_\ell(x):\ell<
\omega\rangle <_{J'_\alpha} g_{\alpha', a'}\mbox{ where }\\
&\qquad \alpha'=\alpha_{a,\omega \xi+m}\mbox{ and }a'=a\cap
\alpha'\big\}.
\end{array}$$ Let $${{\rm inv}}_\alpha(x,a,T)=:\langle\xi^*_{\alpha+n}(x,a,T):n<\omega\rangle,$$ $${{\rm INv}}(a,T)=:\big\{{{\rm inv}}_\alpha(x,a,T):x\in T\ \&\ \alpha<\alpha^* \ \&\
{{\rm lev}}_T(x)=\omega\big\},$$ and $$\begin{array}{rl}
{{\rm INV}}(T)=\big\{c:&\mbox{for every club } E^*\mbox{ of }\lambda\mbox{ for some
} a\in\bigcup\limits_{\delta<\lambda} P_\delta\\
&\mbox{with }{{\rm otp}}(a)=\sum\limits_{n<\omega}\lambda_n\mbox{ for arbitrarily
large }\alpha<\alpha^*,\\
&\mbox{there is }x\in{{\rm lev}}_\omega(T)\mbox{ such that }{{\rm inv}}_\alpha(x,a,T)=
c\big\}.
\end{array}$$ As before, the point is to prove the Main Fact.
[*Proof of the [**Main Fact**]{} in general*]{} Suppose ${{\bf h}}:
T^1\longrightarrow T^2$ and $c\in{{\rm INV}}(T^1)\setminus{{\rm INV}}(T^2)$. Let $E'$ be a club of $\lambda$ which witnesses that $c\notin{{\rm INV}}(T^2)$. We define $E_n,E_\omega$ as before, as well as $E^*$ ($\subseteq E_\omega\cap
E'$). Now let us choose $a\in
\bigcup\limits_{\delta<\lambda} P_\delta$ with $a\subseteq E^*$ and ${{\rm otp}}(a)=
\sum\limits_{n<\omega}\lambda_n$. So $a=\{\alpha_{a,\xi}:\xi<\sum\limits_{n<
\omega}\lambda_n\}$, which we shorten as $a=\{\alpha_\xi:\xi<\sum\limits_{n<
\omega}\lambda_n\}$. For each $\xi<\sum\limits_{n<\omega}\lambda_n$, as before, we know that $$(g_{\alpha_\xi,a\cap\alpha_\xi})^+<_{J^*_\alpha}g_{\alpha_{\xi+1},a\cap
\alpha_{\xi+1}}.$$ Therefore, there are $\beta_{\xi,\ell}<\alpha^*$ ($\ell<\ell_\xi$) such that $$\{\ell:g^+_{\alpha_\xi,a\cap\alpha_\xi}(\ell)\ge g_{\alpha_{\xi+1},a\cap
\alpha_{\xi+1}}(\ell)\}\supseteq\bigcup_{\ell<\ell_\xi} A_{\beta_{\xi,\ell}}.$$ Let $c=\langle\xi_n:n<\omega\rangle$ and let $$\Upsilon=\big\{\beta_{\xi,\ell}:\mbox{for some } n\mbox{ and } m\mbox{ we
have }\xi=\omega\xi_n+m\ \&\ \ell<\omega\big\}.$$ Thus $\Upsilon\subseteq\alpha^*$ is countable. Since ${{\rm cf}}(\alpha^*)>\aleph_0$, the set $\Upsilon$ is bounded in $\alpha^*$. Now we know that $c$ appears as an invariant for $a$ and arbitrarily large $\delta<\alpha^*$, for some $x_{a,\delta}\in{{\rm lev}}_\omega(T_1)$. If $\delta>\sup(\Upsilon)$, $c\in{{\rm INV}}(T^2)$ is exemplified by $a,\delta,{{\bf h}}(x_{\alpha,\delta})$, just as before. $\square$
We still have to prove that every $c=\langle\xi_n:n<\omega\rangle$ appears as an invariant; i.e. the parallel of Fact A.
[*Proof of Fact A in the general case:*]{} Define for each $a\in
\bigcup\limits_{\delta<\lambda} P_\delta$ with ${{\rm otp}}(a)=\sum\limits_{n<\omega}
\lambda_n$ and $\beta<\alpha^*$ $$x_{c,a,\beta}=\langle x_{c,a,\beta}(\ell):\ell<\omega\rangle,$$ where $$x_{c,a,\beta}(\ell)=\left\{
\begin{array}{lll}
\alpha_{a,\omega\xi_n+\delta} &\mbox{\underbar{if}} &\lambda_\ell\in {\frak
a}_{\beta+k}\setminus\bigcup\limits_{k'<k} {\frak a}_{\beta+k'}\\
0 &\mbox{\underbar{if}} &\lambda_\ell\notin {\frak a}_{\beta+k}\mbox{ for
any } k<\omega.
\end{array}\right.$$ Form the tree as before. Now for any club $E$ of $\lambda$, we can find $a\in
\bigcup\limits_{\delta<\lambda} P_\delta$ with ${{\rm otp}}(a)=\sum\limits_{n<\omega}
\lambda_n$, $a\subseteq E$ such that $\langle x_{c,a,\beta}:\beta<\alpha^*
\rangle$ shows that $c\in{{\rm INV}}(T)$. $\square_{\ref{7.1}}$
1. Clearly, this proof shows not only that there is no one $T$ which is universal for ${\frak K}^{tr}_{\bar\lambda}$, but that any sequence of $<\prod\limits_{n<\omega}\lambda_n$ trees will fail. This occurs generally in this paper, as we have tried to mention in each particular case.
2. The case “$\lambda<2^{\aleph_0}$" is included in the theorem, though for the Abelian group application the $\bigwedge\limits_{n<\omega}
\lambda^{\aleph_0}_n<\lambda_{n+1}$ is necessary.
\[7.1A\]
1. If $\mu^+< \lambda={{\rm cf}}(\lambda)<\chi<\mu^{\aleph_0}$ and $\chi^{[\lambda]}<\mu^{\aleph_0}$ (or at least $T_{{{\rm id}}^a(\bar C)}(\chi)<\mu^{\aleph_0}$) we can get the results for “no $M \in {\frak K}^x_\chi$ is universal for ${\frak K}^x_\lambda$", see §8 (and [@Sh:456]).
We can below (and subsequently in §8) use $J^3_{\bar m}$ as in §6.
\[7.2\] Assume that $2^{\aleph_0}<\lambda_0$, $\bar\lambda=\langle\lambda_n:n<\omega
\rangle\char 94\langle\lambda\rangle$, $\mu=\sum\limits_{n<\omega}\lambda_n$, $\lambda_n<\lambda_{n+1}$, $\mu^+<\lambda={{\rm cf}}(\lambda)<\mu^{\aleph_0}$.\
, for simplicity, $\bar m=\langle m_i:i<\omega\rangle=\langle
\omega:i<\omega\rangle$ (actually $m_i\in [2,\omega]$ or even $m_i\in
[2,\lambda_0)$, $\lambda_0<\lambda$ are O.K.) and ${{\bf U}}^{<\mu}_{J^2_{\bar m}}
(\lambda)=\lambda$ (remember $$J^2_{\bar m}=\{A\subseteq\prod_{i<\omega} m_i:A \mbox{ is nowhere dense}\}$$ and definition \[5.3\]),\
in ${\frak K}^{tr}_{\bar\lambda}$ there is no universal member.
1) Let $S\subseteq\lambda$, $\bar C=\langle C_\delta:\delta\in S
\rangle$ be a club guessing sequence on $\lambda$ with ${{\rm otp}}(C_\delta)\ge\sup
\lambda_n$. We assume that we have ${\bar{\frak A}}=\langle {\frak A}_\alpha:
\alpha<\lambda\rangle$, $J^2_{\bar m}$, $T^*\in {\frak A}_0$ ($T^*$ is a candidate for the universal), $\bar C=\langle C_\delta:\delta\in S\rangle\in
{\frak A}_\alpha$, ${\frak A}_\alpha\prec({{\cal}H}(\chi),\in,<^*_\chi)$, $\chi=
\beth_7(\lambda)^+$, $\|{\frak A}_\alpha\|<\lambda$, ${\frak A}_\alpha$ increasingly continuous, $\langle {\frak A}_\beta:\beta\le\alpha\rangle\in
{\frak A}_{\alpha+1}$, ${\frak A}_\alpha\cap\lambda$ is an ordinal, ${\frak A}
=\bigcup\limits_{\alpha<\lambda}{\frak A}_\alpha$ and $$E=:\{\alpha:{\frak A}_\alpha\cap\lambda=\alpha\}.$$ Note: $\prod\bar m\subseteq {\frak A}$ (as $\prod\bar m\in {\frak A}$ and $|\prod\bar m|=2^{\aleph_0}$).
NOTE: By ${{\bf U}}^{<\mu}_{J^2_{\bar m}}(\lambda)=\lambda$,
$(*)$
: $x_\eta\in{{\rm lev}}_\omega(T^*)$ for $\eta\in\prod\bar
m$
for some $A\in (J^2_{\bar m})^+$ the set $\langle(\eta,
x_\eta):\eta\in A\rangle$ belongs to ${\frak A}$. But then for some $\nu\in
\bigcup\limits_k\prod\limits_{i<k} m_i$, the set $A$ is dense above $\nu$ (by the definition of $J^2_{\bar m}$) and hence: if the mapping $\eta\mapsto
x_\eta$ is continuous then $\langle x_\rho:\nu\triangleleft\rho\in\prod\bar m
\rangle\in {\frak A}$.
For $\delta\in S$ such that $C_\delta \subseteq E$ we let $$\begin{array}{rl}
P^0_\delta=P^0_\delta({\frak A})=\biggl\{\bar x:&\bar x=\langle x_\rho:\rho
\in t\rangle\in {\frak A}\mbox{ and } x_\rho\in{{\rm lev}}_{\ell g(\rho)}T^*,\\
&\mbox{the mapping }\rho\mapsto x_\rho\mbox{ preserves all of the
relations:}\\
&\ell g(\rho)=n,\rho_1\triangleleft\rho_2, \neg(\rho_1\triangleleft\rho_2),
\neg(\rho_1=\rho_2),\\
&\rho_1\cap\rho_2=\rho_3\mbox{ (and so }\ell g(\rho_1\cap\rho_2)=n\mbox{ is
preserved});\\
&\mbox{and } t\subseteq\bigcup\limits_{\alpha\le\omega}\prod\limits_{i<
\alpha} m_i\biggr\}.
\end{array}$$ Assume $\bar x=\langle x_\rho:\rho\in t\rangle\in P^0_\delta$. Let $${{\rm inv}}(\bar x,C_\delta,T^*,{\bar {\frak A}})=:\big\{\alpha\in C_\delta:
(\exists\rho\in{{\rm Dom}}(\bar x))(x_\rho\in {\frak A}_{\min(C_\delta
\setminus (\alpha+1))}\setminus {\frak A}_\alpha\big)\}.$$ Let $$\begin{array}{l}
{{\rm Inv}}(C_\delta,T^*,\bar{\frak A})=:\\
\big\{a:\mbox{for some }\bar x\in P^0_\delta,\ \ a\mbox{ is a
countable subset of }{{\rm inv}}(\bar x,C_0,T^*,\bar{\frak A})\big\}.
\end{array}$$ Note: ${{\rm inv}}(\bar x,C_0,T^*,\bar{\frak A})$ has cardinality at most continuum, so ${{\rm Inv}}(C_0,T^*,\bar{\frak A})$ is a family of $\le 2^{\aleph_0}\times
|{\frak A}|=\lambda$ countable subsets of $C_\delta$.
We continue as before. Let $\alpha_{\delta,\varepsilon}$ be the $\varepsilon$-th member of $C_\delta$ for $\varepsilon<\sum\limits_{n<\omega}
\lambda_n$. So as $\lambda< \mu^{\aleph_0}$,$\mu> 2^{\aleph_0}$ clearly $\lambda< {{\rm cf}}([\lambda]^{\aleph_0}, \subseteq)$ (equivalently $\lambda<{{\rm cov}}(\mu,\mu,\aleph_1,2)$) hence we can find $\gamma_n\in
(\bigcup\limits_{\ell<n}\lambda_\ell,\lambda_n)$ limit so such that for each $\delta\in S$, $a\in{{\rm Inv}}(C_\delta,T^*,\bar{\frak A})$ we have $\{\gamma_n+\ell:n<\omega\mbox{ and }\ell<m_i\}\cap a$ is bounded in $\mu$.
Now we can find $T$ such that ${{\rm lev}}_n(T)=\prod\limits_{\ell<n}\lambda_\ell$ and $$\begin{array}{rl}
{{\rm lev}}_\omega(T)=\big\{\bar\beta:
&\bar\beta=\langle\beta_\ell:\ell<\omega
\rangle,\mbox{ and for some }\delta \in S,\mbox{ for every }\ell<\omega\\
&\mbox{we have }\gamma'_\ell\in\{\alpha_{\delta,\gamma_\ell+m}:m<m_i\}
\big\}.
\end{array}$$ So, if $T^*$ is universal there is an embedding $f:T\longrightarrow T^*$, and hence $$E'=\{\alpha\in E:{\frak A}_\alpha\mbox{ is closed under } f\mbox{ and }
f^{-1}\}$$ is a club of $\lambda$. By the choice of $\bar C$ for some $\delta\in S$ we have $C_\delta\subseteq E'$. Now use $(*)$ with $x_\eta=f(\bar\beta^{\delta,
\eta})$, where $\beta^{\delta,\eta}_\ell=\alpha_{\delta,\gamma_\ell+
\eta(\ell)}\in{{\rm lev}}_\omega(T)$. Thus we get $A\in (J^2_{\bar m})^+$ such that $\{(\eta,x_\eta):\eta\in A\}\in{\frak A}$, there is $\nu\in\bigcup\limits_k
\prod\limits_{i<k} m_i$ such that $A$ is dense above $\nu$, hence as $f$ is continuous, $\langle(\eta,x_\eta):\nu\triangleleft\eta\in\prod\bar m\rangle
\in {\frak A}$. So $\langle x_\eta:\eta\in\prod\bar m,\nu\trianglelefteq\eta
\rangle\in P^0_\delta({\frak A})$, and hence the set $$\{\alpha_{\delta,\gamma_\ell+m}:\ell\in [\ell g(\nu),\omega)\mbox{ and } m<
m_\ell\}\cup\{\alpha_{\delta,\gamma_i+\nu(i)}:\ell<\ell g(\nu)\}$$ is ${{\rm inv}}(\bar x,C_\delta,T^*,{\frak A})$. Hence $$a=\{\alpha_{\delta,\gamma_\ell}\!:\ell\in [\ell g(\nu),\omega)\}\in{{\rm Inv}}(
C_\delta,T^*,{\frak A}),$$ contradicting
“$\{\alpha_{\delta,\gamma_\ell}:\ell<\omega\}$ has finite intersection with any $a\in{{\rm Inv}}(C_\delta,T^*,{\frak A})$”.
\[7.3\] We can a priori fix a set of $\aleph_0$ candidates and say more on their order of appearance, so that ${{\rm Inv}}(\bar x,C_\delta,T^*,{\bar{\frak A}})$ has order type $\omega$. This makes it easier to phrase a true invariant, i.e. $\langle
(\eta_n,t_n):n<\omega\rangle$ is as above, $\langle\eta_n:n<\omega\rangle$ lists ${}^{\omega >}\omega$ with no repetition, $\langle t_n\cap {}^\omega
\omega:n<\omega\rangle$ are pairwise disjoint. If $x_\rho\in{{\rm lev}}_\omega(T^*)$ for $\rho\in {}^\omega\omega$, $\bar T^*=\langle\bar T^*_\zeta:\zeta<\lambda
\rangle$ representation $$\begin{array}{l}
{{\rm inv}}(\langle x_\rho:\rho\in {}^\omega\omega\rangle,C_\delta,\bar T^*)=\\
\big\{\alpha\in C_\delta:\mbox{for some } n,\ (\forall\rho)[\rho\in t_n\cap
{}^\omega\omega\quad\Rightarrow\quad x_\rho\in T^*_{\min(C_\delta\setminus
(\alpha+1))}\setminus T^*_\alpha]\big\}.
\end{array}$$
\[7.4\] If we have $\Gamma\in (J^2_{\bar m})^+$, $\Gamma$ non-meagre, $J=J^2_m
\restriction\Gamma$ and ${{\bf U}}^2_J(\lambda)<\lambda^{\aleph_0}$ then we can weaken the cardinal assumptions to: $$\bar\lambda=\langle\lambda_n:n<\omega\rangle\char 94\langle\lambda\rangle,
\qquad \mu=\sum_n\lambda_n,\qquad\lambda_n<\lambda_{n+1},$$ $$\mu^+<\lambda={{\rm cf}}(\lambda)\qquad\mbox{ and }\qquad {{\bf U}}^2_J(\lambda)<{{\rm cov}}(\mu,
\mu,\aleph_1,2) (\mbox{see }\ref{0.4}).$$ The proof is similar.
Universals in singular cardinals
================================
In §3, §5, \[7.2\], we can in fact deal with “many” singular cardinals $\lambda$. This is done by proving a stronger assertion on some regular $\lambda$. Here ${\frak K}$ is a class of models.
\[8.1\]
1. There is no universal member in ${\frak K}_{\mu^*}$ if for some $\lambda
<\mu^*$, $\theta\ge 1$ we have:
: not only there is no universal member in ${\frak K}_\lambda$ but if we assume: $$\langle M_i:i<\theta\rangle\ \mbox{ is given, }\ \|M_i\|\le\mu^*<\prod_n
\lambda_n,\ M_i\in {\frak K},$$ then there is a structure $M$ from ${\frak K}_\lambda$ (in some cases of a simple form) not embeddable in any $M_i$.
2. Assume
$\otimes^\sigma_1$
: $\langle\lambda_n:n<\omega\rangle$ is given, $\lambda^{\aleph_0}_n<\lambda_{n+1}$, $$\mu=\sum_{n<\omega}\lambda_n<\lambda={{\rm cf}}(\lambda)\leq
\mu^*<\prod_{n<\omega}\lambda_n$$ and $\mu^+<\lambda$ or at least there is a club guessing $\bar C$ as in $(**)^1_\lambda$ (ii) of \[3.4\] for $(\lambda,\mu)$.
there is no universal member in ${\frak K}_{\mu^*}$ (and moreover $\otimes_{\lambda,\mu^*,\theta}[{\frak K}]$ holds) in the following cases
$\otimes_2$(a)
: for torsion free groups, i.e. ${\frak K}={\frak
K}^{rtf}_{\bar\lambda}$ if ${{\rm cov}}(\mu^*,\lambda^+,\lambda^+,\lambda)<
\prod\limits_{n<\omega}\lambda_n$, see notation \[0.4\] on ${{\rm cov}}$)
(b)
: for ${\frak K}={\frak K}^{tcf}_{\bar\lambda}$,
(c)
: for ${\frak K}={\frak K}^{tr}_{\bar\lambda}$ as in \[7.2\] - ${{\rm cov}}({{\bf U}}_{J^3_{\bar m}}(\mu^*),\lambda^+,\lambda^+,\lambda)<\prod\limits_{n<
\omega}\lambda_n$,
(d)
: for ${\frak K}^{rs(p)}_{\bar\lambda}$: like case (c) (for appropriate ideals), replacing $tr$ by $rs(p)$.
\[8.1A\]
1. For \[7.2\] as $\bar m=\langle\omega:i<\omega\rangle$ it is clear that the subtrees $t_n$ are isomorphic. We can use $m_i\in [2,\omega)$, and use coding; anyhow it is immaterial since ${}^\omega \omega,{}^\omega 2$ are similar.
2. We can also vary $\bar\lambda$ in \[8.1\] $\otimes_2$, case (c).
3. We can replace ${{\rm cov}}$ in $\otimes_2$(a),(c) by $$\sup{{\rm pp}}_{\Gamma(\lambda)}(\chi):{{\rm cf}}(\chi)=\lambda,\lambda<\chi\le
{{\bf U}}_{J^3_{\bar m}}(\mu^*)\}$$ (see [@Sh:355 5.4], \[2.3\]).
Should be clear, e.g.\
[*Proof of Part 2), Case (c)*]{} Let $\langle T_i:i<i^* \rangle$ be given, $i^*<\prod\limits_{n<\omega}\lambda_n$ such that $$\|T_i\|\le\mu^*\quad\mbox{ and }\quad\mu^\otimes=:{{\rm cov}}({{\bf U}}_{J^3_{\bar
m}}(\mu^*),\lambda^+,\lambda^+,\lambda)<\prod_{n<\omega}\lambda_n.$$ By [@Sh:355 5.4] and ${{\rm pp}}$ calculus ([@Sh:355 2.3]), $\mu^\otimes=
{{\rm cov}}(\mu^\otimes,\lambda^+,\lambda^+,\lambda)$. Let $\chi=\beth_7(\lambda)^+$. For $i<i^*$ choose ${\frak B}_i\prec (H(\chi)\in
<^*_\chi)$, $\|{\frak B}_i\|=\mu^\otimes$, $T_i\in {\frak B}_i$, $\mu^\otimes
+1\subseteq {\frak B}_i$. Let $\langle Y_\alpha: \alpha<\mu^\otimes\rangle$ be a family of subsets of $T_i$ exemplifying the Definition of $\mu^\otimes=
{{\rm cov}}(\mu^\otimes,\lambda^+,\lambda^+,\lambda)$.\
Given $\bar x=\langle x_\eta:\eta\in {}^\omega\omega\rangle$, $x_\eta\in
{{\rm lev}}_\omega(T_i)$, $\eta\mapsto x_\eta$ continuous (in our case this means $\ell g(\eta_1\cap\eta_2)=\ell g(x_{\eta_1}\cap x_{\eta_2})=:\ell g(\max
\{\rho:\rho\triangleleft\eta_1\ \&\ \rho\triangleleft\eta_2\})$. Then for some $\eta\in {}^{\omega>}\omega$, $$\langle x_\rho:\eta\triangleleft\rho\in {}^\omega\omega\rangle\in{\frak
B}.$$ So given $\left<\langle x^\zeta_\eta:\eta\in {}^\omega\omega\rangle:\zeta<
\lambda\right>$, $x^\zeta_\eta\in{{\rm lev}}_\omega(T_i)$ we can find $\langle (\alpha_j,\eta_j):j<j^*<\lambda\rangle$ such that: $$\bigwedge_{\zeta<\lambda}\bigvee_j\langle x^\zeta_\eta:\eta_j\triangleleft
\eta\in {}^\omega\omega\rangle\in Y_\alpha.$$ Closing $Y_\alpha$ enough we can continue as usual. $\square_{\ref{8.1}}$
Metric spaces and implications
==============================
\[9.1\]
1. ${\frak K}^{mt}$ is the class of metric spaces $M$ (i.e. $M=(|M|,d)$, $|M|$ is the set of elements, $d$ is the metric, i.e. a two-place function from $|M|$ to ${\Bbb R}^{\geq 0}$ such that $d(x,y)=0\quad\Leftrightarrow\quad x=0$ and $d(x,z)\le d(x,y)+d(y,z)$ and $d(x,y) = d(y,x)$).
An embedding $f$ of $M$ into $N$ is a one-to-one function from $|M|$ into $|N|$ which is continuous, i.e. such that:
> if in $M$, $\langle x_n:n<\omega\rangle$ converges to $x$
>
> then in $N$, $\langle f(x_n):n<\omega\rangle$ converges to $f(x)$.
2. ${\frak K}^{ms}$ is defined similarly but ${{\rm Rang}}(d)\subseteq\{2^{-n}:n
<\omega\}\cup\{0\}$ and instead of the triangular inequality we require $$d(x,y)=2^{-i},\qquad d(y,z)=2^{-j}\quad \Rightarrow\quad d(x,z) \le
2^{-\min\{i-1,j-1\}}.$$
3. ${\frak K}^{tr[\omega]}$ is like ${\frak K}^{tr}$ but $P^M_\omega=|M|$ and embeddings preserve $x\;E_n\;y$ (not necessarily its negation) are one-to-one, and remember $\bigwedge\limits_n
x\;E_n\;y\quad\Rightarrow\quad x \restriction n = y \restriction n$).
4. ${\frak K}^{mt(c)}$ is the class of semi-metric spaces $M=(|M|,d)$, which means that for the constant $c\in\Bbb R^+$ the triangular inequality is weakened to $d(x,z)\le cd(x,y)+cd(y,z)$ with embedding as in \[9.1\](1) (so for $c=1$ we get ${\frak K}^{mt}$).
5. ${\frak K}^{mt[c]}$ is the class of pairs $(A,d)$ such that $A$ is a non-empty set, $d$ a two-place symmetric function from $A$ to ${\Bbb
R}^{\ge 0}$ such that $[d(x,y)=0\ \Leftrightarrow\ x=y]$ and $$d(x_0,x_n)\le c\sum\limits_{\ell<n} d(x_\ell,x_{\ell+1})\ \ \mbox{ for any
$n<\omega$ and $x_0,\ldots,x_n\in A$.}$$
6. ${\frak K}^{ms(c)}$, ${\frak K}^{ms[c]}$ are defined parallely.
7. ${\frak K}^{rs(p),\mbox{pure}}$ is defined like ${\frak K}^{rs(p)}$ but the embeddings are pure.
There are, of course, other notions of embeddings; isometric embeddings if $d$ is preserved, co-embeddings if the image of an open set is open, bi-continuous means an embedding which is a co-embedding. The isometric embedding is the weakest, its case is essentially equivalent to the ${\frak
K}^{tr}_{\lambda}$ case (as in \[9.3\](3)); for the open case there is a universal: discrete space. The universal for ${\frak K}^{mt}_\lambda$ under bicontinuous case exist in cardinality $\lambda^{\aleph_0}$, see [@Ko57].
\[9.1A\]
1. ${{\rm Univ}}^0({\frak K}^1,{\frak K}^2)=\{(\lambda,\kappa,\theta):$ there are $M_i\in {\frak K}^2_\kappa$ for $i<\theta$ such that any $M\in {\frak
K}^1_\lambda$ can be embedded into some $M_i\}$. We may omit $\theta$ if it is 1. We may omit the superscript 0.
2. ${{\rm Univ}}^1({\frak K}^1,{\frak K}^2)=\{(\lambda,\kappa,\theta):$ there are $M_i\in {\frak K}^2_\kappa$ for $i<\theta$ such that any $M\in {\frak
K}^1_\lambda$ can be represented as the union of $<\lambda$ sets $A_\zeta$ ($\zeta<\zeta^*<\lambda)$ such that each $M\restriction A_\zeta$ can be embedded into some $M_i\}$ and is a $\leq_{{\frak K}^1}$-submodel of $M$.
3. If above ${\frak K}^1={\frak K}^2$ we write it just once; (naturally we usually assume ${\frak K}^1 \subseteq {\frak K}^2$).
<!-- -->
1. We prove our theorems for $Univ^0$, we can get parallel things for ${{\rm Univ}}^1$.
2. Many previous results of this paper can be rephrased using a pair of classes.
3. We can make \[9.2\] below deal with pairs and/or function $H$ changing cardinality.
4. ${{\rm Univ}}^\ell$ has the obvious monotonicity properties.
\[9.2\]
1. Assume ${\frak K}^1,{\frak K}^2$ has the same models as their members and every embedding for ${\frak K}^2$ is an embedding for ${\frak K}^1$.\
Then ${{\rm Univ}}({\frak K}^2)\subseteq{{\rm Univ}}({\frak K}^1)$.
2. Assume there is for $\ell=1,2$ a function $H_\ell$ from ${\frak K}^\ell$ into ${\frak K}^{3-\ell}$ such that:
(a)
: $\|H_1(M_1)\|=\|M_1\|$ for $M_1\in {\frak K}^1$,
(b)
: $\|H_2(M_2)\|=\|M_2\|$ for $M_2\in {\frak K}^2$,
(c)
: if $M_1\in {\frak K}^1$, $M_2\in {\frak K}^2$, $H_1(M_1)\in {\frak
K}^2$ is embeddable into $M_2$ $M_1$ is embeddable into $H_2(M_2)\in {\frak K}^1$.
${{\rm Univ}}({\frak K}^2)\subseteq{{\rm Univ}}({\frak K}^1)$.
\[9.2A\] We say ${\frak K}^1\le {\frak K}^2$ if the assumptions of \[9.2\](2) hold. We say ${\frak K}^1\equiv {\frak K}^2$ if ${\frak K}^1\le {\frak K}^2
\le {\frak K}^1$ (so larger means with fewer cases of universality).
\[9.3\]
1. The relation “${\frak K}^1\le {\frak K}^2$” is a quasi-order (i.e. transitive and reflexive).
2. If $({\frak K}^1,{\frak K}^2)$ are as in \[9.2\](1) then ${\frak K}^1
\le {\frak K}^2$ (use $H_1 = H_2 =$ the identity).
3. For $c_1>1$ we have ${\frak K}^{mt(c_1)}\equiv {\frak K}^{mt[c_1]}\equiv
{\frak K}^{ms[c_1]}\equiv {\frak K}^{ms(c_1)]}$.
4. ${\frak K}^{tr[\omega]} \le {\frak K}^{rs(p)}$.
5. ${\frak K}^{tr[\omega]} \le {\frak K}^{tr(\omega)}$.
6. ${\frak K}^{tr(\omega)} \le {\frak K}^{rs(p),\mbox{pure}}$.
1) Check.\
2) Check.\
3) Choose $n(*)<\omega$ large enough and ${\frak K}^1,{\frak K}^2$ any two of the four. We define $H_1,H_2$ as follows. $H_1$ is the identity. For $(A,d)
\in{\frak K}^\ell$ let $H_\ell((A,d))=(A,d^{[\ell]})$ where $d^{[\ell]}(x,y)=
\inf\{1/(n+n(*)):2^{-n}\ge d(x,y)\}$ (the result is not necessarily a metric space, $n(*)$ is chosen so that the semi-metric inequality holds). The point is to check clause (c) of \[9.2\](2); so assume $f$ is a function which ${\frak K}^2$-embeds $H_1((A_1,d_1))$ into $(A_2,d_2)$; but $$H_1((A_1,d_1))=(A_1,d_1),\quad H_2((A_2,d_2))=(A_2,d^{[2]}_2),$$ so it is enough to check that $f$ is a function which ${\frak K}^1$-embeds $(A_1,d^{[1]}_1)$ into $(A_2,d^{[2]}_2)$ i.e. it is one-to-one (obvious) and preserves limit (check).\
4) For $M=(A,E_n)_{n<\omega}\in {\frak K}^{tr[\omega]}$, without loss of generality $A\subseteq {}^\omega\lambda$ and $$\eta E_n\nu\qquad\Leftrightarrow\qquad\eta\in A\ \&\ \nu\in A\ \&\
\eta\restriction n=\nu\restriction n.$$ Let $B^+=\{\eta\restriction n:\eta\in A\mbox{ and } n<\omega\}$. We define $H_1(M)$ as the (Abelian) group generated by $$\{x_\eta:\eta\in A\cup B\}\cup\{y_{\eta,n}:\eta\in A,n<\omega\}$$ freely except $$\begin{array}{rcl}
p^{n+1}x_\eta=0 &\quad\mbox{\underbar{if}}\quad &\eta\in B, \ell g(\eta)=n\\
y_{\eta,0}=x_\eta &\quad\mbox{\underbar{if}}\quad &\eta\in A\\
py_{\eta,n+1}-y_\eta=x_{\eta\restriction n} &\quad\mbox{\underbar{if}}\quad
&\eta\in A, n<\omega\\
p^{n+1}y_{\eta, n}=0 &\quad\mbox{\underbar{if}}\quad &\eta\in B, n<\omega.
\end{array}$$ For $G\in {\frak K}^{rs(p)}$ let $H_2(G)$ be $(A,E_n)_{n<\omega}$ with: $$A = G,\quad xE_ny\quad\underbar{iff}\quad G\models\mbox{``}p^n\mbox{ divides
}(x-y)\mbox{''.}$$ $H_2(G)\in {\frak K}^{tr[\omega]}$ as “$G$ is separable" implies $(\forall
x)(x \ne 0\ \Rightarrow\ (\exists n)[x\notin p^nG])$. Clearly clauses (a), (b) of Definition \[9.1\](2) hold. As for clause (c), assume $(A,E_n)_{n<\omega}
\in {\frak K}^{tr[\omega]}$. As only the isomorphism type counts without loss of generality $A\subseteq {}^\omega\lambda$. Let $B=\{\eta\restriction n:n<
\omega:\eta\in A\}$ and $G=H_1((A,E_n)_{n<\omega})$ be as above. Suppose that $f$ embeds $G$ into some $G^*\in {\frak K}^{rs(p)}$, and let $(A^*,E^*_n)_{n<
\omega}$ be $H_2(G^*)$. We should prove that $(A,E_n)_{n<\omega}$ is embeddable into $(A^*,E^*_n)$.\
Let $f^*:A\longrightarrow A^*$ be $f^*(\eta)=x_\eta\in A^*$. Clearly $f^*$ is one to one from $A$ to $A^*$; if $\eta E_n \nu$ then $\eta\restriction n=\nu
\restriction n$ hence $G \models p^n \restriction (x_\eta-x_\nu)$ hence $(A^*,A^*_n)_{n<\omega}\models\eta E^*_n\nu$. $\square_{\ref{9.3}}$
\[9.3A\] In \[9.3\](4) we can prove ${\frak K}^{tr[\omega]}_{\bar\lambda}\le{\frak
K}^{rs(p)}_{\bar\lambda}$.
\[9.4\]
1. ${\frak K}^{mt} \equiv {\frak K}^{mt(c)}$ for $c \ge 1$.
2. ${\frak K}^{mt} \equiv {\frak K}^{ms[c]}$ for $c > 1$.
1) Let $H_1:{\frak K}^{mt}\longrightarrow {\frak K}^{mt(c)}$ be the identity. Let $H_2:{\frak K}^{mt(c)}\longrightarrow {\frak K}^{mt}$ be defined as follows:\
$H_2((A,d))=(A,d^{mt})$, where $$\begin{array}{l}
d^{mt}(y,z)=\\
\inf\big\{\sum\limits^n_{\ell=0} d(x_\ell,x_{\ell,n}):n<\omega\ \&\
x_\ell\in A\mbox{ (for $\ell\le n$)}\ \&\ x_0=y\ \&\ x_n=z\big\}.
\end{array}$$ Now
$(*)_1$
: $d^{mt}$ is a two-place function from $A$ to ${\Bbb
R}^{\ge 0}$, is symmetric, $d^{mt}(x,x)=0$ and it satisfies the triangular inequality.
This is true even on ${\frak K}^{mt(c)}$, but here also
$(*)_2$
: $d^{mt}(x,y) = 0 \Leftrightarrow x=0$.
\[Why? As by the Definition of ${\frak K}^{mt[c]},d^{mt}(x,y)\ge{\frac 1c}
d(x,y)$. Clearly clauses (a), (b) of \[9.2\](2) hold.\]\
Next,
$(*)_3$
: If $M_1,N\in {\frak K}^{mt}$, $f$ is an embedding (for ${\frak
K}^{mt}$) of $M_1$ into $N$ then $f$ is an embedding (for ${\frak K}^{mt[c]}$) of $H_1(M)$ into $H_1(N)$
\[why? as $H_1(M)=M$ and $H_2(N)=N$\],
$(*)_4$
: If $M,N\in {\frak K}^{mt[c]}$, $f$ is an embedding (for ${\frak K}^{mt[c]}$) of $M$ into $N$ then $f$ is an embedding (for ${\frak
K}^{mt}$) of $H_2(M)$ into $H_1(M)$
\[why? as $H^*_\ell$ preserves $\lim\limits_{n\to\infty} x_n=x$ and $\lim\limits_{n\to\infty} x_n\ne x$\].
So two applications of \[9.2\] give the equivalence.\
2) We combine $H_2$ from the proof of (1) and the proof of \[9.3\](3). $\square_{\ref{9.4}}$
\[9.6\]
1. If $\bigwedge\limits_n\mu_n=\aleph_0$ let $$\hspace{-0.5cm}\begin{array}{ll}
J^{mt}=J^{mt}_{\bar\mu}=\big\{A\subseteq\prod\limits_{n<\omega}\mu_n:&
\mbox{for every } n \mbox{ large enough, } \\
\ & \mbox{ for every }\eta\in\prod\limits_{\ell
<n}\mu_\ell\\
\ &\mbox{the set }\{\eta'(n):\eta\triangleleft\eta'\in A\}\mbox{ is finite}
\big\}.
\end{array}$$
2. Let $T=\bigcup\limits_{\alpha\le\omega}\prod\limits_{n<\alpha}\mu_n$, $(T,d^*)$ be a metric space such that $$\prod_{\ell<n}\mu_\ell\cap\mbox{closure}\left(\bigcup_{m<n}\prod_{\ell<m}
\mu_\ell\right)=\emptyset;$$ now $$\begin{array}{rl}
I^{mt}_{(T,d^*)}=:\big\{A\subseteq\prod\limits_{n<\omega}\mu_n:&\mbox{ for
some }n,\mbox{ the closure of } A\mbox{ (in $(T,d^*)$)}\\
&\mbox{ is disjoint to }\bigcup\limits_{m\in [n,\omega)}\prod\limits_{\ell
<m} \mu_\ell\big\}.
\end{array}$$
3. Let $H\in {\frak K}^{rs(p)}$, $\bar H=\langle H_n:n<\omega\rangle$, $H_n
\subseteq H$ pure and closed, $n<m\ \Rightarrow\ H_n\subseteq H_m$ and $\bigcup\limits_{n<\omega} H_n$ is dense in $H$. Let $$\begin{array}{ll}
I^{rs(p)}_{H,\bar H}=:\big\{A\subseteq H:&\mbox{for some } n\mbox{ the closure
of }\langle A\rangle_H\mbox{ intersected with}\\
&\bigcup\limits_{\ell<\omega}H_\ell\mbox{ is included in }H_n\big\}.
\end{array}$$
\[9.5\] Suppose that $2^{\aleph_0}<\mu$ and $\mu^+<\lambda={{\rm cf}}(\lambda)<
\mu^{\aleph_0}$ and
$(*)_\lambda$
: ${{\bf U}}_{J^{mt}_{\bar\mu}}(\lambda)=\lambda$ or at least ${{\bf U}}_{J^{mt}_{\bar\mu}}(\lambda)<\lambda^{\aleph_0}$ for some $\bar\mu=\langle
\mu_n:n<\omega\rangle$ such that $\prod\limits_{n<\omega}\mu_n<\lambda$.
Then ${\frak K}^{mt}_\lambda$ has no universal member.
\[9.7\]
1. $J^{mt}$ is $\aleph_1$-based.
2. The minimal cardinality of a set which is not in the $\sigma$-ideal generated by $J^{mt}$ is ${\frak b}$.
3. $I^{mt}_{(T,d^*)},I^{rs(p)}_{H,\bar H}$ are $\aleph_1$-based.
4. $J^{mt}$ is a particular case of $I^{mt}_{(T,d^*)}$ (i.e. for some choice of $(T,d^*)$).
5. $I^0_{\bar \mu}$ is a particular case of $I^{rs(p)}_{H,\bar H}$.
of \[9.5\]. Let $$\begin{array}{ll}
T_\alpha=\{(\eta, \nu)\in{}^\alpha\lambda\times {}^\alpha(\omega+
1):& \mbox{ for every }n\mbox{ such that }n+1< \alpha\\
\ & \mbox{ we have
}\nu(n)< \omega\}
\end{array}$$ and for $\alpha\le\omega$ let $T=\bigcup\limits_{\alpha\le
\omega}T_\alpha$. We define on $T$ the relation $<_T$: $$(\eta_1,\nu_1)\le(\eta_1,\nu_2)\quad\mbox{ iff }\quad\eta_1\trianglelefteq
\eta_2\ \&\ \nu_1\triangleleft\nu_2.$$ We define a metric:\
if $(\eta_1,\nu_1)\ne(\eta_2,\nu_2)\in T$ and $(\eta,\nu)$ is their maximal common initial segment and $(\eta,\nu)\in T$ then necessarily $\alpha=
\ell g((\eta,\nu))<\omega$ and we let:
> if $\eta_1(\alpha)\ne\eta_2(\alpha)$ then $$d\left((\eta_1,\nu_1),(\eta_2,\nu_2)\right)=2^{-\sum\{\nu(\ell):\ell<
> \alpha\}},$$ if $\eta_1(\alpha)=\eta_2(\alpha)$ (so $\nu_1(\alpha)\ne\nu_2(\alpha)$ then $$d\left((\eta_1,\nu_1),(\eta_2,\nu_2)\right)=2^{-\sum\{\nu(\ell):\ell<\alpha
> \}}\times 2^{-\min\{\nu_1(\alpha),\nu_2(\alpha)\}}.$$
Now, for every $S\subseteq\{\delta<\lambda:{{\rm cf}}(\delta)=\aleph_0\}$, and $\bar
\eta=\langle\eta_\delta:\delta\in S\rangle$, $\eta_\delta\in {}^\omega
\delta$, $\eta_\delta$ increasing let $M_\eta$ be $(T,d)\restriction A_{\bar
\eta}$, where $$A_{\bar\eta}=\bigcup_{n<\omega} T_n\cup\{(\eta_\delta,\nu):\delta\in S,\;\nu
\in {}^\omega\omega\}.$$ The rest is as in previous cases (note that $\langle(\eta\char 94\langle
\alpha \rangle,\nu\char 94\langle n\rangle):n<\omega\rangle$ converges to $(\eta\char 94\langle\alpha\rangle,\nu\char 94\langle\omega\rangle)$ and even if $(\eta\char 94\langle \alpha\rangle, \nu\char 94\langle
n\rangle)\leq (\eta_n, \nu_n)\in T_\omega$ then $\langle(\eta_n,
\nu_n): n<\omega\rangle$ converge to $(\eta\char 94 \langle
\alpha\rangle, \nu\char 94\langle \omega\rangle)$). $\square_{\ref{9.7}}$
\[9.8\] If ${{\rm IND}}_{\chi'}(\langle\mu_n:n<\omega\rangle)$, then $\prod\limits_{n<\omega}
\mu_n$ is not the union of $\le\chi$ members of $I^0_{\bar\mu}$ (see Definition \[5.5A\] and Theorem \[5.5\]).
Suppose that $A_\zeta=\{\sum\limits_{n<\omega}
p^nx^n_{\alpha_n}:\langle \alpha_n:n<\omega\rangle\in X_\zeta\}$ and $\alpha_n<\mu_n$ are such that if $\sum
p^nx^n_{\alpha_n}\in A_\zeta$ then for infinitely many $n$ for every $k<\omega$ there is $\langle \beta_n:n<\omega\rangle$, $$(\forall\ell<k)[\alpha_\ell=\beta_\ell\ \ \Leftrightarrow\ \ \ell=n]\qquad
\mbox{ and }\qquad\sum_{n<\omega}p^nx^n_{\beta_n}\in A_\zeta\ \ \mbox{ (see
\S5).}$$ This clearly follows. $\square_{\ref{9.8}}$
On Modules
==========
Here we present the straight generalization of the one prime case like Abelian reduced separable $p$-groups. This will be expanded in [@Sh:622] (including the proof of \[10.4new\]).
\[10.1\]
(A)
: $R$ is a ring, $\bar{\frak e}=\langle{\frak e}_n:n<\omega\rangle$, ${\frak e}_n$ is a definition of an additive subgroup of $R$-modules by an existential positive formula (finitary or infinitary) decreasing with $n$, we write ${\frak e}_n(M)$ for this additive subgroup, ${\frak e}_\omega(M)=
\bigcap\limits_n {\frak e}_n(M)$.
(B)
: ${\frak K}$ is the class of $R$-modules.
(C)
: ${\frak K}^*\subseteq {\frak K}$ is a class of $R$-modules, which is closed under direct summand, direct limit and for which there is $M^*$, $x^*
\in M^*$, $M^*=\bigoplus\limits_{\ell\le n} M^*_\ell\oplus M^{**}_n$, $M^*_n
\in {\frak K}$, $x^*_n\in {\frak e}_n(M^*_n)\setminus {\frak e}_{n+1}(M^*)$, $x^*-\sum\limits_{\ell<n} x^*_\ell\in {\frak e}_n(M^*)$.
\[10.2\] For $M_1,M_2\in {\frak K}$, we say $h$ is a $({\frak K},\bar {\frak
e})$-homomorphism from $M_1$ to $M_2$ if it is a homomorphism and it maps $M_1
\setminus {\frak e}_\omega(M_1)$ into $M_2\setminus {\frak
e}_\omega(M_2)$; we say $h$ is an $\bar {\frak e}$-pure homomorphism if for each $n$ it maps $M_1\setminus {\frak e}_n(M_1)$ into $M_2\setminus {\frak
e}_n(M_2)$.
\[10.3\]
1. Let $H_n \subseteq H_{n+1} \subseteq H$, $\bar H=\langle H_n:n<\omega
\rangle$, $c\ell$ is a closure operation on $H$, $c\ell$ is a function from ${{\cal}P}(H)$ to itself and $$X \subseteq c \ell(X) = c \ell(c \ell(X)).$$ Define $$I_{H,\bar H,c\ell}=\big\{A\subseteq H:\mbox{for some }k<\omega\mbox{ we have
} c\ell(A)\cap\bigcup_{n<\omega} H_n\subseteq H_k\big\}.$$
2. We can replace $\omega$ by any regular $\kappa$ (so $H=\langle H_i:i<
\kappa\rangle$).
\[10.4new\] Assume $|R|+\mu^+< \lambda = {{\rm cf}}(\lambda)< \mu^{\aleph_0}$, then for every $M\in {\frak K}_\lambda$ there is $N\in {\frak K}_\lambda$ with no $\bar {\frak e}$-pure homomorphism from $N$ into $M$.
In the interesting cases $c\ell$ has infinitary character.\
The applications here are for $\kappa=\omega$. For the theory, $pcf$ is nicer for higher $\kappa$.
Open problems
=============
1. If $\mu^{\aleph_0}\ge\lambda$ then any $(A,d)\in {\frak K}^{mt}_\lambda$ can be embedded into some $M'\in {\frak K}^{mt}_\lambda$ with density $\le\mu$.
2. If $\mu^{\aleph_0}\ge\lambda$ then any $(A,d)\in {\frak K}^{ms}_\lambda$ can be embedded into some $M'\in {\frak K}^{ms}_\lambda$ with density $\le\mu$.
<!-- -->
1. Other inclusions on ${{\rm Univ}}({\frak K}^x)$ or show consistency of non inclusions (see §9).
2. Is ${\frak K}^1\le {\frak K}^2$ the right partial order? (see §9).
3. By forcing reduce consistency of ${{\bf U}}_{J_1}(\lambda)>\lambda+
2^{\aleph_0}$ to that of ${{\bf U}}_{J_2}(\lambda)>\lambda+2^{\aleph_0}$.
<!-- -->
1. The cases with the weak ${{\rm pcf}}$ assumptions, can they be resolved in ZFC? (the $pcf$ problems are another matter).
2. Use [@Sh:460], [@Sh:513] to get ZFC results for large enough cardinals.
If $\lambda^{\aleph_0}_n<\lambda_{n+1}$, $\mu=\sum\limits_{n<\omega}
\lambda_n$, $\lambda=\mu^+<\mu^{\aleph_0}$ can $(\lambda,\lambda,1)$ belong to ${{\rm Univ}}({\frak K})$? For ${\frak K}={\frak K}^{tr},{\frak K}^{rs(p)},{\frak
K}^{trf}$?
1. If $\lambda=\mu^+$, $2^{<\mu}=\lambda<2^\mu$ can $(\lambda,\lambda,1)
\in {{\rm Univ}}({\frak K}^{\mbox{or}}=$ class of linear orders)?
2. Similarly for $\lambda=\mu^+$, $\mu$ singular, strong limit, ${{\rm cf}}(\mu)=
\aleph_0$, $\lambda<\mu^{\aleph_0}$.
3. Similarly for $\lambda=\mu^+$, $\mu=2^{<\mu}=\lambda^+ <2^\mu$.
<!-- -->
1. Analyze the existence of universal member from ${\frak
K}^{rs(p)}_\lambda$, $\lambda<2^{\aleph_0}$.
2. §4 for many cardinals, i.e. is it consistent that: $2^{\aleph_0}> \aleph_\omega$ and for every $\lambda< 2^{\aleph_0}$ there is a universal member of ${\frak K}^{rs(p)}_\lambda$?
<!-- -->
1. If there are $A_i\subseteq\mu$ for $i<2^{\aleph_0}$, $|A_i\cap A_j|<
\aleph_0$, $2^\mu=2^{\aleph_0}$ find forcing adding $S\subseteq [{}^\omega
\omega]^\mu$ universal for $\{(B, \vartriangleleft):
{}^{\omega>}\omega \subseteq B\subseteq {}^{\omega\geq}\omega, |B|\leq
\lambda\}$ under (level preserving) natural embedding.
For simple countable $T$, $\kappa=\kappa^{<\kappa}<\lambda\subseteq \kappa$ force existence of universal for $T$ in $\lambda$ still $\kappa=\kappa^{<
\kappa}$ but $2^\kappa=\chi$.
Make [@Sh:457 §4], [@Sh:500 §1] work for a larger class of theories more than simple.
See on some of these problems [@DjSh:614], [@Sh:622].
[^1]: I would like to thank Alice Leonhardt for the beautiful typing.\
The research was supported by the German-Israeli Foundation for Scientific Research & Development, Grant No. G-294.081.06/93 .\
Publication No 552.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Methods derived from the generalized quantum master equation (GQME) framework have provided the basis for elucidating energy and charge transfer in systems ranging from molecular solids to photosynthetic complexes. Recently, the non-perturbative combination of the GQME with quantum-classical methods has resulted in approaches whose accuracy and efficiency exceed those of the original quantum-classical schemes while offering significant accuracy improvements over perturbative expansions of the GQME. Here we show that, while the non-Markovian memory kernel required to propagate the GQME scales quartically with the number of subsystem states, the number of trajectories required scales at most quadratically when using quantum-classical methods to construct the kernel. We then present an algorithm that allows further acceleration of the quantum-classical GQME by providing a way to selectively sample the kernel matrix elements that are most important to the process of interest. We demonstrate the utility of these advances by applying the combination of Ehrenfest mean field theory with the GQME (MF-GQME) to models of the Fenna-Matthews-Olson (FMO) complex and the light harvesting complex II (LHCII), with 7 and 14 states, respectively. This allows us to show that MF-GQME is able to accurately capture all the relevant dynamical time-scales in LHCII: the initial nonequilibrium population transfer on the femtosecond time-scale, the steady state-type trapping on the picosecond time-scale, and the long time population relaxation. Remarkably, all of these physical effects spanning tens of picoseconds can be encoded in a memory kernel that decays after only $\sim$65 fs.'
author:
- 'William C. Pfalzgraff'
- 'Andrés Montoya-Castillo'
- Aaron Kelly
- 'Thomas E. Markland'
bibliography:
- 'mendeley.bib'
- 'other\_references.bib'
title: 'Efficient construction of generalized master equation memory kernels for multi-state systems from nonadiabatic quantum-classical dynamics'
---
Introduction
============
Theoretical modelling of processes ranging from energy transfer in photosynthetic light harvesting complexes and organic electronics to proton coupled electron transfer, require the development of methods that allow for the accurate and efficient calculation of the dynamics of open quantum systems, i.e., a subsystem of interest coupled to an external environment. Generalized quantum master equations (GQMEs) provide a formally exact way of rewriting the equation of motion of a set of degrees of freedom of particular interest via the use of the appropriate projection operator [@Nakajima1958; @Zwanzig1960a; @berne_book]. For example, in open quantum systems, the projection operator is frequently chosen to encompass all the states of the electronic subsystem, although this choice need not align with simple partitions at the level of the Hamiltonian, i.e., the projector could contain only a subset of electronic states or even include some nuclear degrees of freedom. The effect of the remaining degrees of freedom on the dynamics of those of interest is encoded through the memory kernel.[@Nakajima1958; @Zwanzig1960a; @Mori1965] However, evaluating the memory kernel requires overcoming the complications that arise from the presence of the projected propagator.
A common approach to avoid the projected propagator is to treat the memory kernel using second order perturbation theory. This leads to methods including Markovian and non-Markovian Redfield theory and its variants [@Bloch1957; @Redfield1965; @Chang1993; @mukamel_redfield], the noninteracting blip approximation (NIBA) [@Leggett1987; @Dekker1987], Förster theory [@Forster1960], and the polaronic quantum master equation [@Jang2008; @Jang2011a; @Nazir2009; @McCutcheon2011], which have been used to treat a wide variety of multi-state electronic energy transfer systems. In these approaches, the memory kernel depends only on correlation functions of the isolated system and bath, and hence can be straightforwardly generated even for systems with many subsystem states.[@BreuerPetruccione; @NitzanBook] However, these approaches are limited in their applicability to systems where the perturbative parameter remains small.
More recently, it has been shown that the projected propagator can be exactly removed by employing the Dyson identity, leading to a representation of the full (non-perturbative, non-Markovian) memory kernel in terms of *unprojected* correlation functions of the system and bath.[@Shi2003] This allows one to gain accurate results over a much wider range of regimes than perturbative master equations, albeit at the expense of requiring cross-correlation functions of system and bath operators to generate the memory kernel. Evaluating these cross-correlation functions using trajectory-based quantum-classical approaches with the GQME formalism has been shown to give considerable increases in both accuracy and efficiency compared to using the same quantum-classical method alone.[@Shi2004a; @Kelly2013; @Kelly2015; @Pfalzgraff2015; @Montoya2016a; @KellyMontoya2016; @Montoya2017b]
However, while these studies showed the efficacy of such a combination in systems containing two subsystem states, for a system with $N_{s}$ subsystem states, the reduced density matrix (RDM) has $N_{s}^2$ elements and hence the size of the memory kernel needed to propagate it grows as $N_{s}^4$. One might therefore imagine that combining quantum-classical methods with the GQME would be limited to problems with small numbers of subsystem states. Here, we demonstrate that for quantum-classical methods that treat the subsystem using wavefunctions, e.g., Ehrenfest mean field theory (MFT) [@mft; @mft2; @mft3] and fewest switches surface hopping (FSSH) [@tully1; @tully2; @Hammes-Schiffer1996], that the full memory kernel can be obtained using only $\mathcal{O}(N_{s}^2)$ trajectories.
We then show how, in cases where one is interested in a particular (or a small set of) subsystem initial conditions, one can further improve the efficiency by focusing resources into accurately capturing the matrix elements of the memory kernel that most strongly determine the evolution of a given initial condition. We achieve this by noting that, while the fully converged memory kernel contains the information necessary propagate the GQME from any subsystem initial condition contained in the projector, often many of these may not be of interest. Here we introduce an algorithm that exploits this observation to identify and selectively converge the kernel elements that are most important to capture the relaxation from a given subsystem initial condition. This allows one to minimize the computational effort expended on the elements that only exert a minor influence on the process of interest.
We then illustrate the scalability of the quantum-classical GQME to multi-state systems by applying the mean field GQME (MF-GQME) method [@Kelly2015] to treat the 7-state Fenna-Matthews-Olsen (FMO) complex [@FMO1; @Ishizaki2009b] and the 14-state light harvesting complex II (LHCII).[@Novoderezhkin2005; @lhcii_model2; @lhcii_model3] We show that MF-GQME gives quantitative accuracy in regimes where direct MFT and modified Redfield theory fail, while remaining less expensive than a direct application of MFT even in these larger systems. This development thus opens the door to the application of quantum-classical GQME approaches to multi-state systems coupled to atomistic environments that lie beyond linear system-bath coupling and harmonic environments.[@Pfalzgraff2015]
Theory {#sec:theory}
======
To show how the GQME formalism can be efficiently used with trajectory-based quantum dynamics methods in multi-state systems, here we give the general expression for the memory kernel for a system with arbitrary system-bath coupling. By analyzing the form of this general expression, we demonstrate that only $\mathcal{O}(N_s^2)$ trajectories are required to calculate the memory kernel in the MF-GQME method in a system with $N_s$ subsystem states.
To begin, we consider a Hamiltonian of the general form, $$\label{eq:general-open-quantum-system-hamiltonian}
\hat{H} = \hat{H}_s + \hat{H}_b + \hat{H}_{sb},$$ where $\hat{H}_s$ is the isolated subsystem Hamiltonian, $\hat{H}_b$ the isolated bath Hamiltonian, and $\hat{H}_{sb}$ contains all the terms that couple the system and the bath. For any $\hat{H}_{sb}$, one can write it such that it is in the form of a sum of direct products of system and bath operators, $$\hat{H}_{sb} = \sum_{j=1}^{N_{sb}} \hat{S}_j \otimes \hat{\Gamma}_j,
\label{eq:general_H_sb}$$ where $\hat{S}_j$ is a pure subsystem operator, $\hat{\Gamma}_j$ is a pure bath operator, and $N_{sb}$ is the number of terms in the sum.
Here we consider initial conditions for the full system density operator $\hat{\rho}(0)$ that are of the spectroscopic form, $\hat{\rho}(0) = \hat{\rho}^{eq}_b \otimes \hat{\rho}_s(0)$, where $\hat{\rho}_s(0)$ encodes the initial state of the subsystem and $\hat{\rho_b}^{eq}$ is the equilibrium bath density operator, $\hat{\rho}^{eq}_b = e^{-\beta \hat{H}_b} / \mathrm{Tr}_b \left\{ e^{-\beta \hat{H}_b}\right\}$. While this choice of initial condition simplifies the expressions for the memory kernel, we emphasize that the scaling arguments presented here do not depend on this choice and are straightforward to generalize to any factorizable initial condition.
The quantity of interest is the RDM, $\hat{\rho}_s(t) = \mathrm{Tr}_b\{ \hat{\rho} (t) \}$, from which any subsystem observable, such as the population relaxation or electronic spectra, can be generated. In a given basis, one can write the RDM as $$\hat{\rho}_s(t) = \sum_{n=1}^{N_s^2} [\rho_s]_n (t) \hat{A}_n,$$ where $\hat{A}_n \in \left\{ \ket{\alpha}\bra{\alpha^{\prime}} \right\}$, which is a set of $N_s^2$ subsystem operators that span the subsystem Liouville space, and $[\rho_s]_n$ denotes a particular matrix element of the RDM. These matrix elements can be expressed as quantum mechanical expectation values, $$\left[ \hat{\rho}_s(t) \right]_n = \mathrm{Tr} \left\{ \hat{\rho}(0) e^{i\mathcal{L}t/\hbar}\hat{A}_n \right\} = \mathrm{Tr}\left\{\hat{\rho}(t) \hat{A}_n \right\}.$$ In the Mori-Nakajima-Zwanzig formalism, the dynamics of the RDM can be obtained by integrating the GQME [@Zwanzig1960a; @Nakajima1958; @Mori1965; @berne_book; @BreuerPetruccione], $$\dot{ \rho}_s(t)=\rho_s(t)\mathcal{X} -\int^t_0 d\tau \ \rho_s(t-\tau)\mathcal{K}(\tau) ,
\label{eq:rdm_evol}$$ where $\rho_s(t)$ is a row vector with $N_s^2$ time-dependent elements and $\mathcal{X}$ and $\mathcal{K}(t)$ are matrices of size $N_s^2 \times N_s^2$, and $\dot{ \rho}_s(t) = \frac{d}{dt} \rho_s(t)$. The static matrix $\mathcal{X}_{mn}$ = $\frac{i}{\hbar} \mathrm{Tr_s}\left\{ \hat{A}_m^{\dagger}\mathcal{L}_s \hat{A}_n \right\}$ ensures that the GQME exactly recovers the free subsystem evolution when the memory kernel, $\mathcal{K}(t)$, is zero, i.e. when the subsystem and bath do not interact. The memory kernel, which encodes the effect of the bath on the subsystem dynamics, can be constructed using correlation functions in the full system and bath Hilbert space. Using the Dyson decomposition of the propagator[@evans_moriss] one obtains,[@Shi2003; @Montoya2016a] $$\mathcal{K}(t) = \mathcal{K}_1(t) - i\int_0^t d\tau \ \mathcal{K}_3(t-\tau)\mathcal{K}(\tau).
\label{eq:K_K1_K3}$$ Here $\mathcal{K}_3$ and $\mathcal{K}_1$ are auxiliary kernels, which can be obtained using one’s choice of dynamical method, and the subscripts 1 and 3 are used to be consistent with earlier work.[@Shi2003; @Shi2004a; @Kelly2013; @Kelly2015; @Pfalzgraff2015; @Montoya2016a] For any system-bath coupling, $\mathcal{K}_3$ and $\mathcal{K}_1$ can be written as linear combinations of correlation functions of system and bath operators of the form (see Appendix \[app:auxiliary-kernels\]), $$\begin{aligned}
\mathcal{K}_3(t) &= \frac{1}{\hbar}\sum_{j=1}^{N_{sb}} \Big[ c^{(j,-)} q_{3}^{(j,+)}(t) + c^{(j,+)} q_{3}^{(j,-)}(t)\Big],
\label{eq:general_K3}\\
\mathcal{K}_1(t) &= \frac{1}{\hbar^2}\sum_{j,k=1}^{N_{sb}} \Big[ c^{(j,-)} q_{1}^{(jk,+)}(t)c^{(k,-)} \nonumber \\
&\qquad \qquad \qquad \quad + c^{(j,+)} q_{1}^{(jk,-)}(t)c^{(k,-)} \Big].
\label{eq:general_K1}\end{aligned}$$ Here, like the memory kernels, $c^{(j,\pm)}$, $q_{3}^{(j,\pm)}(t)$, and $q_{1}^{(jk,\pm)}(t)$ are matrices of size $N_s^2 \times N_s^2$. The transformation matrices $c^{(j,\pm)}$ are time-independent and simple to evaluate analytically, with their elements given by, $$c_{mn}^{(j,\pm)} = \mathrm{Tr_s} \left\{ \hat{A}_m^{\dagger} \left[\hat{S}_j,\hat{A}_n\right]_{\pm} \right\},
\label{eq:transformation_elements}$$ where $[\cdot, \cdot]_{-}$ denotes the commutator and $[\cdot, \cdot]_{+}$ the anti-commutator. The matrix elements of the correlation functions $q_{3}^{(j,\pm)}(t)$ and $q_{1}^{(jk,\pm)}(t)$ are (see Appendix \[app:auxiliary-kernels\]) $$\begin{aligned}
\label{eq:general_q3}
q_{3,mn}^{(j,\pm)}(t) &= \frac{1}{2}\mathrm{Tr}\left\{ \left[ \hat{\rho}_{b}^{eq}, \hat{\Gamma}_j\right]_{\pm} \hat{A}_m^{\dagger} e^{i\mathcal{L}t/\hbar} \hat{A}_n \right\} ,\\
\label{eq:general_q1}
q_{1,mn}^{(jk,\pm)}(t) &= \frac{1}{2}\mathrm{Tr}\left\{ \left[ \hat{\rho}_{b}^{eq}, \hat{\Gamma}_j\right]_{\pm} \hat{A}_m^{\dagger} e^{i\mathcal{L}t/\hbar} \hat{\Gamma}_{k} \hat{A}_n \right\}. \end{aligned}$$ Equations (\[eq:general\_K3\]) and (\[eq:general\_K1\]) provide a formal basis for the calculation of the memory kernel, in which one obtains the auxiliary kernels $\mathcal{K}_3$(t) and $\mathcal{K}_1$(t) by evaluating $c^{(j,\pm)}$ exactly and calculating $q_{3}^{(j,\pm)}(t)$ and $q_{1}^{(jk,\pm)}(t)$ using one’s choice of exact or approximate dynamical method. These quantities are then used to to obtain the auxiliary kernels $\mathcal{K}_1$ and $\mathcal{K}_3$ using Eqs.(\[eq:general\_K3\])–(\[eq:general\_K1\]), and the full memory kernel is obtained using Eq. (\[eq:K\_K1\_K3\]) (see Appendix \[app:K\_K1\_K3\]) In Sec. \[ssec:scaling\], we use these expressions to analyze the scaling with the number of subsystem states.
Scaling of the MF-GQME method with number of subsystem states {#ssec:scaling}
-------------------------------------------------------------
In the MF-GQME method, one approximates the correlation functions defined in Eqs. (\[eq:general\_q3\]) and (\[eq:general\_q1\]) using MFT (see Appendix \[app:mft\_dynamical\_matrices\]). The memory kernel can then be used to solve the GQME for the dynamics of the RDM, given by Eq. (\[eq:rdm\_evol\]). In practice, the computational cost of generating the full memory kernel using Eqs. (\[eq:K\_K1\_K3\])–(\[eq:general\_K1\]) and integrating the GQME is negligible compared to the cost of generating the MFT trajectories necessary to calculate $\mathcal{K}_3$ and $\mathcal{K}_1$. The computational cost is therefore determined by the number of MFT trajectories required to evaluate $q_{3}^{(j,\pm)}(t)$ and $q_{1}^{(jk,\pm)}(t)$, and how this cost scales with the number of subsystem states.
The number of correlation functions that need to be calculated to evaluate $q_{3}^{(j,\pm)}(t)$ and $q_{1}^{(jk,\pm)}(t)$ using Eqs. (\[eq:general\_q3\]) and (\[eq:general\_q1\]) can at first appear formidable, as the indices $m$ and $n$ in those equations each run over $N_s^2$ elements. This might lead one to erroneously conclude that the MF-GQME method scales as $\mathcal{O}(N_s^4)$, which would severely limit its applicability to treat problems with many subsystem states. However, closer inspection of these expressions in the context of MFT dynamics reveals that one only needs to sample the distinct electronic initial conditions, labelled by the index $m$, to calculate *all* the necessary correlation functions. This is because the MFT trajectories generated from a given electronic initial condition, $\hat{A}_m^{\dagger}$, provide the time-dependent observables necessary to calculate all of the correlation functions corresponding to that initial condition. Consequently, since only $N_s^2$ electronic initial conditions need to be sampled, only $\mathcal{O}(N_s^2)$ quantum-classical trajectories are required. In practice, for MFT the number of initial conditions can be shown to be $\frac{1}{2} N_s(N_s+1)$ by following the procedure outlined in Appendix \[App:off\_diagonal\_inicon\], which exploits the fact that the correlation functions corresponding to a given off-diagonal initial condition and its Hermitian conjugate can be obtained from the same set of trajectories.
In general, evaluating forces dominates the overall cost, especially in atomistic systems or systems with a large number of bath degrees of freedom. However, the calculation of the correlation functions required scales as $\mathcal{O}(N_s^4)$ and integration of the GQME scales as $\mathcal{O}(N_s^6)$ which could dominate the cost for large $N_s$. However, even for the systems studied here which use very simple forms of the bath, the cost of integrating the GQME is negligible.
An additional concern regarding the scalability of the MF-GQME approach comes from amount of memory required to store all elements in $q_{3,mn}^{(j,\pm)}(t)$ and $q_{1,mn}^{(j,\pm)}(t)$. For instance, the number of time-dependent elements in $q_{1,mn}^{(j,\pm)}(t)$ can be calculated as the product of $N_s^2$ initial electronic conditions $\hat{A}_m^{\dagger}$, $N_{sb}$ bath operators $\hat{\Gamma}_j$ measured at $t=0$, and $N_s^2$ electronic operators $\hat{A}_n(t)$ and $N_{sb}$ bath operators $\hat{\Gamma}_j(t)$ measured at finite times. This corresponds to $\mathcal{O}(N_{s}^4 \times N_{sb}^2)$ time dependent elements that need to be stored in memory, and hence one might expect the amount of available memory to be a significant limitation. In practice, however, this problem can be easily circumvented because $q_{3}^{(j,\pm)}(t)$ and $q_{1}^{(jk,\pm)}(t)$ do not themselves need to be stored in memory. Instead, one can evaluate their contributions to $\mathcal{K}_3(t)$ and $\mathcal{K}_1(t)$ on the fly using Eqs. (\[eq:general\_K3\]) and (\[eq:general\_K1\]), so that only $\mathcal{K}_3(t)$ and $\mathcal{K}_1(t)$ need to be stored. In addition, because not all of the matrix elements are independent, there are only $\frac{1}{2}N_s^2(N_s^2+1)$ unique elements in each, with the remainder being complex conjugates of these elements.[@Shi2003] The amount of memory required to store the kernels is therefore proportional to $N_s^2(N_s^2+1) \times N_{steps}$ where $N_{steps}$ is the number of timesteps before the kernel decays. In practice, when the memory kernel decays quickly to zero $N_{steps}$ is a small number, meaning that negligible storage is required.
Accelerating convergence of the GQME using selective sampling of kernel matrix elements {#ssec:selective}
---------------------------------------------------------------------------------------
Once the memory kernel has been calculated, solving the GQME, Eq. (\[eq:rdm\_evol\]), provides access to the subsystem dynamics for any factorizable initial condition of the subsystem RDM. However, in many cases one is only interested in a particular initial condition, or a set of them. For example, one may be interested in the relaxation dynamics resulting from the initial excitation of a particular chromophore in a molecular system consisting of multiple chromophores. Here it is important to make a distinction between the initial condition of interest at the RDM level, and the electronic initial conditions that need to be sampled in the MFT calculation of the memory kernel. At the memory kernel level, all separable initial conditions (e.g., initial excitation of each chromophore and coherence between chromophores) need to be sampled to construct the full memory kernel. It is this complete sampling of the initial conditions at the memory kernel level that ensures that one can use Eq. (\[eq:rdm\_evol\]) to propagate the RDM dynamics from *any* separable initial condition.
However, as in the case mentioned above, suppose that one is only interested in propagating the RDM starting from one particular initial condition (or a small subset of initial conditions) in a system with a large number of subsystem states. In this case, there may be many pathways that do not play a role in the relaxation dynamics from the RDM initial condition of interest. Hence, it is clearly unnecessary to have accurate knowledge of the memory kernel elements that describe these unimportant relaxation pathways.
One can exploit this realization by focusing resources into converging the elements of the memory kernel that are most important to describe the relaxation process of interest. When the memory kernel elements are generated using MFT, this corresponds to identifying and selectively converging the most important correlation functions from the sets $\{q_{3}^{(j,\pm)}(t)\}$ and $\{q_{1}^{(jk,\pm)}(t)\}$ (defined in Eqs. (\[eq:general\_q3\]) and (\[eq:general\_q1\])), which give rise to the memory kernel. As such, here we present an algorithm that allows one to use fewer trajectories to generate the less important correlation functions while minimizing the resulting loss in accuracy for the RDM.
Any algorithm that allows for the selective convergence of the memory kernel elements should satisfy certain requirements. In particular, after specifying the RDM-level initial condition that one wishes to propagate, the algorithm should provide a prescription for apportioning a fixed total number of quantum-classical trajectories, $N_{tot}$, among the different initial conditions that need to be sampled at the the memory kernel level. Specifically, the algorithm should provide a set of normalized weights, $\left\{w\right\}$, that determine the portion of trajectories $w_m N_{tot}$ assigned to a given memory kernel level initial condition, $A_m^{\dagger}$. These weights should be chosen such that, when the time-evolved RDM is propagated from the specified initial condition, the error that arises from the statistical error in the memory kernel is minimized. In addition, the calculation of these weights should take into account two important observations. First, correlation functions with certain memory kernel-level initial conditions may contribute more to the relevant relaxation pathways than others. Second, correlation functions arising from different memory kernel-level initial conditions may require different numbers of trajectories to achieve reasonable levels of convergence.
To calculate the optimal weights, some knowledge of the relaxation pathway and current level of convergence of the memory kernel is crucial. Here we suggest an iterative approach, where increasingly improved estimates for the memory kernel and the RDM dynamics are used to re-compute the weights at each iteration. An overview of the algorithm is shown in Fig. \[algorithm\]. We note that while here we present one particular form for the weights, the approach used here is general and does not depend on this particular choice.
![Overview of an iterative approach to identifying and selectively converging the memory kernel given an initial state or states of interest.[]{data-label="algorithm"}](algorithm){width="\columnwidth"}
To provide a form for the weights, $w_m$, we consider how a given realization of the RDM dynamics, which we call the trial RDM dynamics, deviates from its converged form, $\delta \rho_s(t) = \rho_s^{conv}(t) - \rho_s(t)$. Although in practice one does not know the converged dynamics ${\rho}_s^{conv}(t)$, considering how an unconverged dynamics may deviate from its converged form will allow us to elucidate how the convergence of the dynamics can be accelerated via selective sampling. Our goal is to reduce the absolute value of this difference from the converged dynamics for a given number of trajectories, $N_{tot}$. Because the dynamics $\rho_s(t)$, arise from integrating Eq. (\[eq:rdm\_evol\]), one can minimize the absolute difference between the time-derivative of the trial RDM and its converged form, $\delta \rho_s(t) = \rho_s^{conv}(t) - \rho_s(t)$. Because this error metric is time dependent and different for each element of the density, it is convenient to consider the mean absolute deviation, which is the absolute value of $\delta \rho_s(t)$, averaged over time and elements of the density matrix:
$$\langle~ \left| \delta \rho_s \right| ~\rangle = \frac{1}{t_{max} N_s^2}\sum_{n=1}^{N_s^2} \int_0^{t_{max}} d\tau \left|~ [\delta \rho_s(\tau)]_n ~\right|,
\label{eq:mean_absolute_deviation}$$
where $t_{max}$ is the maximum time that is of interest of the dynamics.
Because there is no error in the initial condition, (i.e., $\delta \rho_s(0) = 0$), one can minimize the error at subsequent times by minimizing the error in the time-derivative of the RDM, $\delta\dot{\rho}_s(t) = \dot{\rho}_s^{conv}(t) - \dot{\rho_s}(t)=\frac{d}{dt} \delta \rho_s (t)$. Again it is convenient to consider the mean absolute deviation in this quantity, which is:
$$\langle~ \left| \delta \dot{\rho}_s \right| ~\rangle = \frac{1}{t_{max} N_s^2}\sum_{n=1}^{N_s^2} \int_0^{t_{max}} d\tau \left|~ [\delta \dot{\rho}_s(\tau)]_n ~\right|,$$
An approximate upper bound of this quantity is (see Appendix \[App:GQME\_error\_analysis\]), $$\abs*{\delta \dot{\rho}_s(t) } \lesssim \frac{1}{t_{max}} \sum_m \int_0^{t_{max}} dt ~ [\epsilon(t)]_m
\label{Eq:RDM_error}$$ where the index $m$ runs over all initial conditions. The selective sampling error metric (see also Appendix \[App:GQME\_error\_analysis\]), is approximately proportional to the error in the dynamics at time $t$ that arises from finite sampling of correlation functions with initial condition $m$. For the models considered here, this quantity is: $$[\epsilon(t)]_m = \frac{1}{\sqrt{N_m}}\int_0^t dt^{\prime} \sum_{\beta=1}^{N_s^2} \left|\vphantom{\int_0^t}[\rho_s(t-t^{\prime})]_m \right| [\sigma_{K_1}(t^{\prime})]_{m \beta},
\label{eq:epsilon_tilde_m}$$ where $N_m$ is the number of trajectories that have been run in all previous iterations of the algorithm with memory kernel level initial condition $m$. The quantity $[\sigma_{K_1}(t)]_{m \beta}$ is given by Eq. (\[eq:sigma\_K1\]) in Appendix \[App:GQME\_error\_analysis\].
Motivated by the above considerations, the computed weights in our algorithm depend on $\rho_s(t)$, as well as the statistical error of the correlation functions that contribute to a specific elements of the memory kernel. By distributing $N_{tot}$ trajectories over $n_{step}$ steps in increments of $\Delta N = N_{tot}/n_{step}$ to different GQME-level initial conditions, we refine our estimate for these weights iteratively at each step, as described below. We repeat this procedure either until the RDM converges to a desired threshold or until $N_{tot}$ trajectories have been run.
Using the definitions provided above, we define the un-normalized weight, $a_m$, as $$a_m = \int_0^{t_{max}} [\epsilon(t)]_m,
\label{eq:un_normalized_weight}$$ The normalized weight $w_m$ is then $$w_m = \frac{a_m}{\sum_k a_k}.
\label{eq:normalized_weight}$$ Using these weights, we then apportion trajectories to each memory kernel level initial condition $\hat{A}_m^{\dagger}$ in the subsequent cycle so that, $$N_m^{(j)} = w_m \times j \times \Delta N.
\label{eq:trajectory_assignment}$$ where $j$ is the index of the current iteration of the algorithm (for example, in the first iteration of the algorithm $j = 1$). $ N_m^{(j)}$ is therefore the number of trajectories that have been assigned to initial condition $m$ at step $j$. The algorithm can be summarized as follows:
1. Initialize the algorithm by setting $\mathcal{K}(t) = 0$ and $ [\sigma_{\mathcal{K}_1}]_{m\beta} = 1$.
2. Use $\mathcal{K}(t)$ to intergrate the GQME (Eq. (\[eq:rdm\_evol\])) for some time range $t_{max}$, yielding $\rho_{s}(t)$
3. Use the current guess for $\mathcal{K}(t)$ and $\rho_s(t)$ to compute the weights, $w_m$, corresponding to a given memory kernel-level initial condition $\hat{A}_m^{\dagger}$ using Eqs. (\[eq:epsilon\_tilde\_m\])–(\[eq:normalized\_weight\]).
4. Distribute a number of MFT trajectories, $\Delta N$, among all memory kernel-level initial conditions according to Eq. (\[eq:trajectory\_assignment\]).
5. Using the correlation functions $q_{3}^{(j,\pm)}(t)$ and $q_{1}^{(jk,\pm)}(t)$ obtained from all MFT trajectories that have been run so far, rebuild $\mathcal{K}(t)$ using Eqs. (\[eq:K\_K1\_K3\])–(\[eq:general\_q1\]).
6. Calculate $\sigma_{\mathcal{K}_1}(t)$ using the standard deviations in the correlation functions $q_{1}^{(jk,\pm)}(t)$ (see Appendix \[App:GQME\_error\_analysis\]).
7. Calculate the mean absolute deviation between the previous and the current iteration of $\rho_s(t)$. If the maximum number of trajectories, $N_{tot}$ is exceeded, or if the mean absolute deviation falls below the predetermined threshold, stop. Otherwise, return to step 2.
A plot of the number of trajectories assigned to each initial condition for the LHCII model discussed in Section \[sec:results\_and\_discussion\] is available in the SI.
Simulation Details
==================
Hamiltonian {#ssec:model_hamiltonian}
-----------
Here, we apply MF-GQME to the Frenkel exciton Hamiltonian. [@FMO1; @Ishizaki2009b] This model Hamiltonian forms the basis of many studies of electronic energy transfer in photosynthetic light harvesting systems [@MayKuhn; @Jang2018]. In addition to providing physical insight into these systems, the Frenkel exciton model’s simplicity, i.e., its bilinear coupling to harmonic baths, allows for exact results to be calculated using, for example, the hierarchical equations of motion (HEOM) approach.[@Ishizaki2009] Although GQMEs, and in particular MF-GQME, are not limited to any specific form of the Hamiltonian, the availability of exact results for this model system allows us to benchmark the accuracy and efficiency of MF-GQME for multi-state problems. The Frenkel exciton Hamiltonian is of the general form given in Eq. (\[eq:general-open-quantum-system-hamiltonian\]), with the subsystem Hamiltonian taking the form, $$\hat{H}_s = \sum_{k=1}^{N} E^{(k)} \ket{k}\bra{k} + \sum_{k \neq k^{\prime}} \Delta_{kk^{\prime}} \ket{k}\bra{k^{\prime}},
\label{eq:FE_H_s}$$ where $E^{(k)}$ is the energy of diabatic state $\ket{k}$ and $\Delta_{kk^{\prime}}$ is the electronic coupling between states $\ket{k}$ and $\ket{k^{\prime}}$. Each diabatic state is bilinearly coupled to a bath of $N_{osc}$ independent harmonic oscillators, $$\label{eq:FE_H_sb}
\begin{split}
\hat{H}_{sb} &= \sum_{k=1}^{N_s}\sum_{j=1}^{N_{osc}} c_{j,k} \hat{R}_{j,k} \otimes \ket{k}\bra{k} \\
&\equiv \sum_{k=1}^{N_s} \hat{\gamma}_{k} \otimes \ket{k}\bra{k},
\end{split}$$ where $j$ indexes the bath degrees of freedom, $c_{j,k}$ is the coupling constant for a given bath oscillator, $\hat{R}_{j,k}$ is the position operator for that oscillator, and in the final equality we define $\hat{\gamma}_{k}$ to be the bath part of the system-bath coupling to state $\ket{k}$. The bath therefore consists of $N_s \times N_{osc}$ independent harmonic oscillators: $$\hat{H}_b = \frac{1}{2} \sum_{k=1}^{N_s}\sum_{j=1}^{N_{osc}}\Big[\hat{P}_{j,k}^2 + \omega^2_{j,k}\hat{R}^2_{j,k}\Big],
\label{eq:FE_H_b}$$ where $\hat{P}_{j,k}$ is the momentum operator of the bath mode, and $\omega_{j,k}$ is its frequency, and the Hamiltonian is given in mass-weighted coordinates and momenta. The coupling of each diabatic state to its local bath is characterized by its spectral density, $$J_k(\omega) = \frac{\pi}{2} \sum_{j=1}^{N_{osc}} \frac{c_{j,k}^2}{\omega_{j,k}} \delta(\omega - \omega_{j,k}).$$ For the FMO model, it is common to assume that all spectral densities are equivalent and take the Debye form, $$J(\omega) = \frac{2 \lambda \omega_c \omega}{\omega_c^2 + \omega^2},$$ where $\lambda$ is the reorganization energy that defines the strength of the system-bath coupling and $\omega_c$ is the inverse of the characteristic timescale of the bath. For the LHCII model, we adopt a sum of shifted Drude-Lorentz peaks [@kreisbeck_heom_lhcii]: $$J(\omega) = \sum_{k=1}^{N_{peaks}} \frac{\omega_{c,k}\lambda_k\omega}{\omega_{c,k}^2 + (\omega + \Omega_k)^2} + \frac{\omega_{c,k}\lambda_k\omega}{\omega_{c,k}^2 + (\omega - \Omega_k)^2}$$ where we use the parameters of Ref. : $N_{peaks} = 7$, $\omega_{c}^{-1} = \left\{ 30, 1400, 1000, 1400, 1000,
600, 1000 \right\} $ fs, $\lambda = \left\{ 130, 6, 18, 6, 16, 48, 17 \right\} \mathrm{cm^{-1}}$ , and $\Omega = \left\{240, 297, 342, 388, 518, 745, 915 \right\} \mathrm{cm^{-1}}$.
Computational Methods
---------------------
For the FMO model, the spectral density was discretized with frequencies $$\omega_j = \omega_c \tan \left[ \frac{\pi}{2 N_{osc}} \left(j - \frac{1}{2} \right) \right].
\label{eq:debye_discretization}$$
For the LHCII model the spectral density was discretized by numerically solving for $\omega_j$ that satisfies $$\begin{split}
j - \frac{1}{2} = \sum_{k=1}^{N_{peaks}} \frac{N_{osc} \lambda_k}{\pi \lambda_{tot}} \Bigg[& \arctan \left( \frac{\omega_j + \Omega_k}{\omega_{c,k}} \right) \\
&\ + \arctan \left( \frac{\omega_j - \Omega_k}{\omega_{c,k}} \right) \Bigg], \label{eq:lhcii_discretization}
\end{split}$$ where $\lambda_{tot} = \sum_k \lambda_k$. For both Eqs. (\[eq:debye\_discretization\]) and (\[eq:lhcii\_discretization\]), $\omega_j$ is obtained by taking $j = 1, 2, \ldots, N_{osc}$. For both models, the couplings of the discrete oscillators are given by: $$c_j = \omega_j \sqrt{\frac{2\lambda_{tot}}{N_{osc}}}.$$
For Ehrenfest trajectories, the equations of motion were solved using a split evolution method, where at each dynamical timestep, the bath is evolved through a half timestep of $\frac{\Delta t}{2}$ using mean field theory (see Appendix \[app:mft\_dynamical\_matrices\]) with the velocity Verlet algorithm. [@Swope1982] The system was then evolved for a full timestep $\Delta t$ by numerically diagonalizing $H_s + H_{sb}(R,P)$, holding the bath positions and momenta constant. The bath was then evolved for another half timestep of $\frac{\Delta t}{2}$ using velocity Verlet, keeping the state of the electronic subsystem constant. For the FMO model, the timestep used was $\Delta t = $0.25 fs, while for LHCII the timestep was $\Delta t = $0.05 fs.
The partial memory kernels were constructed using Eqs. (\[eq:general\_K3\])-(\[eq:general\_q1\]), with the necessary correlation functions in Eqs. (\[eq:general\_q3\]) and (\[eq:general\_q1\]) approximated using mean field theory as described in Appendix \[app:mft\_dynamical\_matrices\]. The memory kernel was calculated at a lower time resolution (0.5 fs) than the direct MFT dynamics used to generate the underlying correlation functions, Eqs. (\[eq:general\_q3\]) and (\[eq:general\_q1\]), which was observed not to change the resulting GQME dynamics. Additional details on the specific form these equations in the context of the Frenkel exciton model are given in Appendix \[app:auxiliary-kernels\]. Plots of select memory kernel matrix elements that relate to transitions between the states that have significant populations for FMO and LHCII are available in the SI.
Typically the memory kernel decays to zero after a short amount of time and hence can be truncated (set equal to zero) at all times greater than this time. For the FMO model, this kernel cutoff time, $\tau_c$, was chosen by increasing the cutoff until the dynamics ceased to change, with $\tau_c$ between 125 and 300 fs giving very similar results. In practice, larger values of $\tau_c$ require more trajectories to converge. For the dynamics of the FMO model shown in this work we used $\tau_c$ = 150 fs. For the LHCII model, the dynamics remain unchanged for $\tau_c$ between 50 and 65 fs. To stably integrate the MF-GQME dynamics with cutoff times greater than 65 fs requires a large number of trajectories to achieve stable dynamics. This is because the LHCII complex is in a more strongly coupled regime, which gives rise to highly oscillatory memory kernels that require a very large number of trajectories to converge. Such oscillatory memory kernels have been observed for the dynamics of the spin-boson model with the Debye spectral density in intermediate to strong coupling regimes [@Montoya2016a]. For all LHCII dynamics shown, we used $\tau_c$ = 65 fs. The instabilities observed for LHCII might be able to be alleviated by employing alternative closures,[@Zhang2006a; @Montoya2016a; @Mulvihill2019] or by using higher level quantum-classical methods to generate the memory kernel.[@Kelly2013; @KellyMontoya2016]
After calculating the partial memory kernels, the GQME, Eq. (\[eq:rdm\_evol\]), was integrated using Heun’s method [@heun_book], with the integral in Eq. (\[eq:rdm\_evol\]) approximated using the trapezoid rule.
Because the sampling of bath initial conditions from which the correlation functions are generated is stochastic, there is variability in how many trajectories are required to converge the memory kernel for a given realization of uniform or selective sampling. To evaluate our selective sampling algorithm on the LHCII model, we performed 50 separate realizations of the selective sampling algorithm, as well as 50 separate realizations of uniform sampling. All GQME dynamics shown are from the realization with the median error in the dynamics, where error in the dynamics is quantified by the mean absolute deviation in the subsystem density matrix from its converged form, Eq. (\[eq:mean\_absolute\_deviation\]). The converged form for LHCII corresponds to 65 million total trajectories, apportioned equally across all initial conditions. A plot that shows the distribution of errors of different realizations of the selective sampling procedure as a function of number of trajectories added can be found in the Supporting Information (SI). The selective sampling procedure was performed by adding a total of $\Delta N = 100$ trajectories at each iteration.
Results and Discussion {#sec:results_and_discussion}
======================
Here we apply the MF-GQME method to electronic energy transfer in Frenkel exciton models of the FMO and LHCII light harvesting complexes. The FMO model consists of $N_s = 7$ subsystem states whereas $N_s = 14$ for LHCII. Therefore, for FMO there are $7^4 =2401$ time dependent matrix elements in the memory kernel, while for LHCII there are $14^{4}=38,416$. However, as described in Sec. \[ssec:scaling\], all of the matrix elements can be obtained using very short simulations starting from 28 initial conditions for FMO or 105 distinct initial conditions for LHCII.
Figure \[fmo\_dynamics\] compares the direct MFT and MF-GQME dynamics to exact HEOM results[@kreisbeck_heom_lhcii] for the population relaxation following an initial excitation for the FMO model at low temperature in a regime of high nonadiabaticity ($\omega_c \gtrsim \Delta_{kk'}$ for most pairs of states). As observed in previous work for the FMO model [@Berkelbach2012b], direct MFT gives qualitatively correct coherent oscillations at short times, but fails to recover the correct long-time populations. This can be explained by the high nonadiabaticity and the low temperature bath, both of which are problematic for MFT. In contrast, MF-GQME produces highly accurate results, consistent with previous observations that this scheme offers significant improvements over direct MFT in nonadiabatic regimes. [@Kelly2015; @Pfalzgraff2015; @Montoya2016a]
For this nonadiabatic problem, the memory kernel decays by a cutoff time of $\tau_c = 0.15$ ps, which is considerably shorter than the 1-2 ps lifetime of the population dynamics. However, because MF-GQME requires sampling 28 distinct initial conditions to generate the memory kernel, uniform sampling of all initial conditions in this regime, i.e., when trajectories are apportioned equally to each initial condition, requires a total of $10^5$ total trajectories ($\sim$3,500 per initial condition) to converge the memory kernel. This is in contrast to the $\sim10^4$ trajectories required to obtain a similar level of convergence with direct MFT. Hence, despite the shorter lifetime of the kernel, the larger number of trajectories required to generate the kernel means that the total cost of generating the MF-GQME dynamics shown in Fig. \[fmo\_dynamics\] is $\sim$1.5 times greater than that of MFT. Moreover, because the memory kernel provides the dynamics for all factorizable initial conditions of the subsystem RDM, the dynamics resulting from the two distinct initial subsystem excitations shown in panels (b.) and (d.) of Fig. \[fmo\_dynamics\] are generated using the same memory kernel. In contrast, panels (a.) and (c.) of Fig. \[fmo\_dynamics\], which correspond to direct MFT dynamics, require separate sets of MFT trajectories initialized from the two different subsystem initial conditions.
![Populations for direct MFT and MF-GQME in the FMO model at T = 77K with 10,000 trajectories per initial condition. The model parameters are $\lambda = 35 \mathrm{cm^{-1}}$, $\omega_c^{-1} = 50 ~\mathrm{fs}$. Filled points are the exact HEOM results from Ref. . Panels (a.) and (b.) correspond to the initial excitation of state 6 while panels (c.) and (d.) correspond to initial excitation of state 1.[]{data-label="fmo_dynamics"}](compare_fmo){width="\columnwidth"}
An additional advantage of the MF-GQME scheme is that, once the memory kernel is converged, one can obtain dynamics for arbitrarily long times by integrating Eq. (\[eq:rdm\_evol\]). In practice, the GQME propagation can be done at negligible computational cost compared to the costs of generating MFT trajectories. Figure \[long\_time\_population\] shows the long time population dynamics for direct MFT and MF-GQME following initial excitation of state 6 in the same parameter regime as Fig. \[fmo\_dynamics\]. Because of the prohibitive cost of converging the exact dynamics for such long times, the arrows in Fig. \[long\_time\_population\] correspond to the Boltzmann equilibrium populations that would be expected based on the subsystem Hamiltonian, $e^{-\beta \hat{H}_s}$. As expected, due to the well-known detailed balance problems in methods that rely on a mean field description of the interaction between the electronic subsystem and the nuclei,[@tully_ehrenfest_balance; @Tao2010; @Kelly2011] MFT incorrectly predicts that all states have nearly equal population at long times, in qualitative disagreement with the expected result. In contrast, MF-GQME accurately captures the expected long-time populations. Furthermore, to converge the MFT dynamics out to the 20 ps shown in Fig. \[long\_time\_population\] requires significant additional computational effort (at least 20 times more than for 1 ps), while in the GQME approach, Eq. (\[eq:rdm\_evol\]) can be integrated for arbitrarily long times using the same memory kernel as Fig. \[fmo\_dynamics\]. Because integrating the GQME for longer times incurs negligible additional cost, the dynamics shown in Fig. \[long\_time\_population\] are $\sim 13$ times cheaper when generated with MF-GQME as compared with direct MFT.
![Long time populations in the FMO model for the parameter regime in Fig. \[fmo\_dynamics\] with initial excitation in state 6 using direct MFT (panel (a.)) and MF-GQME (panel (b.)). The MF-GQME dynamics are generated using the same memory kernel as Fig. \[fmo\_dynamics\]. The arrows show the long-time populations that would be expected from $e^{-\beta \hat{H}_s}$.[]{data-label="long_time_population"}](fmo_long_time_2panel){width="\columnwidth"}
A more stringent test of the MF-GQME method is the LHCII model, which has significantly more subsystem states each of which is also more strongly coupled to the bath degrees of freedom. The combination of a large number of states, stronger system-bath coupling, and the complex form of the spectral density of LHCII presents a significant challenge to exact methods. Hence, very limited benchmarks are available using the HEOM method and these required substantial computational effort and some simplification of the spectral density. [@kreisbeck_heom_lhcii] Here we use the form and parameterization of the LHCII Hamiltonian of Ref. with the J7 spectral density from that reference, and compare against the HEOM results presented therein. These HEOM results relied on the high temperature approximation,[@kreisbeck_heom_lhcii] and hence, while they are expected to be highly accurate, it remains unclear whether they are numerically exact.
Figure \[lhcii\_populations\] (a.) shows the dynamics for the sum of the populations of all Chl b chromophores for the LHCII model following initial excitation of the highest energy eigenstate of the subsystem Hamiltonian. For this model the exact HEOM dynamics exhibit an initial fast (sub picosecond) decay in the population of Chl b chromophores followed by a longer time component which decays over tens of picoseconds. Direct MFT fails to capture the magnitude of the short time drop in the population of Chl b chromophores, but qualitatively reproduces the longer timescale of population transfer out of this set of chromophores. As previously observed, modified Redfield dynamics, which is a major workhorse in the modelling and investigation of the excitation dynamics in photosynthetic systems, shows erroneous rapid decay within the first 1-2 ps[@kreisbeck_heom_lhcii]. MF-GQME, however, yields accurate short and long time behavior, in good agreement with the HEOM result. This behavior is mirrored in the populations of the Chl 604a and Chl 605b states in Fig. \[lhcii\_populations\] (b.) and (c.) which contribute to the Chl b decay shown in (a.). Direct MFT underestimates the transfer into Chl 604a while overestimating that into Chl 605b, while modified Redfield theory leads to spuriously fast population decay of these states. In contrast, MF-GQME accurately captures the HEOM result.
![Population dynamics for the 14-state LHCII model of Ref. . Panel (a.) shows the sum of the populations of all Chl b chromophores and panels (b.) and (c.) show the populations of the Chl604a and Chl605b chromophores, respectively. The direct MFT and MF-GQME results are from this work. The direct MFT results used 100,000 trajectories and the MF-GQME results used 625,000 trajectories per initial condition ($6.5 \times 10^6$ total trajectories). For comparison, HEOM and modified Redfield results for the same model are from Ref. . (a.) Sum of the populations of all Chl b chromophores. (b.) population of the Chl 604a state. (c.) Population of the Chl 605b state. The parameters of the model are those of Ref. []{data-label="lhcii_populations"}](lhcii_combined){width="\columnwidth"}
Figure \[lhcii\_long\_time\] shows the long-time dynamics obtained from MF-GQME using the same initial condition. In addition to capturing the exact HEOM dynamics at short times MF-GQME also obtains the intermediate timescale (picosecond) trapping in Chl 604a and Chl 605b bottleneck states, and the long timescale (tens of picoseconds) relaxation to the Chl 610a-611a-612a trimer, the Chl 613a-614a dimer, and the Chl 602a-603a dimer. [@Novoderezhkin2005]
![Long time populations in the LHCII model of ref. with initial excitation in the highest energy eigenstate of the subsystem Hamiltonian. The dynamics are generated using the same memory kernel as Fig. \[lhcii\_populations\]. The dashed lines show the long-time populations that would be expected from $e^{-\beta \hat{H}_s}$. []{data-label="lhcii_long_time"}](lhcii_long_time_limit){width="0.9\columnwidth"}
Because of the stronger system-bath coupling of the LHCII model in comparison to the FMO complex, converging the memory kernel using uniform sampling of all 105 distinct initial conditions would require approximately 65 million total trajectories. Although this number of trajectories is feasible for simple models, it is intractable when more sophisticated atomistic descriptions of the environment are used. The fact that LHCII consists a large number of subsystem states makes it a prime candidate to benefit from using selective sampling to accelerate convergence. As such, we now consider the computational speedups that can be obtained by applying the selective sampling algorithm introduced in Sec. \[ssec:selective\] to this model. To demonstrate the utility of the selective sampling scheme, we thus compare the convergence of the GQME dynamics for a fixed total number of trajectories when using the selective sampling and the uniform sampling approaches to compute the memory kernel. See the SI for a plot showing the average number of trajectories assigned to each memory-kernel level initial condition at each step of the algorithm.
![MF-GQME population dynamics for the LHCII model using uniform and selective sampling. Filled circles are the fully converged uniform sampling result with 65 million total trajectories. Panels (a.) and (b.) correspond to 100,000 trajectories panels (c.) and (d.) correspond to 200,000 trajectories. The high frequency and large amplitude oscillations seen panel (a.) are caused by numerical instability due to noise in the memory kernel.[]{data-label="ss_populations"}](compare_ss_populations){width="\columnwidth"}
Figure \[ss\_populations\] shows the populations for the median realizations of selective and uniform sampling at 100,000 and 200,000 trajectories. Already at 100,000 trajectories the selective sampling yields qualitatively correct dynamics, while the uniformly sampled memory kernel is not numerically stable, with the populations several chromophores becoming negative by $t = 1.0$ ps, and the unphysically high frequency and large amplitude oscillations seen in panel (a.) caused by numerical instability. By 200,000 trajectories, selective sampling is already in semi-quantitive agreement to the exact result, while uniform sampling produces negative populations after only $\sim$ 0.5 ps. In fact, the uniform sampling approach fully converges only after $\sim$ 65,000,000 trajectories. Hence, in this regime the selective sampling algorithm is approximately 300 times more efficient than uniform sampling.
Conclusions
===========
In this paper, we have shown how the MF-GQME approach can be efficiently extended to treat systems where the electronic subspace consists of many states. By analyzing the symmetries of the memory kernel, we have demonstrated that, for a system consisting of $N_s$ subsystem states, the memory kernel, despite having $N_s^4$ matrix elements, can be generated using only $\mathcal{O}(N_s^2)$ quantum-classical trajectories. This scaling analysis is directly applicable to quantum-classical methods using wavefunction-based treatments of the electronic subsystem dynamics, such as MFT and FSSH. The same scaling arguments also apply to methods of generating the kernel using only the original dynamics [@Kidon2015a; @Montoya2016a; @KellyMontoya2016], and to other methods that involve a time convolution master equation such as the transfer tensor method [@ttm1; @ttm2]. This is in contrast to methods[@Sun1998; @Thoss1999; @Liao2002; @Ananth2010; @Huo2011; @Hsieh2012; @Ananth2013; @Richardson2013] based on the Meyer-Miller-Stock-Thoss (MMST) mapping [@Meyer1979; @Stock1997], where the initial conditions necessary to generate all memory kernel elements could be recovered by an appropriate weighting of the variables drawn from a single Gaussian distribution of MMST variables [@Shi2004a; @KellyMontoya2016].
To further reduce the cost of generating the memory kernel necessary to propagate an RDM subject to specific initial conditions, we have outlined a general procedure for how one can select an optimal set of quantum-classical trajectories to converge the kernel. Employing a specific realization of a selective sampling procedure on the LHCII model that focuses resources on initial conditions that most contribute to the fast the convergence of the memory kernel, we have demonstrated that one can obtain a further $2$ orders of magnitude speed-up. While the specific procedure used here is one possible choice, alternative ways of apportioning trajectories to specific initial conditions could lead to more efficient selective sampling schemes and represents a promising direction for future work.
By applying MF-GQME to models of excitation energy transfer in the FMO and LHCII complexes, which consist of $7$ and $14$ electronic states, respectively, we have shown that this method can provide accurate results for multi-state systems for both short-time dynamics and long-time populations at the computational cost of low-level quantum classical trajectory-based methods. Further, when memory kernels are short-lived in comparison to the electronic dynamics of interest, the MF-GQME approach usually costs less than a direct application of MFT. Moreover, once the memory kernel is generated up to the required cutoff time, $\tau_{c}$, the GQME can be propagated with linear scaling in time, making long-time dynamics easy to access.
To date, MF-GQME has previously proven successful in treating the dynamics of systems coupled to harmonic [@Kelly2015; @Montoya2016a; @Montoya2017b] or fully atomistic [@Pfalzgraff2015] environments, both in[@Montoya2017b] and out [@Kelly2015; @Pfalzgraff2015; @Montoya2016a] of equilibrium. In the present work, we have demonstrated that this approach is also able to accurately and efficiently address the dynamics of systems where the electronic subspace consists of many states. Indeed, we have shown that MF-GQME can provide quantitatively accurate results in LHCII where commonly invoked methods, such as Redfield theory[@Bloch1957; @Redfield1965] and its modified variant[@Chang1993; @mukamel_redfield], are known to fail [@kreisbeck_heom_lhcii]. MF-GQME thus provides a powerful tool to investigate charge and energy transfer in the condensed phase.
Supplementary Material {#supplementary-material .unnumbered}
======================
See supplementary material for plots of select memory kernel matrix elements for FMO and LHCII, a plot of the distribution of errors of different realizations of the selective sampling procedure as a function of number of trajectories added, and a plot of the number of trajectories assigned to each initial condition of the LHCII model using the selective sampling approach.
This material is based upon work supported by the U.S. Department of Energy, Office of Science, Office of Basic Energy Sciences under Award Number DE-SC0014437. T.E.M also acknowledges support from the Camille Dreyfus Teacher-Scholar Awards Program. This research used resources of the National Energy Research Scientific Computing Center, a DOE Office of Science User Facility supported by the Office of Science of the U.S. Department of Energy under Contract No. DE-AC02-05CH11231. W.C.P acknowledges support from the Melvin and Joan Lane Stanford Graduate Fellowship. We would also like to thank Stanford University and the Stanford Research Computing Center for providing computational resources and support that have contributed to these research results.
Auxiliary kernels for a multi-level system {#app:auxiliary-kernels}
==========================================
In this section, we provide more detail for the derivation of the expressions for the auxiliary kernels $\mathcal{K}_1(t)$ and $\mathcal{K}_3(t)$ for the GQME that describes the dynamics of the subsystem RDM subject to a spectroscopic initial condition. These expressions are applicable to a general $N_s$ state system coupled to an arbitrary bath of the type described in Sec. \[sec:theory\]. We subsequently specialize our discussion to the Frenkel exciton model presented in Sec. \[ssec:model\_hamiltonian\] given by Eqs. (\[eq:FE\_H\_s\])-(\[eq:FE\_H\_b\]).
We start from the general form of the Hamiltonian, Eq. (\[eq:general-open-quantum-system-hamiltonian\]), for an $N_s$ state electronic system coupled to a bath, which can take a harmonic (e.g., Eq. (\[eq:FE\_H\_sb\])) or atomistic form, with a general form of the system-bath coupling, given by Eq. (\[eq:general\_H\_sb\]).
As stated in Sec. \[sec:theory\], we are interested in the nonequilibrium RDM dynamics arising from a spectroscopic or Franck-Condon initial condition, $\hat{\rho}(0) = \hat{\rho}_{s} \otimes \hat{\rho}_b^{eq}$, where the initial condition of the electronic subsystem can be expanded in terms of the Liouville basis, $\hat{\rho}_s(0) = \sum_{n} a_n \hat{A}^{\dagger}_n$, and the bath is in local equilibrium, $\hat{\rho}_b^{eq} = \frac{e^{-\beta \hat{H}_b}}{\mathrm{Tr}[e^{-\beta \hat{H}_b}]}$. To derive a particular GQME, it is also necessary to choose a projection operator. Here, we choose the $N_s$ state system generalization of the Redfield-type projection operator, shown in Ref. in the context of a two-level system. In the more conventional Nakajima-Zwanzig form of the GQME for nonequilibrium RDM dynamics [@Nakajima1958; @Zwanzig1960a], this projector corresponds to the commonly used Argyres-Kelley projector [@Argyres1964], $\mathcal{P}$, which performs a partial trace over the bath and multiples by the canonical density of the bath, i.e., $\mathcal{P}(\hat{O}_b\hat{O}_s) = \hat{\rho}_b^{eq} \mathrm{Tr}_{b}[\hat{O}_b\hat{O}_s] = \hat{\rho}_b^{eq}\hat{O}_s \mathrm{Tr}_{b}[\hat{O}_b]$, where $\hat{O}_b$ and $\hat{O}_s$ are bath and system operators, respectively.
With this form of the projection operator, and assuming that $\mathrm{Tr}[\hat{\rho}_b^{eq} \hat{\Gamma}_j] = 0$ (which can always be enforced with no loss of generality by redefining the subsystem and bath parts of the Hamiltonian: $\hat{H}_s \rightarrow \hat{H}_s + \mathrm{Tr}[\hat{\rho}_b^{eq} H_{sb}]$ and $\hat{H}_{sb} \rightarrow \hat{H}_{sb} - \mathrm{Tr}[\hat{\rho}_b^{eq} H_{sb}]$), it is possible to obtain the general forms for the auxiliary kernels. Focusing first on $\mathcal{K}_1(t)$,
$$\begin{split}
[\mathcal{K}_{1}(t)]_{mn} &= - \frac{1}{\hbar^2} \mathrm{Tr}\Big[(\mathcal{L}_{sb}\hat{\rho}_b^{eq}\hat{A}_m^{\dagger}) e^{i\mathcal{L}t/\hbar}(\mathcal{L}_{sb}\hat{A}_n) \Big]\\
&= -\frac{1}{2\hbar^2} \sum_{j,k}^{N_{sb}} \Bigg\{ \mathrm{Tr}\Big[[\hat{\Gamma}_j,\hat{\rho}_b^{eq}]_{-}[\hat{S}_j,\hat{A}_m^{\dagger}]_{+} e^{i\mathcal{L}t/\hbar}[\hat{S}_{k},\hat{A}_n]_{-} \hat{\Gamma}_k \Big] + \mathrm{Tr}\Big[ [\hat{\Gamma}_j,\hat{\rho}_b^{eq}]_{+}[\hat{S}_j,\hat{A}_m^{\dagger}]_{-} e^{i\mathcal{L}t/\hbar}[\hat{S}_{k},\hat{A}_n]_{-} \hat{\Gamma}_k \Big] \Bigg\} \\
&= \frac{1}{2\hbar^2}\sum_{j,k}^{N_{sb}} \sum_{r,s}^{N_s^2} \Bigg\{ c^{(j,+)}_{m,r} \mathrm{Tr}\Big[[\hat{\rho}_b^{eq},\hat{\Gamma}_j]_{-}\hat{A}_{r}^{\dagger} e^{i\mathcal{L}t/\hbar}\hat{A}_{s}\hat{\Gamma}_k \Big] + c_{mr}^{(j,-)}\mathrm{Tr}\Big[ [\hat{\Gamma}_j,\hat{\rho}_b^{eq}]_{+}\hat{A}_r^{\dagger} e^{i\mathcal{L}t/\hbar}\hat{A}_{s}\hat{\Gamma}_k\Big] \Bigg\}\mathbf{c}_{sn}^{(k,-)} \\
&= \frac{1}{\hbar^2}\sum_{j,k}^{N_{sb}} \sum_{r,s}^{N_s^2} \Big\{ c_{mr}^{(j,+)}[q_1^{(jk,-)}]_{rs}(t) - c_{mr}^{(j,-)}[q_1^{(jk,+)}]_{rs}(t) \Big\} c_{sn}^{(k,-)}.
\label{eq:general-K1-appendix}
\end{split}$$
Using similar manipulations, one can show that, $$\label{eq:general-K3-appendix}
\begin{split}
[\mathcal{K}_{3}(t)]_{mn} &= -\frac{1}{\hbar} \mathrm{Tr}\Big[(\mathcal{L}_{sb}\hat{\rho}_b^{eq}\hat{A}_m^{\dagger}) e^{i\mathcal{L}t/\hbar}\hat{A}_n \Big]\\
&= \frac{1}{\hbar}\sum_{j}^{N_{sb}}\sum_{r}^{N_s^2} \Big\{ c_{mr}^{(j,+)}[q_3^{(j,-)}]_{rn}(t) + c_{mr}^{(j,-)}[q_3^{(j,+)}]_{rn}(t) \Big\},
\end{split}$$
where the elements of the dynamical matrices, $q_1^{(jk,\pm)}$ and $q_3^{(j,\pm)}$, are given by Eqs. (\[eq:general\_q1\]) and (\[eq:general\_q3\]), the elements of the static transformation matrix $c^{(j,\pm)}$ take the form, $$c^{(j,\pm)}_{mn} = \mathrm{Tr}_s\Big[\hat{A}_{m}^{\dagger} [\hat{S}_j, \hat{A}_{n}]_{\pm}\Big].$$ and we have used the fact that
$$\begin{aligned}
[\hat{S}_j, \hat{A}_n]_{\pm} &= \sum_{r}^{N_s^2} \hat{A}_{r} c_{rn}^{(j,\pm)},\\
[\hat{S}_j, \hat{A}_{n}^{\dagger}]_{\pm} &= \pm\sum_{r}^{N_s^2}c_{nr}^{(j,\pm)} \hat{A}_r^{\dagger}.\end{aligned}$$
Equations (\[eq:general-K1-appendix\])-(\[eq:general-K3-appendix\]) are general, applicable to an $N_s$ state system coupled to an arbitrary bath with system-bath coupling given by Eq. (\[eq:general\_H\_sb\]), and subject to a spectroscopic initial condition. We emphasize that such systems include electronic states coupled to harmonic or atomistic baths and system-bath interactions beyond linear coupling.
We now specialize our discussion to the Frenkel exciton Hamiltonian, Eqs. (\[eq:FE\_H\_s\])-(\[eq:FE\_H\_b\]), where the bath is harmonic and can be decomposed into contributions that are local, in addition to the coupling being bilinear and local. Given the Frenkel exciton form of the Hamiltonian, the matrix elements of the static transformation matrix are given by, $$c^{(j,\pm)}_{mn} = \delta_{j \alpha} \delta_{j \beta} \delta_{\alpha^{\prime}{\beta}^{\prime}} \pm \delta_{j \alpha^{\prime}} \delta_{j \beta^{\prime}} \delta_{\alpha \beta},$$ where $\delta$ is the Kronecker delta.
Obtaining $\mathcal{K}(t)$ from $\mathcal{K}_1(t)$ and $\mathcal{K}_3(t)$ by solving a system of linear equations {#app:K_K1_K3}
=================================================================================================================
Here we describe a procedure for obtaining $\mathcal{K}$ from $\mathcal{K}_1$ and $\mathcal{K}_3$ by discretizing the self-consistent expansion and solving a system of linear equations, hence avoiding an iterative solution. The self consistency relation for the memory kernel, Eq. (\[eq:K\_K1\_K3\]) can be written in discrete time as, $$\begin{split}
\mathcal{K}(t_n) \approx \mathcal{K}_1(t_n)
- i\frac{\Delta t}{2} \sum_{j=0}^{n-1} & \mathcal{K}_3(t_{n-j})\mathcal{K}(t_j) \\ &+ \mathcal{K}_3(t_{n-j-1})\mathcal{K}(t_{j+1}),
\label{eq:discrete_K_decomp}
\end{split}$$ where we have discretized the integral using the trapezoid rule. Here $t_0, t_1, \ldots, t_n$ are equally spaced discrete times, and hence $t_{j+1} - t_j \equiv \Delta t$, and $t_m - t_j = t_{m-j}$ ($m > j$). Equation (\[eq:discrete\_K\_decomp\]) can be solved to obtain an explicit expression for $\mathcal{K}(t_n)$ in terms of $\mathcal{K}_1$, $\mathcal{K}_3$, and $\mathcal{K}(t_{j})$ ($0 \leq j \leq n-1$),
(t\_0) &= \_1(t\_0)
(t\_n) = \_1(t\_n) - &i ( \_3(t\_1)(t\_[n-1]{}) + \_[j=0]{}\^[n-2]{} \_3(t\_[n-j]{})(t\_j) + \_3(t\_[n-j-1]{})(t\_[j+1]{})) \[eq:onestep\_K1\_K3\]
where $n > 0$ and $\mathbf{1}$ is the identity matrix. The above equation allows $\mathcal{K}(t_n)$ ($n > 0$) to be obtained from the $\mathcal{K}(t_{j})$ ($j < n$) using a single, non-iterative step where one solves for the unknown matrix $\mathcal{K}(t_n)$ in Eq. (\[eq:onestep\_K1\_K3\]). In practice, the system of linear equations implied by Eq. (\[eq:onestep\_K1\_K3\]) was solved by computing and storing the Cholesky factorization of $[\mathbf{1} + i\frac{\Delta t}{2}\mathcal{K}_3(t_0) ]$, which can be used to solve for $\mathcal{K}(t_n)$.
MFT treatment of the dynamical matrices, $q_1^{(jk,\pm)}(t)$ and $q_3^{(j,\pm)}(t)$ {#app:mft_dynamical_matrices}
===================================================================================
Here, we outline how one can calculate the dynamical matrices, $q_1^{(jk,\pm)}(t)$ and $q_3^{(j,\pm)}(t)$ in Eqs. (\[eq:general\_q1\]) and (\[eq:general\_q3\]) using the Ehrenfest MFT method. To do this, we first review the partial transformation to Wigner phase space with respect to bath variables of a general correlation function. We then show the general MFT approximation to the dynamical matrices corresponding to a general $N_s$-state system, and specialize our discussion to the Frenkel exciton model. Finally, we provide the equations of motion for the subsystem and bath variables.
To treat a general correlation function of a system which can be separated into a subsystem coupled to a bath, $$\begin{split}
C_{AB}(t) &= \mathrm{Tr}[\hat{\rho} \mathcal{A} \mathcal{B}(t)]\\
&= \mathrm{Tr}_s\mathrm{Tr}_b[\hat{\rho} \mathcal{A} \mathcal{B}(t)],
\end{split}$$ using the Ehrenfest MFT method, we first perform a partial Wigner transformation [@Wigner1932; @Hillery1984a] with respect to the bath variables. Here, $\hat{\rho}$ encodes the initial condition of the full system, and $\mathcal{A}$ and $\mathcal{B}$ are arbitrary operators. Upon performing the partial Wigner transformation, we can rewrite the correlation function as follows, $$\label{eq:correlation-function-partially-wigner-transformed}
C_{AB}(t) = \int dX\ \mathrm{Tr}_s[(\hat{\rho} \mathcal{A})^W (\mathcal{B}(t))^W],$$ where $X = (R,P)$ are the classical coordinates, $R$, and momenta, $P$, of the bath, and the Wigner transform of an operator, $\hat{O}$, is defined as [@Imre1967; @Hillery1984a] $$\label{eq:wigner-transform-definition}
\hat{O}^W(X) = \int dz\ e^{\frac{iP\cdot z}{\hbar}} \langle R - \frac{z}{2} | \hat{O} | R + \frac{z}{2} \rangle.$$ The Wigner transform of the product of two operators is [@Imre1967; @Hillery1984a] $$\left(\hat{A}\hat{C}\right)^{W} (X)= \hat{A}^W(X) e^{\frac{\hbar \hat{\Lambda}}{2i}}\hat{C}^{W}(X), \label{eq:wigner_transform_op_product}$$ where $$\hat{\Lambda} = \sum_j\left(\frac{\overleftarrow{\partial}}{\partial P_j}\frac{\overrightarrow{\partial}}{\partial R_j} - \frac{\overleftarrow{\partial}}{\partial R_j}\frac{\overrightarrow{\partial}}{\partial P_j}\right)$$ is the Poisson bracket operator and the arrows indicate the directions in which the differential operators act.
The MFT approximation to this correlation function amounts to replacing the time-dependent operator $(\mathcal{B}(t))^W = (e^{iHt/\hbar}\mathcal{B}e^{-iHt/\hbar})^W$ with $\mathcal{B}^W(X_t, t)$, where the classical bath degrees of freedom evolve according to the Poisson bracket subject to an additional time-dependent mean force arising from the subsystem dynamics and the quantum variables evolve according to the time-dependent Schrodinger equation subject to the time-dependent contribution of the bath. This allows one to rewrite the dynamical matrices, $q_1^{(jk,\pm)}(t)$ and $q_3^{(j,\pm)}(t)$, corresponding to a general $N_s$ state system in the MFT approximation as,
$$\begin{aligned}
\label{eq:FE_q3_plus}
q_{3,mn}^{(j,+)}(t) &\approx \int dX\ \frac{([\hat{\rho}_{b}^{eq}, \hat{\Gamma}_{j}]_{+})^W(X_0)}{2} C^{(s)}_{mn}(t), \\
\label{eq:FE_q3_minus}
q_{3,mn}^{(j,-)}(t) &\approx \int dX\ \frac{([\hat{\rho}_{b}^{eq}, \hat{\Gamma}_{j}]_{-})^W(X_0)}{2} C^{(s)}_{mn}(t), \\
\label{eq:FE_q1_plus}
q_{1,mn}^{(jk,+)}(t) &\approx \int dX\ \frac{([\hat{\rho}_{b}^{eq}, \hat{\Gamma}_{j}]_{+})^W(X_0)}{2} \Gamma_{k}^{W}(X_t) C^{(s)}_{mn}(t), \\
\label{eq:FE_q1_minus}
q_{1,mn}^{(jk,-)}(t) &\approx \int dX\ \frac{([\hat{\rho}_{b}^{eq}, \hat{\Gamma}_{j}]_{-})^W(X_0)}{2} \Gamma_{k}^{W}(X_t) C^{(s)}_{mn}(t), \end{aligned}$$
where $$C^{(s)}_{mn}(t) = \mathrm{Tr}_s\left\{\hat{A}_m^{\dagger} \hat{A}_n(X_t, t)\right\}.$$ Here we have explicitly included the dependence of operator $\hat{A}_n(t)$ on the time-dependent bath variables, $X_t$, to emphasize that the value of $C_{mn}^{(s)}(t, X_t)$ is dependent on the realization of the bath.
For the Frenkel exciton model, where the bath is harmonic and the bath part of the system-bath coupling is linear in the bath coordinates and given by $\hat{\Gamma}_j \rightarrow \hat{\gamma}_{j} = \sum_{k} c_{j,k} \hat{R}_{j,k}$, the Wigner transform of the anticommutator and commutator of the bath canonical distribution and $\hat{\gamma}_{k}$ take simple forms,
$$\begin{aligned}
\frac{([\hat{\rho}_{b}^{eq}, \hat{\gamma}_{j}]_{+})^W(X_0)}{2} &= \rho_b^W(X_0) \gamma_j^W(X_0),\\
\frac{([\hat{\rho}_{b}^{eq}, \hat{\gamma}_{j}]_{-})^W(X_0)}{2} &= \rho_b^W(X_0) \xi_j^W(X_0),\end{aligned}$$
where, using Eq. (\[eq:wigner\_transform\_op\_product\]), one obtains
$$\begin{aligned}
\gamma_{j}^W &= \sum_{k} c_{j,k} R_{j,k},\\
\xi_{j}^W &= - i \sum_{k} c_{j,k} P_{j,k}\frac{\tanh(\beta \hbar \omega_{j,k}/2)}{\omega_{j,k}},\end{aligned}$$
and $$\begin{split}
\rho_b^W &= \prod_{j,k}\frac{\tanh({\beta \hbar\omega_{j,k}/2)}}{\pi} \\
&\ \times \exp\Bigg[ - \frac{\tanh(\beta \hbar\omega_{j,k}/2)}{\hbar\omega_{j,k}}\Big[ P_{j,k}^2 + \omega_{j,k}^2 R_{j,k}^2\Big]\Bigg].
\end{split}$$
Application of the Ehrenfest MFT approximation to the dynamics of the Frenkel exciton model leads to the evolution of the subsystem wavefunction subject to the Hamiltonian of the subsystem modulated by the time-dependent system-bath coupling, $$\frac{d}{dt} \rho_s(t, X_t) = -i\Big[ \Big(\hat{H}_s + \sum_{k} \gamma_{k}(X_t) \ket{k}\bra{k}\Big), \rho_s(t, X_t)\Big],$$ where the subsystem initial condition, $\rho_s(0) = \ket{\psi}\bra{\psi}$, is of pure form.
The bath, in turn, follows Hamilton’s equations,
$$\begin{aligned}
\frac{\partial P_{j,k}}{\partial t} &= - \frac{\partial H_{b}^{Eh}}{\partial Q_{j,k}},\\
\frac{\partial Q_{j,k}}{\partial t} &= \frac{\partial H_{b}^{Eh}}{\partial P_{j,k}},\end{aligned}$$
where the Ehrenfest version of the bath Hamiltonian takes the form, $$H^{Eh}_{b}(t) = \frac{1}{2}\sum_{j,k}\Big[ P_{j,k}^2 + \omega_{j,k}^2 R_{j,k}^2 + 2c_{j,k} R_{j,k} \overline{S}_j(t) \Big],$$ where $\overline{S}_j(t) = \mathrm{Tr}_s[\rho_s(t, X_t) \hat{S}_j]$ is the mean potential arising from the subsystem dynamics.
The correlation functions given in Eqs. (\[eq:FE\_q3\_plus\])-(\[eq:FE\_q1\_minus\]) are generated using MFT and the memory kernel is then generated according to Eqs. (\[eq:K\_K1\_K3\])-(\[eq:general\_K1\]).
Correlation functions with off-diagonal initial conditions using Ehrenfest mean field theory {#App:off_diagonal_inicon}
============================================================================================
In this section, we describe how the correlation functions in Eqs. (\[eq:general\_q3\]) and (\[eq:general\_q1\]) are evaluated using MFT for initial conditions that do not correspond to a single wavefunction.
If the subsystem initial condition is diagonal, i.e. $\hat{A}_j = \ket{\alpha}\bra{\alpha^{\prime}}$ where $\alpha = \alpha^{\prime}$, then MFT is straightforward to apply because one can use the subsystem initial condition corresponding to pure state $\ket{\alpha}$, i.e. $\ket{\psi(0)} = \ket{\alpha}$. On the other hand, if $A_j$ is an off-diagonal (if $\alpha \neq \alpha^{\prime}$), the procedure is less straightforward because $A_j$ does not correspond to an initial condition that arises from a single wavefunction. Several possible approaches and subtle issues related to this are discussed in detail in the Appendices of Ref. . A simple and computationally efficient approach for calculating such correlation functions involves an auxiliary wavefunction, $$\ket{\bar{\psi}} = \frac{1}{\sqrt{2}} \left( e^{i\phi}\ket{\alpha} + \ket{\alpha^{\prime}} \right)$$ where $\phi$ is a random number uniformly sampled from the interval $[0,2\pi)$. Then consider the expression
$$\begin{aligned}
\bar{q}(t) &= \int d \phi \int dX ~ \rho_{b,W} \Gamma_{j,W}(0) \braket{\bar{\psi}|\alpha^{\prime}} \braket{\alpha|\bar{\psi}} \bra{\bar{\psi}} \Gamma_{k,W}(t) A_n(t) \ket{\bar{\psi}} \\
&= \frac{1}{2} \int d\phi \int dX ~ \rho_{b,W} \Gamma_{j,W}(0) e^{i \phi} \bra{\bar{\psi}} \Gamma_{k,W}(t) A_n(t) \ket{\bar{\psi}} \\
&= \frac{1}{4} \int dX ~ \rho_{b,W} \Gamma_{j,W}(0) \bra{\alpha} \Gamma_{k,W}(t) A_n(t) \ket{\alpha^{\prime}} \\
&= \frac{1}{4} q(t),
\end{aligned}
\label{eq:off_diagonal_q}$$
where $q(t)$ is a general memory kernel correlation function which is partially Wigner transformed over the bath degrees of freedom, such as those defined in Eqs. (\[eq:FE\_q3\_plus\])-(\[eq:FE\_q1\_minus\]), and we have explicitly used that $\alpha \neq \alpha^{\prime}$. Equation (\[eq:off\_diagonal\_q\]) thus provides a way to access two $q$ correlation functions using the same initial wavefunction $\ket{\bar{\psi}}$, in particular those for which $\hat{A}_m = \ket{\alpha}\bra{\alpha^{\prime}}$ and $\hat{A}_m=\ket{\alpha^{\prime}}\bra{\alpha}$. The explicit procedure for calculating such correlation functions is:
1. Initialize the wavefunction $\ket{\psi(0)} = \ket{\bar{\psi}}$, sampling a new random number $\phi$ from a uniform distribution on the interval $[0,2\pi)$. Sample the bath initial conditions from the appropriate bath distribution, $\rho_{b,W}$, as in the diagonal case.
2. At $t=0$, evaluate and store $e^{i\phi}$ and $e^{-i\phi}$.
3. Run each trajectory using MFT to calculate the desired observables at time $t$ in the same way as for a diagonal initial condition. Multiply the correlation function by $e^{i\phi}$ to obtain $\bar{q}_{+}(t)$ and by $e^{-i\phi}$ to obtain $\bar{q}_{-}(t)$.
4. Repeat steps 1-4, averaging $\bar{q}_{\pm}(t)$ over an ensemble of trajectories with different bath initial conditions and different initial values of $\phi$. As shown in Eq. (\[eq:off\_diagonal\_q\]), the trajectory-averaged $\bar{q}_{+}$ corresponds $q(t)$ with $A_m = \ket{\alpha}\bra{\alpha^{\prime}}$ and $\bar{q}_{-}$ corresponds to $q(t)$ with $A_j = \ket{\alpha^{\prime}}\bra{\alpha}$.
This approach of phase averaging over a superposition state in order to obtain the correlation function of interest is similar to that employed in the recently proposed CFBT method,[@cfbt] which extends Ehrenfest mean field theory.
Convergence analysis of the memory kernel {#App:GQME_error_analysis}
=========================================
Here, we describe how to motivate our selective sampling algorithm from formal analysis of how the error in the memory kernel arising from lack of convergence translates into error in the GQME dynamics.
Our analysis starts from the GQME corresponding to the fully converged RDM, $\rho_s^{conv}(t)$, and the GQME using corresponding to a (possibly unconverged) RDM $\rho_s(t)$. The fully converged dynamics arise from a fully converged memory kernel, $\mathcal{K}^{\mathrm{conv}}(t)$, while the unconverged dynamics arise from an unconverged memory kernel, $\mathcal{K}(t)$. A simple metric for the error in $\rho_s(t)$ is its deviation from the converged RDM, $$\delta \rho_s(t) \equiv \rho_s^{conv}(t) - \rho_s(t). \label{eq:deviation_definition}$$ Since this quantity can be positive or negative, it is convenient to consider its absolute value, $\left| \delta \rho_s(t) \right|$. This deviation is time dependent and is different for each element of the density. A scalar metric for the total error is the mean absolute deviation of the RDM from its converged value, with the mean taken over time and elements of the density, $$\langle \left| \delta \rho_s \right| \rangle = \frac{1}{t_{max} N_s^2}\sum_{n=1}^{N_s^2} \left[\int_0^{t_{max}} dt \left| \delta \rho_s(t) \right|\right]_n,$$ where $t_{max}$ is the total amount of time over which one wants to know the RDM dynamics. Because there is no error in the initial condition, (i.e., $\delta \rho_s(0) = 0$), one can minimize the error at subsequent times by minimizing the error in the time-derivative of the RDM, $$\begin{split}
\delta \dot{\rho}_s(t) &\equiv \dot{\rho}_s^{\mathrm{conv}}(t) - \dot{\rho}_s(t) \\
&= \frac{d}{dt} \left(\vphantom{\frac{d}{dt}}\delta \rho_s(t)\right).
\label{eq:deriv_error}
\end{split}$$
Equation (\[eq:rdm\_evol\]) can be combined with Eq. (\[eq:deriv\_error\]) in order to express the error in the time derivative, $ \delta \dot{\rho}_s(t)$, in terms of $\delta \rho_s(t)$ and $\delta \mathcal{K}(t)=\mathcal{K}^{\mathrm{conv}}(t) - \mathcal{K}(t)$: $$\delta \dot{\rho}_s(t) = E_1 + E_2 + E_3,
\label{eq:error_sum}$$ where $$\begin{aligned}
\label{eq:E1}
E_1(t) &= \delta \rho_s(t) \mathcal{X}, \\
\label{eq:E2}
E_2(t) &= -\int_0^t d\tau \rho_s(t-\tau) \delta \mathcal{K}(\tau),\\
\label{eq:E3}
E_3(t) &= -\int_0^t d\tau \delta \rho_s(t-\tau) \mathcal{K}^{\mathrm{conv}}(\tau).\end{aligned}$$
{width="80.00000%"}
Of the three terms, only $E_2$ can be directly controlled by running more trajectories, since more trajectories lead to a smaller $\delta \mathcal{K}(t)$ as the correlation functions that give rise to $\mathcal{K}(t)$ become more converged. In contrast, $E_1$ and $E_3$ cannot be directly controlled because they depend on $\delta \rho_s(t)$, a quantity which arises from $\delta \mathcal{K}(t)$ and previous values of $\delta \rho_s(t)$ according to Eqs. (\[eq:deriv\_error\])-(\[eq:E3\]). For our choice of selective sampling algorithm we make the ansatz that the mean absolute deviation in the RDM dynamics is proportional to the mean absolute value of $E_2$, i.e., $\langle \left |\delta \rho_s \right| \rangle \propto \langle |E_2| \rangle$. Figure \[fig:error\_analysis\]a, which is generated using data from LHCII, shows strong correlation between the error in the GQME dynamics and $|E_2|$, suggests that this ansatz is valid.
In order to obtain an expression for the selective sampling weights, $w_m$, we consider how $\langle |E_2| \rangle$ is related to the error in specific elements of $\mathcal{K}(t)$,
$$\begin{split}
\langle |E_2| \rangle &= \frac{1}{t_{max} N_s^2} \sum_{\beta=1}^{N_s^2}\int_0^{t_{max}} d\tau \left|\vphantom{\int_0^\tau} \left[ E_2(\tau) \right]_{\beta } \right| \\
&= \frac{1}{t_{max} N_s^2} \sum_{\alpha,\beta=1}^{N_s^2} \int_0^{t_{max}} d\tau \left|\int_0^\tau d\tau^{\prime} [\rho_s(\tau-\tau^{\prime})]_\alpha [\delta K(\tau^{\prime})]_{\alpha \beta} \right| \\
& \leq \frac{1}{t_{max} N_s^2} \sum_{\alpha,\beta=1}^{N_s^2} \int_0^{t_{max}} d\tau \int_0^\tau d\tau^{\prime} \left|\vphantom{\int_0^t}[\rho_s(\tau-\tau^{\prime})]_\alpha \right| \left|\vphantom{\int_0^t} [\delta K(\tau^{\prime})]_{\alpha \beta} \right|.
\end{split}$$
Using this expression to determine the error in the dynamics still requires prior knowledge of the converged kernel. In order to use this expression to obtain an expression for the selective sampling weights, we make the additional approximation that $\left|\delta K(t)\right| \approx \left| \delta K_1(t) \right| \approx \sqrt{\frac{\sigma^2_{K_1}(t)}{N}}$ so that:
$$\langle |E_2| \rangle \lesssim \frac{1}{t_{max} N_s^2} \sum_{\alpha,\beta} \int_0^{t_{max}} d\tau \int_0^\tau d\tau^{\prime} \left|\vphantom{\int_0^t}[\rho_s(\tau-\tau^{\prime})]_\alpha \right| \frac{[\sigma_{K_1}(\tau^{\prime})]_{\alpha \beta}}{\sqrt{N_\alpha}},
\label{eq:E2_inequality}$$
where $\sigma_{\mathcal{K}_1}$ is the standard deviation in $\mathcal{K}_1$: $$\left[ \sigma_{\mathcal{K}_1 (t)} \right]_{\alpha \beta} = \sum_{\substack{j,k,\pm\\ m,n}} [c^{(j,\pm)}]_{ \alpha m} [\sigma^{(jk,\mp)}_{q_1}(t)]_{m n} [c^{(k,\pm)}]_{n \beta}.
\label{eq:sigma_K1}$$
Here it is useful to re-write Eq. (\[eq:E2\_inequality\]) as a sum of terms arising from different initial conditions at the memory kernel level: $$\langle |E_2| \rangle \lesssim \sum_m \int_0^{t_{max}} dt\ [\epsilon(t)]_m
\label{eq:decomposed_E2_inequality}$$ where $m$ indexes the initial conditions at the memory kernel level and $[\epsilon(t)]_m$ is the sum of all the terms in Eq. (\[eq:E2\_inequality\]) arising from initial condition $m$:
$$[\epsilon(t)]_m = \frac{1}{\sqrt{N_m}}\int_0^t dt^{\prime} \sum_{\substack{j,k,\pm\\ \alpha, \beta, n}} \left|\vphantom{\int_0^t}[\rho_s(t-t^{\prime})]_{\alpha} \right| [c^{(j,\pm)}]_{ \alpha m} [\sigma^{(jk,\mp)}_{q_1}(t)]_{m n} [c^{(k,-)}]_{n \beta},
\label{eq:epsilon_m}$$
Equations (\[eq:decomposed\_E2\_inequality\]) and (\[eq:epsilon\_m\]) provide the basis for our choice of weights in our selective sampling algorithm. In particular, the weight $w_m$ associated with initial condition $m$ is proportional to the time average of $[\epsilon(t) ]_m$ $$w_m \propto \int_0^{t_{max}} dt\ [\epsilon(t)]_m,$$ This metric is an upper bound parts of the $|E_2|$ error term that arise from initial condition $m$, as shown in Eq. (\[eq:E2\_inequality\]). As shown in Fig. \[fig:error\_analysis\]b, $|E_2(t)|$ and $|\epsilon(t)|$ track the error in the GQME dynamics, $|\delta \rho_s(t)|$. Hence, a selective sampling algorithm that minimizes this term should also minimize the error in the GQME dynamics. For the models considered here, we note that $[\epsilon(t) ]_m$ takes a somewhat simpler form because $[c^{(j,\pm)}]_{ \alpha m} = 0$ if $\alpha \neq m$ and hence
$$[\epsilon(t)]_m = \frac{1}{\sqrt{N_m}}\int_0^t dt^{\prime} \sum_{n} \left|\vphantom{\int_0^t}[\rho_s(t-t^{\prime})]_{m} \right| [\sigma_{K_1}(t)]_{m n}.
\label{eq:FE_epsilon_m}$$
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Originally, the Hubbard model has been derived for describing the behaviour of strongly-correlated electrons in solids. However, since over a decade now, variations of it are also routinely being implemented with ultracold atoms in optical lattices, allowing their study in a clean, essentially defect-free environment. Here, we review some of the rich literature on this subject, with a focus on more recent non-standard forms of the Hubbard model. After an introduction to standard (fermionic and bosonic) Hubbard models, we discuss briefly common models for mixtures, as well as the so called extended Bose-Hubbard models, that include interactions between neighboring sites, next-neighboring sites, and so on. The main part of the review discusses the importance of additional terms appearing when refining the tight-binding approximation on the original physical Hamiltonian. Even when restricting the models to the lowest Bloch band is justified, the standard approach neglects the density-induced tunneling (which has the same origin as the usual on-site interaction). The importance of these contributions is discussed for both contact and dipolar interactions. For sufficiently strong interactions, also the effects related to higher Bloch bands become important even for deep optical lattices. Different approaches that aim at incorporating these effects, mainly via dressing the basis Wannier functions with interactions, leading to effective, density-dependent Hubbard-type models, are reviewed. We discuss also examples of Hubbard-like models that explicitly involve higher $p$-orbitals, as well as models that couple dynamically spin and orbital degrees of freedom. Finally, we review mean-field nonlinear-Schrödinger models of the Salerno type that share with the non-standard Hubbard models the nonlinear coupling between the adjacent sites. In that part, discrete solitons are the main subject of the consideration. We conclude by listing some future open problems.'
author:
- 'Omjyoti Dutta$^1$, Mariusz Gajda$^{2,3}$, Philipp Hauke$^{4,5}$, Maciej Lewenstein$^{6,7}$, Dirk-Sören Lühmann$^8$, Boris A. Malomed$^{6,9}$, Tomasz Sowiński$^{2,3}$, and Jakub Zakrzewski$^{1,10}$'
bibliography:
- 'references\_v02.bib'
title: 'Non-standard Hubbard models in optical lattices'
---
Introduction
============
Standard Hubbard models in optical lattices
===========================================
Hubbard models in excited bands {#orbitalhub}
===============================
Hubbard models with dynamical spin {#orbitaldipole}
==================================
Mutual interactions of atomic magnets
-------------------------------------
1D and 2D models of the *Salerno type*: the mean-field and quantum versions
===========================================================================
Conclusions
===========
Acknowledgments
===============
Appendix
========
| {
"pile_set_name": "ArXiv"
} |
---
abstract: |
Here we report observations of the two lowest inversion transitions of ammonia () with the 70-m Tidbinbilla radio telescope. The aim of the observations is to determine the kinetic temperatures in the dense clumps of the G333 giant molecular cloud associated with RCW 106 (hereafter known as the G333 GMC) and to examine the effect that accurate measures of temperature have on the calculation of derived quantities such as mass. This project is part of a larger investigation to understand the timescales and evolutionary sequence associated with high-mass star formation, particularly its earliest stages. Assuming that the initial chemical composition of a giant molecular cloud is uniform, any abundance variations within will be due to evolutionary state.
We have identified 63 clumps using SIMBA 1.2-mm dust continuum maps and have calculated gas temperatures for most (78 per cent) of these dense clumps. After using *Spitzer* GLIMPSE 8.0[${\rm \mu m}$]{} emission to separate the sample into IR-bright and IR-faint clumps, we use statistical tests to examine whether our classification shows different populations in terms of mass and temperature. We find that in terms of log clump mass (2.44 – 4.12[M$_\odot$]{}) and log column density (15.3 – 16.6 cm$^{-2}$), that there is no significant population difference between IR-bright and IR-faint clumps, and that kinetic temperature is the best parameter to distinguish between the gravitationally bound state of each clump. The kinetic temperature was the only parameter found to have a significantly low probability of being drawn from the same population. This suggests that clump radii does not have a large effect on the temperature of a clump, so clumps of similar radii may have different internal heating mechanisms. We also find that while the IR-bright clumps have a higher median log virial mass than the IR-faint clumps (IR-bright: 2.88[M$_\odot$]{}; IR-faint: 2.73[M$_\odot$]{}), both samples have a similar range for both virial mass and FWHM (IR-bright: log virial mass = 2.03 – 3.68[M$_\odot$]{}, FWHM = 1.17 – 4.50[kms$^{-1}$]{}; IR-faint: log virial mass = 2.09 – 3.35[M$_\odot$]{}, FWHM = 1.05 – 4.41[kms$^{-1}$]{}). There are 87 per cent (40 of 46) of the clumps with masses larger than the virial mass, suggesting that they will form stars or are already undergoing star formation.
author:
- '\'
date: 'Accepted 2014 March 18. Received 2014 March 18; in original form 2013 December 8'
title: 'Molecular line mapping of the giant molecular cloud associated with RCW 106 - IV. Ammonia towards dust emission'
---
\[firstpage\]
Stars: formation – ISM: clouds – ISM: molecules – ISM: structure – radio lines: ISM.
INTRODUCTION
============
{width="99.00000%"}
The life cycles of high-mass ($\ge$8[M$_\odot$]{}) stars have a major impact on the evolution of galaxies, while in turn, the position of a molecular cloud in the Galaxy has a major impact on the efficiency and type of star formation which occurs therein [@Luna2006]. However, exactly how these stars form, on what timescales and how they shape their environments during this active and energetic phase is poorly understood (see the review by @Zinnecker2007).
The G333 giant molecular cloud (GMC), centred at $l$ = 333.$^{\circ}$2, $b$ = -0.$^{\circ}$4, is a 1.2 $\times$ 0.6 region of the southern Galactic Plane located at a distance of 3.6 kpc [@Lockman1979] and is the fourth most active star forming region in the Galaxy [@Urquhart2014]. The complex forms part of the Galactic Ring of molecular clouds at a Galactocentric radius of 3-5 kpc, it contains the bright [H[ii]{}]{} regions RCW 106 (G332.9-0.6; Rodgers, Campbell & Whiteoak 1960) and G333.6-0.2 [@Manchester1969], as well as high-mass star forming clumps, MSX infrared sources and *Spitzer* “Extended Green Objects” [@Cyganowski2008]. This cloud also contains the very young, high-mass star forming core G333.125$-$0.567 [@Becklin1973; @Garay2004; @Lo2007; @Lo2011] which is an isolated core at an evolutionary stage without any radio continuum or near-infrared detection. We have used the Mopra radio telescope to study the G333 GMC extensively in 20 molecular tracers including $^{13}$CO [@Bains2006], C$^{18}$O [@Wong2008], HCN, HCO$^+$, and N$_2$H$^+$ [@Lo2009].
In order to investigate a large sample of star forming clumps at different stages of evolution, we are conducting a multi-wavelength study of the G333 GMC. This region is within the area of the *Spitzer* GLIMPSE (3.6, 4.5, 5.8, 8.0[${\rm \mu m}$]{}; @Benjamin2003 [@Churchwell2009]) and MIPSGAL (24, 70[${\rm \mu m}$]{}; @Carey2009), and the *Herschel* Hi-GAL (70, 160, 250, 350, 500[${\rm \mu m}$]{}; @Molinari2010) surveys. @Mookerjea2004 have defined multiple millimetre continuum clumps within this one molecular cloud, some associated with maser emission (e.g. @Breen2007), as well as obvious polycyclic aromatic hydrocarbon (PAH) emission. We expect to have a range of evolutionary states within our clumps.
The ammonia () molecule can be used to probe the physical conditions within dense molecular gas. It has a large number of transitions that are sensitive to a variety of excitation conditions and can be detectable in both warm molecular gas and quiescent dark clouds. The rotational temperature and optical depth for each clump can be calculated from multiple ([*J*]{},[*K*]{}) inversion transitions and their hyperfine components, respectively [@Ungerechts1983]. Hence the kinetic temperature can be calculated [@Tafalla2004]; however the lower [(1,1)]{} and [(2,2)]{} transitions are only useful for constraining temperatures lower than 30K [@Danby1988; @Hill2010].
In this paper we report on the selection of dense clumps based on SIMBA dust emission and pointed [(1,1)]{} and [(2,2)]{} observations towards these clumps. Although finding the kinetic temperature is our main concern there are also several other physical parameters that can be determined from our measurements, including virial mass and column density. This is so we will have a sample of clumps within the G333 GMC, with clear selection criteria which will be used to compare the 3-mm molecular transitions and infrared *Herschel* and *Spitzer* dust surveys. We have chosen to study a single molecular cloud as the initial chemical composition of each clump should be very similar. Obtaining a reliable temperature measurement of each clump will allow us, in future papers, to calculate accurate molecular abundances for comparison between clumps.
OBSERVATIONS AND DATA REDUCTION {#sect:observations}
===============================
Selection of clumps {#sect:clumps}
-------------------
We have utilised the 1.2-mm dust continuum emission from the SIMBA instrument on the Swedish-ESO Telescope (SEST; @Mookerjea2004) as a tracer for dense molecular material. To complement the density tracers from @Bains2006, @Wong2008 and @Lo2009, we have convolved the @Mookerjea2004 beam size from 24arcsec to 36arcsec, to match the beam size of the Mopra radio telescope at 100GHz [@Ladd2005]. Using the [starlink]{} tool [cupid]{}, we have chosen the 2D [clumpfind]{} algorithm to automatically, but robustly, detect physically realistic clumps. The [clumpfind]{} algorithm works by identifying local maxima and includes all contiguous pixels down to the user-set contour level. If the contiguous pixels have already been identified in an earlier clump, then those pixels are included in that clump. For contiguous contours a “friends-of-friends” algorithm is used to assign the common pixels to a clump. These steps are repeated until the lowest level is reached. For a complete description of the [clumpfind]{} algorithm, see @Williams1994. Using a 3$\sigma$ detection limit we have identified 63 clumps. The catalogue of clumps found in this paper differ from those in @Mookerjea2004 because of the differing resolutions (24 arcsec versus 36 arcsec). The global positions are shown in Fig. \[fig:overview\], the local environment in Fig \[app:3colour\], and the [clumpfind]{} results can be found in §\[sec:parameters\] and Table \[tbl:cf\_parameters\_jsu\].
Ammonia from the 70-m Tidbinbilla radio telescope {#sec:Tidbinbilla}
-------------------------------------------------
Observations using the 70-m Tidbinbilla radio telescope were conducted in good weather conditions on 2011 July 11 and 12. Pointing errors were measured and corrected every few hours, by observing the quasar 1613-586, within 9$^\circ$ from the target sources, yielding residual pointing errors smaller than 2arcsec. System noise temperatures were in the range of $\sim$45-90K, but typically $\sim$50K. Sensitivity was $\sim$40-100mK, but typically $\sim$50mK. During these observations a left hand circularly polarized signal was recorded across two 64MHz bands centred at 23708 and 23870MHz, covering the [(1,1)]{} and [(2,2)]{} transitions in the first band and the (3,3) transition in the second band. With a bandwidth of 2048channels across 64MHz, each channel width was 31.25kHz (or a velocity channel width of 0.4[kms$^{-1}$]{}). Each source was observed for 2min on-source integration.
The [(1,1)]{} and [(2,2)]{} observations were reduced using the ATNF Spectral Analysis Package[^1] ([asap]{}) and the graphical interface developed by Cormac Purcell[^2] (QuickData.py). The (3,3), if detected, will be used in future work to constrain the temperatures with large uncertainties. The individual on-off scans were subtracted to remove sky emission and visually inspected to identify and remove poor data. Where necessary a low-order polynomial was fitted to the channels free of emission and subtracted to remove baseline anomalies. The remaining scans were combined to produce a single spectrum for both transitions at each observed position. Hanning smoothing was used to improve the signal to noise of these spectra, which reduced the spectral resolution by a factor of 2. An example of the emission detected in a variety of environments within the G333 GMC are shown in Fig. \[fig:3colour\_spectra\] with the full selection available in Fig. \[app:spectra\].
{width="39.00000%"}\
{width="39.00000%"}\
{width="39.00000%"}\
Cool dust from the *Herschel* Space Observatory
-----------------------------------------------
The G333 GMC was observed by *Herschel* [@Pilbratt2010] with the PACS [@Poglitsch2010] and SPIRE [@Griffin2010] instruments in Parallel mode at 70, 160, 250, 350 and 500[${\rm \mu m}$]{} as part of the Open Time Key Programme “Hi-Gal” (PI: S. Molinari; @Molinari2010). The observations comprised of a scan and cross-scan of two 2$^\circ$ fields (field 332.0 with obsid: 1342204046/47 and field 334.0 with obsid: 1342204054/55) taken at a scan speed of 60arcsecs$^{-1}$. These observations were reduced in the Herschel Interactive Processing Environment (HIPE; @Ott2010) to Level 1 (basic calibrated data product), then exported to the [scanamorphos]{} map-making software to create a mosaic of the G333 GMC [@Roussel2013]. The pixel size of the final mosaic was 6.4arcsec per pixel at 160[${\rm \mu m}$]{} (equivalent to the instrument native pixel sizes in the red band) was chosen to decrease the low redundancy of the parallel mode scans (due to the high scan speed) and reduce the effect of correlated noise in the final image.
RESULTS AND DERIVED PARAMETERS {#sect:results}
==============================
We derive masses for the clumps from their integrated flux densities at 1.2-mm under the assumption that the clump is optically thin at this wavelength. The dust mass opacity is also quite uncertain, being based on emission models or laboratory dust measurements that may not reflect the actual material we observe. Each clump identified will be used as a mask and aided by a temperature derived from the observations, accurate abundances will be derived within the clump region for comparing different molecular transitions at 3-mm. For this paper, the peak position within each clump was used to target the [(1,1)]{} and [(2,2)]{} observations. We report the results of the observations and temperature calculations in this paper and will discuss the multi-molecular line abundances in a later paper.
---- ------------------ -------------- --------------- -------------------- ------------ ------------------------ -------------------------- ---------------------------- ----------- --- -- --
ID Clump name Peak Integrated Associations$^\dagger$ RMS Type IR Type Region
RA \[J2000\] Dec \[J2000\] \[Jy beam$^{-1}$\] \[pc\] \[Jy\]
1 G333.604$-$0.210 16:22:08.5 -50:05:53 49.4 0.53 950.8 2wo$_{1,2}$M$_2$ [H[ii]{}]{} region IR-bright A
2 G332.826$-$0.547 16:20:09.6 -50:53:09 15.5 0.46 263.7 2wo$_1$m$_1$b [H[ii]{}]{} region/YSO IR-bright C
3 G333.136$-$0.431 16:21:00.8 -50:35:13 14.5 0.55 409.5 nso$_{1,2}$M$_2$m$_1$ [H[ii]{}]{} region IR-bright B
4 G333.286$-$0.387 16:21:30.3 -50:26:48 9.6 0.49 198.1 nsw [H[ii]{}]{} region IR-bright B
5 G333.309$-$0.369 16:21:31.4 -50:24:42 6.1 0.80 263.5 ns Diffuse [H[ii]{}]{} region IR-bright B
6 G333.465$-$0.160 16:21:19.4 -50:09:32 5.6 0.43 95.1 nspwo$_1$M$_1$m$_1$ $\cdots$ IR-bright A
7 G333.068$-$0.446 16:20:48.6 -50:38:37 5.0 0.49 99.3 nso$_{1,2}$ YSO IR-bright B
8 G332.644$-$0.606 16:19:36.2 -51:03:10 4.6 0.64 173.6 ns Diffuse [H[ii]{}]{} region IR-bright C
9 G333.127$-$0.564 16:21:35.2 -50:41:15 3.9 0.48 87.4 nsp2o$_{1}$M$_1$ $\cdots$ IR-faint B
10 G332.691$-$0.612 16:19:51.6 -51:01:26 2.9 0.43 63.2 nsph $\cdots$ IR-bright C
11 G333.018$-$0.449 16:20:37.9 -50:41:22 2.2 0.69 67.1 nshw Diffuse [H[ii]{}]{} region IR-bright B
12 G332.676$-$0.615 16:19:46.6 -51:02:15 2.1 0.38 38.6 ns $\cdots$ IR-faint C
13 G333.524$-$0.272 16:22:02.7 -50:11:42 1.6 0.38 34.0 nsp $\cdots$ IR-faint A
14 G332.985$-$0.487 16:20:36.6 -50:44:09 1.5 0.45 39.1 nsm$_1$ YSO IR-bright B
15 G333.218$-$0.402 16:21:17.0 -50:30:15 1.5 0.33 27.9 nspw $\cdots$ IR-bright B
16 G333.106$-$0.502 16:21:13.6 -50:39:42 1.4 0.59 60.1 nspm$_1$ YSO IR-bright B
17 G333.566$-$0.295 16:22:20.2 -50:11:22 1.3 0.52 51.6 nsp $\cdots$ IR-faint A
18 G333.074$-$0.399 16:20:37.1 -50:36:41 1.3 0.61 56.8 nsp $\cdots$ IR-faint B
19 G333.722$-$0.209 16:22:38.3 -50:00:43 1.2 0.32 23.4 nsph $\cdots$ IR-bright A
20 G332.585$-$0.559 16:19:08.3 -51:03:50 1.2 0.43 29.8 nsp YSO IR-bright C
21 G333.006$-$0.437 16:20:27.7 -50:40:56 1.2 0.41 30.0 nspm$_1$ Diffuse [H[ii]{}]{} region IR-bright B
22 G332.700$-$0.585 16:19:45.7 -51:00:14 1.0 0.30 16.5 nspm$_1$ YSO and [H[ii]{}]{} region IR-bright C
23 G333.292$-$0.422 16:21:43.1 -50:27:34 0.9 0.63 41.8 nsp $\cdots$ IR-faint B
24 G333.692$-$0.198 16:22:30.1 -50:01:32 0.9 0.40 24.7 ns $\cdots$ IR-faint A
25 G332.773$-$0.582 16:20:04.3 -50:57:03 0.9 0.36 17.0 nsp $\cdots$ IR-faint C
26 G333.772$-$0.260 16:23:05.2 -50:00:40 0.8 0.31 15.8 nsp $\cdots$ IR-faint A
27 G333.539$-$0.245 16:21:59.6 -50:10:03 0.8 0.26 15.2 nsp $\cdots$ IR-faint A
28 G333.483$-$0.225 16:21:41.6 -50:11:59 0.8 0.62 29.7 nsp $\cdots$ IR-faint A
29 G333.633$-$0.257 16:22:25.9 -50:06:11 0.8 0.50 29.0 ns YSO IR-faint A
30 G333.038$-$0.405 16:20:27.4 -50:38:16 0.7 0.43 23.3 ns $\cdots$ IR-faint B
31 G332.794$-$0.597 16:20:14.9 -50:56:28 0.7 0.24 10.8 nsp $\cdots$ IR-bright C
32 G332.670$-$0.644 16:19:54.6 -51:03:39 0.7 0.44 21.0 nsp $\cdots$ IR-faint C
33 G333.012$-$0.520 16:20:51.8 -50:44:01 0.7 0.43 17.3 nsp $\cdots$ IR-faint B
34 G333.568$-$0.168 16:21:47.8 -50:05:36 0.7 0.43 20.4 nsp $\cdots$ IR-faint A
35 G332.741$-$0.620 16:20:03.9 -50:59:52 0.7 0.42 16.0 nsphwM$_1$m$_{1,2}$o$_1$ $\cdots$ IR-faint C
36 G333.754$-$0.230 16:22:53.0 -50:00:06 0.7 0.16 7.7 nspm$_1$h YSO IR-faint A
37 G333.171$-$0.431 16:21:13.0 -50:33:44 0.6 0.31 10.9 nsp$\cdots$ $\cdots$ IR-faint B
38 G333.681$-$0.257 16:22:41.5 -50:04:44 0.6 0.47 18.2 ns $\cdots$ IR-faint A
39 G333.642$-$0.106 16:21:51.3 -49:59:44 0.6 0.40 16.5 nsph $\cdots$ IR-bright A
40 G333.047$-$0.479 16:20:50.2 -50:40:40 0.6 0.40 16.6 nsp $\cdots$ IR-faint B
41 G332.971$-$0.467 16:20:27.0 -50:43:47 0.6 0.41 20.1 ns $\cdots$ IR-bright B
42 G333.206$-$0.366 16:21:01.9 -50:29:08 0.5 0.35 13.5 nsp $\cdots$ IR-bright B
43 G332.809$-$0.700 16:20:46.9 -50:59:57 0.5 0.18 6.8 nspM$_{1,2}$ $\cdots$ IR-bright C
44 G332.558$-$0.591 16:19:08.9 -51:06:07 0.5 0.20 7.2 nsp $\cdots$ IR-bright C
45 G333.183$-$0.396 16:21:05.2 -50:31:34 0.5 0.10 5.0 nsp $\cdots$ IR-faint B
46 G333.521$-$0.239 16:21:53.0 -50:10:48 0.5 0.21 7.3 nsp $\cdots$ IR-faint A
47 G333.074$-$0.558 16:21:19.6 -50:42:48 0.4 0.31 8.1 nsph $\cdots$ IR-bright B
48 G332.759$-$0.467 16:19:29.4 -50:52:27 0.4 0.12 5.1 ns $\cdots$ IR-faint C
49 G333.236$-$0.520 16:21:51.7 -50:34:26 0.4 0.45 12.6 nsp $\cdots$ IR-bright B
50 G332.715$-$0.676 16:20:13.4 -51:02:50 0.4 0.31 8.7 ns $\cdots$ IR-faint C
51 G332.903$-$0.546 16:20:31.3 -50:49:47 0.4 0.21 5.8 nsp $\cdots$ IR-faint C
52 G333.097$-$0.352 16:20:30.7 -50:33:20 0.4 0.23 6.2 ns $\cdots$ IR-bright B
53 G333.563$-$0.145 16:21:38.7 -50:04:40 0.3 0.31 6.1 ns $\cdots$ IR-faint A
54 G333.636$-$0.165 16:22:04.0 -50:02:37 0.3 0.10 4.2 ns $\cdots$ IR-faint A
55 G333.645$-$0.133 16:21:59.0 -50:00:50 0.3 0.30 7.6 nsp $\cdots$ IR-faint A
56 G333.657$-$0.133 16:22:01.4 -50:00:01 0.3 0.16 3.8 nsp $\cdots$ IR-faint A
---- ------------------ -------------- --------------- -------------------- ------------ ------------------------ -------------------------- ---------------------------- ----------- --- -- --
---- ------------------ -------------- --------------- -------------------- -------- ------------------------ ---------- ---------- ----------- --- -- --
ID Clump name Peak Sum Associations$^\dagger$ RMS Type IR Type Region
RA \[J2000\] Dec \[J2000\] \[Jy beam$^{-1}$\] \[pc\] \[Jy\]
57 G333.660$-$0.218 16:22:25.0 -50:03:43 0.3 0.19 4.7 nsp $\cdots$ IR-faint A
58 G333.410$-$0.328 16:21:47.9 -50:19:02 0.3 0.52 10.4 nsph $\cdots$ IR-bright A
59 G333.595$-$0.077 16:21:30.0 -50:00:40 0.3 0.32 5.1 nsp $\cdots$ IR-bright A
60 G332.894$-$0.567 16:20:32.6 -50:51:16 0.2 0.22 4.4 nsp $\cdots$ IR-faint C
61 G333.262$-$0.275 16:20:54.0 -50:23:02 0.2 0.28 4.5 nsp $\cdots$ IR-bright B
62 G332.753$-$0.561 16:19:54.0 -50:56:47 0.2 0.23 4.6 nsp $\cdots$ IR-bright C
63 G332.614$-$0.671 16:19:43.5 -51:07:40 0.2 0.33 6.0 nsp $\cdots$ IR-faint C
---- ------------------ -------------- --------------- -------------------- -------- ------------------------ ---------- ---------- ----------- --- -- --
\
$\dagger$ References – Water (22GHz ; @Braz1982 [@Breen2007; @Breen2010; @Walsh2011]), Class I (44.01GHz; @Slysh1994, @Voronkov2014 and 95.1GHz; @Ellingsen2005) and II (6.7GHz; @Caswell1996 [@Caswell1997; @Caswell2009; @Caswell2011; @Ellingsen1996]; and 12.2GHz; @Breen2012) methanol () and hydroxyl (OH; 1665/1667/1720MHz; @Caswell1980 [@Caswell1995; @Caswell1998] and 6.035 GHz; @Caswell1997) masers.
Clump parameters {#sec:parameters}
----------------
The parameters of the 63 SIMBA 1.2-mm dust clumps identified with [clumpfind]{} can be found in Table \[tbl:cf\_parameters\_jsu\]. The [clumpfind]{} number is listed in Column 1, followed by the identifier for each source using the Galactic longitude and latitude designation for the peak pixel in Column 2. Columns 3 and 4 list the coordinates for the clump centroid. We find that the peak of the dust emission correlates poorly with the centroid of the emission, with a maximum deviation of 39.5arcsec ($\sim$1.3 convolved SIMBA beam). Since the size and shapes of clumps can differ when identified with different algorithms, we have performed our observations towards the peak dust emission within each clump. The peak and integrated flux (in Jy beam$^{-1}$ and Jy, respectively) within the clumps is found in Columns 5 and 7. [CUPID]{} outputs the pixel size of the clumps projected onto the equatorial plane, so the average projected values were then converted into parsecs by assuming a distance of 3.6kpc [@Lockman1979] and are displayed in Column 6. Column 8 lists the nearby ($<$ 30arcsec) infrared and maser associations (see §\[sec:separation\] or Table \[tbl:cf\_parameters\_jsu\] caption for further information). If there is a nearby ($<$ 30arcsec) RMS object, then its type is listed in Column 9. Column 10 identifies whether the clump has been classified, via *Spitzer* GLIMPSE 8.0[${\rm \mu m}$]{} emission, as IR-bright or IR-faint. Column 11 identifies which region the clump has been allocated to (see §\[sec:variations\]).
The distribution of the clumps throughout the G333 GMC can be seen in Fig. \[fig:overview\]. A sample of the dust clumps and their accompanying [(1,1)]{} and [(2,2)]{} emission and Gaussian fits, where possible, can be seen in Fig. \[fig:3colour\_spectra\], with the full version available in Fig. \[app:3colour\] and Fig. \[app:spectra\], for the infrared composites and spectra, respectively.
Derivation of physical parameters {#sect:formulae}
---------------------------------
Parameter Mean 1$\sigma$ Min Max
-------------------------------------------------- --------------- ----------- ----------------- ---------------- --
[$T_{\rm{kin}}$]{}\[K\] 18.1$\pm$0.5 3.6 12.5 35.3
[$T_{\rm{rot}}$]{}\[K\] 22.0$\pm$1.0 7.4 13.6 63.6$^\dagger$
Log\[N()\] \[cm$^{-2}$\] 15.9$\pm$2.3 0.24 15.3 16.6
Log\[n()\] \[$\times 10^5$ cm$^{-3}$\] 4.5$\pm$0.4 2.8 0.7 16.0
[(1,1)]{} [$v_{\rm{LSR}}$]{} \[[kms$^{-1}$]{}\] -52.9$\pm$1.1 8.1 -88.5$^\dagger$ -42.57
[(1,1)]{} $\Delta$V \[[kms$^{-1}$]{}\] 2.2$\pm$0.9 0.87 1.1 4.5
$\tau_{(1,1)}$ 1.9$\pm$0.1 0.86 0.16 3.76
[(2,2)]{} [$v_{\rm{LSR}}$]{}\[[kms$^{-1}$]{}\] -53.1$\pm$1.2 8.2 -88.7$^\dagger$ -42.8
[(2,2)]{} $\Delta$V \[[kms$^{-1}$]{}\] 2.7$\pm$0.4 1.0 1.3 6.0
Log(Clump mass) \[[M$_\odot$]{}\] 3.37$\pm$2.61 3.45 2.44 4.12
Log(Virial mass) \[[M$_\odot$]{}\] 2.99$\pm$2.18 3.03 2.03 3.68
: Summary of calculated physical parameters.[]{data-label="tbl:derived_parameters"}
\
\
$^\dagger$These values are outliers and their corresponding clumps and physical parameters have been excluded from §\[sec:exclusions\] onwards. The next highest [$T_{\rm{kin}}$]{} is Clump 21 (G333.006$-$0.437) with [$T_{\rm{kin}}$]{}=35.6K.
---- ------------------ ---------------- ---------------- ------------------------- ----------------- ---------------- -------------------- -------------------- ----------------------- ---------------- -------------------- -------------------- ------------------- ------------------- --------------------
Spectral
ID Clump Name $T_{\rm{rot}}$ $T_{\rm{kin}}$ Log\[$N({\rm{NH_3}})$\] Log\[n(H$_2$)\] T$_{\rm{A}}^*$ [$v_{\rm{LSR}}$]{} $\Delta$V $\tau_{\rm{(m,1,1)}}$ T$_{\rm{A}}^*$ [$v_{\rm{LSR}}$]{} $\Delta$V Clump Virial features$^\dagger$
\[K\] \[K\] \[cm$^{-2}$\] \[cm$^{-3}$\] \[K\] \[[kms$^{-1}$]{}\] \[[kms$^{-1}$]{}\] \[K\] \[[kms$^{-1}$]{}\] \[[kms$^{-1}$]{}\] \[[M$_\odot$]{}\] \[[M$_\odot$]{}\]
1 G333.604$-$0.210 $\cdots$ $\cdots$ $\cdots$ $\cdots$ $<$0.2 $\cdots$ $\cdots$ $\cdots$ $\cdots$ $\cdots$ $\cdots$ $\cdots$ $\cdots$ Aa
2 G332.826$-$0.547 22.0 28.4 16.2 5.75 0.9 $-$57.62 4.50 1.16 0.6 $-$57.16 5.96 3.99 3.62 Hbhgb
3 G333.136$-$0.431 23.8 31.9 15.5 4.97 1.1 $-$53.56 4.45 0.25 0.7 $-$53.79 4.23 4.12 3.68 HABhgab
4 G333.286$-$0.387 22.5 29.2 16.1 5.62 0.9 $-$51.35 2.87 1.52 0.7 $-$51.46 3.50 3.85 3.25 Hbhgb
5 G333.309$-$0.369 20.5 25.5 15.8 5.11 0.9 $-$50.29 2.62 0.85 0.5 $-$50.36 3.10 4.04 3.39 Hg
6 G333.465$-$0.160 19.8 24.4 16.0 5.58 0.9 $-$42.57 2.62 1.38 0.5 $-$42.79 3.10 3.62 3.13 Hg
7 G333.068$-$0.446 20.6 25.7 16.1 5.62 2.5 $-$53.01 2.02 2.25 1.8 $-$53.12 2.66 3.62 2.95 Hhg
8 G332.644$-$0.606 20.1 24.9 16.1 5.50 0.9 $-$49.30 3.01 1.64 0.6 $-$49.43 3.93 3.87 3.41 Hbgb
9 G333.127$-$0.564 17.3 20.3 16.5 6.03 1.6 $-$57.17 3.26 3.76 1.1 $-$57.49 4.31 3.69 3.35 Hbhgb
10 G332.691$-$0.612 18.3 22.0 16.2 5.78 0.6 $-$47.89 2.68 2.51 0.4 $-$47.98 3.23 3.51 3.13 Hbhgb
11 G333.018$-$0.449 25.6 35.6 16.1 5.47 0.4 $-$54.34 3.08 1.29 0.3 $-$54.52 3.25 3.27 3.46 Hbg
12 G332.676$-$0.615 16.3 18.8 16.2 5.83 1.4 $-$57.46 1.89 3.51 0.9 $-$57.62 2.40 3.38 2.78 Hhg
13 G333.524$-$0.272 17.3 20.3 16.0 5.63 1.3 $-$49.34 2.15 1.95 0.7 $-$49.45 2.76 3.28 2.90 Hhg
14 G332.985$-$0.487 18.4 22.1 15.9 5.46 1.6 $-$52.56 1.84 1.92 1.0 $-$52.61 2.31 3.29 2.83 Hhg
15 G333.218$-$0.402 20.1 24.8 15.9 5.59 1.1 $-$51.92 1.76 1.86 0.8 $-$52.06 2.29 3.08 2.65 Hhg
16 G333.106$-$0.502 18.1 21.5 16.0 5.44 1.0 $-$56.12 1.95 2.18 0.6 $-$56.16 2.69 3.50 3.00 Hhg
17 G333.566$-$0.295 15.1 17.1 15.9 5.39 1.1 $-$46.09 1.05 3.27 0.6 $-$46.12 1.34 3.57 2.41 Hahga
18 G333.074$-$0.558 15.9 18.2 16.0 5.42 1.0 $-$55.77 1.74 2.75 0.6 $-$55.95 2.30 3.52 3.31 Hg
19 G333.722$-$0.209 15.6 17.8 16.0 5.70 1.5 $-$46.58 1.41 2.85 0.8 $-$46.69 1.92 3.20 2.46 Hhg
20 G332.585$-$0.559 15.6 17.8 16.0 5.58 0.8 $-$50.99 1.47 2.80 0.4 $-$51.03 2.07 3.30 2.61 Hhg
21 G333.006$-$0.437 35.3 63.6 16.6 6.20 0.2 $-$56.20 4.46 1.80 0.2 $-$55.85 5.17 2.64 3.56 Hbgb
22 G332.700$-$0.585 19.1 23.2 15.9 5.63 0.5 $-$59.15 2.40 1.46 0.3 $-$59.76 2.46 2.89 2.88 Hbg
23 G333.292$-$0.422 18.0 21.4 15.9 5.31 0.9 $-$49.94 1.96 1.56 0.5 $-$49.95 2.50 3.34 3.03 Hg\*
24 G333.692$-$0.198 12.5 13.6 16.0 5.61 0.6 $-$50.37 1.25 3.53 0.2 $-$50.36 1.95 3.39 2.44 Hg
25 G332.773$-$0.582 18.7 22.6 15.6 5.25 0.8 $-$55.86 1.72 0.95 0.4 $-$55.92 2.12 2.92 2.68 Hg
26 G333.772$-$0.260 16.6 19.3 15.9 5.62 0.7 $-$48.95 2.43 1.51 0.3 $-$49.02 2.74 2.98 2.91 Hg
27 G333.539$-$0.245 16.5 19.2 15.7 5.49 0.6 $-$48.20 1.39 1.51 0.3 $-$48.32 1.64 2.97 2.35 Hbgb
28 G333.483$-$0.225 14.3 16.0 16.0 5.42 0.8 $-$48.58 1.47 2.89 0.4 $-$48.62 1.99 3.36 2.77 Hg
30 G333.038$-$0.405 17.5 20.6 16.3 5.88 0.3 $-$51.75 4.41 1.85 0.2 $-$52.96 5.49 3.03 2.74 Hbgb
31 G332.794$-$0.597 20.7 26.0 15.9 5.73 0.3 $-$54.89 2.41 1.19 0.2 $-$55.10 2.71 2.65 2.78 Hbgb
32 G332.670$-$0.644 18.0 21.5 15.7 5.27 0.4 $-$49.58 1.96 1.01 0.2 $-$49.58 2.15 3.04 2.89 Hbb
33 G333.012$-$0.520 16.9 19.7 15.8 5.38 0.8 $-$52.94 2.25 1.18 0.4 $-$52.94 2.41 3.01 2.99 Hg
34 G333.568$-$0.168 13.1 14.3 16.0 5.58 0.9 $-$88.53 1.29 3.46 0.4 $-$88.69 1.81 3.27 2.50 Hg
35 G332.741$-$0.620 15.4 17.6 15.8 5.39 0.8 $-$49.81 1.48 2.05 0.4 $-$49.88 2.10 3.04 2.60 Hg
36 G333.754$-$0.230 16.0 18.4 16.0 6.01 0.6 $-$49.70 1.69 2.70 0.3 $-$49.96 2.35 2.69 2.30 Hg
37 G333.171$-$0.431 19.3 23.5 16.0 5.72 1.1 $-$51.07 2.51 1.53 0.7 $-$51.07 3.14 2.70 2.93 Hg
38 G333.681$-$0.257 19.4 23.7 15.3 4.84 0.2 $-$47.04 1.11 0.77 0.1 $-$47.69 2.34 2.93 2.39 Hagb
39 G333.642$-$0.106 13.3 14.6 16.1 5.71 0.8 $-$87.95 1.71 3.06 0.3 $-$87.98 2.11 3.17 2.72 Hg
40 G333.047$-$0.479 16.8 19.5 15.8 5.41 0.6 $-$52.55 1.99 1.40 0.3 $-$52.79 2.76 2.99 2.85 Hg
42 G333.206$-$0.366 $<$9.3 $<$9.5 $<$14.0 $<$3.67 0.2 $-$48.49 2.21 0.16 $\cdots$ $\cdots$ $\cdots$ $\cdots$ 2.87 H
43 G332.809$-$0.700 17.7 21.0 15.8 5.75 0.5 $-$52.98 1.17 2.44 0.3 $-$53.17 1.87 2.56 2.03 Hagb
44 G332.558$-$0.591 19.0 23.0 16.0 5.91 0.7 $-$49.64 2.28 1.93 0.4 $-$49.61 2.49 2.54 2.66 Hg
45 G333.183$-$0.396 17.3 20.4 15.8 6.01 0.7 $-$49.95 1.73 1.70 0.4 $-$49.96 2.57 2.44 2.11 Hg
46 G333.521$-$0.239 16.1 18.5 15.8 5.69 0.9 $-$48.16 1.54 1.79 0.4 $-$48.34 2.04 2.67 2.35 Hg
47 G333.074$-$0.399 17.0 19.8 16.1 5.82 1.2 $-$53.69 2.75 1.80 0.6 $-$53.78 3.44 2.72 2.62 Hhg
48 G332.759$-$0.467 17.0 19.9 15.6 5.73 0.6 $-$52.87 1.82 0.93 0.2 $-$52.80 2.44 2.47 2.25 Hg\*
50 G332.715$-$0.676 17.0 19.8 16.0 5.72 0.2 $-$45.64 1.92 2.37 0.1 $-$45.38 1.52 2.68 2.79 Hbgb
51 G332.903$-$0.546 15.0 16.9 15.7 5.59 0.6 $-$55.37 1.15 1.91 0.3 $-$55.61 1.37 2.62 2.09 Hbg
52 G333.097$-$0.352 17.1 20.0 15.6 5.45 0.5 $-$51.60 1.71 1.05 0.2 $-$51.59 1.85 2.55 2.48 Hg
53 G333.563$-$0.145 15.9 18.2 15.4 5.12 0.2 $-$44.76 1.90 0.52 0.1 $-$44.60 4.10 2.58 2.72 Hbgb
60 G332.894$-$0.567 15.2 17.2 15.9 5.77 0.5 $-$56.42 1.67 1.96 0.2 $-$56.40 2.07 2.49 2.44 Hg
---- ------------------ ---------------- ---------------- ------------------------- ----------------- ---------------- -------------------- -------------------- ----------------------- ---------------- -------------------- -------------------- ------------------- ------------------- --------------------
\
$\dagger$ The presence of a ‘H’ denotes a [(1,1)]{} hyperfine, ‘A/a’ denotes an absorption feature, ‘B/b’ denotes a blended spectra, ‘g’ denotes a Gaussian was fit to the [(2,2)]{}, ‘hg’ denotes that a [(2,2)]{} hyperfine structure could be seen but since the signal-to-noise was below 3$\sigma$, a Gaussian was used to model the fit, and ‘g\*’ denotes a [(2,2)]{} characteristic not included in any of the other cases (see §\[sect:formulae\] for further details).
The spectra presented here are calibrated to the corrected antenna temperature scale ($T_{\rm A}^*$) and, since the main beam efficiency has not been well determined at this frequency for the 70-m Tidbinbilla radio telescope, we do not convert these to the telescope-independent main-beam temperature scale ($T_{\rm mb}$). However, all our analysis relies on the ratio of the two ammonia transitions, which are essentially independent of calibration as they probe the same volume of gas, so the choice of temperature scale is not important.
We have followed the same fitting routine as described in @Urquhart2011 and include an overview as follows. For the sources with a detectable [(1,1)]{} hyperfine structure ($\sim$75 per cent), we simultaneously fit all 18 hyperfine components and derive the optical depth and FWHM. For the remaining [(1,1)]{} sources and all the [(2,2)]{} sources, a single Gaussian profile was fitted to the main line to obtain a corrected antenna temperature. The [(1,1)]{} and [(2,2)]{} FWHM were obtained for all sources by fitting the hyperfine components of their spectra with their respective main line emission in order to remove the effect of line broadening due to the optical depth. An example of the resulting fits are shown in Fig. \[fig:3colour\_spectra\] and the full selection of reduced [(1,1)]{} and [(2,2)]{} inversion spectra and the resulting fits can be seen in Fig. \[app:spectra\]. Emission from the ammonia inversion transition is detected towards 50 (79 per cent) and 49 (78 per cent) of the 63 SIMBA 1.2-mm dust continuum sources, for [(1,1)]{} and [(2,2)]{}, respectively. There were 13 (21 per cent) clumps without any detectable . We present a summary of the detection rates in Table \[tbl:derived\_parameters\]. The standard error of the mean ($s/\sqrt{n}$) is included with the mean. The results of the calculations and spectral fits are given in Table \[tbl:nh3\_parameters\]. The presence of a ‘H’ denotes a [(1,1)]{} hyperfine, ‘g’ denotes a Gaussian was fit to the [(2,2)]{}, and ‘hg’ denotes that a [(2,2)]{} hyperfine structure could be seen but since the signal-to-noise was below 3$\sigma$, a Gaussian was used to model the fit. For two sources, Clumps 23 (G333.292$-$0.422) and 48 (G332.759$-$0.467), a ‘g\*’ is used to indicate that the clump has [(2,2)]{} emission at the G333 [$v_{\rm{LSR}}$]{} as well as at -76[kms$^{-1}$]{} and -69[kms$^{-1}$]{}, respectively. Features associated with the [(1,1)]{} and [(2,2)]{} were shown with an upper and lower case, respectively: ‘A/a’ denotes an absorption feature and ‘B/b’ denotes a blended spectra.
In this section we will describe the method used to determine the physical properties of the millimetre clumps identified within the G333 GMC. The fitted ([$v_{\rm{LSR}}$]{}, FWHM) and calculated (optical depth and corrected antenna temperatures) parameters are presented in Table \[tbl:nh3\_parameters\]. The distribution of various derived parameters are shown in Fig. \[fig:histo\_SFNSF\]. The uncertainties in the detection rates have been calculated using binomial statistics.
### Optical depth, $\tau$
The total optical depth of the transition, $\tau_{\rm{(1,1)}}$, was obtained numerically from the ratio of the satellite to main line intensity of the hyperfine fits described in §\[sec:Tidbinbilla\] [@Ho1983]: $$\label{eq:tau}
\frac{\Delta T^{*}_{a}\left(J,K,m\right)}{\Delta T^{*}_{a}\left(J,K,s\right)} = \frac{1-e^{-\tau\left(J,K,m\right)}}{1-e^{-\alpha\tau\left(J,K,m\right)}}$$ where *J* and *K* are the principal quantum numbers corresponding to the total angular momentum and its projection along the molecular axis, $\Delta T^{*}_{a}$ is the observed brightness temperature, m and s refer to the main and satellite hyperfine components, $\tau$ is the optical depth, and $\alpha$ is the line intensity ratio of the satellite-to-main line \[$\alpha$ = 0.28 and 0.22 for the (1,1) satellites\]. In using this equation, we have assumed that the hyperfine components have the same beam-filling factor and excitation temperature. This is an acceptable assumption as the energy separations and probability of special excitation mechanisms that differentiate between the hyperfine components are small.
### Rotation temperature, [$T_{\rm{rot}}$]{}
The rotational temperature, [$T_{\rm{rot}}$]{}, for the [(1,1)]{} and [(2,2)]{} transitions can be calculated using the main line intensities and the [(1,1)]{} main quadrupole transition optical depth [@Ho1983]: $$T_\mathrm{rot} = \frac{-T_0}{\ln \left\{\frac{-0.282}{\tau _{(1,1,m)}} \ln \left[1-\frac{\Delta T_{\mathrm{a}}^{*}\left(2,2\right)}{\Delta T_{\mathrm{a}}^{*}\left(1,1,m\right)}\left(1-\mathrm{e}^{-\tau _(1,1,m)} \right) \right] \right\}}~[{\rm{K}}]$$ where $T_0=\frac{E_{(2,2)}-E_{(1,1)}}{k_B}\approx 41.5$K is the temperature of the energy difference between the [(1,1)]{} and [(2,2)]{} levels, $T_{\mathrm{a}}^{*}$ is the observed brightness temperature and $\tau$ is the optical depth. Fourteen clumps did not have a detectable [(2,2)]{} emission, whilst thirteen clumps did not have a detectable [(1,1)]{} emission. Clump 42 (G333.206$-$0.366), has a detectable [(1,1)]{}, using the root mean square of the [(2,2)]{} band, we calculate an upper limit of 9.3K and 9.5K for [$T_{\rm{rot}}$]{} and [$T_{\rm{kin}}$]{}, respectively. For all other clumps (1, 29, 41, 49, 54-59, 61-63), fitting the [(1,1)]{} and/or [(2,2)]{} spectra was not possible so no physical parameters have been calculated for these clumps.
### Kinetic temperature, [$T_{\rm{kin}}$]{}
For [$T_{\rm{kin}}$]{}$<$ $T_0 \approx 41.5$K, an empirical relationship between the rotational and kinetic temperature can be calculated [@Swift2005]: $$T_\mathrm{rot} = \frac{T_\mathrm{kin}} {1+\frac{T_\mathrm{kin}}{T_0}\ln\left[ 1+0.6\times \exp\left(\frac{-15.7}{T_\mathrm{kin}}\right)\right]}~[{\rm{K}}]$$ however, at high kinetic temperatures this expression underestimates the [$T_{\rm{rot}}$]{}. From the original sample of 63 SIMBA clumps, 49 had detectable [(2,2)]{} emission. However, one clump displays a [$T_{\rm{rot}}$]{} which has large uncertainties \[Clump 21 (G333.006$-$0.437)\].
### Column density, N
By assuming that all hyperfine lines have the same excitation temperature and that the excitation conditions are homogeneous along the beam, then the column density at a given transition (e.g. N$_{(1,1)}$) can be written [@Mangum1992]: $$\label{eq:coldensity}
N_{\rm{(1,1)}} = 6.60 \times 10^{14} \Delta v_{\rm{(1,1)}}\tau_{\rm{(1,1)}}\frac{T_\mathrm{rot}}{\nu_{\rm{(1,1)}}} ~~[{\rm{cm}}^{-2}]$$ where $\Delta v_{\rm{(1,1)}}$ is the FWHM of the (1,1) transition in [kms$^{-1}$]{} and $\nu_{\rm{(1,1)}}$ is the transition frequency in GHz. We estimate the total ammonia column density following [@Li2003]: $$\begin{aligned}
\label{eq:coldensity_total}
N_{\rm{NH_3}} &= N_{\rm{(1,1)}} \left[ 1+\frac{1}{3}{\rm{exp}}\left(\frac{23.1}{T_\mathrm{rot}}\right)+\frac{5}{3}{\rm{exp}}\left(\frac{-41.2}{T_\mathrm{rot}}\right)\right.\nonumber \\
&\left.+\frac{14}{3}{\rm{exp}}\left(\frac{-99.4}{T_\mathrm{rot}}\right) \right] ~~[{\rm{cm}}^{-2}]\end{aligned}$$
### Volume density, n
By assuming a fixed /H$_2$ abundance of 10$^{-8}$ [@Johnstone2010], the volume density, $n$, can be calculated: $$\label{eq:voldensity}
n = \frac{N}{2R} ~~[{\rm{cm}}^{-3}]$$ where $N$ is the column density in cm$^{-2}$ and $R$ is the clump radius in cm. The mean volume density within the G333 GMC is 4.5 $\times~10^5$cm$^{-3}$, with a standard deviation of 2.8 $\times~10^5$cm$^{-3}$.
### Clump (M$_{clump}$) and virial (M$_{virial}$) masses
Assuming the dust is optically thin and using the $T_{kin}$ derived from , we calculate the clump mass, $M_{clump}$, as [@Hildebrand1983]: $$\label{eqn:mclump}
M_{clump}=\frac{F_\nu D^2}{\kappa_\nu B_\nu\left(T_{kin}\right)} \left[ M_{\sun}\right]$$ where $F_{\nu}$ is the 1.2-mm flux density, $D$ is the distance, $\kappa_\nu$ is the dust mass opacity coefficient [@Ossenkopf1994] and $B_{\nu}\left(T_{kin}\right)$ is the Planck function at the dust temperature, $T_{kin}$.
Assuming that we have a Gaussian volume density distribution throughout the clump and that the gravitational energy within a clump is balanced by the turbulent energy, we calculate the virial mass, $M_{virial}$, following [@Protheroe2008]: $$M_{virial}= 444 \left(\frac{R}{1\, \mathrm{pc}}\right)\left(\frac{v_{\mathrm{FWHM}}}{1\,\mathrm{km s}^{-1}}\right)^2 \left[ M_{\sun}\right]$$ where $R$ is the radius of the SIMBA 1.2-mm dust clump and $v_{\mathrm{FWHM}}$ is the FWHM.
-------------------------------------------------- -------- ----------------- ----------- -------- -------- -------- -------- ----------------- ----------- -------- -------- --------
Parameter Number Mean 1$\sigma$ Median Min Max Number Mean 1$\sigma$ Median Min Max
[$T_{\rm{rot}}$]{} \[K\] 20 19.6$\pm$0.6 2.6 19.5 15.6 25.6 26 16.6$\pm$0.3 1.5 16.7 12.5 19.4
[$T_{\rm{kin}}$]{} \[K\] 20 24.2$\pm$1.0 f 4.5 23.8 17.8 35.6 26 19.3$\pm$0.4 2.3 19.4 13.6 23.7
Log\[N()\] \[cm$^{-2}$\] 20 16.0$\pm$0.04 0.2 16.0 15.5 16.2 26 15.9$\pm$0.05 0.3 15.9 15.3 16.5
Log\[n()\] \[$\times 10^2$ cm$^{-3}$\] 20 5.57$\pm$0.05 0.23 5.61 4.97 5.91 26 5.56$\pm$0.06 0.29 5.60 4.84 6.03
[(1,1)]{} [$v_{\rm{LSR}}$]{} \[[kms$^{-1}$]{}\] 21 -51.84$\pm$0.82 3.78 -51.92 -59.15 -42.57 26 -50.97$\pm$0.71 3.63 -49.95 -57.46 -44.76
[(1,1)]{} FWHM \[[kms$^{-1}$]{}\] 21 2.44$\pm$0.19 0.86 2.40 1.17 4.50 26 1.90$\pm$0.14 0.70 1.78 1.05 4.41
$\tau_{(1,1)}$ 21 1.64$\pm$0.16 0.74 1.64 0.16 2.85 26 1.96$\pm$0.18 0.90 1.82 0.52 3.76
[(2,2)]{} [$v_{\rm{LSR}}$]{} \[[kms$^{-1}$]{}\] 20 -52.11$\pm$0.85 3.80 -52.34 -59.76 -42.79 26 -51.11$\pm$0.72 3.68 -49.96 -57.62 -44.60
[(2,2)]{} FWHM \[[kms$^{-1}$]{}\] 20 2.95$\pm$0.22 0.98 2.70 1.85 5.96 26 2.50$\pm$0.18 0.92 2.35 1.34 5.49
Log($M_{clump}$) \[[M$_\odot$]{}\] 20 3.31$\pm$0.12 0.53 3.30 2.54 4.12 26 2.99$\pm$0.07 0.36 2.99 2.44 3.69
Log($M_{virial}$) \[[M$_\odot$]{}\] 21 2.95$\pm$0.09 0.42 2.88 2.03 3.68 26 2.67$\pm$0.07 0.34 2.73 2.09 3.35
-------------------------------------------------- -------- ----------------- ----------- -------- -------- -------- -------- ----------------- ----------- -------- -------- --------
\
ANALYSIS AND DISCUSSION {#sect:discussion}
=======================
{width="98.00000%"}
Separation into clumps with and without 8[${\rm \mu m}$]{} emission {#sec:separation}
-------------------------------------------------------------------
Through visual inspection of *Spitzer* GLIMPSE 8[${\rm \mu m}$]{} emission we separated our sample of 63 SIMBA 1.2-mm clumps into two categories: clumps with (IR-bright) or without (IR-faint) significant 8[${\rm \mu m}$]{} emission (Fig. \[fig:distro\_SFNSF\]), since the 8[${\rm \mu m}$]{} band is sensitive to ionised PAH (PAH$^+$) emission which is excited by the ultraviolet radiation from high-mass stars in photon dominated regions (PDRs). MIPSGAL 24[${\rm \mu m}$]{} emission, which represents heated very small dust grains was not used due to saturation in the 24[${\rm \mu m}$]{} band within the G333 GMC.
We identified 30 IR-bright and 33 IR-faint clumps. Column 8 of Table \[tbl:cf\_parameters\_jsu\] lists the nearby ($<$ 30arcsec) star forming signatures (see table caption for details). The Red MSX Source (RMS) Survey uses a multi-wavelength follow-up to disentangle their colour-selected MSX and 2MASS point sources into stages of star formation [@Lumsden2013]. There are RMS objects associated within 30 arcsec of the SIMBA dust peak for 15 of 63 clumps (24 per cent). The type of RMS object is listed in Column 9 of Table \[tbl:cf\_parameters\_jsu\]. Possible types include YSO, [H[ii]{}]{} region/YSO, [H[ii]{}]{} region and diffuse [H[ii]{}]{} region. Masers are known to be signposts for star formation so the water (), Class I and II (6.7GHz and 12.2GHz) methanol () and hydroxyl (OH; 1665/1667/1720MHz and 6.035GHz) masers have been overlaid to show potential sites of star formation. The different classes of masers indicate different pumping mechanisms. Class I masers are collisionally pumped and associated with regions with outflows and shocks. Class II masers are radiatively pumped and exclusively associated with high-mass star forming regions [@Minier2003; @Breen2013]. The infrared (4.5, 8.0, 160[${\rm \mu m}$]{}) dust emission has also been included to show regions of shocked gas, PDRs, and cool dust, respectively.
We find that our SIMBA dust clumps are associated with a variety of environments within the G333 GMC. Each of our clumps are associated with varying amounts of infrared emission, as seen in Fig. \[fig:3colour\_spectra\] and \[app:3colour\]. Emission at 8.0[${\rm \mu m}$]{} is common, which indicates that the G333 GMC is undergoing star formation, but there are also pockets of 160[${\rm \mu m}$]{} emission at sites of cool dust which may show regions where the embedded protostar has not significantly warmed its surroundings. It is known that modelling the spectral energy distribution of the dust can be used to identify clumps at different stages of star formation and that is a good tracer of the dense cold gas which traces the molecular gas involved in star formation (e.g. @Dunham2011 [@Urquhart2011; @Wienen2012]). By combining the information from both the and dust emission for the G333 GMC, we will develop a better understanding of the star formation occurring therein.
### Main cloud clumps {#sec:exclusions}
From our original sample of 63 SIMBA dust clumps, we have detected [(1,1)]{} towards 50 clumps (79 per cent) and [(2,2)]{} towards 49 clumps (78 per cent). No was detected towards Clump 1 (G333.604-0.210) due to self-absorption against a strong continuum emission generated by the [H[ii]{}]{} region, RCW 106 [@Rodgers1960]. The majority of clumps with no detectable amounts of are associated with the weakest millimetre continuum sources within the G333 GMC. The [(1,1)]{} and [(2,2)]{} detection rates for the IR-bright and IR-faint clumps are not obviously different. Of the 30 IR-bright clumps, 23 (77 per cent) and 22 (73 per cent) of the clumps had detectable [(1,1)]{} and [(2,2)]{} emission, respectively. Of the 33 IR-faint clumps, 27 clumps (82 per cent) had both [(1,1)]{} and [(2,2)]{} emission.
Two clumps (one clump from each category) was identified via [(1,1)]{} velocity information to be outside of the [$v_{\rm{LSR}}$]{} range of the main G333 GMC and are excluded from the analysis in the later portions of this paper unless explicitly included. These clumps are Clump 34 (G333.586$-$0.168) with a [$v_{\rm{LSR}}$]{} of -88.53[kms$^{-1}$]{} (a clump with trace amounts of diffuse 8[${\rm \mu m}$]{} emission), and Clump 39 (G333.642$-$0.106) with a [$v_{\rm{LSR}}$]{} of -87.95[kms$^{-1}$]{} (a YSO which has cleared its surrounding [H[ii]{}]{} region). The [$v_{\rm{LSR}}$]{} for the excluded clumps corresponds with the Norma-Cygnus arm and the included clumps corresponds with the Norma-Cygnus and/or Scutum-Crux arms (fig. 3 of @Vallee2008). At temperatures much greater than 30K, [(1,1)]{} and [(2,2)]{} cannot accurately be used as a temperature probe and any physical parameters using this kinetic temperature calculation would also be inaccurate [@Danby1988; @Hill2010]. As Clump 21 (G333.006$-$0.437) had a derived kinetic temperature of 63.6K, it has been removed from the following analysis. We note that Clump 3 (G333.136$-$0.431) has a [$T_{\rm{kin}}$]{} of 31.9K and Clump 11 (G333.018$-$0.449) has a [$T_{\rm{kin}}$]{} of 35.6K, but has been retained for the following analysis.
From the original sample of 63 SIMBA 1.2mm dust clumps, we are left with 29 IR-bright and 31 IR-faint clumps (Table \[tbl:SFNSF\_parameters\]). Their distribution (with the excluded three clumps) for the derived parameters is shown in Fig. \[fig:histo\_SFNSF\]. We note that derived parameters could not be calculated for all clumps due to the non-detection of [(1,1)]{} and/or [(2,2)]{}. The black shaded histogram shows the total number of clumps with that derived parameter. IR-bright and IR-faint clumps are shown by the solid blue and dashed red line, respectively. The distributions in these histograms generally have skewed Gaussian-like profiles. The kinetic temperature, the FWHM of the [(1,1)]{} transition and the dust and virial masses show that the distribution is skewed in favour of higher values for all these parameters for the IR-bright clumps, while the IR-faint clumps cluster at the lower end. These are not wholly independent parameters, as we have used the same spectrum to measure the FWHM and temperature for the calculations of the virial and clump mass, respectively. However, these histograms show a qualitative difference between the two populations. We investigate these variations further with statistical methods in the next section.
Importance of accurate kinetic temperatures {#sec:tkin}
-------------------------------------------
It is common to assume 20K for isolated clumps and 40K for [H[ii]{}]{} regions (e.g. @Mookerjea2004 for the G333 GMC). Since [(1,1)]{} and [(2,2)]{} are closely separated ($\Delta\nu\approx 28\,\mathrm{MHz}$), we are able to obtain the [$T_{\rm{rot}}$]{} and [$T_{\rm{kin}}$]{} with the same observation per clump. In Fig. \[fig:tkin\], we display the effect that kinetic temperature has on the calculation of clump mass \[Eqn. (\[eqn:mclump\])\]. We find that clumps with [$T_{\rm{kin}}$]{}=20K are more massive, whereas clumps with [$T_{\rm{kin}}$]{}= 40K are less massive. From the dust mass for clumps derived using [$T_{\rm{kin}}$]{}, we found a mean of 2390 [M$_\odot$]{}, a median of 1089[M$_\odot$]{}, a standard deviation of 2920[M$_\odot$]{} and a range between 278–13036[M$_\odot$]{}. For the masses calculated assuming 40K, we found a mean of 1285 [M$_\odot$]{}, a median of 483[M$_\odot$]{}, a standard deviation of 1992[M$_\odot$]{} and a range between 105–9951[M$_\odot$]{}. For the masses calculated using 20K, we found a mean of 3064 [M$_\odot$]{}, a median of 1151[M$_\odot$]{}, a standard deviation of 4750[M$_\odot$]{} and a range between 251–23723[M$_\odot$]{}. We find that T=20K overestimates the mass whereas T=40K underestimates it. Therefore, an accurate temperature is important for determining if a clump is bound and will collapse onto itself, or if it is unbound and remains a starless core.
![\[fig:tkin\] The distribution of clump mass as a function of temperature. The solid black line was calculated from the temperature found from , whilst the dotted red and dashed blue were calculated from the assumption of 20 and 40K, respectively.](FIGURES/Fig5){width="49.00000%"}
Variations in physical parameters throughout the cloud {#sec:variations}
------------------------------------------------------
{width="99.00000%"}
{width="99.00000%"}
The majority of clumps can be grouped into three regions (A, B, C; Fig. \[fig:distro\_SFNSF\]) and the distribution of physical parameters throughout the G333 GMC is shown in Table \[tbl:grouping\] and Fig. \[fig:distro\_vlsr11\], \[fig:distro\_fwhm11\] and \[fig:distro\_tkin\]. For Fig. \[fig:distro\_SFNSF\], each clump was classified as IR-bright (blue pluses) or IR-faint (red pluses) clumps using the presence of *Spitzer* GLIMPSE 8.0$\mu m$ emission (grey-scale image). The excluded clumps are shown with green pluses. For the [(1,1)]{} FWHM (Fig. \[fig:distro\_fwhm11\]) and [$T_{\rm{kin}}$]{} (Fig. \[fig:distro\_tkin\]), the samples were separated into non-detections and detections, with the detections further split into four sub-samples (Q1, Q2, Q3, Q4) at the 25th, 50th and 75th percentile by taking the medians of the [(1,1)]{} FWHM and [$T_{\rm{kin}}$]{}, respectively. The [(1,1)]{} [$v_{\rm{LSR}}$]{} was separated in a similar fashion but the absolute difference between the [$v_{\rm{LSR}}$]{} of the G333 GMC ($\sim$50[kms$^{-1}$]{}; @Bains2006) and the [(1,1)]{} [$v_{\rm{LSR}}$]{} was used. The values of these subdivisions can be seen in Table \[tbl:grouping\]. The symbol size correlates with the percentile sub-sample, and IR-bright and IR-faint clumps are shown as blue circles and red squares, respectively. The IR-bright and IR-faint clumps which were excluded, as per §\[sec:exclusions\], are shown as green circles and squares, respectively, and the symbol size for these clumps corresponds with the percentile calculations for the included clumps.
Region A contains one of the brightest infrared sources in the southern Galactic Plane (G333.6$-$0.2; @Becklin1973), however for the associated Clump 1 (G333.604$-$0.210) was not fit due to self-absorption, so we were unable to extract any physical parameters from the largest and most active clump within the G333 GMC. It also appears to be surrounded by IR-faint clumps. Region A contains 5 and 15 IR-bright and IR-faint clumps, respectively. This asymmetry could be due to the destructive interaction of the [H[ii]{}]{} region with the GMC. There are two other [H[ii]{}]{} regions \[G333.6$-$0.1 [@Goss1970] and IRAS 16175-5002 [@Ellingsen1996]\] within this region and both are associated with IR-bright clumps. However, only IRAS 16175-5002 has detectable emission. The [$v_{\rm{LSR}}$]{} of the clumps within Region A does not significantly differ from the [$v_{\rm{LSR}}$]{} of the G333 GMC. However, the two clumps which were excluded, as per §\[sec:exclusions\], due to having a [$v_{\rm{LSR}}$]{} different to the main G333 cloud, were found in this region.
Region A shows relatively small [$v_{\rm{LSR}}$]{} deviations ($<$ 1.5 [kms$^{-1}$]{} from the [$v_{\rm{LSR}}$]{} of the G333 GMC) for IR-faint clumps. However, with over half the included IR-bright clumps showing non-detections (with, at least, one due to self-absorption), we cannot reliably comment further on the [$v_{\rm{LSR}}$]{} of the IR-bright clumps in Region A. Clump 6 (G333.465$-$0.160) contains the largest [(1,1)]{} FWHM for IR-bright clumps in this region and is associated with the [UCH[ii]{}]{} region, IRAS 16175$-$5002 (e.g. @Walsh1998). However, there are two IR-faint clumps \[Clumps 13 (G333.524$-$0.272) and 26 (G333.772$-$0.260)\] in the highest percentile. Both of these clumps have an excess of 160[${\rm \mu m}$]{} emission with Clump 13 (G333.524$-$0.272) showing possible compression between two expanding 8[${\rm \mu m}$]{} bubbles. Most (60%) of the [$T_{\rm{kin}}$]{} values for IR-bright clumps are non-detections. Half (53%) of the IR-faint clumps are below the 50th percentile, and a third are non-detections. Clump 13 (G333.524$-$0.272) shows the second highest [$T_{\rm{kin}}$]{} (20.3K) for the IR-faint clumps and Clump 38 (G333.681$-$0.106) shows the highest (measured) [$T_{\rm{kin}}$]{} (23.7K) of the region.
Region B appears to be the opposite and contains 14 and 8 IR-bright and IR-faint clumps, respectively. There are three clusters of [H[ii]{}]{} regions and they correspond with IR-bright clumps; however there are also IR-bright clumps that are not [H[ii]{}]{} regions. The IR-faint clumps appear to be on the outskirts of [H[ii]{}]{} regions. The IR-bright clumps in Region B show smaller deviations from the [$v_{\rm{LSR}}$]{} towards the north of the GMC and the clumps with larger deviations clustered towards the south. The IR-faint clumps have larger [$v_{\rm{LSR}}$]{} deviations and are closer to the south. This region shows the largest [(1,1)]{} FWHM for IR-faint clumps, whereas the IR-bright clumps are somewhat equally spaced throughout all quartiles. The majority of high [$T_{\rm{kin}}$]{} appear to be associated with, or adjacent to, the [H[ii]{}]{} regions, however Clump 9 (G333.127$-$0.564) has a [$T_{\rm{kin}}$]{} of 20.3K and appears to have excess 160[${\rm \mu m}$]{} emission coincident with the 1.2-mm flux peak, and Class I and II masers.
Region C also contains multiple [H[ii]{}]{} regions (including RCW 106) but the number of IR-bright and IR-faint clumps are equal (with 9 clumps in each category). @Rodgers1960 identified two bright areas in RCW 106: one at G332.9$-$0.6 and the other at G333.7$-$0.5, with their extent covering 20arcmin $\times$ 7arcmin and 12arcmin $\times$ 12arcmin, respectively. This covers the majority of SIMBA clumps within Region C. For Region B and C, we note that the IR-faint clumps are often towards the edge of each region. This suggests that star formation on these scales are propagating outwards from the central and more evolved parts of the regions, which suggests sequentially triggered star formation. Region C has a larger proportion of IR-faint clumps showing larger [$v_{\rm{LSR}}$]{} deviations, with the largest deviations seen towards the boundary of Regions B with C. The largest values for the [(1,1)]{} FWHM are seen towards the IR-bright clumps that are associated with [H[ii]{}]{} regions. For the IR-faint clumps, the [$T_{\rm{kin}}$]{} tend towards the larger values. Clump 25 (G332.773$-$0.582) and Clump 32 (G332.670$-$0.644) are IR-faint clumps however they show high [$T_{\rm{kin}}$]{} (22.6 and 21.5K, respectively) but have low values for [(1,1)]{} FWHM (1.72 and 1.96[kms$^{-1}$]{}). These clumps are also adjacent to IR-bright clumps so appear to be sites of future or currently unseen star formation.
The 63 clumps within the G333 GMC can be grouped into three regions. The physical parameters throughout the GMC are influenced by the numerous [H[ii]{}]{} regions. Region A contains the most number of IR-faint clumps (15), whilst Region B contains the most number of IR-bright clumps (14), and Region C contains the same number (nine) of IR-bright and IR-faint clumps. Region A and B contain the most number of non-detections, with eight and three, respectively. This may be because of the effect that the [H[ii]{}]{} regions, G333.6$-$0.2 and RCW 106, have on the surrounding molecular gas. Region B contains seven clumps with large (Q4; 2.15$-$4.41[kms$^{-1}$]{}) [(1,1)]{} FWHM, whilst both Regions A and C have two each. This could indicate that the [H[ii]{}]{} regions in Region A and C are sweeping molecular material into Region B. Region A and C have the most clumps with low (Q1; 13.6$-$18.2K) [$T_{\rm{kin}}$]{}, whereas Region B has the warmest \[Q3-Q4; 13 clumps with [$T_{\rm{kin}}$]{} above the median (19.4K)\]. Region C also contains the most numbers of clumps with the largest (Q4; 5.37$-$7.46[kms$^{-1}$]{}) deviation from the [$v_{\rm{LSR}}$]{} of the G333 GMC.
{width="99.00000%"}
------------------------------- ----------- ---------- ----------- ---------- ----------- ----------
FWHM
$[$[kms$^{-1}$]{}\] IR-bright IR-faint IR-bright IR-faint IR-bright IR-faint
2.15–4.41 0 2 3+ 4 2 0
1.82–1.99 1 1 2 2 2 4
1.54–1.74 0 2 3 2 2 2
1.11–1.48 1+ 5+ 3 0 2 2
non-detection 3 5 3 0 1 1
[$T_{\rm{kin}}$]{}
$[$K\] IR-bright IR-faint IR-bright IR-faint IR-bright IR-faint
20.6–23.7 0 1 3+ 3 2 2
19.5–20.4 1 1 3 4 1 2
18.4–19.3 0 4 2 0 3 1
13.6–18.2 1+ 4+ 2 1 2 3
non-detection 3 5 4 0 1 1
[(1,1)]{} [$v_{\rm{LSR}}$]{}
$[$[kms$^{-1}$]{}\] IR-bright IR-faint IR-bright IR-faint IR-bright IR-faint
5.37–7.46 1+ 0+ 1+ 2 3 4
2.55–5.24 1 3 4 2 0 2
1.05–1.84 0 4 3 2 2 0
0.05–0.66 0 3 3 2 3 2
non-detection 3 5 3 0 1 1
Total 5+ 15+ 14+ 8 9 9
------------------------------- ----------- ---------- ----------- ---------- ----------- ----------
: Variations in physical parameters within the regions in the G333 GMC. The samples were separated into non-detections and detections, with the detections further split into four sub-samples (Q1, Q2, Q3, Q4) at the 25th, 50th and 75th percentile. The [(1,1)]{} [$v_{\rm{LSR}}$]{} was separated in a similar fashion but the absolute difference between the [$v_{\rm{LSR}}$]{} of the G333 GMC ($\sim$50 [kms$^{-1}$]{}; @Bains2006) and the [(1,1)]{} [$v_{\rm{LSR}}$]{} was used. Excluded clumps as per §\[sec:exclusions\] were not used to calculate the quartiles but a + indicates the quartile it would belong to.[]{data-label="tbl:grouping"}
Correlations of physical parameters {#sec:correlations}
-----------------------------------
We have calculated the linear Pearson correlation coefficient and the linear fit to test the level of correlation between physical parameters for the IR-bright and IR-faint clumps.
The linear Pearson correlation coefficient, $r$, for different pairs of physical parameters, were found with the [correlate]{} function found in IDL. For each $\left(x_i,y_i\right)$ pair, with $i=1,....,N$, $r$ can be written following @Press1992 [chap. 14] as: $$\label{eq:correlate}
r = \frac{\sum_{i}\left(x_i-\bar{x}\right)\left(y_i-\bar{y}\right)}{{\sqrt{\sum_{i}\left(x_i-\bar{x}\right)^2}}{\sqrt{\sum_i\left(y_i-\bar{y}\right)^2}}}$$
The linear fits between different physical parameters were found with the [linfit]{} function found in IDL. It fits a linear model ($y = a+bx$), while minimising the chi-square error statistic. The chi-square error statistic, $\chi^2$, can be written following @Press1992 [chap. 15] as: $$\label{eq:chisq}
\chi^2(a,b) = \displaystyle\sum_{i=1}^{N} \left(y_i - a - bx_i\right)^2$$
The relationship between clump radius and FWHM can be seen in Fig. \[fig:correlation\](a). There is significant scatter in the data, as seen by the medium positive correlation for IR-bright clumps ($r_\mathrm{IR-bright}~=~0.43$) and low positive correlation for IR-faint clumps ($r_\mathrm{IR-faint}~=~0.10$), which shows a poor linear fit to the data. This implies that before the onset of 8[${\rm \mu m}$]{} emission, is not correlated with the physical parameters of the clump but once they become IR-bright, the FWHM has some correlation with 8[${\rm \mu m}$]{} emission. There are four clumps with [(1,1)]{} FWHM $\sim$4.5[kms$^{-1}$]{} that deviate from the other clumps. Three of these clumps are IR-bright, however one (Clump 21 G333.006$-$0.437) has been excluded from our analysis as per §\[sec:exclusions\]. Two of the included clumps (Clump 3 G333.136$-$0.431 and Clump 30 G333.038$-$0.405) belong to Region B but one is IR-bright and the other is IR-faint. The third included clump belongs to Region C.
The relationship between the virial and clump mass can be seen in Fig. \[fig:correlation\](b). However, the G333 GMC contains numerous [H[ii]{}]{} regions which have free-free emission which can contribute to the 1.2mm emission. This may contaminate surrounding clumps resulting in smaller clump masses \[Eqn. (\[eqn:mclump\])\]. Of these 32 clumps, 72 per cent (23 of 32) were classified as IR-faint clumps. The masses for 14 clumps could not be calculated because no emission was detected in 12 clumps; the emission profile was too complex, due to self-absorption, for Clump 1 (G333.604$-$0.210); and [(2,2)]{} was not detected in Clump 42 (G333.206$-$0.366). There are high positive correlations for both IR-bright and IR-faint clumps, with $r_\mathrm{IR-bright}=0.87$ and $r_\mathrm{IR-faint}=0.67$. The three clumps which have been previously excluded in §\[sec:exclusions\], were included in Fig. \[fig:correlation\] as unfilled blue triangles and unfilled red squares to represent IR-bright and IR-faint clumps, respectively; however, they were not included in the fits. We can see that the Clump 21 (G333.006$-$0.437) with a [$T_{\rm{kin}}$]{} of 63.6K contains a significantly larger virial mass and would skew further analysis. However, the clumps with a [$v_{\rm{LSR}}$]{} different to the G333 GMC do not deviate much from the scatter of the other clumps and would not greatly influence our later results. The clumps that are gravitationally bound are shown below the black solid 1:1 line. Three clumps, with a clump mass smaller than the virial mass, have been classified as clumps with signs of star formation. For the six clumps with a smaller clump mass, four clumps have masses within 30 per cent of their corresponding virial mass.
The virial ratio ($= M_{clump}/M_{virial}$) versus the clump mass is shown in Fig. \[fig:correlation\](c). The green dotted horizontal line marks the line of gravitational instability (i.e. $M_{clump} = M_{virial}$). Clumps below this line are unbound and will not collapse under gravity. There are 87 per cent (40 of 46) of the clumps with masses larger than the virial mass, suggesting that they will form stars or are already undergoing star formation. The IR-bright and IR-faint clumps have a mean virial ratio of 2.7 and 2.8, respectively, and a median virial ratio of 2.8 and 2.1, respectively. However, magnetic fields can prevent the collapse of a clump and by assuming the equipartition of kinetic and magnetic energy, the relationship between clump and virial masses can be found to be $M_{clump}=2M_{virial}$ (e.g. @Bertoldi1992) and is shown by the dashed green horizontal line. By including magnetic fields, the number of clumps which are gravitationally unstable becomes 70 per cent (14 of 20) and 54 per cent (14 of 26), for IR-bright and IR-faint clumps, respectively.
![Correlations between physical parameters. IR-bright clumps are represented by blue filled triangles. IR-faint clumps are represented by red filled squares. The excluded IR-bright and IR-faint clumps are represented by blue hollow triangles and red hollow squares, respectively. For (b): The blue dotted line displays the fit to the IR-bright clumps. The red dashed line displays the fits to the IR-faint clumps. The black dashed dot line displays the fit to both samples. These fits do not include the clumps that have been excluded, as per §\[sec:exclusions\]. The black solid line displays the 1:1 line. For (c): The green dotted line displays the equilibrium between clump and virial masses, while the green dashed line shows the equilibrium between kinetic and magnetic energy.[]{data-label="fig:correlation"}](FIGURES/Fig9a "fig:"){width="49.00000%"} ![Correlations between physical parameters. IR-bright clumps are represented by blue filled triangles. IR-faint clumps are represented by red filled squares. The excluded IR-bright and IR-faint clumps are represented by blue hollow triangles and red hollow squares, respectively. For (b): The blue dotted line displays the fit to the IR-bright clumps. The red dashed line displays the fits to the IR-faint clumps. The black dashed dot line displays the fit to both samples. These fits do not include the clumps that have been excluded, as per §\[sec:exclusions\]. The black solid line displays the 1:1 line. For (c): The green dotted line displays the equilibrium between clump and virial masses, while the green dashed line shows the equilibrium between kinetic and magnetic energy.[]{data-label="fig:correlation"}](FIGURES/Fig9b "fig:"){width="49.00000%"} ![Correlations between physical parameters. IR-bright clumps are represented by blue filled triangles. IR-faint clumps are represented by red filled squares. The excluded IR-bright and IR-faint clumps are represented by blue hollow triangles and red hollow squares, respectively. For (b): The blue dotted line displays the fit to the IR-bright clumps. The red dashed line displays the fits to the IR-faint clumps. The black dashed dot line displays the fit to both samples. These fits do not include the clumps that have been excluded, as per §\[sec:exclusions\]. The black solid line displays the 1:1 line. For (c): The green dotted line displays the equilibrium between clump and virial masses, while the green dashed line shows the equilibrium between kinetic and magnetic energy.[]{data-label="fig:correlation"}](FIGURES/Fig9c "fig:"){width="49.00000%"}
Kolmogorov–Smirnov test {#sec:kstest}
-----------------------
The Kolmogorov–Smirnov (K-S) test is used to test whether our IR-bright and IR-faint clumps are likely to be drawn from the same population. We perform a K-S test on the physical parameters derived from the [(1,1)]{} and [(2,2)]{} spectra. To compare the cumulative distribution functions of the physical parameters for the IR-bright and IR-faint clumps, $S_\mathrm{IR-bright}(x)$ and $S_\mathrm{IR-faint}(x)$, the K-S statistic, can be written following @Press1992 [chap. 14] as: $$\mathrm{D} = \max_{-\infty < x < \infty}|S_\mathrm{IR-bright}(x) - S_\mathrm{IR-faint}(x)|$$ To calculate the significance of the KS test: $$\label{eq:Q_KS}
\mathrm{Q}_{KS} (\lambda) = 2 \sum_{j=1}^\infty (-1)^{j-1}\mathrm{e}^{-2j^2\lambda^2}$$ where $$\lambda = \left[{\sqrt{N_{eff}} + \dfrac{0.11}{\sqrt{N_{eff}}}}\right] D$$ and $N_{eff}$ is the effective number of data points, with $$N_{eff} = \dfrac{N_1N_2}{N_1+N_2}$$ where $N_1$ and $N_2$ are the two distributions to be tested.
The significance and K-S statistic for the clump and derived parameters can be seen in Table \[tbl:ks\]. The significances were calculated to be 0.0078, 0.3747, 0.0162, 0.0002, 0.0446 and 0.0056 for peak flux, clump radius, [(1,1)]{} FWHM, kinetic temperature, clump mass, and virial mass, respectively. We note that the significance of the cumulative distribution functions between the two populations for the kinetic temperature are over the 0.003 significance level (3$\sigma$) than if they were drawn from the same population [@Wall2012]. Through accurate kinetic temperatures, we confirm that 8[${\rm \mu m}$]{} emission traces the warm dust and gas. In Fig. \[fig:histo\_SFNSF\], the [(1,1)]{} FWHM displays a difference between the two populations, however the significance in the K-S test is below the 3$\sigma$ level. The same outcome is not apparent through the clump radius or either the clump or virial masses. The effect of turbulence is important to high-mass star formation as it can encourage the development of stars above the mass threshold. However, if the amount of turbulence is too high, it can discourage star formation [@Padoan1995]. As the NH3 (1,1) linewidths (1.1–4.5[kms$^{-1}$]{}) are larger than the thermal component ($\sim$0.28[kms$^{-1}$]{} for [$T_{\rm{kin}}$]{}=30K), all our clumps show varying amounts of turbulence. However, the K-S test was unable to find a significant difference between the IR-bright and IR-faint clumps, so we are unable to find a link between the amount of turbulence and 8[${\rm \mu m}$]{} emission. The K-S test was unable to distinguish a significant difference for the clump radius. This is important because a significant difference would indicate that different algorithms and input settings would drastically affect the parameters derived from the clumps.
[lcc]{}\
Parameter & Q$_{KS}$ & D\
Peak flux & 0.0059 & 0.42\
Clump radius & 0.4694 & 0.21\
Total flux & 0.0436 & 0.34\
[(1,1)]{} FWHM & 0.0136 & 0.43\
Column density & 0.1917 & 0.30\
Kinetic temperature & 0.0002 & 0.59\
Clump mass & 0.0595 & 0.35\
Virial mass & 0.0041 & 0.48\
\
Parameter & Q$_{KS}$ & D\
Peak flux & 0.0078 & 0.42\
Clump radius & 0.3742 & 0.23\
Total flux & 0.0297 & 0.36\
[(1,1)]{} FWHM & 0.0162 & 0.44\
Column density & 0.3667 & 0.26\
Kinetic temperature & 0.0002 & 0.61\
Clump mass & 0.0446 & 0.39\
Virial mass & 0.0056 & 0.48\
\
$\dagger$The clumps designated as “main cloud” are those with calculated parameters from the spectra but exclude those as per §\[sec:exclusions\].
Environments around interesting sources {#sec:sources}
---------------------------------------
Here we discuss sources that deserve further detail than those provided by Columns 8, 9 and 10 of Table \[tbl:cf\_parameters\_jsu\]. Infrared composite images centered on the peak emission of the dust clumps identified by [clumpfind]{} can be found in Fig. \[app:3colour\]. [(1,1)]{} and [(2,2)]{} emission detected towards the peak emission of each clump can be found in Fig. \[app:spectra\].
*Clump 1 (G333.604$-$0.210).* This is the brightest IR source in the southern Galactic Plane [@Becklin1973]. @Fujiyoshi2005 [@Fujiyoshi2006] found that this [H[ii]{}]{} region is excited by a cluster of O and B stars. @Fujiyoshi2006 also found a complex H90$\alpha$ radio recombination line spectrum which they suggest may be caused by champagne outflows. emission is weak and shows absorption but both [(1,1)]{} and [(2,2)]{} show similar intensities.
*Clump 2 (G332.826$-$0.547) and Clump 60 (G332.894$-$0.567).* These clumps appear to be related to a 8[${\rm \mu m}$]{} bubble. Clump 2 (G332.894$-$0.567) coincides with an [H[ii]{}]{} region and YSO [@Urquhart2008] which is within the extent of RCW 106 [@Rodgers1960]. There are , Class II and OH masers. There is bifurcated 8[${\rm \mu m}$]{} emission which may explain the broad [(1,1)]{} and [(2,2)]{} spectrum. The [(1,1)]{} spectrum appears to be blended with no clear baseline between the hyperfine transitions. The [H[ii]{}]{} region and YSO also coincides with a 160 [${\rm \mu m}$]{} diffuse source. Clump 60 (G332.894$-$0.567) is not associated with any known maser activity and there is no detectable emission.
*Clump 3 (G333.136$-$0.431).* This clump contains an assortment of masers however the peak of the SIMBA dust clump coincides with a 160[${\rm \mu m}$]{} diffuse source as well as OH and 6.7GHz Class I masers. The peak of this SIMBA dust clump is surrounded by 8[${\rm \mu m}$]{} emission with an assortment of maser activity. There are four clusters of maser activity around the peak with two clusters containing 6.7GHz Class I and masers as well as 44.07 and 95.1GHz class II masers. It is rare to see the two classes of masers in the same cluster. Class II masers are closely associated with YSOs, while Class I masers are associated with shocks and cloud-cloud collisions. There may be triggered star formation occurring in this 8[${\rm \mu m}$]{} arc which has allowed for the two classes of masers to be seen at the same time. A collision between clouds would create the Class I masers but also trigger star formation which leads to outflows, YSOs and the Class II masers.
*Clump 4 (G333.286$-$0.387) and Clump 5 (G333.309$-$0.369).* These clumps appear to be local peaks in a larger clump. Clump 4 contains multiple point sources with a water maser and was classified by the RMS survey as a [H[ii]{}]{} region. Clump 5 was classified by the RMS survey as a diffuse [H[ii]{}]{} region.
*Clump 6 (G333.465$-$0.160).* This clump contains a mixture of Class I and II masers as well as a water maser. It has a lot of 8[${\rm \mu m}$]{} emission around a diffuse 160[${\rm \mu m}$]{} source. This looks like Clump 2 (G332.826$-$0.547) but with significantly more 8[${\rm \mu m}$]{} emission. This may be a later evolutionary stage but the [(1,1)]{} and [(2,2)]{} do not show blended spectra, which could indicate turbulence, like Clump 2 (G332.826$-$0.547).
*Clump 7 (G333.068$-$0.446).* Satellite lines can be seen in both the [(1,1)]{} and [(2,2)]{}. There are Class II masers associated with the SIMBA dust peak and a Class II and water maser offset by $\sim$1arcmin on either side of the SIMBA dust peak.
*Clump 8 (G332.644$-$0.606), Clump 10 (G332.691$-$0.612), Clump 12 (G332.676$-$0.615) and Clump 22 (G332.700$-$0.585).* There are two 160[${\rm \mu m}$]{} sources within 8[${\rm \mu m}$]{} arcs on either side of another 160[${\rm \mu m}$]{} source. There is a 6.7GHz Class I and water maser along the west 8[${\rm \mu m}$]{} arc, however neither is coincident with the SIMBA dust peak. The ionising source of this bubble is unknown; however, further analysis is required to show if it is expanding outwards from Clump 12 (G332.676$-$0.615). These clumps are also within the region classified by @Rodgers1960 as RCW 106.
*Clump 9 (G333.127$-$0.564).* This is a relatively isolated, cold dust clump which coincides with the second brightest SiO source within this GMC [@Lo2007]. There are low levels of the [(2,2)]{} satellite lines.
*Clump 10 (G332.691$-$0.612).* This is an elongated [H[ii]{}]{} region and could be star formation triggered by a bow shock of a nearby protostellar object. This clump could be associated with nearby Clump 12 (G332.676$-$0.615). The maser at the top corresponds to a YSO and [H[ii]{}]{} region.
*Clump 11 (G333.018$-$0.449) and Clump 21 (G333.006$-$0.437).* These two clumps are on the edge of the same 8[${\rm \mu m}$]{} arc. The SIMBA dust clumps appear to coincide with 160[${\rm \mu m}$]{} emission. Both clumps have masers on the far side of the 8[${\rm \mu m}$]{} arc. Both clumps have high [$T_{\rm{kin}}$]{}. Clump 11 (G333.018$-$0.449) has a [$T_{\rm{kin}}$]{} of 35.6K and Clump 21 (G333.006$-$0.437) has a [$T_{\rm{kin}}$]{} of 63.6K.
*Clump 13 (G333.524$-$0.272), Clump 27 (G333.539$-$0.245) and Clump 46 (G333.521$-$0.239).* Clump 13 (G333.524$-$0.272) appears to follow the 160[${\rm \mu m}$]{} emission which appears to be compressed between two 8[${\rm \mu m}$]{} arcs. There is no known maser emission in this region. Clump 27 (G333.539$-$0.245) also appear to follow the compressed 160[${\rm \mu m}$]{} emission however the intensity is lower. Clump 46 (G333.521$-$0.239) is also associated but the 160[${\rm \mu m}$]{} emission is the weakest of the three.
*Clump 14 (G332.985$-$0.487), Clump 18 (G333.074$-$0.558), Clump 19 (G333.722$-$0.209), Clump 39 (G333.642$-$0.106), Clump 47 (G333.074$-$0.399), Clump 51 (G332.903$-$0.546) and Clump 58 (G333.410$-$0.328).* These clumps contain a compact 160[${\rm \mu m}$]{} object towards the position of the peak SIMBA position. There is no 4.5 or 8[${\rm \mu m}$]{} emission nearby. It is likely to be the earliest stage of star formation.
From the environment around select clumps, there are a variety of star formation stages within the G333 GMC. This ranges from isolated, compact 160[${\rm \mu m}$]{} objects, to areas with complex infrared emission and multiple masers. There also appears to be possible triggered star formation \[e.g. Clump 3 (G333.136$-$0.431)\]. In future papers, we will be combining the molecular and dust information for all 63 clumps identified in this paper, to characterise the clumps in relation to the evolution of star formation within the G333 GMC.
Summary and Conclusions {#sect:conclusion}
=======================
In this paper we have presented data from the two lowest inversion transitions of in the G333 giant molecular cloud with the 70-m Tidbinbilla radio telescope. Our main conclusions are as follows:
\(1) There were 63 SIMBA 1.2-mm dust clumps identified with [clumpfind]{} and pointed observations were conducted towards the peak of each clump. The peak of each clump was offset from the clump centroid by between 1.6 and 39.5 arcsec.
\(2) A variety of environments were identified within the G333 GMC and the sample was separated into 30 IR-bright and 33 IR-faint clumps. One clump from each category was identified via [(2,2)]{} emission, to not be associated with the [$v_{\rm{LSR}}$]{} of the G333 GMC, and one clump was identified to have an inaccurate [$T_{\rm{kin}}$]{}; resulting in 29 IR-bright and 31 IR-faint clumps. The [(1,1)]{} and [(2,2)]{} were fitted for 20 IR-bright and 26 IR-faint clumps, with mean [$T_{\rm{kin}}$]{} values of 24.2K and 19.3K, respectively.
\(3) We found that the clumps appeared to cluster into three regions. The IR-bright clumps were found predominately towards the centre of each region, especially for Regions B and C. This may indicate sequentially triggered star formation in the vicinity around IR-faint clumps.
\(4) Using [(1,1)]{} and [(2,2)]{}, the clump masses were found to be between 278 and 13036 [M$_\odot$]{}. The dust temperature is often assumed: 20K for isolated clumps and 40K for [H[ii]{}]{} regions; however this assumption can significantly shift the masses of the clumps (e.g. for 20K the masses of our clumps were 251–23723[M$_\odot$]{}, and for 40K the masses of our clumps were 105–9951[M$_\odot$]{}) and can affect whether the clump is determined to be bound and will collapse onto itself, or if it is unbound and remains a starless core. Hence accurate measurements of kinetic temperatures are needed for studies of the gravitationally bound state of clumps.
\(5) The largest clumps are also the most unstable. We find that the clumps with signs of star formation are predominately associated with the least stable clumps. This may be due to feedback from an internal heating source. We found that 70 per cent of the clumps had masses larger than the virial mass (187–6848[M$_\odot$]{}), suggesting that they will form stars or are already undergoing star formation.
\(6) A K-S test showed that the [$T_{\rm{kin}}$]{} for IR-bright and IR-faint clumps have a much smaller probability of being drawn from the same population. A tentative difference between the two populations in terms of FWHM could indicate a contribution due to feedback. No difference was seen for the other physical parameters. We conclude that IR-bright and IR-faint clumps can have the same masses but that the warmer regions are more evolved.
This is the fourth paper in a series of papers utilising molecular line observations of the G333 giant molecular cloud associated with RCW 106 (G333 GMC). In this paper we reported on the selection of dense clumps based on SIMBA dust emission and pointed [(1,1)]{} and [(2,2)]{} observations towards these clumps. Future papers in the series will analyse the individual clumps and combine the information presented in this paper with the molecular 3-mm tracers presented in @Bains2006, @Wong2008 and @Lo2009, and dust from the *Spitzer* Space Telescope and the *Herschel* Space Observatory.
Acknowledgments {#acknowledgments .unnumbered}
===============
We would like to thank the anonymous referee for their comments that have greatly improved this paper. Partial financial support for this work was provided by the Australian Research Council. JPM is supported by Spanish grant AYA 2011-26202. JSU was supported by a CSIRO OCE postdoctoral grant. NL is supported by a CONICYT/FONDECYT postdoctorado, under project no. 3130540. NL acknowledges partial support from the ALMA-CONICYT Fund for the Development of Chilean Astronomy Project 31090013, Center of Excellence in Astrophysics and Associated Technologies (PFB 06) and Centro de Astrofísica FONDAP15010003. The Tidbinbilla 70-m Radio Telescope is part of the NASA Deep Space Network and is operated by CSIRO Astronomy and Space Science. *Herschel* is an ESA space observatory with science instruments provided by European-led Principal Investigator consortia and with important participation from NASA. This research has made use of NASA’s Astrophysics Data System. This research has made use of the SIMBAD database, operated at CDS, Strasbourg, France. This research has made use of the NASA/ IPAC Infrared Science Archive, which is operated by the Jet Propulsion Laboratory, California Institute of Technology, under contract with the National Aeronautics and Space Administration. This work is based in part on archival data obtained with the Spitzer Space Telescope, which is operated by the Jet Propulsion Laboratory, California Institute of Technology under a contract with NASA. Support for this work was provided by an award issued by JPL/Caltech. This paper made use of information from the Red MSX Source survey database at www.ast.leeds.ac.uk/RMS which was constructed with support from the Science and Technology Facilities Council of the UK.
I., [Wong]{} T., [Cunningham]{} M., [Sparks]{} P., [Brisbin]{} D., [Calisse]{} P., [Dempsey]{} J. T., [Deragopian]{} G., et al., 2006, MNRAS, 367, 1609 Becklin E. E., Frogel J. A., Neugebauer G., Persson S. E., Wynn-Williams C. G., 1973, ApJ, 182, L125 Benjamin R. A., et al., 2003, PASP, 115, 953 Bertoldi F., McKee C. F., 1992, ApJ, 395, 140 Braz M. A., Scalise E., Jr., 1982, A&A, 107, 272 Breen S. L., et al., 2007, MNRAS, 377, 491 Breen S. L., Caswell J. L., Ellingsen S. P., Phillips C. J., 2010, MNRAS, 406, 1487 Breen S. L., Ellingsen S. P., Caswell J. L., Green J. A., Voronkov M. A., Fuller G. A., Quinn L. J., Avison A., 2012, MNRAS, 421, 1703 Breen S. L., Ellingsen S. P., Contreras Y., Green J. A., Caswell J. L., Stevens J. B., Dawson J. R., Voronkov M. A., 2013, MNRAS, 435, 524 Carey S. J., et al., 2009, PASP, 121, 76 Caswell J. L., et al., 2011, MNRAS, 417, 1964 Caswell J. L., Haynes R. F., Goss W. M., 1980, AuJPh, 33, 639 Caswell J. L., Vaile R. A., Forster J. R., 1995, MNRAS, 277, 210 Caswell J. L., 1996, MNRAS, 279, 79 Caswell J. L., 1997, MNRAS, 289, 203 Caswell J. L., 1998, MNRAS, 297, 215 Caswell J. L., 2009, PASA, 26, 454 Churchwell E., et al., 2009, PASP, 121, 213 Cyganowski C. J., et al., 2008, AJ, 136, 2391 Danby G., Flower D. R., Valiron P., Schilke P., Walmsley C. M., 1988, MNRAS, 235, 229 Dunham M. K., Rosolowsky E., Evans N. J., II, Cyganowski C., Urquhart J. S., 2011, ApJ, 741, 110 Ellingsen S. P., von Bibra M. L., McCulloch P. M., Norris R. P., Deshpande A. A., Phillips C. J., 1996, MNRAS, 280, 378 Ellingsen S. P., 2005, MNRAS, 359, 1498 G., [Fa[ú]{}ndez]{} S., [Mardones]{} D., [Bronfman]{} L., [Chini]{}R., [Nyman]{} L.-[Å]{}., 2004, ApJ, 610, 313 Fujiyoshi T., Smith C. H., Moore T. J. T., Lumsden S. L., Aitken D. K., Roche P. F., 2005, MNRAS, 356, 801 Fujiyoshi T., Smith C. H., Caswell J. L., Moore T. J. T., Lumsden S. L., Aitken D. K., Roche P. F., 2006, MNRAS, 368, 1843 Goss W. M., Shaver P. A., 1970, AuJPA, 14, 1 Griffin M. J., et al., 2010, A&A, 518, L3 Hildebrand R. H., 1983, QJRAS, 24, 267 Hill T., Longmore S. N., Pinte C., Cunningham M. R., Burton M. G., Minier V., 2010, MNRAS, 402, 2682 Ho P. T. P., Townes C. H., 1983, ARA&A, 21, 239 Johnstone D., Rosolowsky E., Tafalla M., Kirk H., 2010, ApJ, 711, 655 N., [Purcell]{} C., [Wong]{} T., [Robertson]{} S., 2005, Publications of the Astronomical Society of Australia, 22, 62 Li D., Goldsmith P. F., Menten K., 2003, ApJ, 587, 262 N., [Cunningham]{} M., [Bains]{} I., [Burton]{} M. G., [Garay]{} G., 2007, MNRAS, 381, L30 N., [Cunningham]{} M. R., [Jones]{} P. A., [Bains]{} I., [Burton]{} M. G., [Wong]{} T., [Muller]{} E., [Kramer]{} C., et al., 2009, MNRAS, 395, 1021 Lo N., Redman M. P., Jones P. A., Cunningham M. R., Chhetri R., Bains I., Burton M. G., 2011, MNRAS, 415, 525 F. J., 1979, ApJ, 232, 761 Lumsden S. L., Hoare M. G., Urquhart J. S., Oudmaijer R. D., Davies B., Mottram J. C., Cooper H. D. B., Moore T. J. T., 2013, ApJS, 208, 11 Luna A., Bronfman L., Carrasco L., May J., 2006, ApJ, 641, 938 Manchester B. A., 1969, AuJPA, 12, 3 J. G., [Wootten]{} A., [Mundy]{} L. G., 1992, ApJ, 388, 467 Minier V., Ellingsen S. P., Norris R. P., Booth R. S., 2003, A&A, 403, 1095 Molinari S., et al., 2010, PASP, 122, 314 Mookerjea B., Kramer C., Nielbock M., Nyman L.-[Å]{}., 2004, A&A, 426, 119 Ott S., 2010, ASPC, 434, 139 Ossenkopf V., Henning T., 1994, A&A, 291, 943 Padoan P., 1995, MNRAS, 277, 377 Pilbratt G. L., et al., 2010, A&A, 518, L1 Poglitsch A., et al., 2010, A&A, 518, L2 Press, W. H., et al., 1992, Numerical Recipes in C: The Art of Scientific Computing. Cambridge University Press, New York. Protheroe R. J., Ott J., Ekers R. D., Jones D. I., Crocker R. M., 2008, MNRAS, 390, 683 Rodgers A. W., Campbell C. T., Whiteoak J. B., 1960, MNRAS, 121, 103 Roussel H., 2013, PASP, 125, 1126 Slysh V. I., Kalenskii S. V., Valtts I. E., Otrupcek R., 1994, MNRAS, 268, 464 Spitzer Science C., 2009, yCat, 2293, 0 J. J., [Welch]{} W. J., [Di Francesco]{} J., 2005, ApJ, 620, 823 Tafalla M., Myers P. C., Caselli P., Walmsley C. M., 2004, A&A, 416, 191 Urquhart J. S., Hoare M. G., Lumsden S. L., Oudmaijer R. D., Moore T. J. T., 2008, ASPC, 387, 381 Urquhart J. S., et al., 2011, MNRAS, 418, 1689 Urquhart J. S., Figura C. C., Moore T. J. T., Hoare M. G., Lumsden S. L., Mottram J. C., Thompson M. A., Oudmaijer R. D., 2014, MNRAS, 437, 1791 Vall[é]{}e J. P., 2008, AJ, 135, 1301 Voronkov M. A., Caswell J. L., Ellingsen S. P., Green J. A., Breen S. L., 2014, MNRAS, 334 Wall J. V., Jenkins C. R., 2012, psa..book, Walsh A. J., et al., 2011, MNRAS, 416, 1764 Walmsley C. M., Ungerechts H., 1983, A&A, 122, 164 Walsh A. J., Burton M. G., Hyland A. R., Robinson G., 1998, MNRAS, 301, 640 Wienen M., Wyrowski F., Schuller F., Menten K. M., Walmsley C. M., Bronfman L., Motte F., 2012, A&A, 544, A146 Williams J. P., de Geus E. J., Blitz L., 1994, ApJ, 428, 693 T., [Ladd]{} E. F., [Brisbin]{} D., [Burton]{} M. G., [Bains]{} I., [Cunningham]{} M. R., [Lo]{} N., [Jones]{} P. A., et al., 2008, MNRAS, 386, 1069 Zinnecker H., Yorke H. W., 2007, ARA&A, 45, 481
SUPPORTING INFORMATION {#supporting-information .unnumbered}
======================
Additional Supporting Information may be found in the online version of this article.\
\
[**Figure \[app:3colour\].**]{} Infrared composite images centred on the peak of the 63 SIMBA dust clumps.\
[**Figure \[app:spectra\].**]{} [(1,1)]{} and [(2,2)]{} spectra towards the peak of the 63 SIMBA dust clumps.\
{width="32.00000%"} {width="32.00000%"} {width="32.00000%"}\
{width="32.00000%"} {width="32.00000%"} {width="32.00000%"}\
{width="32.00000%"} {width="32.00000%"} {width="32.00000%"}\
{width="32.00000%"} {width="32.00000%"} {width="32.00000%"}\
{width="32.00000%"} {width="32.00000%"} {width="32.00000%"}\
{width="32.00000%"} {width="32.00000%"} {width="32.00000%"}\
{width="32.00000%"} {width="32.00000%"} {width="32.00000%"}\
{width="32.00000%"} {width="32.00000%"} {width="32.00000%"}\
{width="32.00000%"} {width="32.00000%"} {width="32.00000%"}\
{width="32.00000%"} {width="32.00000%"} {width="32.00000%"}\
{width="32.00000%"} {width="32.00000%"} {width="32.00000%"}\
{width="32.00000%"} {width="32.00000%"} {width="32.00000%"}\
{width="32.00000%"} {width="32.00000%"} {width="32.00000%"}\
{width="32.00000%"} {width="32.00000%"} {width="32.00000%"}\
{width="32.00000%"} {width="32.00000%"} {width="32.00000%"}\
{width="32.00000%"} {width="32.00000%"} {width="32.00000%"}\
{width="32.00000%"} {width="32.00000%"} {width="32.00000%"}\
{width="32.00000%"} {width="32.00000%"} {width="32.00000%"}\
{width="32.00000%"} {width="32.00000%"} {width="32.00000%"}\
{width="32.00000%"} {width="32.00000%"} {width="32.00000%"}\
{width="32.00000%"} {width="32.00000%"} {width="32.00000%"}\
\
\
\
\
\
\
\
\
\
\
\
\
\
\
\
\
\
\
\
\
\
\
\
\
\
\
\
\
\
\
\
[^1]: http://svn.atnf.csiro.au/trac/asap
[^2]: http://www.narrabri.atnf.csiro.au/mopra/SoftwareDepot/QuickData.py
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We consider the use of a Kinetic Monte Carlo approach for the description of non-equilibrium bosonic systems, taking non-resonantly excited exciton-polariton condensates and bosonic cascade lasers as examples. In the former case, the considered approach allows the study of the cross-over between incoherent and coherent regimes, which represents the formation of a quasi-condensate that forms purely from the action of energy relaxation processes rather than interactions between the condensing particles themselves. In the latter case, we show that a bosonic cascade can theoretically develop an output coherent state.'
address:
- 'Russian Quantum Center, Novaya 100, 143025 Skolkovo, Moscow Region, Russia'
- 'Departamento de Física Teórica de la Materia Condensada and Condensed Matter Physics Center (IFIMAC), Universidad Autónoma de Madrid, E-28049, Spain'
author:
- 'T. C. H. Liew'
- 'H. Flayac'
- 'D. Poletti'
- 'I. G. Savenko'
- 'F. P. Laussy'
title: 'Kinetic Monte Carlo Approach to Non-equilibrium Bosonic Systems'
---
Introduction
============
Although Bose-Einstein condensation (BEC) was originally defined as an effect taking place in thermal equilibrium, it is striking to see that the concept has been generalized to nonequilibrium systems. For example, the BEC of photons in a cavity [@Klaers2010] has been reported and several groups have studied the BEC of exciton-polaritons (hybrid light-matter quasiparticles) appearing in semiconductor microcavities [@Kasprzak2006; @Balili2007; @Lai2007]. Here BEC is characterized [@Byrnes2014] by the spontaneous formation of coherence, typically measured by the transition of the second order coherence function with increasing particle density, as reported by several experimental groups [@Love2008; @Kasprzak2008; @Horikiri2010; @Tempel2012; @RahimiIman2012; @Amthor2015; @Kim2016].
The physics of nonequilibrium condensates has been shown to be radically different from that of equilibrium systems, where condensates may form in non-ground states [@Maragkou2010], multiple states [@Vorberg2013], and have distributions undescribable by a single temperature.
The theoretical description of nonequilibrium condensates typically requires an explicit treatment of energy relaxation processes. Such processes compete with dissipative processes, which cause particles to be lost from the system before reaching the ground state. The dynamical interplay of relaxation and dissipation ultimately determines the steady-state of the system. Energy relaxation mechanisms have been handled previously in exciton-polariton systems using semiclassical Boltzmann equations [@Porras2002; @Doan2005; @Kasprzak2008b; @Cao2008] or introduced phenomenologically into mean-field equations [@Read2009; @Wouters2010; @Wouters2012; @Sieberer2013]. Methods treating energy relaxation from first principles have also been developed based on stochastic sampling of mean-field equations [@Savenko2013] or their hybridization with the Boltzmann equations [@Solnyshkov2014]. However, these methods do not account for quantum fluctuations, which are needed for the unified treatment of nonequilibrium condensation below and above threshold. In principle, density matrix approaches [@Racine2014] (possibly supplemented with Monte Carlo techniques [@Molmer1993]) are applicable to this task, however, in practice they are only feasible for systems with small numbers of particles and modes [@Flayac2015]. Bosonic cascade lasers [@Liew2013; @Tzimis2015], which may operate with millions of particles have been treated with stochastic sampling of the positive-P distribution [@Chaturvedi1977; @Liew2016], however, such a method is only accurate in the presence of an initial coherent state.
In the present work we employ a kinetic Monte Carlo approach based on quantum Boltzmann equations for the description of non-equilibrium multimode open quantum systems. Kinetic Monte Carlo has been developed under different names in different fields, from vacancy migration in binary ordered alloys [@Young1966], the Ising model [@Bortz1975] and chemical reactions [@Gillespie1976]. A good overview of the method can be found in Ref. . The approach allows stochastic sampling of the quantum particle distribution function and allows the treatment of systems with up to hundreds of modes with possibly thousands of particles each. From the particle distribution functions we have full access to the coherence statistics, as characterized by the second order correlation function. We apply the technique to two specific examples: polariton condensation in one-dimensional microwires and terahertz lasing in bosonic cascade lasers.
In the former case we are able to describe the gradual cross-over from incoherent population of excited states to partial coherence in non-ground states and the formation of a fully coherent BEC with increasing particle density. The responsible energy relaxation processes are described from first principles, accounting for polariton-phonon scattering and the scattering of polaritons with hot exciton states [@Porras2002]. Aside these interaction processes, it is notable that additional interactions between the condensing particles themselves are not required for the formation of a condensate (which is consistent with the original equilibrium theory of BEC of the ideal gas).
In the case of bosonic cascade lasers, we access for the first time theoretically the coherence of the lasing mode and show that it can be useful for terahertz lasing with high quantum efficiency.
Generic Kinetic Monte Carlo Approach
====================================
We start with the consideration of a set of $M$ discrete modes with populations $n_1$, $n_2$, $\ldots$, $n_M$. The probability of the system being in any particular state at time $t$ is $P_{n_1,n_2,\ldots,n_M}(t)$. The probability distribution contains sufficient information to calculate the quantum expectation values of a variety of quantities, in particular those with operators that commute with the number operator. For example, one can calculate: $$\begin{aligned}
\langle n_i(t) \rangle &= \sum P_{n_1,n_2,\ldots,n_i,\ldots,n_M}(t) n_i\\
\langle n_i^2(t) \rangle &= \sum P_{n_1,n_2,\ldots,n_i,\ldots,n_M}(t) n_i^2\end{aligned}$$ Of course the probability distribution $P_{n_1,n_2,\ldots,n_M}(t)$ does not contain all the information on the state of the system, which would require the full quantum density matrix. However, from the above we can gain access to the second order correlation function $g_{2,n_i}(t)=\langle n_i^2(t) \rangle/\langle n_i(t) \rangle^2$, which is the parameter typically used to measure the coherence of a given mode. The above prescription can also be easily generalized to the case of non-zero time delay and cross-correlations between different modes.
The calculation of the quantum probability distribution can be based on the quantum Boltzmann master equation, with generic form:
$$\begin{aligned}
&\frac{dP_{n_1,n_2,\ldots,n_M}(t)}{dt}\notag\\
&\hspace{5mm}=\sum_{ij} W_{i\rightarrow j}P_{n_1,n_2,\ldots,n_j-1,n_i+1,\ldots,n_M}(t)\left(n_i+1\right)n_j\notag\\
&\hspace{5mm}+\sum_i\frac{1}{\tau_i}\left[P_{n_1,n_2,\ldots,n_i+1,\ldots,n_M}(t)\left(n_i+1\right)\right.\notag\\
&\hspace{30mm}\left.-P_{n_1,n_2,\ldots,n_i,\ldots,n_M}(t)n_i\right]\notag\\
&\hspace{5mm}+\sum_i\Gamma_i\left[P_{n_1,n_2,\ldots,n_i-1,\ldots,n_M}(t)n_i\right.\notag\\
&\hspace{30mm}\left.-P_{n_1,n_2,\ldots,n_i,\ldots,n_M}(t)\left(n_i+1\right)\right]\label{eq:QuantumBoltzmann}\end{aligned}$$
The first term represents stimulated scattering processes between modes, where $W_{i\rightarrow j}$ is the bare (spontaneous) scattering rate from mode $i$ to mode $j$. In an exciton-polariton system, this term would include phonon emission (or absorption) processes as well as scattering processes involving hot excitons [@Porras2002]. These processes introduce a temperature dependence of the system via the temperatures of phonon or exciton baths.
The second and third terms represent decay and incoherent/non-resonant pumping of the modes, at rates $\tau_i$ and $\Gamma_i$, respectively. Their form is consistent with the Liouvillian operator for the full quantum density matrix, written in Ref. for the case of incoherent/non-resonant pumping. In principle other scattering processes (e.g., parametric scattering processes [@Savvidis2000]) can also be included, where the generic form is the coupling of one probability in the distribution to another.
Equation \[eq:QuantumBoltzmann\] can be simulated numerically given $W_{i\rightarrow j}$, $\tau_i$, and $\Gamma_i$ using kinetic Monte Carlo. This approach is based on first defining an initial state: $$\left(n_1,n_2,\ldots,n_M\right)$$ Equation \[eq:QuantumBoltzmann\] defines the scattering rates to other possible states to which the above state can jump to. A probability distribution of possible jumps to other states is associated to the scattering rates and a random quantum jump is selected from the probability distribution. The jump time defines the amount of time the system spends in the original state, from which the calculation of expectation values can be updated. Then the process is repeated until the end of the time range of the calculation. The process is then further repeated sampling over different quantum trajectories characterized by different stochastic quantum jumps. The system is able to attain a steady state, characterized by constant average expectation values. This is because nonlinear loss processes [@Keeling2008] have effectively been accounted for in the quantum Boltzmann equations; when a given mode becomes highly occupied, the probability for it to lose particles increases such that its occupation is bounded.
Since we neglect off-diagonal elements in the density matrix, we note that our approach is strictly speaking valid only when the system is not too far above the condensation threshold and the various scattering processes can be obtained accurately from single-particle wavefunctions. Far above threshold, polaritons are typically modelled with the mean-field Gross-Pitaevskii equation, where coherence is assumed. Here, we are interested in the behavior crossing the threshold. In principle, polariton-polariton scattering could renormalize the energy dispersion and alter the various energy relaxation rates, however, this does not affect the general trend of relaxing to the lowest available state, which is why we will obtain results consistent with experiments even above threshold.
Given that the range of validity of the quantum Boltzmann approach is the same as that for standard classical Boltzmann equations, the quantum Boltzmann equations can always be reduced to classical ones, depending on the quantities of interest. Our motivation for working with quantum Boltzmann equations is that they give access to second order correlations.
Non-Equilibrium Condensation in Polariton Microwires
====================================================
Exciton-polariton systems are short-lived bosonic quantum systems that are subject to weak energy relaxation processes. Consequently they are an exemplary non-equilibrium quantum system. They have been experimentally shown to form Bose-Einstein condensates [@Kasprzak2006; @Balili2007; @Lai2007], yet they may also become trapped in non-ground states [@Richard2005] or form non-ground state condensates [@Maragkou2010].
For simplicity, we will consider a one-dimensional exciton-polariton system or microwire [@Wertz2010]. The study of partial energy relaxation processes in such systems is particularly relevant to the study of polariton condensate transistors [@Anton2012; @Anton2013] and the control of spin currents for spintronics [@Anton2015; @Gao2015]. The main energy relaxation mechanisms in this system arise from polariton scattering with acoustic phonons [@Piermarocchi1996] and the scattering of polaritons with high momentum exciton states that can be considered as a reservoir [@Porras2002] (provided that we are not interested in the coherence statistics of these excitons). The scattering processes are illustrated in Fig. \[fig:ScatteringMechanisms\].
![(color online) Polariton energy relaxation mechanisms illustrated on the low-momentum polariton dispersion (blue/solid). The green/dashed curve illustrates the high-momentum exciton dispersion. a) Polariton-phonon scattering. b) Polariton-exciton scattering. c) Polariton-polariton to polariton-exciton scattering. d) Polariton pumping.[]{data-label="fig:ScatteringMechanisms"}](F1){width="\columnwidth"}
The calculation of the polariton-phonon scattering rates is shown in the Appendix. For typical parameters we find the result shown in Fig. \[fig:PhononScattering\] for a temperature of $5$K. Due to the reduced density of states in a one-dimensional system, as compared to planar two-dimensional microcavities, we find that the polariton-phonon scattering rates are small, below neV. Given that typical polariton decay rates are at least $10\mu$eV in typical microcavities, [*polariton-phonon scattering alone is insufficient to describe the relaxation of polaritons in one-dimensional systems*]{}. Even accounting for bosonic stimulation, very large polariton occupation numbers or very high lifetime microcavities [@Sun2017] would be needed to make polariton-phonon scattering dominant.
![(color online) Polariton-phonon scattering rates from mode $k_1$ to $k_2$. a) Polariton relaxation processes. The scattering to $k=0$ modes is larger than that of other modes due to the larger density of states at $k=0$ in a 1D system. b) Polariton excitation processes, requiring non-zero temperature.[]{data-label="fig:PhononScattering"}](F2){width="\columnwidth"}
For this reason, it is important to account for the polariton-exciton scattering processes illustrated in Figs. \[fig:ScatteringMechanisms\]b-d. The calculation of these rates is outlined in the Appendix. The process in Fig. \[fig:ScatteringMechanisms\]b adds to the scattering rates $W_{ij}$ introduced in Eq. \[eq:QuantumBoltzmann\]. The processes in Fig. \[fig:ScatteringMechanisms\]c require the addition of new terms in the quantum Boltzmann equation (Eq. \[eq:QuantumBoltzmann\]) of the form: $$W_{ij\rightarrow l}P_{n_1,\ldots,n_l-1,\ldots,n_i+1,\ldots,n_j+1,\ldots,n_M}(t)(n_i+1)(n_j+1)n_l$$ We assume that the incoherent pumping processes of the system can be derived primarily from the process in Fig. \[fig:ScatteringMechanisms\]d to provide $\Gamma_i$. It should be noted that the rates of the pumping processes and the scattering process illustrated in Fig. \[fig:ScatteringMechanisms\]b are proportional to the occupation of exciton reservoir states, which is modelled with a Boltzmann distribution of the form: $$n_{ex,k}=n_{ex}e^{-\hbar^2k^2/(2m_Xk_BT)},$$ where $m_X$ is the exciton effective mass (taken as 0.22 times the free electron mass in GaAs based microcavities) and $n_{ex}$ is a parameter representing the maximum occupation of a state in the thermal exciton reservoir. This parameter can be taken as a measure of the strength of incoherent pumping in the system, which would be controlled experimentally via the intensity of a non-resonant laser or current from an electrical injection mechanism.
Given the aforementioned scattering rates for typical microcavity parameters (given in the Appendix) we obtain the time dependence of the average occupation of the ground state below threshold shown in Fig. \[fig:MicrowireTimeDependence\]a. Here the system is evolved from an initial vacuum state. The average occupation remains below unity and the state is that of an incoherent state with $g_2=2$ (Fig. \[fig:MicrowireTimeDependence\]b). The shown quantities are here obtained after averaging over more than $10^9$ quantum jumps. By studying the statistical variation over different runs, one can obtain an estimate for the error in the obtained quantities. Since for larger occupations one has to sample a larger number of different states, the statistical error in the occupations is larger.
![(color online) a) Time dependence of $\langle n(t)\rangle$ of the ground state below threshold. b) Corresponding time dependence of $g_2(t)$. c) Time dependence of $\langle n(t)\rangle$ for the ground state ($k=0$) above threshold. d) Corresponding time dependence of $g_2(t)$. In each panel the light shaded region represents the statistical error, corresponding to one standard deviation.[]{data-label="fig:MicrowireTimeDependence"}](F3v4){width="\columnwidth"}
Above threshold, we find a large occupation of the ground state developing after an initial stabilization time, as shown in Fig. \[fig:MicrowireTimeDependence\]c. This is accompanied by the formation of coherence characterized by $g_2=1$, as shown in Fig. \[fig:MicrowireTimeDependence\]d. While the statistical error in $g_2$ is very large when the occupation numbers are small, which correspond to a small denominator in calculating $g_2$ and consequently large effects of small fluctuations in $\langle n\rangle$, the statistical error above threshold becomes small and indistinguishable in the plot.
We stress that while we have accounted for polariton-phonon scattering, it does not affect significantly our results, in which polariton-exciton scattering is the dominant and sufficient energy relaxation mechanism. This is consistent with earlier works [@Porras2002; @Flayac2015].
In addition to describing the behaviour of polaritons below and above threshold, the kinetic Monte Carlo theory is able to access the cross-over between the incoherent and condensed regimes. Figure \[fig:MicrowirePowerDependence\] illustrates the change in the momentum distribution of polaritons on the polariton dispersion. As the pumping intensity is increased the various energy relaxing scattering mechanisms become more and more stimulated. This is captured as a gradual overcoming of the bottleneck region [@Richard2005], before full condensation is obtained at large pumping strength. Here the large majority of polaritons collect in the system ground state and full coherence is characterized by $g_2=1$.
![(color online) a-g) Momentum distribution of polaritons in the long-time (continuous wave excitation) limit for increasing pumping intensity, represented by the maximum reservoir occupation number, $n_{ex}$. h) Dependence of the second order correlation function, $g_2$ (circles and left-hand scale), and the average occupation number, $\langle n\rangle$ (squares and right-hand scale), for the highest occupied state as a function of the reservoir occupation number. While there is a significant statistical error below threshold, the error bars narrow upon condensation.[]{data-label="fig:MicrowirePowerDependence"}](F4){width="\columnwidth"}
It is worth noting that the obtained phenomenon of condensation is obtained here [*without direct interactions between condensing particles*]{}. The interactions that we introduce are only to provide a physical mechanism of energy relaxation, but in principle any mechanism of energy relaxation would generate similar behaviour. The considered system is thus a nonequilibrium analogue of the non-interacting ideal gas. It should be noted though that here we are considering a confined system, in which the quantized modes in $k$ correspond to the levels discretely modelled in our approach. While we have focused on the second order coherence, polariton condensates are typically also characterized by the appearance of a first order spatial coherence, which decays exponentially with distance [@Mouchliadis2008; @Belykh2013]. As our technique neglects off-diagonal terms in the density matrix we are unable to access this property, which could be treated with other techniques [@Sarchi2007; @Doan2008].
Coherence Formation in Bosonic Cascade Lasers
=============================================
A bosonic cascade laser is composed of a series of equidistant energy levels and was originally proposed for the high efficiency generation of terahertz (THz) frequency radiation [@Liew2013]. Potential realizations making use of parabolic quantum wells are in experimental development [@Tzimis2015]. When a particle is excited in a particular level of the cascade it is assumed that it can undergo a radiative transition to the next level in the cascade. Thus, in the case that the radiative transition is at THz frequency, one can have a high quantum efficiency process where an optical quantum of energy injected into the system undergoes multiple energy relaxing processes resulting in the emission of many THz frequency photons. The THz emission processes are typically weak in strength, but they can become enhanced by bosonic final state stimulation at high occupation numbers. When the system is placed inside a THz cavity, it has been assumed that the result will be the generation of a coherent THz mode although the theory of such a process has not been attempted.
The bosonic cascade laser can be described by the quantum Boltzmann rate equations: $$\begin{aligned}
&\frac{dP_{n_1,\ldots,n_M,n_T}}{dt}=\notag\\
&\hspace{5mm}P_0\sum_{\lambda=1}^M\left[P_{n_1,\ldots,n_\lambda-1,\ldots,n_M,n_T}-P_{n_1,\ldots,n_\lambda,\ldots,n_M,n_T}\right]\notag\\
&\hspace{2mm}+W\sum_{\lambda=2}^M\left[-P_{n_1,\ldots,n_{\lambda-1},n_\lambda,\ldots,n_M,n_T}n_{\lambda-1}(n_\lambda+1)n_T\right.\notag\\
&\hspace{5mm}+P_{n_1,\ldots,n_{\lambda-1}-1,n_\lambda+1,\ldots,n_M,n_T-1}n_{\lambda-1}(n_\lambda+1)n_T\notag\\
&\hspace{5mm}+P_{n_1,\ldots,n_{\lambda-1}+1,n_\lambda-1,\ldots,n_M,n_T+1}(n_{\lambda-1}+1)n_\lambda(n_T+1)\notag\\
&\hspace{5mm}\left.-P_{n_1,\ldots,n_{\lambda-1},n_\lambda,\ldots,n_M,n_T}(n_{\lambda-1}+1)n_\lambda(n_T+1)\right]\notag\\
&\hspace{2mm}+\frac{1}{\tau}\sum_{\lambda=1}^M\left[P_{n_1,\ldots,n_\lambda+1,\ldots,n_M,n_T}(n_\lambda+1)\right.\notag\\
&\hspace{10mm}-\left.P_{n_1,\ldots,n_\lambda,\ldots,n_M,n_T}n_\lambda\right]\notag\\
&\hspace{2mm}+\frac{1}{\tau_T}\left[P_{n_1,\ldots,n_M,n_T+1}(n_T+1)-n_TP_{n_1,\ldots,n_M,n_T}\right]\end{aligned}$$ where $M$ bosonic levels have populations $n_1$, $n_2$, $\ldots$, $n_M$, and $n_T$ is the number of THz photons in the THz cavity. We assume for simplicity an equal pumping rate of all levels in the cascade by a coherent driving of strength $P_0$. $\tau$ is the decay rate of bosons in each level, $\tau_T$ the decay rate of THz photons, and $W$ is the nearest neighbour level scattering rate, which is assumed here independent of the level index for simplicity. Rather than repeating the detailed calculation of the scattering rate, we take the value of $W\tau=8.3\times10^{-7}$ consistent with Refs. . Since our main objective is to derive the formation of coherence, that is, lasing in the THz mode we will consider a fixed value of $\tau_T=0.1\tau$ rather than presenting a detailed dependence on all parameters.
Figure \[fig:Quantum Cascade\] shows results from the kinetic Monte Carlo modelling of a bosonic cascade of $M=10$ levels. The mode occupations agree fully with the result from the corresponding classical Boltzmann rate equations, however, the kinetic Monte Carlo approach provides an additional access to the second order correlation function. This reveals the smooth transition from an incoherent to a coherent state of THz photons in the THz cavity.
![(color online) Kinetic Monte Carlo simulation of a bosonic cascade laser with $M=10$ levels. a) Power dependence of the level occupations. b) Power dependence of the THz mode occupation. c) Power dependence of the second order correlation function of the THz mode. d) Power dependence of the quantum efficiency of THz generation. In all panels the points show results from kinetic Monte Carlo simulation, while the solid curves in (a), (b), and (d) show the result of solution of the classical Boltzmann equations (see Appendix).[]{data-label="fig:Quantum Cascade"}](F5v2){width="\columnwidth"}
For completeness we also calculated the quantum efficiency of THz emission, which, we define as the ratio of the number of THz photons emitted by the system to the number of optical frequency photons put into the system [@Liew2013]. This is equivalent to the ratio of the THz photon emission rate ($n_T/\tau_T$) to the total system pumping rate ($MP_0$): $$Q=\frac{n_T}{M\tau_TP_0}$$ The cascade geometry allows THz lasing to appear with quantum efficiency exceeding unity, in the regime of stimulated scattering despite the weak spontaneous scattering rates in the system.
Conclusion
==========
The kinetic Monte Carlo approach offers an efficient numerical simulation of quantum Boltzmann equations, suitable for the description of non-equilibrium bosonic systems. Such an approach offers access to the second order correlation function and allows to study its development across non-equilibrium phase transitions. As an example, we considered the behaviour of the second order correlation function in one-dimensional exciton-polariton microwires. Here we are able to treat the cross-over from an incoherent state, to a non-ground state, to a ground quasi-condensate. While we required interactions between polaritons and hot excitons to provide a mechanism of energy relaxation from first principles, interactions between the condensing polaritons themselves were not required to generate the condensate. We also studied the formation of coherence in a bosonic cascade laser geometry, verifying the possibility of coherent terahertz emission with quantum efficiency exceeding unity.
We hope that the kinetic Monte Carlo approach can also serve in the description of partial energy relaxation and coherence formation in a variety of other exciton-polariton systems. In particular, we anticipate that the quantum optics of geometries confined with static potentials [@Lai2007; @Kim2013; @Ostrovskaya2013; @Zhang2015; @Winkler2016] and self-induced traps [@Askitopoulos2015; @Gao2015b], where transitions were observed between non-ground and ground quasi condensates would be accessible.
[*Acknowledgements:*]{} T.L. was supported by the Singaporean Ministry of Education – Singapore Academic Research Fund Tier-2 project (Project No. 2015-T2-1-055) and Tier-1 project (Project No. 2016-T1-1-084). D.P. was supported by the Singaporean Ministry of Education – Singapore Academic Research Fund Tier-2 project (Project No. MOE2014-T2-2-119, with WBS No. R-144-000-350-112). I.G.S. was supported by IBS-R024-D1, the Australian Research Council’s Discovery Projects funding scheme (project DE160100167), and the President of Russian Federation (project MK-5903.2016.2).
Polariton Relaxation Mediated by Acoustic Phonons
=================================================
The phonon-assisted sctattering rate between two polariton states involving excitons of wavevector ${\bf{k}}_1$ and ${\bf{k}}_2$ is computed following Ref.[@Piermarocchi1996] as $$\begin{aligned}
\label{eq:Wk1k2}
{W_{{{\bf{k}}_1} \to {{\bf{k}}_2}}} &=& \frac{{{L_z}}}{{\rho uV}}\frac{{{{\left| {\Delta {\bf{k}}} \right|}^2} + q_z^2}}{{\left| {\hbar u{q_z}} \right|}}{\left| {X\left( {{{\bf{k}}_1}} \right)X\left( {{{\bf{k}}_2}} \right)} \right|^2}\\
\nonumber &\times& {\left[ {{a_e}I_e^{||}\left( {\left| {\Delta {\bf{k}}} \right|} \right)I_e^ \bot \left( {{q_z}} \right) - {a_h}I_h^{||}\left( {\left| {\Delta {\bf{k}}} \right|} \right)I_h^ \bot \left( {{q_z}} \right)} \right]^2}\end{aligned}$$ where $\Delta {\bf{k}} = {{\bf{k}}_1} - {{\bf{k}}_2}$ and $q_z$ is the projection of the phonon momentum on the vertical axis. We use here typical $GaAs$ material paramters: $\rho=5318$ kg/m$^3$ is the material density, $L_z=10$ nm and $V=\pi R^2 L_z$ are the quantum well thickness and volume (for a microcavity radius $R$) respectively. $u=3350$ m/s is the speed of sound, $a_{\rm{e}}=-7$ eV and $a_{\rm{h}}=2.7$ eV are the lattice deformation potentials of the induced by phonons at the locations of electrons and holes. ${I_{e,h}^{||}}$ and ${I_{e,h}^{\bot}}$ are the overlap integrals of the phonon wave functions with the electron and hole wave functions, respectively, in the in-plane and growth directions. They are expressed as $$\begin{aligned}
I_{e,h}^{||}\left( {\left| {\Delta {\bf{k}}} \right|} \right) = {\left[ {1 + {{\left( {\frac{{{m_{e,h}}}}{{{m_{\rm{e}}} + {m_{\rm{h}}}}}|\Delta {\bf{k}}|{a_{\rm{B}}}} \right)}^2}} \right]^{ - 3/2}}\\
I_{e,h}^ \bot \left( {{q_z}} \right) = \frac{{{\pi ^2}}}{{\frac{{{q_z}{L_z}}}{2}\left[ {{\pi ^2} - {{\left( {\frac{{{q_z}{L_z}}}{2}} \right)}^2}} \right]}}{\rm{sin}}\left( {\frac{{{q_z}{L_z}}}{2}} \right)\end{aligned}$$ where $m_\textrm{e}=0.067m_0$ and $m_\textrm{h}0.18m_0$ are the effective masses of electrons and holes in terms of the free electron mass $m_0$ and $a_\textrm{B}=10$ nm is the exciton Bohr radius. Finally $X({\bf k})$ is the excitonic fraction defined as $$X\left( {\bf{k}} \right) = \frac{2}{{\sqrt {4 + {{\left| {{E_{\rm p}}\left( {\bf{k}} \right)/{\Omega _R}} \right|}^2}} }}$$ where ${E_{\rm p}}\left( {\bf{k}} \right)$ is the polariton dispersion relation and $\Omega_R=10$ meV is the Rabi splitting.
Polariton Relaxation Mediated by Hot Excitons
=============================================
The matrix elements of scattering between polariton and hot-exciton states are obtained from the Fermi Golden rule following Ref. . The scattering rate of the process illustrated in Fig. \[fig:ScatteringMechanisms\]b, from a polariton state of wavevector $k_1$ and exciton state of wavevector $k_3$ to a polariton state of wavevector $k_2$ and an exciton state of wavevector $k_4$, is given by: $$W_{k_1,k_2}=\frac{2\pi}{\hbar}\left(\frac{Lm_X}{\pi\hbar^2\sqrt{|k_3k_4|}}\right)\left(\frac{6E_Ba_B^2}{S}\right)^2n_{ex,k_3}$$ Here the factor in the first parenthesis on the right-hand side is an average of the initial and final exciton density of states in one-dimension (assuming a parabolic exciton dispersion). The factor in the second parenthesis is the matrix element of exciton-exciton scattering [@Tassone1999], with $E_B$ the exciton binding energy, $a_B$ the exciton Bohr radius and $S$ a normalization area. $n_{ex,k_3}$ is the occupation of excitons in initial state $k_3$. For each combination of wavevectors $k_1$ and $k_2$, $k_3$ and $k_4$ are obtained from energy and momentum conservation: $$\begin{aligned}
k_3&=\frac{2m_X}{\hbar^2}\frac{E_{k_1}-E_{k_2}}{2(k_1-k_2)}-\frac{k_1-k_2}{2}\\
k_4&=k_1-k_2+k_3\end{aligned}$$ where $E_k$ represents the polariton dispersion.
Similar expressions can be used for the processes in Figs. \[fig:ScatteringMechanisms\]c and d. We note that in the case of Fig. \[fig:ScatteringMechanisms\]c one should sum over a few different processes that can satisfy the energy and momentum (phase matching) conditions.
Classical Boltzmann Equations for the Bosonic Cascade
=====================================================
In the classical regime, the quantum cascade can be modelled by a set of classical rate equations [@Liew2013] for the mode occupations, $n_\lambda$ $$\begin{aligned}
\frac{dn_M}{dt}&=P_0+W\left(n_{M-1}(n_M+1)n_T\right.\notag\\
&\hspace{20mm}\left.-n_M(n_{M-1}+1)(n_T+1)\right)-\frac{n_M}{\tau}\\
\frac{dn_\lambda}{dt}&=P_0+W\left(n_{\lambda-1}(n_\lambda+1)n_T\right.\notag\\
&\hspace{20mm}\left.-n_\lambda(n_{\lambda-1}+1)(n_T+1)\right.\notag\\
&\hspace{20mm}\left.+n_{\lambda+1}(n_\lambda+1)(n_T+1)\right.\notag\\
&\hspace{20mm}\left.-n_\lambda(n_{\lambda+1}+1)n_T\right)-\frac{n_\lambda}{\tau}\\
\frac{dn_1}{dt}&=P_0+W\left(n_2(n_1+1)(n_T+1)\right.\notag\\
&\hspace{20mm}\left.-n_1(n_2+1)n_T\right)-\frac{n_1}{\tau}\end{aligned}$$ where $1<\lambda<M$, and the THz mode occupation $n_T$, $$\begin{aligned}
\frac{dn_T}{dt}&=W\sum_\lambda^2\left(n_\lambda(n_{\lambda-1}+1)(n_T+1)\right.\notag\\
&\hspace{20mm}\left.-n_{\lambda-1}(n_\lambda+1)n_T\right)-\frac{n_T}{\tau_T}\end{aligned}$$ Equations for the steady state are readily obtained by setting the time derivatives to zero: $$\begin{aligned}
n_M&=\frac{P_0\tau+W\tau n_{M-1}n_T}{W\tau\left(n_T+1+n_{M-1}\right)+1}\\
n_\lambda&=\frac{P_0\tau+W\tau\left((n_{\lambda-1}+n_{\lambda+1})n_T+n_{\lambda+1}\right)}{W\tau\left(2n_T+1+n_{\lambda-1}-n_{\lambda+1}\right)+1}\\
n_1&=\frac{P_0\tau+W\tau\left(n_T+1\right)n_2}{W\left(n_T-n_2\right)+1}\\
n_T&=\frac{W\tau\sum_{\lambda=2}^M\left(1+n_{\lambda-1}\right)n_\lambda}{W\tau\left(n_1-n_M\right)+\frac{\tau}{\tau_T}}\end{aligned}$$ A simultaneous solution to this set of equations can be easily found by starting from an initially unoccupied state and evaluating the quantities $n_\lambda$ and $n_T$ iteratively until the equations become consistent. The result of this procedure gives rise to the solid curves in Figs. \[fig:Quantum Cascade\]a, b, and d, which are in agreement with the result of full kinetic Monte Carlo modelling.
[10]{}
J Klaers, J Schmitt, F Vewinger, & M Weitz, Nature, [**468**]{}, 545 (2010).
J Kasprzak, M Richard, S Kundermann, A Baas, P Jeambrun, J M J Keeling, F M Marchetti, M H Szymańska, R André, J L Staehli, V Savona, P B Littlewood, B Deveaud, & Le Si Dang, Nature, [**443**]{}, 409 (2006).
R Balili, V Hartwell, D Snoke, L Pfeiffer, & K West, Science, [**316**]{}, 1007 (2007).
C W Lai, N Y Kim, S Utsunomiya, G Roumpos, H Deng, M D Fraser, T Byrnes, P Recher, N Kumada, T Fujisawa, & Y Yamamoto, Nature, [**450**]{}, 529 (2007).
T Byrnes, N Y Kim, & Y Yamamoto, Nature Phys., [**10**]{}, 803 (2014).
A P D Love, D N Krizhanovskii, D M Whittaker, R Bouchekioua, D Sanvitto, S Al Rizeiqi, R Bradley, M S Skolnick, P R Eastham, R Andre, & L S Dang, Phys. Rev. Lett., [**101**]{}, 067404 (2008).
J Kasprzak, M Richard, A Baas, B Deveaud, R André, J-Ph Poizat, & L S Dang, Phys. Rev. Lett., [**100**]{}, 067402 (2008).
T Horikiri, P Schwendimann, A Quattropani, S Höfling, A Forchel, & Y Yamamoto, Phys. Rev. B, [**81**]{}, 033307 (2010).
J-S Tempel, F Veit, M Aßmann, L E Kreilkamp, A Rahimi-Iman, A Löffler, S Höfling, S Reitzenstein, L Worschech, A Forchel, & M Bayer, Phys. Rev. B, [**85**]{}, 075318 (2012).
A Rahimi-Iman, A V Chernenko, J Fischer, S Brodbeck, M Amthor, C Schneider, A Forchel, S Höfling, S Reitzenstein, & M Kamp, Phys. Rev. B, [**86**]{}, 155308 (2012).
M Amthor, H Flayac, I G Savenko, S Brodbeck, M Kamp, T Ala-Nissila, C Schneider, & S Höfling, arXiv: 1511.00878 (2015).
S Kim, B Zhang, Z Wang, J Fischer, S Brodbeck, M Kamp, C Schneider, S Hofling, & H Deng, Phys. Rev. X, [**6**]{}, 011026 (2016).
M Maragkou, A. J. D. Grundy, E. Wertz, A. Lemaître, I. Sagnes, P. Senellart, J. Bloch, & P. G. Lagoudakis, Phys. Rev. B, [**81**]{}, 081307(R) (2010).
D Vorberg, W Wustmann, R Ketzmerick, & A Eckardt, Phys. Rev. Lett., [**111**]{}, 240405 (2013).
D Porras, C Ciuti, J J Baumberg, & C Tejedor, Phys. Rev. B, [**66**]{}, 085304 (2002).
T D Doan, H T Cao, D B T Thoai, & H Haug, Phys. Rev. B, [**72**]{}, 085301 (2005).
J Kasprzak, D D Solnyshkov, R André, Le Si Dang, & G Malpuech, Phys. Rev. Lett., [**101**]{}, 146404 (2008).
H T Cao, T D Doan, D B T Thoai, & H Haug, Phys. Rev. B, [**77**]{}, 075320 (2008).
D Read, T C H Liew, Yu G. Rubo, & A V Kavokin, Phys. Rev. B, [**80**]{}, 195309 (2009).
M Wouters, T C H Liew, & V Savona, Phys. Rev. B, [**82**]{}, 245315 (2010).
M Wouters, New J. Phys., [**14**]{}, 075020 (2012).
L M Sieberer, S D Huber, E Altman, & S Diehl, Phys. Rev. Lett., [**110**]{}, 195301 (2013).
I G Savenko, T C H Liew, & I A Shelykh, Phys. Rev. Lett., [**110**]{}, 127402 (2013).
D D Solnyshkov, H Terças, K Dini, & G Malpuech, Phys. Rev. A, [**89**]{}, 033626 (2014).
D Racine & P R Eastham, Phys. Rev. B, [**90**]{}, 085308 (2014).
K Molmer, Y Castin, & J Dalibard, J. Opt. Soc. Am. B, [**10**]{}, 524 (1993).
H Flayac, I G Savenko, M Möttönen, T Ala-Nissila, Phys. Rev. B, [**92**]{}, 115117 (2015).
T C H Liew, M M Glazov, K V Kavokin, I A Shelykh, M A Kaliteevski, & A V Kavokin, Phys. Rev. Lett., [**110**]{}, 047402 (2013).
A Tzimis, A V Trifonov, G Christmann, S I Tsintzos, Z Hatzopoulos, I V Ignatiev, A V Kavokin, & P G Savvidis, Appl. Phys. Lett., [**107**]{}, 101101 (2015).
S Chaturvedi, C W Gardiner, I S Matheson, & D F Walls, J. Stat. Phys., [**17**]{}, 469 (1977).
T C H Liew, Y G Rubo, A S Sheremet, S De Liberato, I A Shelykh, F P Laussy, & A V Kavokin, New J. Phys., [**18**]{}, 023041 (2016).
W M Young & E W Elcock, P Phys. Soc., [**89**]{}, 735 (1966).
A B Bortz, M H Kalos, & J L Lebowitz, J. Comput. Phys., [**17**]{}, 10 (1975).
D T Gillespie, J. Comput. Phys., [**22**]{}, 403 (1976).
K A Fichthorn & W H Weinberg, J. Chem. Phys., [**95**]{}, 1090 (1991).
F P Laussy, E del Valle, & C Tejedor, Phys. Rev. Lett., [**101**]{}, 083601 (2008).
P G Savvidis, J J Baumberg, R M Stevenson, M S Skolnick, D M Whittaker, & J S Roberts, Phys. Rev. Lett., [**84**]{}, 1547 (2000).
J Keeling & N G Berloff, Phys. Rev. Lett., [**100**]{}, 250401 (2008).
M Richard, J Kasprzak, R André, R Romestain, L S Dang, G Malpuech, & A Kavokin, Phys. Rev. B, [**72**]{}, 201301 (2005).
E Wertz, L Ferrier, D Solnyshkov, R Johne, D Sanvitto, A Lemaître, I Sagnes, R Grousson, A V Kavokin, P Senellart, G Malpuech, & J Bloch, Nature Phys., [**6**]{}, 860 (2010).
C Anton, T C H Liew, G Tosi, M D Martín, T Gao, Z Hatzopoulos, P S Eldridge, P G Savvidis, & L Viña, Appl. Phys. Lett., [**101**]{}, 261116 (2012).
C Anton, T C H Liew, G Tosi, M D Martín, T Gao, Z Hatzopoulos, P S Eldridge, P G Savvidis, & L Viña, Phys. Rev. B, [**88**]{}, 245307 (2013)
C Antón, S Morina, T Gao, P S Eldrdige, T C H Liew, M D Martín, Z Hatzopoulos, P G Savvidis, I A Shelykh, & L Viña, Phys. Rev. B, [**91**]{}, 075305 (2015).
T Gao, C Antón, T C H Liew, M D Martín, Z Hatzopoulos, L Viña, P S Eldridge, & P G Savvidis, Appl. Phys. Lett., [**107**]{}, 011106 (2015).
C Piermarocchi, F Tassone, V Savona, A Quattropani, & P Schwendimann, Phys. Rev. B, [**53**]{}, 15834 (1996).
Y Sun, P Wen, Y Yoon, G Liu, M Steger, L N Pfeiffer, K West, D W Snoke, & K A Nelson, Phys. Rev. Lett., [**118**]{}, 016602 (2017).
L Mouchliadis & A L Ivanov, Phys. Rev. B, [**78**]{}, 033306 (2008).
V V Belykh, N N Sibeldin, V D Kulakovskii, M M Glazov, M A Semina, C Schneider, S Höfling, M Kamp, & A Forchel, Phys. Rev. Lett., [**110**]{}, 137402 (2013).
D Sarchi & V Savona, Phys. Rev. B, [**75**]{}, 115326 (2007).
T D Doan, H T Cao, D B T Thoai, & H Haug, Phys. Rev. B, [**78**]{}, 205306 (2008).
N Y Kim, K Kusudo, A Loffler, S Hofling, A Forchel, & Y Yamamoto, New J. Phys., [**15**]{}, 035032 (2013).
E A Ostrovskaya, J Abdullaev, M D Fraser, A S Desyatnikov, & Y S Kivshar, Phys. Rev. Lett., [**110**]{}, 170407 (2013).
L Zhang, W Xie, J Wang, A Poddubny, J Lu, Y Wang, J Gu, W Liu, D Xu, X Shen, Y G Rubo, B L Altshuler, A V Kavokin, & Z Chen, Proc. Natl. Acad. Sci. U. S. A., [**112**]{}, E1516 (2015).
K Winkler, O A Egorov, I G Savenko, X Ma, E Estecho,T Gao, S Müller, M Kamp, T C H Liew, E A Ostrovskaya, S öfling, & C Schneider, Phys. Rev. B, [**93**]{}, 121303(R) (2016).
Askitopoulos, T C H Liew, H Ohadi, Z Hatzopoulos, P G Savvidis, & P G Lagoudakis, Phys. Rev. B, [**92**]{}, 035305 (2015).
T Gao, E Estrecho, K Y Bliokh, T C H Liew, M D Fraser, S Brodbeck, M Kamp, C Schneider, S Höfling, Y Yamamoto, F Nori, Y S Kivshar, A G Truscott, R G Dall, & E A Ostrovskaya, Nature, [**526**]{}, 554 (2015).
F Tassone & Y Yamamoto, Phys. Rev. B, [**59**]{}, 10830 (1999).
| {
"pile_set_name": "ArXiv"
} |
---
author:
- 'Z.-Y. Lin'
- 'I.-L. Lai'
- 'C.-C. Su'
- 'W.-H. Ip'
- 'J.-C. Lee'
- 'J.-S., Wu'
- 'J.-B. Vincent'
- 'F. La Forgia'
- 'H. Sierks'
- 'C. Barbieri'
- 'P. L. Lamy'
- 'R. Rodrigo'
- 'D. Koschny'
- 'H. Rickman'
- 'H. U. Keller'
- 'J. Agarwal'
- 'M. F. A’Hearn'
- 'M. A. Barucci'
- 'J.–L. Bertaux'
- 'I. Bertini'
- 'D. Bodewits'
- 'G. Cremonese'
- 'V. Da Deppo'
- 'B. Davidsson'
- 'S. Debei'
- 'M. De Cecco'
- 'S. Fornasier'
- 'M. Fulle'
- 'O. Groussin'
- 'P. J. Gutiérrez'
- 'C. Güttler'
- 'S. F. Hviid'
- 'L. Jorda'
- 'J. Knollenberg'
- 'G. Kovacs'
- 'J.-R. Kramm'
- 'E. Kührt'
- 'M. Küppers'
- 'L.M. Lara'
- 'M. Lazzarin'
- 'J. J. López-Moreno'
- 'S. Lowry'
- 'F. Marzari'
- 'H. Michalik'
- 'S. Mottola'
- 'G. Naletto'
- 'N. Oklay'
- 'M. Pajola'
- 'A. Rożek'
- 'N. Thomas'
- 'C. Tubiana'
bibliography:
- 'OSIRIS\_CurvedJet.bib'
date: 'Received:...; accepted: ... '
title: 'Observations and analysis of a curved jet in the coma of comet 67P/Churyumov-Gerasimenko'
---
[Analysis of the physical properties and dynamical origin of a curved jet of comet 67P/Churyumov-Gerasimenko that was observed repeatedly in several nucleus rotations starting on May 30 and persisting until the early-August, 2015. ]{} [Simulation of the motion of dust grains ejected from the nucleus surface under the influence of the gravity and viscous drag effect of the expanding gas flow from the rotating nucleus.]{} [The formation of the curved jet is a combination of the size of the dust particles ($\sim$0.1-1 mm) and the location of the source region near the nucleus equator, hence enhancing the spiral feature of the collimated dust stream after being accelerated to a terminal speed of the order of m s$^{-1}$.\
]{}
Introduction {#intro}
============
Comets, as the most primitive bodies from the outer region of the solar system, are known to be very abundant in volatile ices and refractory dust grains. For a comet with orbital parameters inside 3-4 au, solar radiation raises the surface temperature to such an extent that ice sublimation initiates, as indicated by the appearance of a coma containing expanding gas and small dust particles. Therefore, the outgassing activity increases as the comet approaches perihelion. The behavior of comet 67P Churymov-Gerasimenko (67P hereafter) follows this pattern closely (; ; ). In addition to the nearly spherically symmetric coma, anisotropic structures in the form of collimated jets have been identified (; ). The identification of the source mechanism and acceleration process of these dust jets are one of the main focus of the Rosetta mission.
The close-up observations of the OSIRIS scientific camera on the Rosetta spacecraft ([@2007SSRv..128..433K]) have provided an unprecedented view of the morphology of the near-nucleus coma ([@Sierks2015]; [@Thomas2015]). The jet feature were visible in the early phases of the rendezvous mission (; ), and their development was closely monitored since then. Due to the orientation of its rotational axis and its complex shape, and a rotation period of 12.4 hours (Mottola et al. 2015), significant diurnal variations and seasonal effect of the gas flow and dust coma structure were detected ([@Gulkis2015]; [@2015Sci...347a0276H]).
![Gradual growth of the dust coma and dust jets of comet 67P: (a) image taken on 5 August, 2014, when the heliocentric distance r$_h$= 3.60 au, only a small dust jet can be seen emanating from the Hapi region; (b) 9 February, 2015 for r$_h$ = 2.35 au, the formation of a main jet is accompanied by a few fainter jets originating from other regions; (c) 10 May, 2015 for r$_h$ = 1.67 au, a system of bright jets appearing on the sunward side of the coma. The image contrast level is adjusted to log scale, ranging from -4 to -7.[]{data-label="JetsTime"}](JetvsTime3e.eps){width="50.00000%"}
Figure 1 shows the time development of the dust coma and jets as 67P approaches the Sun. From August 2014 ($r_h$ $\sim$3.60 au) to May 2015 ($r_h$ $\sim$1.67 au), the dust coma became more dense with the same image contrast level. At the beginning, the Hapi region located in the neck between the two lobes appeared to be the main source of the water gas flow ([@Gulkis2015]; ; [@2015Sci...347a0276H]) and dust jets ([@Sierks2015]; ; ; [@Vincent2015a]). All the dust jets have very straight configurations perpendicular to the surface, suggesting efficient acceleration of the embedded solid grains to radial speeds far exceeding the angular velocity ( $V_r$$\sim$1-2 m s$^{-1}$) of the nucleus due to its rotation. It was therefore surprising that a jet structure with large curvature appeared in late May (Fig. 2). This is the first time that a spiral structure was seen in the near-coma region of a comet made possible because of the close distance of the Rosetta spacecraft to the comet nucleus. Nevertheless, from ground-based observations ([@1991Icar...93..194S], , , [@2013AJ....146....4L]), we did see the repeatability of a curved appearance of the gaseous and dust jets related to the rotation of the nucleus or of the dust jets forming to dust tail due to the solar gravity and radiation pressure. However, these ground-based observations have much larger scales than Rosetta observations and the important physical processes might not be the same as the curved jets observed by ROSETTA. The curved jets persisted about two months and disappeared in early-August, 2015. In this work, we will examine the observed properties and dynamics of the curved jet as well as the localization of its possible source region.
The paper is organized as follows. Section 2 describes the morphology and time variation of the curved jet in comparison to other collimated dust jet features. The results of a set of computations making use of the gravity field model of the comet nucleus and of the Direct Simulation Monte Carlo (DSMC) simulation are presented in Section 3. A discussion on the theoretical results and the OSIRIS imaging data is given in Section 4.
Observations and data analysis {#obs}
==============================
Observations on May 30-31
-------------------------
Figure 2 shows the time sequence of the dust coma of 67P in approximately one nucleus rotation on May 30-31, 2015. The sun is towards the top and that is the reason why a clear shadow was cast behind the nucleus. The spin axis was pointing away from the projection plane with the rotation in the clockwise direction. According to the shape model ([@Sierks2015]; ; [@Jorda2015]), the sub-solar point during this observation was close to the equator of comet 67P. We note that three different components can be identified in Fig.2.a. The brightest one (J1) can be traced to the middle of the Hapi region. On its left side an array of jets (e.g., J2) became clearly visible in Fig. 2.b because of the sun-lit effect on Imhotep. Last but not least, on the right-hand side a jet (J3) in the form of a spiral arm came into view in Fig. 2.b. It is less discernable in subsequent images as a result of the orientation of the spacecraft relative to the rotating nucleus and/or decrease of the dust production. The spiral jet reappeared again in Fig. 2.h just one rotation period later.
![Jet structures obtained with the wide-angle camera from 13:03 UT on May 30 to 07:28 UT on May 31, 2015. Sub-panels (a) to (h) are separated by about two hours between two frames. The spatial scales and field of view range from 6.72 m/px,13.75 km (fig. a) to 7.01 m/px, 14.36 km (fig. h). []{data-label="CurvedJetAll"}](CurvedJet_May_All.eps){width="50.00000%"}
Source region of the Curved jet
-------------------------------
In order to find the source region of the curved jet, we both used the method described in Lin et al (2015) and the jet inversion method tracing the orientation of the colliminated beams back to their emission points ([@Vincent2015b], ). The shape model constructed by Jorda et al (2015) and current SPICE kernels have been used for this purpose. Nonetheless, it is difficult to obtain a precise location of the source region from the OSIRIS images taken from late-May to early-June, 2015 because of the diffusive structure of the jet. The most probable source region is located between Nut and Serget (see Figure 3 for a context image of where these regions are located on the nucleus). At closer scrutiny, it can be seen that this region is covered by smooth deposits of fine materials which are likely the result of airfall of low velocity particles, not being able to escape from the nucleus surface ([@Thomas2015], and reference in). However when looking closely at Nut region there is a more granular deposit including boulders with diameters up to a few tens of meters. A part of this area contains “pit-like” features with sub-meter to 4-5 meter diameter (see fig. 7 in ) that might be remnants formed by wind erosion or sublimation of the volatile-rich blocks. The curved jet appeared for two more months since the first detection. As a consequence, the features of the source region surface might have changed. A study of possible changes in localized areas of the identified source region would however be possible only in the later phase of the Rosetta mission, when high resolution images at spatial scales as small as 0.5 m/px will be available again.
![Geological map of 67P observed from the top of the head lobe and nomenclature of the geological regions (modified from Fig. 2 in and the possible source regions, red-circles, of the curved jet. The latitude of the possible source region is about 6 degree at the northern hemisphere. The subsolar point at the time of the present observation (31, May, 2015) is 9 degree at the southern hemisphere and moves toward to higher latitude until the beginning of September, 2015. The right-panel is the NAC image obtained on September 19.5, 2014 with a resolution about 0.53 m/px.[]{data-label="SR_CurvedJet"}](SR_curvedJet2.eps){width="50.00000%"}
Numerical Simulation
====================
To examine the origin and dynamical evolution of the curved jet, we need to consider the trajectories of different dust particles with different sizes under the influence of the gravitational attraction of the nucleus. Because of its highly irregular shape, the gravitational field have been computed by dividing the whole object into 33681 elements - according to the shape model with homogeneous structure and a bulk density of 532 kg m$^{-3}$ ([@Jorda2015]). Figure 4.a shows the contour plot of the gravitational field in the vicinity of the nucleus. At distances larger than three nucleus radii ($\sim$6 km) the gravitational field can be reasonably approximated by that from a point mass. However, close to the nucleus surface, the field distribution is far from spherical symmetry.
Figure 4.b shows the flow field of the coma gas expanding from the nucleus surface. To treat the transition from a collisional region close to the central nucleus to the collisionless coma at large distance, the Direct Simulation Monte Carlo (DSMC) method is required ([@Bird1994]). The basic structure of the DSMC code used for obtaining this result has been described in detail in [@Wu2004166], [@Su20101136], [@Lai2016] et al, and [@Liao2015] and will not be repeated here. For this simulation, we assume the water production rate is 10$^{27}$ molecules s$^{-1}$ and the sunlit portion of the nucleus surface at the time of consideration has been assumed to be all active in outgassing. Thus, a uniform gas production rate in sunlit side is Z = 4.8$\times$10$^{19}$ H$_2$O molecules m$^{-2}$ s$^{-1}$. The initial velocity distribution of the gas is described by a half-Maxwellian distribution with a thermal temperature of 228 K. The sunlit side is assumed to be free of gas outflow even though we know this is not necessarily true according to both the Rosina measurements ([@2015Sci...347a0276H]) and the recent discovery of the so-called night-side outbursts or Sunset jet activities by the OSIRIS camera team ([@Knollenberg2015] submitted and [@Xian2016]). It is clear that a certain level of weak outgassing activity existed on the nightside of the nucleus. Also, there can be activity driven by other gases (CO$_2$) - as may be the case for both the sunrise jets and for the sunset jets, and as was seen by Haessig (CO$_2$/H$_2$O = 4). However, even under such circumstances, the number distribution of the gas molecules should be highly non-isotropic within 5-10 nucleus radii covered by the simulation box. In a different study ([@Lai2016] et al), it has been demonstrated that the global gas flows tend to follow streamlines perpendicular to the surface of their source regions if the gas emission rate (Z) on the illuminated side is assumed to be proportional to square root $\cos \theta$ where $\theta$ is the solar zenith angle where $\theta$ $>$0. For the whole surface, Z = 0.1$\times$ Z$_0$ where Z$_0$ is the peak sublimation rate at $\theta$= 0. It is interesting to note that, even for localized outburst events, the dust jets appeared to be highly collimated (J-B.Vincent, private communication, 2016).
![ (a) A contour plot of the gravitational field of comet 67P in the XY plane containing the long and short axes. The rotational axis is in the perpendicular direction. The assumed bulk density is $\rho$= 532 kg m$^{-3}$. (b) The flow field of the expanding gas outflow. []{data-label="CurvedJet_GG"}](Gravity_GasFlow1.eps){width="40.00000%"}
In the simulation code, the motion of a dust particle of mass m with initial zero velocity is determined by the viscous drag effect of the expanding gas flow and the gravitational attraction of the nucleus. Note that the effects of solar radiation pressure on curved jet are not considerable in our simulation but will be involved in the future. The used equation of motion is shown below. ([@Gombosi1986]; [@1999Icar..140..173S]; ).\
$$\label{color_measure}
m\frac{dv}{dt} = m\vec{g}+\frac{1}{2}A\rho_gC_dv_r^2$$
where m is the mass of dust particle, and g is the “local” gravity. The second term on the right-hand side represents the gas drag effect. The dust particle of cross section A is assumed to be spherical and its density ($\rho$) is taken to be 1000 kg m$^{-3}$ ([@2015ApJ...802L..12F]). In Equation 1, v$_r$ is the relative velocity between the gas molecules and the dust particle, $\rho$$_g$ is the mass density of gas flow, and C$_d$ is the drag coefficient. Note that C$_d$ = 2 ([@Wallis1982]).
To demonstrate the combined effect of the gas drag and nucleus gravity, a source region is tentatively chosen in the Nut region (see Figure 3) with zero initial velocity. The general idea is to examine the dependence of particle size, gas production rate and emission location on the jet dynamics.
Result and conclusions
======================
Figure 5.a illustrates the obtained velocity profiles of dust grains with radius ranging between 1 $\mu$m and 1 mm for the gas sublimation rate of Z = 4.8$\times$10$^{19}$ H$_2$O molecules m$^{-2}$ s$^{-1}$. The bigger dust grains will be accelerated to lower radial velocity as compared with the particles of smaller sizes. This effect can be understood in terms of the size dependence of the gas drag force under the influence of the gravitational attraction of the comet nucleus. In addition, the dust grains are accelerated to their terminal speeds within a distance of the order of 2 km or slightly more in all the considered cases. The small micron-sized dust could reach an outflow speed as high as 30 m s$^{-1}$ while those of mm-size have terminal speed of the order of 0.5 m s$^{-1}$ (or less) which is comparable to the rotational speed of the nucleus at its surface. This immediately suggests that the grains in the curved jet must be relatively large (i.e, d $\sim$0.1-1 mm). The result is consistent with previous findings ([@2015Sci...347a3905R]) that the optically dominant particles in 67P coma are exactly those of size of 0.1-1 mm. In additional to the curved jet, the straight jets (Fig. 2 J1 and J2) might consist of relative smaller particles. The identification of the exact sizes would depend on the characteristics of the surface material and the effective outgassing rate which controls the gas drag acceleration.
![(a)The velocity profiles of dust grains of different sizes with sublimation rate Z = 4.8$\times$10$^{19}$ H$_2$O molecules m$^{-2}$ s$^{-1}$. (b) Trajectories of dust grains of different radii (from 1$\mu$m to 1mm) as the nucleus rotates. []{data-label="CurvedJet_V"}](velocity4.eps){width="40.00000%"}
It is perhaps not an accident that, the curved jet was observed to be emitted near the equatorial region of the head of 67P (Nut, Serqet and Ma’at ), i.e. where the nucleus rotation speed is the largest. As mentioned before in Section 2, the collimated jets of linear configuration seemed to have been all emanated from the Hapi region that is more or less along the spin axis, where the centrifugal force is minimal.
From our data analysis and preliminary numerical simulation, it is now understood that the appearance of a curved jet in May and June 2015 is caused by a combination of the ejection of mm-sized dust grains from the equatorial source region in the vicinity of Nut, Serqet and Ma’at. This unique set of OSIRIS observations provides important information on the physical properties of the dust grains and on the acceleration process.
OSIRIS was built by a consortium led by the Max-Planck-Institut für Sonnensystemforschung, Göttingen, Germany, in collaboration with CISAS, University of Padova, Italy, the Laboratoire d’Astrophysique de Marseille, France, the Instituto de Astrofísica de Andalucía, CSIC, Granada, Spain, the Scientific Support Office of the European Space Agency, Noordwijk, Netherlands, the Instituto Nacional de Técnica Aeroespacial, Madrid, Spain, the Universidad Politéchnica de Madrid, Spain, the Department of Physics and Astronomy of Uppsala University, Sweden, and the Institut für Datentechnik und Kommunikationsnetze der Technischen Universität Braunschweig, Germany. The support of the national funding agencies of Germany (Deutschen Zentrums für Luft- und Raumfahrt), France (Centre National d’Etudes Spatales), Italy (Agenzia Spaziale Italiana), Spain (Ministerio de Educación, Cultura y Deporte), Sweden (Swedish National Space Board; grant no. 74/10:2), and the ESA Technical Directorate is gratefully acknowledged. This work was also supported by grant number NSC 102-2112-M-008-013-MY3 and NSC 101-2111-M-008-016 from the Ministry of Science and Technology of Taiwan. We are indebted to the whole Rosetta mission team, Science Ground Segment, and Rosetta Mission Operation Control for their hard work making this mission possible.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We propose the use of Bayesian networks, which provide both a mean value and an uncertainty estimate as output, to enhance the safety of learned control policies under circumstances in which a test-time input differs significantly from the training set. Our algorithm combines reinforcement learning and end-to-end imitation learning to simultaneously learn a control policy as well as a threshold over the predictive uncertainty of the learned model, with no hand-tuning required. Corrective action, such as a return of control to the model predictive controller or human expert, is taken when the uncertainty threshold is exceeded. We validate our method on fully-observable and vision-based partially-observable systems using cart-pole and autonomous driving simulations using deep convolutional Bayesian neural networks. We demonstrate that our method is robust to uncertainty resulting from varying system dynamics as well as from partial state observability.'
author:
- 'Keuntaek Lee, Kamil Saigol and Evangelos A. Theodorou [^1] [^2]'
bibliography:
- 'references.bib'
title: 'Safe end-to-end imitation learning for model predictive control'
---
Introduction {#sec:introduction}
============
As the deployment of deep neural networks as controllers for physical robotic systems becomes more prevalent, the issue of safety within artificial intelligence becomes an increasingly important concern. Recently the use of end-to-end imitation learning to develop neural network control policies has surged in popularity, due in large part to the ease with which deep models can learn complex dynamics and infer global state from local data while bypassing the need for significant parameter tuning. In contrast, traditional approaches to vision-based control rely on methods such image segmentation and object detection, classification, labeling, and filtering; often, these methods require significant engineering and tuning. In addition, collecting the large datasets required by these methods is a difficult process, and it is often impossible to collect and process a dataset in a manner that is complete enough to account for every possible scenario.
The challenge, therefore, is to design a system capable of taking an appropriate action or making an appropriate prediction given novel data at test time, ideally with as little hand-tuning as possible. The ability to effectively handle novel data is key to introducing safety in intelligent control and is a particular weakness of current end-to-end methods of control.
It is easy to imagine situations in which novel data could be presented to a learned model at test time. For example, malfunctioning sensors could pass unusual signals to a control system; changes in the environment could significantly affect system dynamics, such as in the case of robots that move from a smooth floor to a carpeted area; or, as is our motivation in this paper, a camera-based controller could encounter obstacles of a heretofore unseen type.
Bayesian models – that is, models that output predictive distributions, rather than point estimates – have long been considered invaluable in machine learning for a number of reasons. An estimate of model uncertainty can help to minimize risk, develop rejection criteria, compensate for extreme priors, and combine models [@Bishop2006]. In the context of end-to-end learning, especially for autonomous vehicles and mobile robots, uncertainty estimates provide an easily interpreted safety measure. By training models that can provide predictive uncertainty, we obtain a valuable indication of whether the model is able to produce a usable output; if it is not – that is, if the uncertainty is high (Fig. \[fig:uncertainty increase\]) – we can take an appropriate action to compensate for the model’s lack of knowledge. In the context of end-to-end learning for autonomous driving such an action could include, for example, the return of control to a human driver.
Deep neural networks (DNN), popular due to their ability to model extremely complex functions, have enjoyed great success in a number of fields, including model-free reinforcement learning and model-based control. However, while deep networks have been a revolutionary change to many engineering applications, their particular weakness is that they provide point estimates instead of a predictive distribution. While undeniably effective, the use of DNNs results in a loss of the properties that make Bayesian methods so powerful. As described in detail below, recent work on efficient deep Bayesian neural networks has helped to bridge this gap.
In this paper, we present an approach that combines deep end-to-end learning with Bayesian networks to learn high-speed model predictive control (MPC) while exploiting the predictive uncertainty of a Bayesian convolutional network to provide a measure of safety. Our approach combines MPC with imitation learning and reinforcement learning (RL) methods to successfully perform complex tasks in both fully-observable and vision-based partially-observable scenarios while effectively handling novel data at test time. Our approach is fully end-to-end; aside from the initial selection of network structure, comparatively little engineering or hand-tuning is required for success.
The rest of the paper is organized as follows: In Section \[sec:related\_work\], we discuss work related to our method that appears in the literature. Section \[sec:background\] briefly introduces key background material. In Section \[sec:safe\_imlearn\_control\], we define the problem of safe learning for MPC and present our solution. Results of experiments are presented in Section \[sec:experiments\], followed by a discussion of future directions in Section \[sec:discussion\]. We conclude with Section \[sec:conclusion\].
Related Work and Contributions {#sec:related_work}
==============================
Imitation and end-to-end learning are popular in the literature for a variety of tasks including the control of autonomous robots. The DAgger meta-learning algorithm and its variants are popular due to the fact that these methods provide a regret bound that is linear in the time horizon [@DAgger], contrary to earlier approaches with larger worst-case regret. Earlier work, based on DAgger, specifically addressed the problem of learning a MPC policy for high-speed autonomous navigation [@Pan-NIPSWS-17], but did not account for safety and robustness in the learned model. In [@Pereira2018], the authors develop an extension to the DAgger algorithm for end-to-end learning of MPC sequence policies with structured recurrent neural networks but do not address the problem of safe MPC. In addition, partially-observable states and vision-based tasks are not considered. In contrast, our work imposes no algorithmic structure on the neural network, avoids recurrent networks and the resulting difficulties in training, and successfully completes complex tasks without full state observability. The NVIDIA PilotNet project [@Bojarski2017] does consider vision-based vehicle control with end-to-end learning and learns to recognize diverse obstacles and objects, but is not a probabilistic method and does not seem to explicitly handle truly novel test inputs. Combining uncertainty measurement into the autonomous driving task, [@Strickland-17] investigated a predictive approach for collision risk assessment; however, this is a classification task rather than a regression task and cannot provide a learned controller for continuous action spaces.
The work most similar to our own is the DropoutDAgger approach [@DropoutDAgger], a probabilistic extension to DAgger which uses the Dropout method of Bayesian approximation first presented in [@Gal2016Bayesian] to obtain a predictive distribution from a neural network. The output variance is then used as a threshold to determine whether to return control to an expert controller, similar to the method presented herein. However, in [@DropoutDAgger] both the dropout probability and the uncertainty threshold are selected manually; the tasks and experts considered are based on RL; and the case of partially observable environments and vision-based tasks is not considered at all. In contrast, our work uses more principled methods to select dropout probabilities and the uncertainty threshold work. In addition, we work in the MPC task domain rather than in the RL domain, which is more challenging due to the requirements for reactive, high-frequency control but is well-suited to deployment on real autonomous vehicles. Finally, our method specifically addresses vision-based tasks using Bayesian convolutional neural networks.
Given the prior work in the area of end-to-end learning for control the contributions of this paper are summarized as follows:
1. We introduce the DropoutVGG Net, which combines VGG Net [@VGG], the dropout technique for Bayesian approximation [@Gal2016Bayesian], and Concrete Dropout [@Gal2017Concrete] for selection of the dropout probabilities in our neural network.
2. We use Cross-Entropy RL to optimally select an uncertainty threshold for return of control from a learned model to a MPC expert policy.
3. Our formulation considers both the aleatoric and epistemic uncertainty of our neural network policy.
4. We demonstrate that our method is robust to uncertainty from varying system dynamics as well as uncertainty from partial state observability.
5. Use a model predictive controller as expert for a demonstration and successfully learn safe MPC using end-to-end learning with a Bayesian convolutional neural networks.
6. We demonstrate our results on cart-pole balancing and autonomous driving tasks in both fully and partially observable cases.
Technical Background {#sec:background}
====================
In this section we briefly review key ideas used in our method. Our formulation consists of an autonomous expert model predictive controller, a deep Bayesian neural network policy trained with imitation learning, and a reinforcement learning module to optimize the uncertainty threshold.
Differential Dynamic Programming for Trajectory Optimization
------------------------------------------------------------
Iterative LQG/Differential Dynamic Programing (iLQG/DDP) [@MPCDDP] is a dynamic programming method which provides locally optimal solutions to the optimal control problem. When used as the underlying optimizer for a model predictive controller, DDP provides real-time performance for complex, high-speed tasks; it also provides improved robustness relative to other methods due to the ability to calculate and return explicit trajectory-local feedback and feedforward gains.
Let the state of the system under consideration be ${\mathbf{x}}\in \mathbb{R}^n$ with controls ${\mathbf{u}}\in \mathcal{U} \subset \mathbb{R}^m$. We then consider the following optimal control problem.
$$\begin{aligned}
\label{eq:optcontrolprob}
&V({\mathbf{x}}(t_0), t_0) = \min_{{\mathbf{u}}(t)}\Big[\phi({\mathbf{x}}(t_f), t_f) + \int_{t_0}^{t_f}l({\mathbf{x}}(t), {\mathbf{u}}(t), t){\mathrm{d}}t\Big] \\
&\mbox{subject to dynamics}~ {\frac{{\mathrm{d}}{\mathbf{x}}(t)}{{\mathrm{d}}t}}= f({\mathbf{x}}(t), {\mathbf{u}}(t), t)\end{aligned}$$
Here $V$ is the value function, the time horizon is $[t_0, t_f]$, and we have terminal cost $\phi$ with running cost $l$. Let nominal state and control sequences be given by $\bar{{\mathbf{x}}}({\cdot}$) and $\bar{{\mathbf{u}}}(\cdot)$, respectively, so that we can define the state and control sequence updates $\delta{\mathbf{x}}$ and $\delta{\mathbf{u}}$ such that $ {\mathbf{x}}= \bar{{\mathbf{x}}} + \delta{\mathbf{x}}$ and $ {\mathbf{u}}= \bar{{\mathbf{u}}} + \delta{\mathbf{u}}$. DDP provides an efficient iterative method to compute the trajectory updates. To begin, note that discretization and linearization of the dynamics about the nominal trajectory results in the discrete, time-varying linear system
$$\begin{aligned}
\label{eq:linear_expansion}
\delta{\mathbf{x}}(t_{k+1})
&= \Phi\delta{\mathbf{x}}(t_k) + B\delta{\mathbf{u}}(t_k),\end{aligned}$$
where we have defined $\Phi = \Big[I + \frac{\partial f}{\partial {\mathbf{x}}}(\bar{{\mathbf{x}}}(t_k), \bar{{\mathbf{u}}}(t_k), t_k)\Delta t\Big]$ and $B = \frac{\partial f}{\partial {\mathbf{u}}}(\bar{{\mathbf{x}}}(t_k), \bar{{\mathbf{u}}}(t), t_k)\Delta t$. Applying the Bellman principle to \[eq:optcontrolprob\] and discretizing, we obtain
$$\begin{aligned}
\label{eq:discrete_bellman}
V({\mathbf{x}}(t_k), t_k) = \min_{{\mathbf{u}}(t_k)} \Big [\underbrace{l({\mathbf{x}}(t_k), {\mathbf{u}}(t_k))\Delta t + V({\mathbf{x}}(t_{k+1}), t_{k+1})}_{Q} \Big ]. \end{aligned}$$
The prime notation on $V$ and derivatives indicates evaluation at $t_{k+1}$. Taking a quadratic expansion of the right hand side of \[eq:discrete\_bellman\] and then noting that ${\mathbf{u}}$ and $\delta{\mathbf{u}}$ have a linear relationship so that we can perform the minimization in \[eq:discrete\_bellman\] over $\delta{\mathbf{u}}$, we obtain the optimal control update law
$$\label{eq:opt_control_update}
\delta{\mathbf{u}}(t_k) = -Q_{uu}^{-1}(Q_u + Q_{ux}\delta{\mathbf{x}}(t_k)),$$
where the first and second derivatives of $Q$ are
$$\begin{aligned}
\label{eq:Q_and_derivs}\begin{split}
Q_0 &= l\Delta t + V', Q_x = l_x\Delta t + \Phi^T V'_x,~ Q_u = l_u \Delta t + B^T V'_x \\
Q_{xu} &= l_{xu}\Delta t + \Phi^TV'_{xx}B,~ Q_{ux} = Q_{xu}^T \\
Q_{xx} &= l_{xx} \Delta t + \Phi^TV'_{xx} \Phi,~ Q_{uu} = l_{uu} \Delta t + B^TV'_{xx}B.
\end{split}\end{aligned}$$
It is important to note that the optimal control update is given by a feedback control law with feedback gain $K_{fb} = -Q_{uu}^{-1}Q_{ux}$ and feedforward gain $K_{ff} = -Q_{uu}^{-1}Q_u$.
For each iteration of DDP, updates along a trajectory are computed using the dynamic programming methodology, starting at the terminal time and computing optimal updates backward along the trajectory. The equations for this backward pass are computed by substituting \[eq:opt\_control\_update\] into the quadratic expansion of \[eq:discrete\_bellman\] and grouping terms of corresponding order to obtain
$$\begin{aligned}
V(\bar{{\mathbf{x}}}(t_k), t_k) &= Q_0 - \frac{1}{2} Q_u^T Q_{uu}^{-1} Q_u \label{eq:bw_first}\\
V_x(\bar{{\mathbf{x}}}(t_k), t_k) &= Q_x - Q_{xu} Q_{uu}^{-1} Q_u \\
V_{xx}(\bar{{\mathbf{x}}}(t_k), t_k) &= Q_{xx} - Q_{xu} Q_{uu}^{-1} Q_{ux}. \label{eq:bw_last}\end{aligned}$$
After the backward pass is complete, a forward pass finds the new nominal control and state trajectories using the computed gains and the nonlinear dynamics, respectively. The backward and forward passes are iterated until the maximum number of iterations is exceeded or until convergence. DDP is summarized in Algorithm \[alg:algorithm\_ddp\] . In our work, we apply DDP as part of the model predictive control algorithm by performing a small number of DDP iterations at each time step, applying the first control of the resulting optimal sequence to the plant, obtaining a new state estimate, and repeating until the task is complete.
Epistemic and Aleatoric Uncertainty
-----------------------------------
As indicated in [@Kendall2017], uncertainty in machine learning models can come from two sources: incomplete data and incomplete knowledge of the environment. The former, called *epistemic uncertainty* is the result of a data set that is either not large enough or does not cover enough of the “search space”. Such uncertainty could be explained away by access to unlimited data and can be reduced by the collection of additional data. The latter source, called *aleatoric uncertainty*, is the result of our inability to completely sense all aspects of the environment. While such uncertainty could not be explained away with access to more data – or even with access to unlimited data – it could be explained away with access to unlimited sensing.
Note that in both cases, completely explaining away the uncertainty requires access to something we cannot possible obtain, i.e., either unlimited data or unlimited sensing capability. It is therefore useful to consider using these sources of uncertainty in an engineering or machine learning context. Epistemic uncertainty measures what our model doesn’t know; given data, we should be able to train a model that can output this type of uncertainty. Aleatoric uncertainty measures noise that is inherent in the environment, i.e., it measures what cannot be explained by data. Again, given data, we should be able to train a model to output this type of uncertainty as well. In principle, the total predictive uncertainty would be the combination of the two. As shown in [@Kendall2017], it is sometimes possible to use just one of these measures to develop a reasonable model; however, the more principled approach of combining these predictive uncertainty estimates often results in a superior model.
Bayesian Approximation with Dropout
-----------------------------------
Bayesian learning methodologies in and with neural networks have a long history [@Neal1995]. Currently, at least two popular methods of Bayesian learning with neural networks appear frequently in literature. The first, which essentially produces predictive distributions resulting from a distribution over neural network structure, is the Dropout approach introduced in [@Gal2016Bayesian]. The second, in which predictive distributions are obtained due to probability distributions over network weights, is the Weight Uncertainty approach introduced in [@pmlr-v37-blundell15]. The two approaches are similar in a number of ways, not least of which is the need to impose a variational approximating distribution over network weights and perform Monte Carlo sampling. We discuss the dropout technique, which we elect to use in our work; the alternative approach requires doubling the number of parameters in the network, making it difficult to use for real-time MPC for vehicles with only onboard computation given very large convolutional neural networks.
Dropout has traditionally been used as a regularization method to prevent overfitting while training DNNs [@Hinton2012]. It does so by effectively annihilating connections between neurons, resulting in a network that is much simpler, i.e., has fewer effective model parameters and is thus less prone to overfitting training data. This is achieved by randomly sampling a binary *dropout mask* for each layer $i$ from a Bernoulli distribution with probability of success $p_i$ and applying this mask to the weights of layer $i$. For example, a simple two-layer feedforward network is represented by the equation
$$\begin{aligned}
\mathbf{y} = \mathbf{W}_2 \sigma(\mathbf{W}_1 \mathbf{x} + \mathbf{b_1}),\end{aligned}$$
where $\sigma$ is an component-wise activation function and the network has outputs $\mathbf{y} \in \mathbb{R}^K$, inputs $\mathbf{x} \in \mathbb{R}^D$, weight matrices $W_1 \in \mathbb{R}^{N \times D}$ and $W_2 \in \mathbb{R}^{K \times N}$, and bias $\mathbf{b}_1 \in \mathbb{R}^N$ (output bias has been neglected for clarity). Binary dropout masks $\mathbf{z}_1 \in \{0, 1\}^D$ and $\mathbf{z}_2 \in \{0, 1\}^N$ can be sampled component-wise from Bernoulli distributions with success probability $p_1$ and $p_2$ for layers $1$ and $2$, respectively, and applied to the network resulting in a structurally simpler model with some connections removed.
$$\begin{aligned}
\mathbf{\hat{y}} = \big[\mathbf{W}_2 diag(\mathbf{z}_2)\big] \sigma(\big[\mathbf{W}_1 diag(\mathbf{z}_1)\big] \mathbf{x} + \mathbf{b_1})\end{aligned}$$
Dropout as a method of Bayesian approximation differs from the traditional approach in two respects. First, as presented in [@Gal2016Bayesian], dropout in this context is randomly applied at test time as well as during training. This means that the predictive output for a single forward pass with a single input is the result of a network that is structurally simpler than the original. Since the random nature of the dropout masks means that the same test-time input will result in a different output each time the input is passed through the network, several samples of the output corresponding to the same input result in a distribution that estimates epistemic uncertainty. In addition, the dropout probability $p$ for each layer can be learned along with the network weights during training using the Concrete Dropout approach [@Gal2017Concrete].
Aleatoric uncertainty is not learned via sampling, but instead learned by using the so-called *heteroscedastic loss* function (described below) and splitting the output layer of the network (Fig. \[fig:DropoutVGG\]) such that one channel is trained to output a predictive mean and the other the aleatoric uncertainty [@Kendall2017]. By combining the uncertainty obtained via sampling with that obtained during training, an accurate representation of total model uncertainty is obtained.
Batch Imitation Learning
------------------------
Let $\mathcal{S}, \mathcal{O}, \mathcal{U}$ be the state, observation, and admissable control spaces of the task under consideration, respectively. Our goal is to determine a neural network policy $\pi: \mathcal{O} \rightarrow \mathcal{U}$ that is able to minimize the expected accumulated cost over a discrete finite time horizon $H$:
$$\begin{aligned}
\min_{\pi} \mathbb{E}_{p_{\pi}} \Big[ \sum_{t=0}^{H-1} l({\mathbf{x}}_t, {\mathbf{u}}_t) \Big],\end{aligned}$$
where ${\mathbf{x}}_t \in \mathcal{S}$, ${\mathbf{u}}_t \in \mathcal{U}$, and running (immediate) cost $l$ for consistency with the notation introduced earlier. $p_{\pi}$ is the joint distribution of ${\mathbf{x}}_t$, ${\mathbf{u}}_t$, and $\mathbf{o}_t \in \mathcal{O}$ for the policy $\pi$ for $t = 0, \ldots, H-1$.
This cost may be unavailable or impossible for a learned model to optimize directly. Instead, imitation learning assumes that an expert capable of optimizing this cost is available. The goal of imitation learning is to learn a policy that minimizes the deviation between the cost of the learned model and that incurred by the expert policy $\pi^*$
$$\begin{aligned}
\pi_{NN} = \arg\min_\pi \mathbb{E}_{p_{\pi^*}} \Big[l({\mathbf{x}}_t, {\mathbf{u}}_t) \Big],\end{aligned}$$
where ${\mathbf{u}}_t = \pi(\mathbf{o}_t)$ and $\pi_{NN}$ denotes the selected neural network policy. In batch imitation learning, as used in this work, this is reduced to a supervised learning problem in which the model is trained to learn from data, consisting of observation-action or state-action pairs, generated by the expert.
End-to-end learning offers a number of advantages; our primary motivation in this work is that end-to-end methods avoid the need for burdensome manual engineering and unprincipled parameter tuning.
Safe End-to-End Learning for Model Predictive Control of Autonomous Vehicles {#sec:safe_imlearn_control}
============================================================================
We wish exploit end-to-end learning to train a neural network policy to perform MPC for autonomous vehicle control. The trained model must be safe; in our framework, safety is introduced by returning control to an expert controller when the learned model is unable to produce a suitable output. Such a framework is useful for long-term missions in which it is infeasible or impossible to continuously use an expert. This could be the case, for example, when the power and computational requirements of an iterative optimal controller or the natural limitations of a human expert become a burden on the system.
As noted above, our probabilistic policy is a deep Bayesian neural network using the dropout technique. We apply dropout to all weight-containing layers in our models. At test time we perform Monte Carlo sampling of the output by repeatedly passing in the same input, thereby estimating epistemic uncertainty by calculation of the sample distribution. The linear output layers of our models are split, with one channel trained to calculate aleatoric uncertainty. In order to do so, we train the model with the heteroscedastic loss introduced in [@Kendall2017]
$$\label{eq:loss function}
Loss = \sum_i \frac{{\left\lVerty_i - \mu\right\rVert}_2^2}{\sigma^2} + \log \sigma^2,$$
where the network predicts $\mu, \sigma^2$ and the true target is $y$. We recognize this as being essentially equivalent to the scalar Gaussian log likelihood of independent data samples.
Our expert controller in all tasks considered herein is a real-time MPC controller operating between 40Hz and 60Hz which applies the DDP trajectory optimization algorithm in a receding horizon fashion. In all cases, the MPC-DDP expert has access to full state information, i.e., is assumed to have perfect knowledge of the state during the generation of training data. The learned model may have only partial state information depending on the specific task, each of which is discussed in subsequent sections.
In keeping with the motivation to avoid manual selection of model parameters and hyperparameters, the dropout probability for each layer is trained along with network weights using the Concrete Dropout approach. In contrast to competing methods, our framework also avoids hand-tuned specification of the uncertainty threshold above which the MPC-DDP expert regains control of the system.
Reinforcement Learning of the Uncertainty Threshold
---------------------------------------------------
High predictive uncertainty generated by the Bayesian network is indicative of a novel test-time input, i.e., data of a type very different from the distribution of the training data. When the output uncertainty is high, the learned model is incapable of producing a useful control; we therefore return control of the system to the MPC-DDP controller. The threshold above which the switch from learned model to expert controller is manually selected in earlier works; here, we propose selecting the threshold in a principled manner with reinforcement learning.
In imitation learning, an expert will generally be more successful and obtain a larger reward (or lower cost) than a learned model for the same task. As a result, naive application of RL to learn the uncertainty threshold will result in a policy that immediately hands control to the expert, i.e., it will learn a threshold value of 0 so that the expert’s action is always queried for maximum reward. Our approach therefore encourages the use of the learned model for as long as possible.
The choice of RL algorithm is important; our choice is motivated by the observation that the switch from the Bayesian neural network policy to the MPC-DDP expert is best viewed as a rare discrete event. The Cross-Entropy Method (CEM) [@DeBoer2005] applied to reinforcement learning is well-suited to such cases. We use CEM to attempt optimization of the uncertainty threshold for switching from the neural network policy to the MPC-DDP expert as summarized in Algorithm \[alg:cem\].
Our results, discussed below, demonstrate that our framework converges to a threshold that allows the learned model to perform the task while switching back to the expert while still allowing the expert enough time to recover from or handle the unusual inputs.
return $\mu$, $R_{sum}$
Experiments {#sec:experiments}
===========
We validate our framework for safe learning of model predictive control on two different tasks:
1. Cart-pole swing-up with full and partial state observability
2. Vision-based rally car driving with obstacle avoidance
For vision-based tasks, we introduce the DropoutVGG Network. This Bayesian extension to VGG Net [@VGG] uses the dropout technique as described above. In addition, we split the output layer and train the network to output aleatoric uncertainty as shown in Figure \[fig:DropoutVGG\].
For all tasks, we provide positive terminal reward in the CEM algorithm if the task was completed successfully by the Bayesian network model or by the MPC-DDP controller.
$$\begin{aligned}
\label{eq:terminal_reward}
\phi_T =
\begin{cases}
T, & \mbox{if success} \\
0, & \mbox{otherwise}
\end{cases}\end{aligned}$$
In order to encourage use of the learned model, we provide a positive immediate reward for every time step for which the learned model is used.
$$\begin{aligned}
r_t =
\begin{cases}
1, & \mbox{if using learner} \\
0, & \mbox{if using DDP}
\end{cases}\end{aligned}$$
Cart-Pole Swing-Up
------------------
For the cart-pole swing-up task, the MPC-DDP expert was provided a system dynamics model identical to that used in the popular OpenAI Gym simulator [@gym]. This system consists of a weighted pole balanced on a cart. The goal is to swing up and balance the pole only by applying a horizontal force to the cart. The terminal reward is typical of that used in OpenAI Gym; that is, we provide the reward in \[eq:terminal\_reward\] if at the terminal time $T$:
$$-2.4 < x_{T} < 2.4 \enspace \text{and} \enspace -15 \degree <\theta_{T} < 15\degree,$$
where $x$ is the horizontal position of the cart and $\theta$ is the pole angle.
### Fully Observable Case
In fully observable case, both the expert and the Bayesian neural network have perfect information regarding the complete system state $[x, \dot{x}, \theta, \dot{\theta}]^T$. A convolutional model is not required as no vision-based control is performed; we instead use a simple feedforward Bayesian network (Fig. \[fig:cartpolenn\]).
After training the neural network with the DDP expert, we allow the learned model to attempt the task. The mass of the system is changed at an arbitrarily selected time to create a disturbance; the learner must revert control to the DDP controller before the angle of the pole falls outside the required limits and must do so in time for the DDP controller to recover.
### Partially Observable Case
In the partially observable case, the neural network can estimate $x$ and $\theta$ but velocities $\dot{x}$ and $\dot{\theta}$ are not directly accessible. DropoutVGG is trained with inputs consisting of (third-person view) gray-scale images of the cart-pole system and observations of the expert control; the learned model must use image information to infer the velocity components of the state in order to produce a usable control signal.
The goal and limits of the task are as the same as in the fully observable case. In this case, however, a disturbance is created when the input image is randomly corrupted with the addition of a new object that was not present in the training data.
Interestingly, we found that the predictive uncertainty of the neural network grows when larger objects corrupt the input image. Tests with color images also showed extreme sensitivity to the color of the new object; we hope to investigate this further, but for our current results the image was converted to grayscale before being passed to the neural network.
Autonomous Navigation and Obstacle Avoidance
--------------------------------------------
Autonomous driving is a more realistic and significantly more complex task than cart-pole swing-up. Here, we demonstrate the success of our framework on the autonomous navigation and obstacle avoidance task using the simulator provided by the AutoRally project [@AutoRallyGithub]. This software \[fig:gazebo\] simulates driving a 1/5th scale autonomous vehicle on a track and is capable of providing simulated color images from onboard cameras. We again use DDP as the expert controller; in this experiment, the DDP controller uses a data-driven adaptive dynamics model based on Sparse Spectrum Gaussian Processes (SSGP) [@Pan-NIPS-16].
DropoutVGG is trained with observations consisting of $80\times160$ pixel RGB images from the vehicle’s left camera as well as the corresponding steering angle produced by the MPC-DDP expert as it navigates around the track at a speed of 5 m/s. The goal is to train the network to learn to steer the vehicle; we do not learn the throttle (acceleration) in this experiment. The learned model is not provided with any direct information about the vehicle state; it may only access the camera image. As usual, no disturbances or unusual inputs are given to the neural network in the training phase.
In the absence of unusual test-time inputs, the Bayesian network easily learns to steer the vehicle based only on vision. However, when a previously unseen obstacle comes into view (Fig. \[fig:uncertainty increase\]) the predictive uncertainty increases significantly, indicating potential failure of the learned model. Our method successfully trains the learned model to pass control back to the DDP expert in time to avoid the obstacle.
Future Work {#sec:discussion}
===========
As demonstrated above, the dropout approach for Bayesian approximation is a good fit for safe end-to-end learning for control. However, sampling the network enough times to obtain accurate epistemic uncertainty estimates is time-consuming if done sequentially. It may be parallelized but this is difficult for networks with a large number of weights. We plan to investigate schemes for efficient sampling through Bayesian networks in a future work. On a more theoretical note, we found during our experiments with the AutoRally simulator that the predictive uncertainty increases far later than would be ideal; a promising avenue of research is to investigate recurrent or generative models capable of propagating predictive uncertainty into the future in order to obtain better behavior, especially in high-speed tasks. Another future direction is the extension of the current work to online imitation learning. As noted above, our current framework uses batch imitation learning, essentially reducing the problem to a simple supervised learning problem. However, as shown in [@DAgger], the DAgger algorithm and its variants can significantly reduce the regret bounds relative to this type of algorithm. We continue to investigate the best way to combine our framework – especially the RL component – with the DAgger methodology.
Conclusion {#sec:conclusion}
==========
We have presented herein a fully end-to-end framework for safe learning of model predictive control. Our approach successfully trains a Bayesian neural network to perform complex vision-based MPC with access to a dataset containing image observations and corresponding controls from an expert based on differential dynamic programming. This approach allows the learned model to effectively know when it will be unable to safely complete the task and return control to an expert capable of doing so. Our framework requires no manual parameter tuning; instead we propose the use of a reinforcement learning method to learn uncertainty thresholds in a principled manner. We validate our method on cart-pole swing-up and autonomous driving tasks.
[^1]: : Equal contribution. School of Electrical and Computer Engineering, the Institute for Robotics and Intelligent Machines, and the D. Guggenheim School of Aerospace Engineering, Georgia Institute of Technology, 270 Ferst Drive Atlanta GA 30332-0150, USA. Lyft Inc., San Francisco, CA, USA. Email: [email protected], [email protected], evangelos.theodorou@
[^2]: gatech.edu
| {
"pile_set_name": "ArXiv"
} |
---
author:
- '[Alireza Dirafzoon, Alper Bozkurt, and Edgar Lobaton [^1] ]{}'
bibliography:
- 'qual.bib'
- 'cinema.bib'
title: 'Dynamic Topological Mapping with Biobotic Swarms [^2]'
---
[^1]: Department of Electrical Engineering, [North Carolina State University]{}. Email: {adirafz, alper.bozkurt, ejlobato}@ncsu.edu
[^2]: This work was supported by the National Science Foundation under award CNS-1239243.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'In addition to finding meaningful clusters, centroid-based clustering algorithms such as $K$-means or mean-shift should ideally find centroids that are valid patterns in the input space, representative of data in their cluster. This is challenging with data having a nonconvex or manifold structure, as with images or text. We introduce a new algorithm, Laplacian $K$-modes, which naturally combines three powerful ideas in clustering: the explicit use of assignment variables (as in $K$-means); the estimation of cluster centroids which are modes of each cluster’s density estimate (as in mean-shift); and the regularizing effect of the graph Laplacian, which encourages similar assignments for nearby points (as in spectral clustering). The optimization algorithm alternates an assignment step, which is a convex quadratic program, and a mean-shift step, which separates for each cluster centroid. The algorithm finds meaningful density estimates for each cluster, even with challenging problems where the clusters have manifold structure, are highly nonconvex or in high dimension. It also provides centroids that are valid patterns, truly representative of their cluster (unlike $K$-means), and an out-of-sample mapping that predicts soft assignments for a new point.'
author:
- |
Weiran Wang Miguel Á. Carreira-Perpiñán\
Electrical Engineering and Computer Science, University of California, Merced\
[<http://eecs.ucmerced.edu>]{}
date: 'Jun 15, 2014'
title: 'The Laplacian $K$-Modes Algorithm for Clustering'
---
Introduction {#s:intro}
============
Given a dataset $\x_1,\dots,\x_N \in \bbR^D$, centroid-based clustering algorithms such as $K$-means [@Bishop06a] and mean-shift [@FukunagHostet75a; @Cheng95a; @Carreir00b; @ComanicMeer02a] estimate a representative $\c_k \in \bbR^D$ of each cluster $k$ in addition to assigning data points to clusters. Besides finding meaningful clusters, we would ideally like to find centroids that are valid patterns in the input space, representative of data in their cluster. This is challenging with data having a nonconvex or manifold structure, as with images or text. Fig. \[f:rotated1\] illustrates this with a single cluster consisting of continuously rotated digit-1 images. Since these images represent a nonconvex cluster in the high-dimensional pixel space, their mean (which averages all orientations) is not a valid digit-1 image, which makes the centroid not interpretable and hardly representative of a digit 1. Mean-shift does not work well either: to produce a single mode, a large bandwidth is required, which makes the mode lie far from the manifold; a smaller bandwidth does produce valid digit-1 images, but then multiple modes arise for the same cluster, and under mean-shift they define each a cluster.
Forcing the centroids to be exemplars is often regarded as a way to ensure the centroids are valid patterns. Although there exist exemplar-based or $K$-medoids clustering algorithms [@KaufmanRousseeuw90a; @Bishop06a; @Hastie_09a] which constrain centroids to be points from the dataset (“exemplars”) and often minimize a $K$-means type of objective function with a possibly non-Euclidean distance, such algorithms are typically slow because updating centroid $\c_k$ requires testing all pairs of points in cluster $k$. Besides, the exemplars themselves are often noisy and thus not that representative of their neighborhood.
------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- --------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- --------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- -------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
$K$-means $K$-modes GMS
![A cluster of 7 rotated-1 USPS digit images and the centroids found by $K$-means, $K$-modes (both with $K=1$) and mean-shift (with $\sigma$ so there is one mode). Example taken from @CarreirWang13a.[]{data-label="f:rotated1"}](rotated1_data_001.eps "fig:"){width="0.10\linewidth"} ![A cluster of 7 rotated-1 USPS digit images and the centroids found by $K$-means, $K$-modes (both with $K=1$) and mean-shift (with $\sigma$ so there is one mode). Example taken from @CarreirWang13a.[]{data-label="f:rotated1"}](rotated1_data_002.eps "fig:"){width="0.10\linewidth"} ![A cluster of 7 rotated-1 USPS digit images and the centroids found by $K$-means, $K$-modes (both with $K=1$) and mean-shift (with $\sigma$ so there is one mode). Example taken from @CarreirWang13a.[]{data-label="f:rotated1"}](rotated1_data_003.eps "fig:"){width="0.10\linewidth"} ![A cluster of 7 rotated-1 USPS digit images and the centroids found by $K$-means, $K$-modes (both with $K=1$) and mean-shift (with $\sigma$ so there is one mode). Example taken from @CarreirWang13a.[]{data-label="f:rotated1"}](rotated1_data_004.eps "fig:"){width="0.10\linewidth"} ![A cluster of 7 rotated-1 USPS digit images and the centroids found by $K$-means, $K$-modes (both with $K=1$) and mean-shift (with $\sigma$ so there is one mode). Example taken from @CarreirWang13a.[]{data-label="f:rotated1"}](rotated1_data_005.eps "fig:"){width="0.10\linewidth"} ![A cluster of 7 rotated-1 USPS digit images and the centroids found by $K$-means, $K$-modes (both with $K=1$) and mean-shift (with $\sigma$ so there is one mode). Example taken from @CarreirWang13a.[]{data-label="f:rotated1"}](rotated1_data_006.eps "fig:"){width="0.10\linewidth"} ![A cluster of 7 rotated-1 USPS digit images and the centroids found by $K$-means, $K$-modes (both with $K=1$) and mean-shift (with $\sigma$ so there is one mode). Example taken from @CarreirWang13a.[]{data-label="f:rotated1"}](rotated1_data_007.eps "fig:"){width="0.10\linewidth"} ![A cluster of 7 rotated-1 USPS digit images and the centroids found by $K$-means, $K$-modes (both with $K=1$) and mean-shift (with $\sigma$ so there is one mode). Example taken from @CarreirWang13a.[]{data-label="f:rotated1"}](rotated1_center_000.eps "fig:"){width="0.10\linewidth"} ![A cluster of 7 rotated-1 USPS digit images and the centroids found by $K$-means, $K$-modes (both with $K=1$) and mean-shift (with $\sigma$ so there is one mode). Example taken from @CarreirWang13a.[]{data-label="f:rotated1"}](rotated1_center_030.eps "fig:"){width="0.10\linewidth"} ![A cluster of 7 rotated-1 USPS digit images and the centroids found by $K$-means, $K$-modes (both with $K=1$) and mean-shift (with $\sigma$ so there is one mode). Example taken from @CarreirWang13a.[]{data-label="f:rotated1"}](rotated1_ms.eps "fig:"){width="0.10\linewidth"}
------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- --------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- --------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- -------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
Remarkably, no algorithm that partitions the data using exactly $K$ meaningful modes existed in the literature until the *$K$-modes algorithm* proposed by @CarreirWang13a. This combines the idea of clustering through binary assignment variables with the idea that high-density points are representative of a cluster. Each centroid found by the $K$-modes algorithm is the mode of a kernel density estimate defined by data points in each cluster. As a result, the centroids average out noise or idiosyncrasies that exist in individual data points and are representative of their cluster and neighborhood. This can be seen from the $K$-modes centroid for the rotated digit-1 problem in Fig. \[f:rotated1\]. $K$-modes was also shown to have nice properties such as being more robust to mis-specification of the bandwidth and to outliers. Its optimization procedure is also very efficient.
One important disadvantage of $K$-modes is that it uses the same assignment rule as $K$-means (each point is assigned to its closest centroid in Euclidean distance), so it can only find convex clusters (a Voronoi tessellation). Therefore, like $K$-means, it cannot handle clusters with nonconvex shapes or manifold structure, unlike mean-shift or spectral clustering [@ShiMalik00a]. The main contribution of this paper is to solve this issue, while keeping the nice properties that $K$-modes does have. The key idea is to modify the $K$-modes objective function such that the assignment rule becomes much more flexible. We then give an alternating optimization procedure to find the assignments and the modes. The resulting *Laplacian $K$-modes* algorithm is able to produce for each cluster a nonparametric density and a mode as valid representative (like $K$-modes), to separate nonconvex shaped clusters (like mean-shift and spectral clustering), and to give soft assignment of data points to each cluster. Yet, all of these merits are achieved at a reasonable computational cost, and the algorithm works well with high-dimensional data.
Related work {#s:related}
============
Centroids-based algorithms
--------------------------
Given the number of clusters $K$, $K$-means [@Bishop06a] minimizes the objective $$\begin{aligned}
\label{e:kmeans-objfcn}
\min_{\Z,\C} &\quad { \sum\nolimits^K_{k=1}{\sum\nolimits^N_{n=1}{z_{nk} \norm{\x_n-\c_k}^2}} } \\
\text{s.t.} &\quad z_{nk} \in\{0,1\},\; \sum\nolimits^K_{k=1}{z_{nk}} = 1,\; n=1,\dots,N\notag\end{aligned}$$ where $\Z = (z_{nk})$ are binary assignment variables (of point $n$ to cluster $k$) and $\C = (\c_1,\dots,\c_K)$ are centroids living in $\bbR^D$. At an optimum, centroid $\c_k$ is the mean of the points in its cluster.
Given a bandwidth $\sigma > 0$, Gaussian mean-shift [@FukunagHostet75a; @Cheng95a; @Carreir00b; @ComanicMeer02a] defines a kernel density estimate (kde) $$\label{e:kde}
p(\x) = \frac{1}{N} \sum^N_{n=1}{G\bigl(\norm{(\x-\x_n)/\sigma}^2\bigr)}$$ with kernel $G(t) \propto \smash{e^{-t/2}}$, and applies the iteration (started from each data point): $$\label{e:GMS}
p(n|\x) = \frac{\exp{\bigl(-\frac{1}{2} \norm{(\x-\x_n)/\sigma}^2\bigr)}}{\sum^N_{n'=1}{\exp{\bigl(-\frac{1}{2} \norm{(\x-\x_{n'})/\sigma}^2\bigr)}}}, \quad \x \leftarrow \f(\x) = \sum\nolimits^N_{n=1}{p(n|\x) \x_n}$$ which converges to a mode (local maximum) of $p$ from nearly any initial [@Carreir07a]. Each mode is the centroid for one cluster, which contains all data points that converge to its mode. The user parameter is the bandwidth $\sigma$, which determines the resulting number of clusters implicitly.
Both algorithms have well-known pros and cons. $K$-means tends to define round clusters; mean-shift can obtain clusters of arbitrary shapes and has been very popular in low-dimensional clustering applications such as image segmentation [@ComanicMeer02a], but does not work well in high dimension due to the scarcity of data and difficulty in obtaining a good kde. Both algorithms suffer from outliers, which can move centroids outside their cluster in $K$-means or create singleton modes in mean-shift. Computationally, $K$-means is much faster than mean-shift, at $\calO(KND)$ and $\calO(N^2D)$ per iteration, respectively, particularly with large datasets (in fact, accelerating mean-shift has been a topic of active research, e.g. @Carreir06a [@Yuan_10a]). Mean-shift does not require a value of $K$, which is sometimes convenient, but it is often desirable to force an algorithm to produce exactly $K$ clusters (e.g. in figure-ground separation ($K=2$) or in medical image analysis, where one may know or estimate the number of organs sought), which is not straightforward for mean-shift to achieve.
### The $K$-modes algorithm
We now briefly summarize the original $K$-modes algorithm [@CarreirWang13a]. Its objective function $$\begin{aligned}
\label{e:kmodes}
\max_{\Z,\C} &\quad \sum\nolimits_{n=1}^N \sum\nolimits_{k=1}^K z_{nk} {G\biggl( \norm{\frac{\x_n-\c_k}{\sigma}}^2 \biggr)} \\
\text{ s.t. } &\quad z_{nk} \in\{0,1\},\; \sum\nolimits^K_{k=1}{z_{nk}} = 1,\; n=1,\dots,N \notag\end{aligned}$$ can be seen as the sum of a kde defined by each cluster separately. It is convenient to solve this problem with alternating optimization over $\Z$ and $\C$. For fixed $\C$, the optimization over $\Z$ decouples over each point, and $\x_n$ is assigned to cluster $l=\smash{ \arg\max\nolimits_k\; {G( \norm{(\x_n-\c_k)/{\sigma}}^2 )} }= \arg\min\nolimits_k\; \norm{\x_n-\c_k}$ due to the discrete constraints of the problem. For fixed $\Z$, the optimization over $\C$ decouples over each cluster and we have a separate unconstrained maximization for each centroid, of the form $ L(\c_k) = \smash{\sum^N_{n=1}{ z_{nk} {G( \norm{(\x_n-\c_k)/{\sigma}}^2 )} }}$, which is proportional to the cluster’s kde (this is why each centroid is truly a mode), and can be done with mean-shift updates as in eq. . The cost per outer iteration of this procedure is $\calO(KND)$, which mainly comes from computing distances between data points and centroids. Since each step is strictly feasible and decreases the objective or leaves it unchanged, this converges to a local optimum in a finite number of outer-loop steps if the -step is exact.
Laplacian smoothing and learning soft assignments
-------------------------------------------------
Obtaining hard assignments by optimizing over a discrete cluster indicator matrix is usually difficult, because interesting objective functions are typically NP-hard. Spectral clustering algorithms [@ShiMalik00a; @YuShi03a] avoid this difficulty by first approximating the solution using eigenvectors of the normalized graph Laplacian. However, since the eigenvectors do not readily provide valid assignments, these algorithms need to run another clustering algorithm (usually $K$-means) on the eigenvectors to obtain actual partitions of the data—a post-processing step that is somewhat artificial and can introduce multiple local optima. In Laplacian $K$-modes, we relax $\Z$ to be a stochastic matrix, so our $\Z$-step results from a convex QP and provides soft assignments of points to clusters, which may also be used as posterior probabilities.
Laplacian smoothing has also been used in combination with nonnegative matrix factorization (NMF) for clustering [@Cai_11b]. NMF learns a decomposition of the input data matrix where both basis and coefficients are nonnegative, and tends to produce a parts-based representation of the data [@LeeSeung99a], though this is not always so. @Cai_11b add a Laplacian smoothing term regarding the coefficient matrix to the NMF objective function, so that data points that are close in input space are encouraged to have similar representations using the basis set. $K$-means is then applied to the learned coefficients matrix to obtain a final partition of the data. Like spectral clustering, this algorithm does not directly optimize over the assignments, but obtains them in a post-processing step.
There has been recent work in clustering that directly optimizes over a stochastic assignment matrix. @Arora_11a optimize over a stochastic matrix $\PP$ such that $\PP \PP^\top$ best approximates a rescaled similarity matrix. However, the optimization problem has multiple solutions which are related by rotations. Therefore, they propose to exploit the geometry of the problem using a rotation-based algorithm, which is straightforward for up to $K=4$ clusters but requires an optimization procedure to computing the projection onto the probability simplex for $K>4$ clusters. The idea of AnchorGraphs [@Liu_10c] is used by @YangOja12a to approximate the affinities between data points through a two-step Data-Cluster-Data (DCD) random walk. They then minimize the generalized KL divergence between a given sparse affinity matrix and the affinities obtained from DCD. In this formulation, the matrix containing probabilities of points moving to the (augmented) cluster nodes is stochastic. These approaches are related to Laplacian $K$-modes in that they optimize some objective over the assignment probabilities. But our Laplacian $K$-modes algorithm also obtains prototypical centroids, does not have the issue of rotational equivalence of @Arora_11a, and makes use of the efficient projection onto the probability simplex to deal with any number of clusters.
Algorithm {#s:alg}
=========
The Laplacian $K$-modes algorithm
---------------------------------
We change the assignment rule of $K$-modes to handle more complex shaped clusters based on two ideas: (1) the observation that *nearby data points should have similar assignments*; and (2) the use of *soft assignments*, which allows more flexibility in the clusters and simplifies the optimization. We first build a graph (e.g. $k$-nearest-neighbor graph) on the dataset, and let $w_{mn}$ be an affinity (e.g. binary, heat kernel) between $\x_m$ and $\x_n$. We then add to the $K$-modes objective function a Laplacian smoothing term $\smash{\frac{\lambda}{2}\sum_{m=1}^N \sum_{n=1}^N w_{mn} \norm{\z_m-\z_n}^2}$ to be minimized, where $\z_n=[z_{n1},\dots,z_{nK}]^\top$ is the assignment vector of $\x_n$, $n=1,\dots,N$, to each of the $K$ clusters, and $\lambda \ge 0$ is a trade-off parameter. The assignments are now continuous variables, but constrained to be positive and sum to $1$. Thus, $z_{nk}$ can be considered as the probability of assigning $\x_n$ to cluster $k$ (soft assignment). Thus, the *Laplacian $K$-modes* objective function is: $$\begin{aligned}
\min_{\Z,\C} &\quad \frac{\lambda}{2}\sum_{m=1}^N\sum_{n=1}^N w_{mn} \norm{\z_m-\z_n}^2 - \sum_{n=1}^N \sum_{k=1}^K z_{nk} {G\biggl( \norm{\frac{\x_n-\c_k}{\sigma}}^2 \biggr)} \label{e:lapkmodes1} \\
\mbox{s.t.} &\quad \sum_{k=1}^K z_{nk}=1, \; n=1,\dots,N, \notag \\
&\quad z_{nk}\ge 0, \; n=1,\dots,N,\; k=1,\dots,K. \notag
\end{aligned}$$ We can rewrite this objective in matrix form: $$\begin{aligned}
\label{e:lapkmodes2}
\min_{\Z,\C} &\quad \lambda \trace{\Z^\top \LL \Z} - \trace{\B^\top \Z} \\
\mbox{s.t.} &\quad \Z \1_K = \1_N,\quad \Z \ge \0 \notag\end{aligned}$$ where $\LL = \D - \W$ is the graph Laplacian for the affinity matrix $\W = (w_{nm})$ and degree matrix $\D = \diag{\smash{\sum_{n=1}^N w_{mn}}}$, $\B = (b_{nk})$ is an $N\times K$ matrix containing data-centroid affinities $b_{nk}= \smash{ {G( \norm{(\x_n-\c_k)/{\sigma}}^2 )} }$, $n=1,\dots,N$, $k=1,\dots,K$, $\1_K$ is a $K$ dimensional vector of $1$s and $\ge$ means elementwise comparison. Other variations of the graph Laplacian can also be used (e.g. the normalized Laplacian), see @Luxbur07a. The constraint on $\Z$ shows it is a stochastic matrix. We can obtain a hard clustering if desired by assigning each point to the cluster with highest assignment value.
### Special cases of the hyperparameters $(\lambda,\sigma)$
In Laplacian $K$-modes, in addition to $K$ there are two user parameters: $\lambda$ controls the smoothness of the assignment, and $\sigma$ controls the smoothness of the kde defined on each cluster. Consider first the case of $\lambda=0$, where Laplacian $K$-modes becomes the original $K$-modes algorithm. @CarreirWang13a already noted that the $K$-modes algorithm has two interesting limit cases: it becomes $K$-means when $\sigma\rightarrow\infty$, and a form of $K$-medoids when $\sigma\rightarrow 0$, since the centroids are driven towards data points. In both cases the assignments are hard (1-out-of-$K$ coding). The case when $\lambda\rightarrow\infty$ makes the first term in eq. dominant and forces all connected points to have identical assignments, which is not interesting for the purpose of clustering. Therefore, the most interesting behavior of the algorithm is for intermediate $\lambda$. Finally, another interesting special case of Laplacian $K$-modes corresponds to $\lambda>0$ and $\sigma\rightarrow\infty$, which we call *Laplacian $K$-means*, and which seems to be a new algorithm as well.
Optimization procedure for Laplacian $K$-modes {#s:opt}
----------------------------------------------
To solve , we use alternating optimization over $\C$ and $\Z$, which takes advantage of the problem’s structure.
#### $\C$-step
For fixed $\Z$, we are only concerned with the second term of which is the $K$-modes objective. Therefore, our step over $\C$ is identical to that of $K$-modes: it decouples over clusters and we apply mean-shift to solve for each $\c_k$ separately. The cost of this step is $\calO(KND)$.
#### $\Z$-step
Unlike in $K$-modes, our $\Z$-step no longer decouples, which means we have to solve for $NK$ variables all together. Since the graph Laplacian is positive semidefinite, the problem over is a convex quadratic program (QP). While we could apply a standard QP algorithm, such as an interior point method, we provide here an algorithm that is very simple (no parameters to set), efficient and that scales well to real problems where the number of points $N$ or the number of clusters $K$ is very large. The solution is based on the gradient proximal algorithm used by @BeckTeboul09a. Their general framework solves convex problems of the form $\min\nolimits_\x f(\x) = g(\x)+h(\x)$, where $g$ is convex and has Lipschitz continuous gradient (with constant $L$), and $h$ is convex but not necessarily differentiable. The gradient proximal algorithm iteratively updates the variables by first taking a gradient step of the first function and then projecting it with the second function, i.e., $\x_{\tau+1} = { \arg\min\nolimits_\y \; {\frac{L}{2} \norm{\smash{\y-(\x_{\tau}-\frac{1}{L}\nabla g(\x_{\tau}))}}^2 + h(\y)} }$. It can be proven that the algorithm converges in objective function value with rate $\calO(1/\tau)$ (where $\tau$ is the iteration counter) with a *constant stepsize* $\frac{1}{L}$, and using Nesterov’s acceleration scheme improves the rate to $\calO(1/\tau^2)$.
To apply this framework to our $\Z$-step, we make the identification that $g$ is our smooth quadratic objective function, which has continuous gradient with $L=2\lambda M$ being the (smallest) Lipschitz constant, where $M$ is the largest eigenvalue of $\LL$, and $h$ is the indicator function of the probability simplex. Consequently, our proximal step is computing the Euclidean projection of the gradient step onto the probability simplex. Note that computing the Euclidean projection onto the $K$-dimensional simplex is itself a quadratic program. Fortunately, there exists an efficient algorithm which computes the exact projection with $\calO(K\log K)$ time complexity [@Duchi_08a; @WangCarreir13a].
Initial $\Z_0 \in\RR^{N\times K}$, $s=\frac{1}{2\lambda M}$, $M=$ largest eigenvalue of graph Laplacian $\LL$. Set $\Y_1 = \Z_0,\ t_1 = 1,\ \tau=1$. Compute gradient at $\Y_\tau$: $\G_\tau = 2\lambda\LL\Y_\tau-\B$ $\Z_\tau =$ simplex projection of each row of $\Y_\tau-s\G_\tau$ $t_{\tau+1} = (1+\sqrt{1+4t_\tau^2})/2$ $\Y_{\tau+1} = \Z_\tau + \big(\frac{t_\tau-1}{t_{\tau+1}}\big)(\Z_\tau-\Z_{\tau-1})$ $\tau=\tau+1$ convergence $\Z_\tau$ is the solution of the $\Z$-step.
We provide the accelerated gradient projection algorithm for our $\Z$-step in Algorithm \[alg:gradproj\]. Notice the graph Laplacian is sparse and its largest eigenvalue $M$ can be obtained efficiently (e.g. by power iterations). Therefore the constant stepsize $s$ can be easily determined right after constructing the graph Laplacian. Compared to a pure gradient projection algorithm, the additional computational effort of the acceleration scheme in maintaining an auxiliary sequence (lines 5–6 of Algorithm \[alg:gradproj\]) is minimal, and we clearly observe an improved convergence behavior in experiments.
Each iteration of Algorithm \[alg:gradproj\] costs $\calO(NK\rho+NK\log K)$, where $\rho$ is the neighborhood size in constructing $\LL$ (or the number of nonzero entries in each row). The first term accounts for computing the gradient and the second term accounts for projecting each row of $\Z$ onto the probability simplex. Notice how, although it is solving a large QP, the cost per iteration of our $\Z$-step is independent of the input dimensionality $D$ (in contrast, the $\C$-step has time complexity $\calO(KND)$). Despite its sublinear convergence rate, the algorithm has a clear advantage in its simplicity: it does not require any line search or costly matrix operation, and it is very easy to implement.
### Convergence properties
In the $\C$-step, each mean-shift update increases the density of the cluster kde (or leaves it unchanged) and its convergence rate to a mode is linear in general [@Carreir07a]. In the $\Z$-step, the accelerated gradient projection converges theoretically at $\calO(1/\tau^2)$ rate where $\tau$ is the iteration counter, although this algorithm seems to perform much better than the theoretical guarantee in practice [@BeckTeboul09a]. We alternate the and steps until a convergence criterion is satisfied (e.g. the change to the variables is below some threshold). Notice both steps use iterative procedures, so the number of iterations depends on the convergence accuracy. Let $\epsilon_1$ and $\epsilon_2$ be the optimization precision for the $\C$-step and $\Z$-step, respectively, then one alternation of our algorithm costs $(D\log(1/\epsilon_1)+\rho/\sqrt{\epsilon_2}) KN$. We found empirically that moderate accuracy and few iterations suffice for good clustering results, and our algorithm scales well due to its low per-iteration cost. In an efficient implementation, both steps can be inexact (e.g. each could run for a fixed, small number of iterations). Since the -step algorithm is feasible, exiting it early produces valid assignments.
### Homotopy algorithm
As with $K$-means and $K$-modes, the Laplacian $K$-modes objective function has local optima, which are caused by the nonlinear, kde term. One strategy to find a good optimum consists of first finding a good optimum for $K$-means and then run a homotopy algorithm initialized there. We can construct a homotopy by varying continuously $\lambda$ from $0$ and $\sigma$ from $\infty$, which corresponds to $K$-means, to their target values $(\lambda^*,\sigma^*)$. In practice, we follow this path approximately, by running some iterations of the fixed-($\lambda$,$\sigma$) Laplacian $K$-modes algorithm for each value of ($\lambda$,$\sigma$). As is well known with homotopy techniques, this tends to find better optima than starting directly at the target value $(\lambda^*,\sigma^*)$. A good optimum for $K$-means can be obtained by picking the best of several random restarts, or by using the $K$-means++ initialization strategy, which has approximation guarantees [@ArthurVassil07a].
### Setting the hyperparameters
We believe that, in an unsupervised setting, the user should be able to explore different scenarios and so the algorithm should have a small number of intuitive hyperparameters to control this—rather than automatically guessing, say, the number of clusters, which often is not uniquely defined for a dataset. Laplacian $K$-modes has 3 intuitive hyperparameters, which allow a user to explore different scenarios: more or less clusters ($K$), different scales ($\sigma$), and degree of membership smoothness ($\lambda$). Typical values are around $\lambda=1$, which allows some amount of propagation among neighbors and thus nonconvex clusters, and $\sigma$ obtained from a kde bandwidth formula [@WandJones94a], such as the average distance to the 7th nearest neighbor gives a reasonable density [@ZelnikPerona05a]. (An even better option is to use an adaptive kde, where each point has a different bandwidth, obtained using “entropic affinities” [@HintonRoweis03a; @VladymCarreir13a]. Here, the bandwidth of each data point is computed so as to produce an effective number of neighbors $k$ set by the user.) These hyperparameters values usually produce good clustering results and provide a starting point for improvement. In the homotopy algorithm, the path of $\sigma$ or $\lambda$ values should be followed slowly so we end in a good minimum. In practice, one changes the parameter geometrically in as many steps as one can afford computationally.
In a supervised setting, the hyperparameters can be selected with a validation set using the out-of-sample mapping for Laplacian $K$-modes (section \[s:out-of-sample\]).
Out-of-sample problem {#s:out-of-sample}
---------------------
We now consider the out-of-sample problem, that is, given an unseen test point $\x \in \bbR^D$, we wish to find a meaningful assignment $\z(\x)$ to the clusters found during training. A natural and efficient way to do this is to solve a problem of the same form as with a dataset consisting of the original training set augmented with , but keeping and fixed to the values obtained during training (this avoids having to solve for all points again). After dropping constant terms, this is equivalent to the following problem: $$\begin{aligned}
\min_{\z} &\quad \lambda \sum\nolimits_{n=1}^N w_{n} \norm{\z-\z_n}^2 - \sum\nolimits_{k=1}^K z_{k} {G\biggl( \norm{\frac{\x-\c_k}{\sigma}}^2 \biggr)} \\
\mbox{s.t.} &\quad z_{k}\ge 0,\ k=1,\dots,K, \; \sum\nolimits_{k=1}^K{z_k} = 1\end{aligned}$$ where $w_n$ is the affinity between test point $\x$ and training point $\x_n$. The above problem can be further reduced to the following quadratic program: $$\label{e:obj-osp}
\textstyle
\min_{\z} \quad \frac{1}{2} \norm{\z-(\bar{\z}+\gamma\q)}^2 \;\text{ s.t.}\quad \z^\top \1_K = 1, \; \z \ge \0$$ where the expressions for $\bar{\z}$, $\q = [q_1,\dots,q_K]^\top$ and $\gamma$ are as follows: $$\begin{aligned}
\bar{\z} = \sum_{n=1}^N \frac{w_n}{\sum\limits_{n'=1}^N{w_{n'}}} \z_n, \ q_k = \frac{G ( \norm{(\x-\c_k)/\sigma}^2 )}{\sum\limits_{k'=1}^K G ( \norm{(\x-\c_{k'})/\sigma}^2 ) }, \ \gamma = \frac{ \sum\limits_{k=1}^K{ G ( \norm{(\x-\c_{k})/\sigma}^2 ) } }{2\lambda \sum\limits_{n=1}^N{w_n}}.\end{aligned}$$ Thus, the out-of-sample solution is the projection of the $K$-dimensional vector $\bar{\z}+\gamma\q$ onto the probability simplex. The computational cost is $\calO(ND)$, dominated by the cost of $\bar{\z}$, since the simplex projection costs $\calO(K \log{K})$.
The solution has an intuitive interpretation, consisting of the linear combination of two terms, each a valid assignment vector (having positive elements that sum to $1$). The Laplacian term, $\bar{\z}$, is the weighted average of the neighboring training points’ assignments, and results in nonconvex clusters. The kde term, $\q$, assigns a point based on its distances (posterior probabilities) to the centroids, and results in convex clusters. These two distinct assignment rules are combined using a weight $\gamma$ to give the final assignment. Essentially, $\x$ is assigned to cluster $k$ with high probability if its nearby points are assigned to it ($\bar{z}_k$ is large) or if it is close to $\c_k$ ($q_k$ is large). Although defined variationally, the out-of-sample mapping is just as useful as a closed-form expression: computationally it does not require an iterative procedure, and the interpretation above also illuminates the meaning of the Laplacian in the training objective . In fact, iterating the out-of-sample mapping sequentially over the training points gives another (slower) way to solve the -step, i.e., alternating optimization over $\z_1,\dots,\z_N$.
Finally, Table \[t:clustercomp\] compares Laplacian $K$-modes with other popular clustering algorithms (see section \[s:opt\] for the complexity analysis).
$K$-means $K$-medoids Mean-shift $K$-modes
-------------------- ----------- ------------- ------------ --------------- ----------- -------
Centroids “valid” “valid” N/A valid valid
Nonconvex clusters no depends yes yes no yes
Density no no yes no yes yes
Assignment hard hard hard hard hard soft
Cost per iteration $KND$ $KN^2D$ $N^2D$ $N^2\sim N^3$ $KND$
: Comparison of properties of different clustering algorithms.[]{data-label="t:clustercomp"}
Experiments {#s:expts}
===========
Illustrative experiments
------------------------
### Spirals dataset
\[r\]\[r\]
-------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- ----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- ---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
$K$-modes
![Synthetic dataset of $5$-spirals. From left to right: $K$-modes clustering ($\lambda=0$, $\sigma=0.2$, the circle at the top right corner has a radius of $\sigma$); Laplacian $K$-modes clustering ($\lambda=100$, $\sigma=0.2$); Laplacian $K$-modes assignment probabilities; contours of the kde of the “red” cluster.[]{data-label="f:5spirals"}](5spirals_kmodes.eps "fig:"){width="0.25\linewidth"} ![Synthetic dataset of $5$-spirals. From left to right: $K$-modes clustering ($\lambda=0$, $\sigma=0.2$, the circle at the top right corner has a radius of $\sigma$); Laplacian $K$-modes clustering ($\lambda=100$, $\sigma=0.2$); Laplacian $K$-modes assignment probabilities; contours of the kde of the “red” cluster.[]{data-label="f:5spirals"}](5spirals_Lapkmodes.eps "fig:"){width="0.25\linewidth"} ![Synthetic dataset of $5$-spirals. From left to right: $K$-modes clustering ($\lambda=0$, $\sigma=0.2$, the circle at the top right corner has a radius of $\sigma$); Laplacian $K$-modes clustering ($\lambda=100$, $\sigma=0.2$); Laplacian $K$-modes assignment probabilities; contours of the kde of the “red” cluster.[]{data-label="f:5spirals"}](5spirals_Lapkmodes_P.eps "fig:"){width="0.25\linewidth"} ![Synthetic dataset of $5$-spirals. From left to right: $K$-modes clustering ($\lambda=0$, $\sigma=0.2$, the circle at the top right corner has a radius of $\sigma$); Laplacian $K$-modes clustering ($\lambda=100$, $\sigma=0.2$); Laplacian $K$-modes assignment probabilities; contours of the kde of the “red” cluster.[]{data-label="f:5spirals"}](5spirals_Lapkmodes_kde.eps "fig:"){width="0.25\linewidth"}
-------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- ----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- ---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
We first demonstrate the power of Laplacian smoothing. The 2D dataset in Fig. \[f:5spirals\] consist of $5$ spirals where each spiral contains $400$ points (denoted by $\circ$). The natural way of partitioning this dataset into $K=5$ groups is to assign points of each spiral into a separate cluster. Due to the nonconvex shape of the spirals, the ideal result can not be possibly achieved by $K$-modes (plot 1, we color each point differently according to the cluster it is assigned to) or $K$-means (not shown, result similar to $K$-modes), even though the $K$-modes centroids (denoted by $\blacktriangleright$) are lying on each spiral and are valid representatives of the dataset. We then build a $5$-nearest-neighbor graph on this dataset using heat kernel weighting and run Laplacian $K$-modes using the $K$-means result as initialization. We achieve a perfect separation of the spirals and one valid centroid for each spiral in a few steps of our alternating optimization scheme, as shown in plot 2. We show the assignment probabilities $\Z$ in plot 3, where each data point $\x_n$ is colored using a mixture of the 5 clusters’ colors with its assignment probability $\z_n$ being the mixing coefficient. We show the contours of the kde defined on the “red” cluster in plot 4. In spite of the nonconvex, 1D manifold nature of the cluster, the kde is localized to the cluster and represents its shape and density well. This illustrates why using soft assignments gives more flexibility: the weights in the kde of each point vary, which allows for a more flexible kde. It is obvious that running mean-shift on this dataset with the same $\sigma$ will result in a large number of modes and therefore clusters. In contrast, the number of modes is fixed in Laplacian $K$-modes and the algorithm will track one of the major modes in each cluster.
It is interesting to notice that because the kernel width $\sigma$ we use is quite small, only a small proportion of data points are close enough to centroids to have nonzero affinity. This implies that the $\B$ matrix in (5) of the main paper is quite sparse. Nonetheless, we achieve good assignment probabilities using the graph Laplacian, which propagates the sparse “label” information in $\B$ throughout the graph. This also partly explains the success of Laplacian smoothing in spectral clustering [@ShiMalik00a] and semi-supervised learning algorithms [@Zhu_03a; @Belkin_06a].
### Noisy two moons
----------------------------------------------------- ------------------------------------------------------------ ---------------------------------------------------------- ----------------------------------------------------
$K$-means
{width="0.25\linewidth"} {width="0.25\linewidth"} {width="0.25\linewidth"} {width="0.25\linewidth"}
----------------------------------------------------- ------------------------------------------------------------ ---------------------------------------------------------- ----------------------------------------------------
We demonstrate Laplacian $K$-modes on the “two-moons” dataset in Fig. \[f:2moons\]. The dataset has two nonconvex, interleaved clusters (each has $400$ points) and we add many outliers ($200$ points) around them. The “moons” cannot be perfectly separated by either $K$-means (results shown in plot 1) or $K$-modes, since both define Voronoi tessellations. This problem is also difficult for hierarchical clustering because, as is well known, its major problem is that it creates connections between different clusters as the merging occurs. We build a $5$-nearest-neighbor graph on this dataset using heat kernel weighting, and run Laplacian $K$-modes from the $K$-means initialization. We run the homotopy version and reduce $\sigma$ from $5$ to $0.1$ in $10$ steps while fixing $\lambda=1$. The hard partition obtained, along with the two centroids and kde’s for each cluster at $\sigma=0.1$ are given in plot 2. Even with the heavy noise and outliers, the “inliers” are still perfectly separated, the modes lie in high density areas and we obtain a good density estimate for each cluster. We show the assignment probabilities in plot 3, colored using the same scheme as in the spirals example, where each data point $\x_n$ is colored using a mixture of the clusters’ colors (red or blue) with its assignment probability $\z_n$ being the mixing coefficient. The assignment is certain near the centroids (purer color) and less crisp at boundaries and outliers (mixed color). Finally, the out-of-sample mapping in input space is shown in plot 4, where we compute out-of-sample assignments for a fine grid and color each grid point using the same scheme as in plot 3. We see clearly that the assignment rule is very different from the hard assignment of $K$-means. The mapping at each point combines the average assignment of nearby training points and the assignment from centroids, being able to model a complex shape.
---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- ----------------------------------------------------------------------------------------- ---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- ---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
Original image Classification error (%) Normalized cut Laplacian $K$-modes
![Occluder segmentation result: original image; classification errors over the range of $\sigma$; segmentations of normalized cut and Laplacian $K$-modes at $\sigma=0.2$.[]{data-label="f:occluder"}](random.213440.eps "fig:"){width="0.24\linewidth"} \[\]\[\][$\sigma$]{} \[\]\[\] \[l\]\[l\][Normalized cut]{} \[l\]\[l\][Lap. $K$-modes]{} ![Occluder segmentation result: original image; classification errors over the range of $\sigma$; segmentations of normalized cut and Laplacian $K$-modes at $\sigma=0.2$.[]{data-label="f:occluder"}](occluder_ncut.eps "fig:"){width="0.23\linewidth"} ![Occluder segmentation result: original image; classification errors over the range of $\sigma$; segmentations of normalized cut and Laplacian $K$-modes at $\sigma=0.2$.[]{data-label="f:occluder"}](occluder_lapkmodes.eps "fig:"){width="0.23\linewidth"}
---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- ----------------------------------------------------------------------------------------- ---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- ---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
Figure-ground segmentation
--------------------------
We consider the problem of segmenting an occluder from a textured background in a grayscale image. This problem has been shown to be difficult for spectral clustering [@ChennubJepson03a; @CarreirZemel05a], because of the intensity gradients between the occluder and the background (and within the background itself), which cause many graph edges to connect them, see the example in fig. \[f:occluder\]. We formalize it as a clustering problem and partition the pixels into $K=5$ clusters. We use for each pixel its 2D location and intensity value as features, and build a graph where each pixel is connected to the eight nearby pixels, with edges weighted using a heat kernel of width $\sigma$ ($\sigma$’s value equals that of the kde bandwidth for Laplacian $K$-modes). The goal is to have one of the clusters extract the occluder, which can then be separated from the background. To measure the performance, we choose the cluster that overlaps most with the occluder as positive prediction (the rest of the pixels are considered as background/negative prediction) and compute the classification error. Fig. \[f:occluder\] shows normalized cut [@YuShi03a] performs well for a narrow range of $\sigma$, while Laplacian $K$-modes (with fixed $\lambda=0.1$) has a much more stable performance when using the same graph: the range of good $\sigma$ values that produce a perfect segmentation is much wider. This shows the difference between our algorithm and spectral clustering: even though both algorithms impose smoothness on assignments, the graph Laplacian is only a regularization term in our model, and the kde term makes our algorithm more robust to the graph construction. As a more powerful algorithm, Laplacian $K$-modes has inherited the robustness properties from the original $K$-modes algorithm [@CarreirWang13a].
Clustering analysis
-------------------
dataset $N$ $D$ $K$
--------- ------ ------- -----
MNIST 2000 784 10
COIL–20 1440 1024 20
TDT2 9394 36771 30
: Statistics (size, dimensionality, \# of classes) of three real world datasets.[]{data-label="t:datasets"}
[@c@|c||c|c|c|c|c|c|c|c|@]{} & dataset & $K$-means & $K$-modes & GMS & NCut & GNMF & DCD & Laplacian $K$-modes\
& MNIST & 58.2 & 59.2 & 15.9 & 65.5 & 66.2 & 69.4 & **70.5**\
ACC \[0pt\]\[0pt\][$\Bigg\{\Bigg.$]{} & COIL–20& 66.5 & 67.2 & 27.2 & 79.0 & 75.3 & 71.5 & **81.0** **(81.5)**\
& TDT2 & 68.9 & 70.0 & N/A & 88.4 & 88.6 & 55.1 & **91.4**\
& MNIST & 53.3 & 53.6 & 6.51 & 66.9 & 64.9 & 65.6 & **68.8**\
NMI \[0pt\]\[0pt\][$\Bigg\{\Bigg.$]{} & COIL–20& 75.3 & 75.9 & 38.9 & **88.0** & 87.5 & 77.6 & 87.3 **(88.0)**\
& TDT2 & 75.3 & 75.8 & N/A & 83.7 & 83.7 & 68.6 & **88.8**\
\[t:expt-acc-nmi\]
We report clustering statistics in datasets with known pattern class labels (which the algorithms did not use): (1) MNIST [@Lecun_98a], which contains $28 \times 28$ grayscale handwritten digit images (we randomly sample 200 of each digit); (2) COIL–20 [@Nene_96a], which contains $32\times 32$ grayscale images of 20 objects viewed from varying angles; (3) the NIST Topic Detection and Tracking (TDT2) corpus, which contains on-topic documents of different semantic categories (documents appearing in more than one category are removed and only the largest 30 categories are kept). Statistics of the datasets are collected in table \[t:datasets\]. Datasets (2) and (3) are the same as used by @Cai_11b, and we also use the same features: pixel values for (1) and (2), and TFIDF for (3).
We compare the following algorithms: $K$-means, initialized randomly; $K$-modes, a special case of Laplacian $K$-modes with $\lambda=0$; Gaussian mean-shift (GMS), we search for $\sigma$ that produces exactly $K$ modes; Normalized cut (NCut), one typical spectral clustering algorithm, we use the implementation of @YuShi03a; Graph regularized NMF (GNMF) proposed by @Cai_11b; Data-Cluster-Data random walk (DCD) proposed by @YangOja12a; and Laplacian $K$-modes, initialized from $K$-means. $K$ is set to the number of classes in the ground truth. All the datasets are normalized to have unit norm per sample.
Several algorithms use the graph Laplacian: for NCut, GNMF, and Laplacian $K$-modes, we build a 5-nearest-neighbor graph and use a binary weighting scheme to compute the graph Laplacian (as in [@Cai_11b]); for DCD, we find that it achieves better performance using a graph built with a larger neighborhood size, so we let DCD select its optimal size in $\{5,10,20,30\}$. We run each algorithm with $20$ random restarts, letting them use optimal values for their respective hyperparameters (if they have any) based on a grid search, and report the best performance from different random restarts. Notice we do not use the out-of-sample mapping here because its does not exist for all algorithms we compare with. Clustering accuracy (ACC) and normalized mutual information (NMI), two widely used criteria [@Cai_11b; @Arora_11a], are used for evaluation. The results are given in table \[t:expt-acc-nmi\] (N/A means our GMS code ran out of memory).
It is clear that algorithms using Laplacian smoothing are in general superior than algorithms not using it, which demonstrates the importance of the graph Laplacian in separating nonconvex and manifold clusters. GMS performs poorly for the reasons described earlier. On all datasets, Laplacian $K$-modes achieves the best or close to best performance under both criteria. We find there exists a wide range of hyperparameters with which our algorithm gives very competitive performance. We are able to further improve our performance on COIL–20 using the homotopy technique described earlier: we fix $\lambda$ at $0.01$, and decrease $\sigma$ from $0.45$ to $0.1$ gradually in $7$ steps, initializing the algorithm for the current $\sigma$ value from the solution for the previous $\sigma$ value. This improved result is shown in parenthesis in table \[t:expt-acc-nmi\].
---------- ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
![Centroids found by different algorithms on MNIST. *First row*: $K$-means ($\lambda=0$, $\sigma\rightarrow\infty$, ACC: 55.2%, NMI: 50.2%). *Second row*: $K$-modes ($\lambda=0$, $\sigma=0.35$, ACC: 56.0%, NMI: 50.6%). *Third row*: Laplacian $K$-modes ($\lambda=0.07$, $\sigma=0.35$, ACC: 70.5%, NMI: 68.8%). *Fourth row*: mean-shift ($\sigma=0.2485$).[]{data-label="f:mnist"}](MNIST/MNIST_centers_kmeans_01.eps "fig:"){width="0.095\linewidth"} ![Centroids found by different algorithms on MNIST. *First row*: $K$-means ($\lambda=0$, $\sigma\rightarrow\infty$, ACC: 55.2%, NMI: 50.2%). *Second row*: $K$-modes ($\lambda=0$, $\sigma=0.35$, ACC: 56.0%, NMI: 50.6%). *Third row*: Laplacian $K$-modes ($\lambda=0.07$, $\sigma=0.35$, ACC: 70.5%, NMI: 68.8%). *Fourth row*: mean-shift ($\sigma=0.2485$).[]{data-label="f:mnist"}](MNIST/MNIST_centers_kmeans_02.eps "fig:"){width="0.095\linewidth"} ![Centroids found by different algorithms on MNIST. *First row*: $K$-means ($\lambda=0$, $\sigma\rightarrow\infty$, ACC: 55.2%, NMI: 50.2%). *Second row*: $K$-modes ($\lambda=0$, $\sigma=0.35$, ACC: 56.0%, NMI: 50.6%). *Third row*: Laplacian $K$-modes ($\lambda=0.07$, $\sigma=0.35$, ACC: 70.5%, NMI: 68.8%). *Fourth row*: mean-shift ($\sigma=0.2485$).[]{data-label="f:mnist"}](MNIST/MNIST_centers_kmeans_03.eps "fig:"){width="0.095\linewidth"} ![Centroids found by different algorithms on MNIST. *First row*: $K$-means ($\lambda=0$, $\sigma\rightarrow\infty$, ACC: 55.2%, NMI: 50.2%). *Second row*: $K$-modes ($\lambda=0$, $\sigma=0.35$, ACC: 56.0%, NMI: 50.6%). *Third row*: Laplacian $K$-modes ($\lambda=0.07$, $\sigma=0.35$, ACC: 70.5%, NMI: 68.8%). *Fourth row*: mean-shift ($\sigma=0.2485$).[]{data-label="f:mnist"}](MNIST/MNIST_centers_kmeans_04.eps "fig:"){width="0.095\linewidth"} ![Centroids found by different algorithms on MNIST. *First row*: $K$-means ($\lambda=0$, $\sigma\rightarrow\infty$, ACC: 55.2%, NMI: 50.2%). *Second row*: $K$-modes ($\lambda=0$, $\sigma=0.35$, ACC: 56.0%, NMI: 50.6%). *Third row*: Laplacian $K$-modes ($\lambda=0.07$, $\sigma=0.35$, ACC: 70.5%, NMI: 68.8%). *Fourth row*: mean-shift ($\sigma=0.2485$).[]{data-label="f:mnist"}](MNIST/MNIST_centers_kmeans_05.eps "fig:"){width="0.095\linewidth"} ![Centroids found by different algorithms on MNIST. *First row*: $K$-means ($\lambda=0$, $\sigma\rightarrow\infty$, ACC: 55.2%, NMI: 50.2%). *Second row*: $K$-modes ($\lambda=0$, $\sigma=0.35$, ACC: 56.0%, NMI: 50.6%). *Third row*: Laplacian $K$-modes ($\lambda=0.07$, $\sigma=0.35$, ACC: 70.5%, NMI: 68.8%). *Fourth row*: mean-shift ($\sigma=0.2485$).[]{data-label="f:mnist"}](MNIST/MNIST_centers_kmeans_06.eps "fig:"){width="0.095\linewidth"} ![Centroids found by different algorithms on MNIST. *First row*: $K$-means ($\lambda=0$, $\sigma\rightarrow\infty$, ACC: 55.2%, NMI: 50.2%). *Second row*: $K$-modes ($\lambda=0$, $\sigma=0.35$, ACC: 56.0%, NMI: 50.6%). *Third row*: Laplacian $K$-modes ($\lambda=0.07$, $\sigma=0.35$, ACC: 70.5%, NMI: 68.8%). *Fourth row*: mean-shift ($\sigma=0.2485$).[]{data-label="f:mnist"}](MNIST/MNIST_centers_kmeans_07.eps "fig:"){width="0.095\linewidth"} ![Centroids found by different algorithms on MNIST. *First row*: $K$-means ($\lambda=0$, $\sigma\rightarrow\infty$, ACC: 55.2%, NMI: 50.2%). *Second row*: $K$-modes ($\lambda=0$, $\sigma=0.35$, ACC: 56.0%, NMI: 50.6%). *Third row*: Laplacian $K$-modes ($\lambda=0.07$, $\sigma=0.35$, ACC: 70.5%, NMI: 68.8%). *Fourth row*: mean-shift ($\sigma=0.2485$).[]{data-label="f:mnist"}](MNIST/MNIST_centers_kmeans_08.eps "fig:"){width="0.095\linewidth"} ![Centroids found by different algorithms on MNIST. *First row*: $K$-means ($\lambda=0$, $\sigma\rightarrow\infty$, ACC: 55.2%, NMI: 50.2%). *Second row*: $K$-modes ($\lambda=0$, $\sigma=0.35$, ACC: 56.0%, NMI: 50.6%). *Third row*: Laplacian $K$-modes ($\lambda=0.07$, $\sigma=0.35$, ACC: 70.5%, NMI: 68.8%). *Fourth row*: mean-shift ($\sigma=0.2485$).[]{data-label="f:mnist"}](MNIST/MNIST_centers_kmeans_09.eps "fig:"){width="0.095\linewidth"} ![Centroids found by different algorithms on MNIST. *First row*: $K$-means ($\lambda=0$, $\sigma\rightarrow\infty$, ACC: 55.2%, NMI: 50.2%). *Second row*: $K$-modes ($\lambda=0$, $\sigma=0.35$, ACC: 56.0%, NMI: 50.6%). *Third row*: Laplacian $K$-modes ($\lambda=0.07$, $\sigma=0.35$, ACC: 70.5%, NMI: 68.8%). *Fourth row*: mean-shift ($\sigma=0.2485$).[]{data-label="f:mnist"}](MNIST/MNIST_centers_kmeans_10.eps "fig:"){width="0.095\linewidth"}
\[-1ex\] ![Centroids found by different algorithms on MNIST. *First row*: $K$-means ($\lambda=0$, $\sigma\rightarrow\infty$, ACC: 55.2%, NMI: 50.2%). *Second row*: $K$-modes ($\lambda=0$, $\sigma=0.35$, ACC: 56.0%, NMI: 50.6%). *Third row*: Laplacian $K$-modes ($\lambda=0.07$, $\sigma=0.35$, ACC: 70.5%, NMI: 68.8%). *Fourth row*: mean-shift ($\sigma=0.2485$).[]{data-label="f:mnist"}](MNIST/MNIST_centers_kmodes_01.eps "fig:"){width="0.095\linewidth"} ![Centroids found by different algorithms on MNIST. *First row*: $K$-means ($\lambda=0$, $\sigma\rightarrow\infty$, ACC: 55.2%, NMI: 50.2%). *Second row*: $K$-modes ($\lambda=0$, $\sigma=0.35$, ACC: 56.0%, NMI: 50.6%). *Third row*: Laplacian $K$-modes ($\lambda=0.07$, $\sigma=0.35$, ACC: 70.5%, NMI: 68.8%). *Fourth row*: mean-shift ($\sigma=0.2485$).[]{data-label="f:mnist"}](MNIST/MNIST_centers_kmodes_02.eps "fig:"){width="0.095\linewidth"} ![Centroids found by different algorithms on MNIST. *First row*: $K$-means ($\lambda=0$, $\sigma\rightarrow\infty$, ACC: 55.2%, NMI: 50.2%). *Second row*: $K$-modes ($\lambda=0$, $\sigma=0.35$, ACC: 56.0%, NMI: 50.6%). *Third row*: Laplacian $K$-modes ($\lambda=0.07$, $\sigma=0.35$, ACC: 70.5%, NMI: 68.8%). *Fourth row*: mean-shift ($\sigma=0.2485$).[]{data-label="f:mnist"}](MNIST/MNIST_centers_kmodes_03.eps "fig:"){width="0.095\linewidth"} ![Centroids found by different algorithms on MNIST. *First row*: $K$-means ($\lambda=0$, $\sigma\rightarrow\infty$, ACC: 55.2%, NMI: 50.2%). *Second row*: $K$-modes ($\lambda=0$, $\sigma=0.35$, ACC: 56.0%, NMI: 50.6%). *Third row*: Laplacian $K$-modes ($\lambda=0.07$, $\sigma=0.35$, ACC: 70.5%, NMI: 68.8%). *Fourth row*: mean-shift ($\sigma=0.2485$).[]{data-label="f:mnist"}](MNIST/MNIST_centers_kmodes_04.eps "fig:"){width="0.095\linewidth"} ![Centroids found by different algorithms on MNIST. *First row*: $K$-means ($\lambda=0$, $\sigma\rightarrow\infty$, ACC: 55.2%, NMI: 50.2%). *Second row*: $K$-modes ($\lambda=0$, $\sigma=0.35$, ACC: 56.0%, NMI: 50.6%). *Third row*: Laplacian $K$-modes ($\lambda=0.07$, $\sigma=0.35$, ACC: 70.5%, NMI: 68.8%). *Fourth row*: mean-shift ($\sigma=0.2485$).[]{data-label="f:mnist"}](MNIST/MNIST_centers_kmodes_05.eps "fig:"){width="0.095\linewidth"} ![Centroids found by different algorithms on MNIST. *First row*: $K$-means ($\lambda=0$, $\sigma\rightarrow\infty$, ACC: 55.2%, NMI: 50.2%). *Second row*: $K$-modes ($\lambda=0$, $\sigma=0.35$, ACC: 56.0%, NMI: 50.6%). *Third row*: Laplacian $K$-modes ($\lambda=0.07$, $\sigma=0.35$, ACC: 70.5%, NMI: 68.8%). *Fourth row*: mean-shift ($\sigma=0.2485$).[]{data-label="f:mnist"}](MNIST/MNIST_centers_kmodes_06.eps "fig:"){width="0.095\linewidth"} ![Centroids found by different algorithms on MNIST. *First row*: $K$-means ($\lambda=0$, $\sigma\rightarrow\infty$, ACC: 55.2%, NMI: 50.2%). *Second row*: $K$-modes ($\lambda=0$, $\sigma=0.35$, ACC: 56.0%, NMI: 50.6%). *Third row*: Laplacian $K$-modes ($\lambda=0.07$, $\sigma=0.35$, ACC: 70.5%, NMI: 68.8%). *Fourth row*: mean-shift ($\sigma=0.2485$).[]{data-label="f:mnist"}](MNIST/MNIST_centers_kmodes_07.eps "fig:"){width="0.095\linewidth"} ![Centroids found by different algorithms on MNIST. *First row*: $K$-means ($\lambda=0$, $\sigma\rightarrow\infty$, ACC: 55.2%, NMI: 50.2%). *Second row*: $K$-modes ($\lambda=0$, $\sigma=0.35$, ACC: 56.0%, NMI: 50.6%). *Third row*: Laplacian $K$-modes ($\lambda=0.07$, $\sigma=0.35$, ACC: 70.5%, NMI: 68.8%). *Fourth row*: mean-shift ($\sigma=0.2485$).[]{data-label="f:mnist"}](MNIST/MNIST_centers_kmodes_08.eps "fig:"){width="0.095\linewidth"} ![Centroids found by different algorithms on MNIST. *First row*: $K$-means ($\lambda=0$, $\sigma\rightarrow\infty$, ACC: 55.2%, NMI: 50.2%). *Second row*: $K$-modes ($\lambda=0$, $\sigma=0.35$, ACC: 56.0%, NMI: 50.6%). *Third row*: Laplacian $K$-modes ($\lambda=0.07$, $\sigma=0.35$, ACC: 70.5%, NMI: 68.8%). *Fourth row*: mean-shift ($\sigma=0.2485$).[]{data-label="f:mnist"}](MNIST/MNIST_centers_kmodes_09.eps "fig:"){width="0.095\linewidth"} ![Centroids found by different algorithms on MNIST. *First row*: $K$-means ($\lambda=0$, $\sigma\rightarrow\infty$, ACC: 55.2%, NMI: 50.2%). *Second row*: $K$-modes ($\lambda=0$, $\sigma=0.35$, ACC: 56.0%, NMI: 50.6%). *Third row*: Laplacian $K$-modes ($\lambda=0.07$, $\sigma=0.35$, ACC: 70.5%, NMI: 68.8%). *Fourth row*: mean-shift ($\sigma=0.2485$).[]{data-label="f:mnist"}](MNIST/MNIST_centers_kmodes_10.eps "fig:"){width="0.095\linewidth"}
\[-1ex\] ![Centroids found by different algorithms on MNIST. *First row*: $K$-means ($\lambda=0$, $\sigma\rightarrow\infty$, ACC: 55.2%, NMI: 50.2%). *Second row*: $K$-modes ($\lambda=0$, $\sigma=0.35$, ACC: 56.0%, NMI: 50.6%). *Third row*: Laplacian $K$-modes ($\lambda=0.07$, $\sigma=0.35$, ACC: 70.5%, NMI: 68.8%). *Fourth row*: mean-shift ($\sigma=0.2485$).[]{data-label="f:mnist"}](MNIST/MNIST_centers_manikmodes_01.eps "fig:"){width="0.095\linewidth"} ![Centroids found by different algorithms on MNIST. *First row*: $K$-means ($\lambda=0$, $\sigma\rightarrow\infty$, ACC: 55.2%, NMI: 50.2%). *Second row*: $K$-modes ($\lambda=0$, $\sigma=0.35$, ACC: 56.0%, NMI: 50.6%). *Third row*: Laplacian $K$-modes ($\lambda=0.07$, $\sigma=0.35$, ACC: 70.5%, NMI: 68.8%). *Fourth row*: mean-shift ($\sigma=0.2485$).[]{data-label="f:mnist"}](MNIST/MNIST_centers_manikmodes_02.eps "fig:"){width="0.095\linewidth"} ![Centroids found by different algorithms on MNIST. *First row*: $K$-means ($\lambda=0$, $\sigma\rightarrow\infty$, ACC: 55.2%, NMI: 50.2%). *Second row*: $K$-modes ($\lambda=0$, $\sigma=0.35$, ACC: 56.0%, NMI: 50.6%). *Third row*: Laplacian $K$-modes ($\lambda=0.07$, $\sigma=0.35$, ACC: 70.5%, NMI: 68.8%). *Fourth row*: mean-shift ($\sigma=0.2485$).[]{data-label="f:mnist"}](MNIST/MNIST_centers_manikmodes_03.eps "fig:"){width="0.095\linewidth"} ![Centroids found by different algorithms on MNIST. *First row*: $K$-means ($\lambda=0$, $\sigma\rightarrow\infty$, ACC: 55.2%, NMI: 50.2%). *Second row*: $K$-modes ($\lambda=0$, $\sigma=0.35$, ACC: 56.0%, NMI: 50.6%). *Third row*: Laplacian $K$-modes ($\lambda=0.07$, $\sigma=0.35$, ACC: 70.5%, NMI: 68.8%). *Fourth row*: mean-shift ($\sigma=0.2485$).[]{data-label="f:mnist"}](MNIST/MNIST_centers_manikmodes_04.eps "fig:"){width="0.095\linewidth"} ![Centroids found by different algorithms on MNIST. *First row*: $K$-means ($\lambda=0$, $\sigma\rightarrow\infty$, ACC: 55.2%, NMI: 50.2%). *Second row*: $K$-modes ($\lambda=0$, $\sigma=0.35$, ACC: 56.0%, NMI: 50.6%). *Third row*: Laplacian $K$-modes ($\lambda=0.07$, $\sigma=0.35$, ACC: 70.5%, NMI: 68.8%). *Fourth row*: mean-shift ($\sigma=0.2485$).[]{data-label="f:mnist"}](MNIST/MNIST_centers_manikmodes_05.eps "fig:"){width="0.095\linewidth"} ![Centroids found by different algorithms on MNIST. *First row*: $K$-means ($\lambda=0$, $\sigma\rightarrow\infty$, ACC: 55.2%, NMI: 50.2%). *Second row*: $K$-modes ($\lambda=0$, $\sigma=0.35$, ACC: 56.0%, NMI: 50.6%). *Third row*: Laplacian $K$-modes ($\lambda=0.07$, $\sigma=0.35$, ACC: 70.5%, NMI: 68.8%). *Fourth row*: mean-shift ($\sigma=0.2485$).[]{data-label="f:mnist"}](MNIST/MNIST_centers_manikmodes_06.eps "fig:"){width="0.095\linewidth"} ![Centroids found by different algorithms on MNIST. *First row*: $K$-means ($\lambda=0$, $\sigma\rightarrow\infty$, ACC: 55.2%, NMI: 50.2%). *Second row*: $K$-modes ($\lambda=0$, $\sigma=0.35$, ACC: 56.0%, NMI: 50.6%). *Third row*: Laplacian $K$-modes ($\lambda=0.07$, $\sigma=0.35$, ACC: 70.5%, NMI: 68.8%). *Fourth row*: mean-shift ($\sigma=0.2485$).[]{data-label="f:mnist"}](MNIST/MNIST_centers_manikmodes_07.eps "fig:"){width="0.095\linewidth"} ![Centroids found by different algorithms on MNIST. *First row*: $K$-means ($\lambda=0$, $\sigma\rightarrow\infty$, ACC: 55.2%, NMI: 50.2%). *Second row*: $K$-modes ($\lambda=0$, $\sigma=0.35$, ACC: 56.0%, NMI: 50.6%). *Third row*: Laplacian $K$-modes ($\lambda=0.07$, $\sigma=0.35$, ACC: 70.5%, NMI: 68.8%). *Fourth row*: mean-shift ($\sigma=0.2485$).[]{data-label="f:mnist"}](MNIST/MNIST_centers_manikmodes_08.eps "fig:"){width="0.095\linewidth"} ![Centroids found by different algorithms on MNIST. *First row*: $K$-means ($\lambda=0$, $\sigma\rightarrow\infty$, ACC: 55.2%, NMI: 50.2%). *Second row*: $K$-modes ($\lambda=0$, $\sigma=0.35$, ACC: 56.0%, NMI: 50.6%). *Third row*: Laplacian $K$-modes ($\lambda=0.07$, $\sigma=0.35$, ACC: 70.5%, NMI: 68.8%). *Fourth row*: mean-shift ($\sigma=0.2485$).[]{data-label="f:mnist"}](MNIST/MNIST_centers_manikmodes_09.eps "fig:"){width="0.095\linewidth"} ![Centroids found by different algorithms on MNIST. *First row*: $K$-means ($\lambda=0$, $\sigma\rightarrow\infty$, ACC: 55.2%, NMI: 50.2%). *Second row*: $K$-modes ($\lambda=0$, $\sigma=0.35$, ACC: 56.0%, NMI: 50.6%). *Third row*: Laplacian $K$-modes ($\lambda=0.07$, $\sigma=0.35$, ACC: 70.5%, NMI: 68.8%). *Fourth row*: mean-shift ($\sigma=0.2485$).[]{data-label="f:mnist"}](MNIST/MNIST_centers_manikmodes_10.eps "fig:"){width="0.095\linewidth"}
\[-1ex\] ![Centroids found by different algorithms on MNIST. *First row*: $K$-means ($\lambda=0$, $\sigma\rightarrow\infty$, ACC: 55.2%, NMI: 50.2%). *Second row*: $K$-modes ($\lambda=0$, $\sigma=0.35$, ACC: 56.0%, NMI: 50.6%). *Third row*: Laplacian $K$-modes ($\lambda=0.07$, $\sigma=0.35$, ACC: 70.5%, NMI: 68.8%). *Fourth row*: mean-shift ($\sigma=0.2485$).[]{data-label="f:mnist"}](MNIST/MNIST_centers_ms_01.eps "fig:"){width="0.095\linewidth"} ![Centroids found by different algorithms on MNIST. *First row*: $K$-means ($\lambda=0$, $\sigma\rightarrow\infty$, ACC: 55.2%, NMI: 50.2%). *Second row*: $K$-modes ($\lambda=0$, $\sigma=0.35$, ACC: 56.0%, NMI: 50.6%). *Third row*: Laplacian $K$-modes ($\lambda=0.07$, $\sigma=0.35$, ACC: 70.5%, NMI: 68.8%). *Fourth row*: mean-shift ($\sigma=0.2485$).[]{data-label="f:mnist"}](MNIST/MNIST_centers_ms_02.eps "fig:"){width="0.095\linewidth"} ![Centroids found by different algorithms on MNIST. *First row*: $K$-means ($\lambda=0$, $\sigma\rightarrow\infty$, ACC: 55.2%, NMI: 50.2%). *Second row*: $K$-modes ($\lambda=0$, $\sigma=0.35$, ACC: 56.0%, NMI: 50.6%). *Third row*: Laplacian $K$-modes ($\lambda=0.07$, $\sigma=0.35$, ACC: 70.5%, NMI: 68.8%). *Fourth row*: mean-shift ($\sigma=0.2485$).[]{data-label="f:mnist"}](MNIST/MNIST_centers_ms_03.eps "fig:"){width="0.095\linewidth"} ![Centroids found by different algorithms on MNIST. *First row*: $K$-means ($\lambda=0$, $\sigma\rightarrow\infty$, ACC: 55.2%, NMI: 50.2%). *Second row*: $K$-modes ($\lambda=0$, $\sigma=0.35$, ACC: 56.0%, NMI: 50.6%). *Third row*: Laplacian $K$-modes ($\lambda=0.07$, $\sigma=0.35$, ACC: 70.5%, NMI: 68.8%). *Fourth row*: mean-shift ($\sigma=0.2485$).[]{data-label="f:mnist"}](MNIST/MNIST_centers_ms_04.eps "fig:"){width="0.095\linewidth"} ![Centroids found by different algorithms on MNIST. *First row*: $K$-means ($\lambda=0$, $\sigma\rightarrow\infty$, ACC: 55.2%, NMI: 50.2%). *Second row*: $K$-modes ($\lambda=0$, $\sigma=0.35$, ACC: 56.0%, NMI: 50.6%). *Third row*: Laplacian $K$-modes ($\lambda=0.07$, $\sigma=0.35$, ACC: 70.5%, NMI: 68.8%). *Fourth row*: mean-shift ($\sigma=0.2485$).[]{data-label="f:mnist"}](MNIST/MNIST_centers_ms_05.eps "fig:"){width="0.095\linewidth"} ![Centroids found by different algorithms on MNIST. *First row*: $K$-means ($\lambda=0$, $\sigma\rightarrow\infty$, ACC: 55.2%, NMI: 50.2%). *Second row*: $K$-modes ($\lambda=0$, $\sigma=0.35$, ACC: 56.0%, NMI: 50.6%). *Third row*: Laplacian $K$-modes ($\lambda=0.07$, $\sigma=0.35$, ACC: 70.5%, NMI: 68.8%). *Fourth row*: mean-shift ($\sigma=0.2485$).[]{data-label="f:mnist"}](MNIST/MNIST_centers_ms_06.eps "fig:"){width="0.095\linewidth"} ![Centroids found by different algorithms on MNIST. *First row*: $K$-means ($\lambda=0$, $\sigma\rightarrow\infty$, ACC: 55.2%, NMI: 50.2%). *Second row*: $K$-modes ($\lambda=0$, $\sigma=0.35$, ACC: 56.0%, NMI: 50.6%). *Third row*: Laplacian $K$-modes ($\lambda=0.07$, $\sigma=0.35$, ACC: 70.5%, NMI: 68.8%). *Fourth row*: mean-shift ($\sigma=0.2485$).[]{data-label="f:mnist"}](MNIST/MNIST_centers_ms_07.eps "fig:"){width="0.095\linewidth"} ![Centroids found by different algorithms on MNIST. *First row*: $K$-means ($\lambda=0$, $\sigma\rightarrow\infty$, ACC: 55.2%, NMI: 50.2%). *Second row*: $K$-modes ($\lambda=0$, $\sigma=0.35$, ACC: 56.0%, NMI: 50.6%). *Third row*: Laplacian $K$-modes ($\lambda=0.07$, $\sigma=0.35$, ACC: 70.5%, NMI: 68.8%). *Fourth row*: mean-shift ($\sigma=0.2485$).[]{data-label="f:mnist"}](MNIST/MNIST_centers_ms_08.eps "fig:"){width="0.095\linewidth"} ![Centroids found by different algorithms on MNIST. *First row*: $K$-means ($\lambda=0$, $\sigma\rightarrow\infty$, ACC: 55.2%, NMI: 50.2%). *Second row*: $K$-modes ($\lambda=0$, $\sigma=0.35$, ACC: 56.0%, NMI: 50.6%). *Third row*: Laplacian $K$-modes ($\lambda=0.07$, $\sigma=0.35$, ACC: 70.5%, NMI: 68.8%). *Fourth row*: mean-shift ($\sigma=0.2485$).[]{data-label="f:mnist"}](MNIST/MNIST_centers_ms_09.eps "fig:"){width="0.095\linewidth"} ![Centroids found by different algorithms on MNIST. *First row*: $K$-means ($\lambda=0$, $\sigma\rightarrow\infty$, ACC: 55.2%, NMI: 50.2%). *Second row*: $K$-modes ($\lambda=0$, $\sigma=0.35$, ACC: 56.0%, NMI: 50.6%). *Third row*: Laplacian $K$-modes ($\lambda=0.07$, $\sigma=0.35$, ACC: 70.5%, NMI: 68.8%). *Fourth row*: mean-shift ($\sigma=0.2485$).[]{data-label="f:mnist"}](MNIST/MNIST_centers_ms_10.eps "fig:"){width="0.095\linewidth"}
---------- ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------ ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
---------- -------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- -------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- -------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- -------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- -------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- -------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- -------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- -------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- -------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- -------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- -------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- -------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- -------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- -------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- -------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- -------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- -------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- -------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- -------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- --------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
![Centroids found by different algorithms on COIL–20. *First row*: $K$-means ($\lambda=0$, $\sigma\rightarrow\infty$, ACC: 64.8%, NMI: 73.5%). *Second row*: $K$-modes ($\lambda=0$, $\sigma=0.3$, ACC: 65.5%, NMI: 73.0%). *Third row*: Laplacian $K$-modes ($\lambda=0.01$, $\sigma=0.1$, ACC: 73.2%, NMI: 83.8%). *Fourth row*: Laplacian $K$-modes with homotopy ($\lambda=0.01$, ACC: 81.5%, NMI: 88.0%) in $\sigma$. *Fourth row*: mean-shift ($\sigma=0.1591$).[]{data-label="f:coil"}](COIL/COIL_centers_kmeans_01.eps "fig:"){width="0.048\linewidth"} ![Centroids found by different algorithms on COIL–20. *First row*: $K$-means ($\lambda=0$, $\sigma\rightarrow\infty$, ACC: 64.8%, NMI: 73.5%). *Second row*: $K$-modes ($\lambda=0$, $\sigma=0.3$, ACC: 65.5%, NMI: 73.0%). *Third row*: Laplacian $K$-modes ($\lambda=0.01$, $\sigma=0.1$, ACC: 73.2%, NMI: 83.8%). *Fourth row*: Laplacian $K$-modes with homotopy ($\lambda=0.01$, ACC: 81.5%, NMI: 88.0%) in $\sigma$. *Fourth row*: mean-shift ($\sigma=0.1591$).[]{data-label="f:coil"}](COIL/COIL_centers_kmeans_02.eps "fig:"){width="0.048\linewidth"} ![Centroids found by different algorithms on COIL–20. *First row*: $K$-means ($\lambda=0$, $\sigma\rightarrow\infty$, ACC: 64.8%, NMI: 73.5%). *Second row*: $K$-modes ($\lambda=0$, $\sigma=0.3$, ACC: 65.5%, NMI: 73.0%). *Third row*: Laplacian $K$-modes ($\lambda=0.01$, $\sigma=0.1$, ACC: 73.2%, NMI: 83.8%). *Fourth row*: Laplacian $K$-modes with homotopy ($\lambda=0.01$, ACC: 81.5%, NMI: 88.0%) in $\sigma$. *Fourth row*: mean-shift ($\sigma=0.1591$).[]{data-label="f:coil"}](COIL/COIL_centers_kmeans_03.eps "fig:"){width="0.048\linewidth"} ![Centroids found by different algorithms on COIL–20. *First row*: $K$-means ($\lambda=0$, $\sigma\rightarrow\infty$, ACC: 64.8%, NMI: 73.5%). *Second row*: $K$-modes ($\lambda=0$, $\sigma=0.3$, ACC: 65.5%, NMI: 73.0%). *Third row*: Laplacian $K$-modes ($\lambda=0.01$, $\sigma=0.1$, ACC: 73.2%, NMI: 83.8%). *Fourth row*: Laplacian $K$-modes with homotopy ($\lambda=0.01$, ACC: 81.5%, NMI: 88.0%) in $\sigma$. *Fourth row*: mean-shift ($\sigma=0.1591$).[]{data-label="f:coil"}](COIL/COIL_centers_kmeans_04.eps "fig:"){width="0.048\linewidth"} ![Centroids found by different algorithms on COIL–20. *First row*: $K$-means ($\lambda=0$, $\sigma\rightarrow\infty$, ACC: 64.8%, NMI: 73.5%). *Second row*: $K$-modes ($\lambda=0$, $\sigma=0.3$, ACC: 65.5%, NMI: 73.0%). *Third row*: Laplacian $K$-modes ($\lambda=0.01$, $\sigma=0.1$, ACC: 73.2%, NMI: 83.8%). *Fourth row*: Laplacian $K$-modes with homotopy ($\lambda=0.01$, ACC: 81.5%, NMI: 88.0%) in $\sigma$. *Fourth row*: mean-shift ($\sigma=0.1591$).[]{data-label="f:coil"}](COIL/COIL_centers_kmeans_05.eps "fig:"){width="0.048\linewidth"} ![Centroids found by different algorithms on COIL–20. *First row*: $K$-means ($\lambda=0$, $\sigma\rightarrow\infty$, ACC: 64.8%, NMI: 73.5%). *Second row*: $K$-modes ($\lambda=0$, $\sigma=0.3$, ACC: 65.5%, NMI: 73.0%). *Third row*: Laplacian $K$-modes ($\lambda=0.01$, $\sigma=0.1$, ACC: 73.2%, NMI: 83.8%). *Fourth row*: Laplacian $K$-modes with homotopy ($\lambda=0.01$, ACC: 81.5%, NMI: 88.0%) in $\sigma$. *Fourth row*: mean-shift ($\sigma=0.1591$).[]{data-label="f:coil"}](COIL/COIL_centers_kmeans_06.eps "fig:"){width="0.048\linewidth"} ![Centroids found by different algorithms on COIL–20. *First row*: $K$-means ($\lambda=0$, $\sigma\rightarrow\infty$, ACC: 64.8%, NMI: 73.5%). *Second row*: $K$-modes ($\lambda=0$, $\sigma=0.3$, ACC: 65.5%, NMI: 73.0%). *Third row*: Laplacian $K$-modes ($\lambda=0.01$, $\sigma=0.1$, ACC: 73.2%, NMI: 83.8%). *Fourth row*: Laplacian $K$-modes with homotopy ($\lambda=0.01$, ACC: 81.5%, NMI: 88.0%) in $\sigma$. *Fourth row*: mean-shift ($\sigma=0.1591$).[]{data-label="f:coil"}](COIL/COIL_centers_kmeans_07.eps "fig:"){width="0.048\linewidth"} ![Centroids found by different algorithms on COIL–20. *First row*: $K$-means ($\lambda=0$, $\sigma\rightarrow\infty$, ACC: 64.8%, NMI: 73.5%). *Second row*: $K$-modes ($\lambda=0$, $\sigma=0.3$, ACC: 65.5%, NMI: 73.0%). *Third row*: Laplacian $K$-modes ($\lambda=0.01$, $\sigma=0.1$, ACC: 73.2%, NMI: 83.8%). *Fourth row*: Laplacian $K$-modes with homotopy ($\lambda=0.01$, ACC: 81.5%, NMI: 88.0%) in $\sigma$. *Fourth row*: mean-shift ($\sigma=0.1591$).[]{data-label="f:coil"}](COIL/COIL_centers_kmeans_08.eps "fig:"){width="0.048\linewidth"} ![Centroids found by different algorithms on COIL–20. *First row*: $K$-means ($\lambda=0$, $\sigma\rightarrow\infty$, ACC: 64.8%, NMI: 73.5%). *Second row*: $K$-modes ($\lambda=0$, $\sigma=0.3$, ACC: 65.5%, NMI: 73.0%). *Third row*: Laplacian $K$-modes ($\lambda=0.01$, $\sigma=0.1$, ACC: 73.2%, NMI: 83.8%). *Fourth row*: Laplacian $K$-modes with homotopy ($\lambda=0.01$, ACC: 81.5%, NMI: 88.0%) in $\sigma$. *Fourth row*: mean-shift ($\sigma=0.1591$).[]{data-label="f:coil"}](COIL/COIL_centers_kmeans_09.eps "fig:"){width="0.048\linewidth"} ![Centroids found by different algorithms on COIL–20. *First row*: $K$-means ($\lambda=0$, $\sigma\rightarrow\infty$, ACC: 64.8%, NMI: 73.5%). *Second row*: $K$-modes ($\lambda=0$, $\sigma=0.3$, ACC: 65.5%, NMI: 73.0%). *Third row*: Laplacian $K$-modes ($\lambda=0.01$, $\sigma=0.1$, ACC: 73.2%, NMI: 83.8%). *Fourth row*: Laplacian $K$-modes with homotopy ($\lambda=0.01$, ACC: 81.5%, NMI: 88.0%) in $\sigma$. *Fourth row*: mean-shift ($\sigma=0.1591$).[]{data-label="f:coil"}](COIL/COIL_centers_kmeans_10.eps "fig:"){width="0.048\linewidth"} ![Centroids found by different algorithms on COIL–20. *First row*: $K$-means ($\lambda=0$, $\sigma\rightarrow\infty$, ACC: 64.8%, NMI: 73.5%). *Second row*: $K$-modes ($\lambda=0$, $\sigma=0.3$, ACC: 65.5%, NMI: 73.0%). *Third row*: Laplacian $K$-modes ($\lambda=0.01$, $\sigma=0.1$, ACC: 73.2%, NMI: 83.8%). *Fourth row*: Laplacian $K$-modes with homotopy ($\lambda=0.01$, ACC: 81.5%, NMI: 88.0%) in $\sigma$. *Fourth row*: mean-shift ($\sigma=0.1591$).[]{data-label="f:coil"}](COIL/COIL_centers_kmeans_11.eps "fig:"){width="0.048\linewidth"} ![Centroids found by different algorithms on COIL–20. *First row*: $K$-means ($\lambda=0$, $\sigma\rightarrow\infty$, ACC: 64.8%, NMI: 73.5%). *Second row*: $K$-modes ($\lambda=0$, $\sigma=0.3$, ACC: 65.5%, NMI: 73.0%). *Third row*: Laplacian $K$-modes ($\lambda=0.01$, $\sigma=0.1$, ACC: 73.2%, NMI: 83.8%). *Fourth row*: Laplacian $K$-modes with homotopy ($\lambda=0.01$, ACC: 81.5%, NMI: 88.0%) in $\sigma$. *Fourth row*: mean-shift ($\sigma=0.1591$).[]{data-label="f:coil"}](COIL/COIL_centers_kmeans_12.eps "fig:"){width="0.048\linewidth"} ![Centroids found by different algorithms on COIL–20. *First row*: $K$-means ($\lambda=0$, $\sigma\rightarrow\infty$, ACC: 64.8%, NMI: 73.5%). *Second row*: $K$-modes ($\lambda=0$, $\sigma=0.3$, ACC: 65.5%, NMI: 73.0%). *Third row*: Laplacian $K$-modes ($\lambda=0.01$, $\sigma=0.1$, ACC: 73.2%, NMI: 83.8%). *Fourth row*: Laplacian $K$-modes with homotopy ($\lambda=0.01$, ACC: 81.5%, NMI: 88.0%) in $\sigma$. *Fourth row*: mean-shift ($\sigma=0.1591$).[]{data-label="f:coil"}](COIL/COIL_centers_kmeans_13.eps "fig:"){width="0.048\linewidth"} ![Centroids found by different algorithms on COIL–20. *First row*: $K$-means ($\lambda=0$, $\sigma\rightarrow\infty$, ACC: 64.8%, NMI: 73.5%). *Second row*: $K$-modes ($\lambda=0$, $\sigma=0.3$, ACC: 65.5%, NMI: 73.0%). *Third row*: Laplacian $K$-modes ($\lambda=0.01$, $\sigma=0.1$, ACC: 73.2%, NMI: 83.8%). *Fourth row*: Laplacian $K$-modes with homotopy ($\lambda=0.01$, ACC: 81.5%, NMI: 88.0%) in $\sigma$. *Fourth row*: mean-shift ($\sigma=0.1591$).[]{data-label="f:coil"}](COIL/COIL_centers_kmeans_14.eps "fig:"){width="0.048\linewidth"} ![Centroids found by different algorithms on COIL–20. *First row*: $K$-means ($\lambda=0$, $\sigma\rightarrow\infty$, ACC: 64.8%, NMI: 73.5%). *Second row*: $K$-modes ($\lambda=0$, $\sigma=0.3$, ACC: 65.5%, NMI: 73.0%). *Third row*: Laplacian $K$-modes ($\lambda=0.01$, $\sigma=0.1$, ACC: 73.2%, NMI: 83.8%). *Fourth row*: Laplacian $K$-modes with homotopy ($\lambda=0.01$, ACC: 81.5%, NMI: 88.0%) in $\sigma$. *Fourth row*: mean-shift ($\sigma=0.1591$).[]{data-label="f:coil"}](COIL/COIL_centers_kmeans_15.eps "fig:"){width="0.048\linewidth"} ![Centroids found by different algorithms on COIL–20. *First row*: $K$-means ($\lambda=0$, $\sigma\rightarrow\infty$, ACC: 64.8%, NMI: 73.5%). *Second row*: $K$-modes ($\lambda=0$, $\sigma=0.3$, ACC: 65.5%, NMI: 73.0%). *Third row*: Laplacian $K$-modes ($\lambda=0.01$, $\sigma=0.1$, ACC: 73.2%, NMI: 83.8%). *Fourth row*: Laplacian $K$-modes with homotopy ($\lambda=0.01$, ACC: 81.5%, NMI: 88.0%) in $\sigma$. *Fourth row*: mean-shift ($\sigma=0.1591$).[]{data-label="f:coil"}](COIL/COIL_centers_kmeans_16.eps "fig:"){width="0.048\linewidth"} ![Centroids found by different algorithms on COIL–20. *First row*: $K$-means ($\lambda=0$, $\sigma\rightarrow\infty$, ACC: 64.8%, NMI: 73.5%). *Second row*: $K$-modes ($\lambda=0$, $\sigma=0.3$, ACC: 65.5%, NMI: 73.0%). *Third row*: Laplacian $K$-modes ($\lambda=0.01$, $\sigma=0.1$, ACC: 73.2%, NMI: 83.8%). *Fourth row*: Laplacian $K$-modes with homotopy ($\lambda=0.01$, ACC: 81.5%, NMI: 88.0%) in $\sigma$. *Fourth row*: mean-shift ($\sigma=0.1591$).[]{data-label="f:coil"}](COIL/COIL_centers_kmeans_17.eps "fig:"){width="0.048\linewidth"} ![Centroids found by different algorithms on COIL–20. *First row*: $K$-means ($\lambda=0$, $\sigma\rightarrow\infty$, ACC: 64.8%, NMI: 73.5%). *Second row*: $K$-modes ($\lambda=0$, $\sigma=0.3$, ACC: 65.5%, NMI: 73.0%). *Third row*: Laplacian $K$-modes ($\lambda=0.01$, $\sigma=0.1$, ACC: 73.2%, NMI: 83.8%). *Fourth row*: Laplacian $K$-modes with homotopy ($\lambda=0.01$, ACC: 81.5%, NMI: 88.0%) in $\sigma$. *Fourth row*: mean-shift ($\sigma=0.1591$).[]{data-label="f:coil"}](COIL/COIL_centers_kmeans_18.eps "fig:"){width="0.048\linewidth"} ![Centroids found by different algorithms on COIL–20. *First row*: $K$-means ($\lambda=0$, $\sigma\rightarrow\infty$, ACC: 64.8%, NMI: 73.5%). *Second row*: $K$-modes ($\lambda=0$, $\sigma=0.3$, ACC: 65.5%, NMI: 73.0%). *Third row*: Laplacian $K$-modes ($\lambda=0.01$, $\sigma=0.1$, ACC: 73.2%, NMI: 83.8%). *Fourth row*: Laplacian $K$-modes with homotopy ($\lambda=0.01$, ACC: 81.5%, NMI: 88.0%) in $\sigma$. *Fourth row*: mean-shift ($\sigma=0.1591$).[]{data-label="f:coil"}](COIL/COIL_centers_kmeans_19.eps "fig:"){width="0.048\linewidth"} ![Centroids found by different algorithms on COIL–20. *First row*: $K$-means ($\lambda=0$, $\sigma\rightarrow\infty$, ACC: 64.8%, NMI: 73.5%). *Second row*: $K$-modes ($\lambda=0$, $\sigma=0.3$, ACC: 65.5%, NMI: 73.0%). *Third row*: Laplacian $K$-modes ($\lambda=0.01$, $\sigma=0.1$, ACC: 73.2%, NMI: 83.8%). *Fourth row*: Laplacian $K$-modes with homotopy ($\lambda=0.01$, ACC: 81.5%, NMI: 88.0%) in $\sigma$. *Fourth row*: mean-shift ($\sigma=0.1591$).[]{data-label="f:coil"}](COIL/COIL_centers_kmeans_20.eps "fig:"){width="0.048\linewidth"}
\[-1ex\] ![Centroids found by different algorithms on COIL–20. *First row*: $K$-means ($\lambda=0$, $\sigma\rightarrow\infty$, ACC: 64.8%, NMI: 73.5%). *Second row*: $K$-modes ($\lambda=0$, $\sigma=0.3$, ACC: 65.5%, NMI: 73.0%). *Third row*: Laplacian $K$-modes ($\lambda=0.01$, $\sigma=0.1$, ACC: 73.2%, NMI: 83.8%). *Fourth row*: Laplacian $K$-modes with homotopy ($\lambda=0.01$, ACC: 81.5%, NMI: 88.0%) in $\sigma$. *Fourth row*: mean-shift ($\sigma=0.1591$).[]{data-label="f:coil"}](COIL/COIL_centers_kmodes_01.eps "fig:"){width="0.048\linewidth"} ![Centroids found by different algorithms on COIL–20. *First row*: $K$-means ($\lambda=0$, $\sigma\rightarrow\infty$, ACC: 64.8%, NMI: 73.5%). *Second row*: $K$-modes ($\lambda=0$, $\sigma=0.3$, ACC: 65.5%, NMI: 73.0%). *Third row*: Laplacian $K$-modes ($\lambda=0.01$, $\sigma=0.1$, ACC: 73.2%, NMI: 83.8%). *Fourth row*: Laplacian $K$-modes with homotopy ($\lambda=0.01$, ACC: 81.5%, NMI: 88.0%) in $\sigma$. *Fourth row*: mean-shift ($\sigma=0.1591$).[]{data-label="f:coil"}](COIL/COIL_centers_kmodes_02.eps "fig:"){width="0.048\linewidth"} ![Centroids found by different algorithms on COIL–20. *First row*: $K$-means ($\lambda=0$, $\sigma\rightarrow\infty$, ACC: 64.8%, NMI: 73.5%). *Second row*: $K$-modes ($\lambda=0$, $\sigma=0.3$, ACC: 65.5%, NMI: 73.0%). *Third row*: Laplacian $K$-modes ($\lambda=0.01$, $\sigma=0.1$, ACC: 73.2%, NMI: 83.8%). *Fourth row*: Laplacian $K$-modes with homotopy ($\lambda=0.01$, ACC: 81.5%, NMI: 88.0%) in $\sigma$. *Fourth row*: mean-shift ($\sigma=0.1591$).[]{data-label="f:coil"}](COIL/COIL_centers_kmodes_03.eps "fig:"){width="0.048\linewidth"} ![Centroids found by different algorithms on COIL–20. *First row*: $K$-means ($\lambda=0$, $\sigma\rightarrow\infty$, ACC: 64.8%, NMI: 73.5%). *Second row*: $K$-modes ($\lambda=0$, $\sigma=0.3$, ACC: 65.5%, NMI: 73.0%). *Third row*: Laplacian $K$-modes ($\lambda=0.01$, $\sigma=0.1$, ACC: 73.2%, NMI: 83.8%). *Fourth row*: Laplacian $K$-modes with homotopy ($\lambda=0.01$, ACC: 81.5%, NMI: 88.0%) in $\sigma$. *Fourth row*: mean-shift ($\sigma=0.1591$).[]{data-label="f:coil"}](COIL/COIL_centers_kmodes_04.eps "fig:"){width="0.048\linewidth"} ![Centroids found by different algorithms on COIL–20. *First row*: $K$-means ($\lambda=0$, $\sigma\rightarrow\infty$, ACC: 64.8%, NMI: 73.5%). *Second row*: $K$-modes ($\lambda=0$, $\sigma=0.3$, ACC: 65.5%, NMI: 73.0%). *Third row*: Laplacian $K$-modes ($\lambda=0.01$, $\sigma=0.1$, ACC: 73.2%, NMI: 83.8%). *Fourth row*: Laplacian $K$-modes with homotopy ($\lambda=0.01$, ACC: 81.5%, NMI: 88.0%) in $\sigma$. *Fourth row*: mean-shift ($\sigma=0.1591$).[]{data-label="f:coil"}](COIL/COIL_centers_kmodes_05.eps "fig:"){width="0.048\linewidth"} ![Centroids found by different algorithms on COIL–20. *First row*: $K$-means ($\lambda=0$, $\sigma\rightarrow\infty$, ACC: 64.8%, NMI: 73.5%). *Second row*: $K$-modes ($\lambda=0$, $\sigma=0.3$, ACC: 65.5%, NMI: 73.0%). *Third row*: Laplacian $K$-modes ($\lambda=0.01$, $\sigma=0.1$, ACC: 73.2%, NMI: 83.8%). *Fourth row*: Laplacian $K$-modes with homotopy ($\lambda=0.01$, ACC: 81.5%, NMI: 88.0%) in $\sigma$. *Fourth row*: mean-shift ($\sigma=0.1591$).[]{data-label="f:coil"}](COIL/COIL_centers_kmodes_06.eps "fig:"){width="0.048\linewidth"} ![Centroids found by different algorithms on COIL–20. *First row*: $K$-means ($\lambda=0$, $\sigma\rightarrow\infty$, ACC: 64.8%, NMI: 73.5%). *Second row*: $K$-modes ($\lambda=0$, $\sigma=0.3$, ACC: 65.5%, NMI: 73.0%). *Third row*: Laplacian $K$-modes ($\lambda=0.01$, $\sigma=0.1$, ACC: 73.2%, NMI: 83.8%). *Fourth row*: Laplacian $K$-modes with homotopy ($\lambda=0.01$, ACC: 81.5%, NMI: 88.0%) in $\sigma$. *Fourth row*: mean-shift ($\sigma=0.1591$).[]{data-label="f:coil"}](COIL/COIL_centers_kmodes_07.eps "fig:"){width="0.048\linewidth"} ![Centroids found by different algorithms on COIL–20. *First row*: $K$-means ($\lambda=0$, $\sigma\rightarrow\infty$, ACC: 64.8%, NMI: 73.5%). *Second row*: $K$-modes ($\lambda=0$, $\sigma=0.3$, ACC: 65.5%, NMI: 73.0%). *Third row*: Laplacian $K$-modes ($\lambda=0.01$, $\sigma=0.1$, ACC: 73.2%, NMI: 83.8%). *Fourth row*: Laplacian $K$-modes with homotopy ($\lambda=0.01$, ACC: 81.5%, NMI: 88.0%) in $\sigma$. *Fourth row*: mean-shift ($\sigma=0.1591$).[]{data-label="f:coil"}](COIL/COIL_centers_kmodes_08.eps "fig:"){width="0.048\linewidth"} ![Centroids found by different algorithms on COIL–20. *First row*: $K$-means ($\lambda=0$, $\sigma\rightarrow\infty$, ACC: 64.8%, NMI: 73.5%). *Second row*: $K$-modes ($\lambda=0$, $\sigma=0.3$, ACC: 65.5%, NMI: 73.0%). *Third row*: Laplacian $K$-modes ($\lambda=0.01$, $\sigma=0.1$, ACC: 73.2%, NMI: 83.8%). *Fourth row*: Laplacian $K$-modes with homotopy ($\lambda=0.01$, ACC: 81.5%, NMI: 88.0%) in $\sigma$. *Fourth row*: mean-shift ($\sigma=0.1591$).[]{data-label="f:coil"}](COIL/COIL_centers_kmodes_09.eps "fig:"){width="0.048\linewidth"} ![Centroids found by different algorithms on COIL–20. *First row*: $K$-means ($\lambda=0$, $\sigma\rightarrow\infty$, ACC: 64.8%, NMI: 73.5%). *Second row*: $K$-modes ($\lambda=0$, $\sigma=0.3$, ACC: 65.5%, NMI: 73.0%). *Third row*: Laplacian $K$-modes ($\lambda=0.01$, $\sigma=0.1$, ACC: 73.2%, NMI: 83.8%). *Fourth row*: Laplacian $K$-modes with homotopy ($\lambda=0.01$, ACC: 81.5%, NMI: 88.0%) in $\sigma$. *Fourth row*: mean-shift ($\sigma=0.1591$).[]{data-label="f:coil"}](COIL/COIL_centers_kmodes_10.eps "fig:"){width="0.048\linewidth"} ![Centroids found by different algorithms on COIL–20. *First row*: $K$-means ($\lambda=0$, $\sigma\rightarrow\infty$, ACC: 64.8%, NMI: 73.5%). *Second row*: $K$-modes ($\lambda=0$, $\sigma=0.3$, ACC: 65.5%, NMI: 73.0%). *Third row*: Laplacian $K$-modes ($\lambda=0.01$, $\sigma=0.1$, ACC: 73.2%, NMI: 83.8%). *Fourth row*: Laplacian $K$-modes with homotopy ($\lambda=0.01$, ACC: 81.5%, NMI: 88.0%) in $\sigma$. *Fourth row*: mean-shift ($\sigma=0.1591$).[]{data-label="f:coil"}](COIL/COIL_centers_kmodes_11.eps "fig:"){width="0.048\linewidth"} ![Centroids found by different algorithms on COIL–20. *First row*: $K$-means ($\lambda=0$, $\sigma\rightarrow\infty$, ACC: 64.8%, NMI: 73.5%). *Second row*: $K$-modes ($\lambda=0$, $\sigma=0.3$, ACC: 65.5%, NMI: 73.0%). *Third row*: Laplacian $K$-modes ($\lambda=0.01$, $\sigma=0.1$, ACC: 73.2%, NMI: 83.8%). *Fourth row*: Laplacian $K$-modes with homotopy ($\lambda=0.01$, ACC: 81.5%, NMI: 88.0%) in $\sigma$. *Fourth row*: mean-shift ($\sigma=0.1591$).[]{data-label="f:coil"}](COIL/COIL_centers_kmodes_12.eps "fig:"){width="0.048\linewidth"} ![Centroids found by different algorithms on COIL–20. *First row*: $K$-means ($\lambda=0$, $\sigma\rightarrow\infty$, ACC: 64.8%, NMI: 73.5%). *Second row*: $K$-modes ($\lambda=0$, $\sigma=0.3$, ACC: 65.5%, NMI: 73.0%). *Third row*: Laplacian $K$-modes ($\lambda=0.01$, $\sigma=0.1$, ACC: 73.2%, NMI: 83.8%). *Fourth row*: Laplacian $K$-modes with homotopy ($\lambda=0.01$, ACC: 81.5%, NMI: 88.0%) in $\sigma$. *Fourth row*: mean-shift ($\sigma=0.1591$).[]{data-label="f:coil"}](COIL/COIL_centers_kmodes_13.eps "fig:"){width="0.048\linewidth"} ![Centroids found by different algorithms on COIL–20. *First row*: $K$-means ($\lambda=0$, $\sigma\rightarrow\infty$, ACC: 64.8%, NMI: 73.5%). *Second row*: $K$-modes ($\lambda=0$, $\sigma=0.3$, ACC: 65.5%, NMI: 73.0%). *Third row*: Laplacian $K$-modes ($\lambda=0.01$, $\sigma=0.1$, ACC: 73.2%, NMI: 83.8%). *Fourth row*: Laplacian $K$-modes with homotopy ($\lambda=0.01$, ACC: 81.5%, NMI: 88.0%) in $\sigma$. *Fourth row*: mean-shift ($\sigma=0.1591$).[]{data-label="f:coil"}](COIL/COIL_centers_kmodes_14.eps "fig:"){width="0.048\linewidth"} ![Centroids found by different algorithms on COIL–20. *First row*: $K$-means ($\lambda=0$, $\sigma\rightarrow\infty$, ACC: 64.8%, NMI: 73.5%). *Second row*: $K$-modes ($\lambda=0$, $\sigma=0.3$, ACC: 65.5%, NMI: 73.0%). *Third row*: Laplacian $K$-modes ($\lambda=0.01$, $\sigma=0.1$, ACC: 73.2%, NMI: 83.8%). *Fourth row*: Laplacian $K$-modes with homotopy ($\lambda=0.01$, ACC: 81.5%, NMI: 88.0%) in $\sigma$. *Fourth row*: mean-shift ($\sigma=0.1591$).[]{data-label="f:coil"}](COIL/COIL_centers_kmodes_15.eps "fig:"){width="0.048\linewidth"} ![Centroids found by different algorithms on COIL–20. *First row*: $K$-means ($\lambda=0$, $\sigma\rightarrow\infty$, ACC: 64.8%, NMI: 73.5%). *Second row*: $K$-modes ($\lambda=0$, $\sigma=0.3$, ACC: 65.5%, NMI: 73.0%). *Third row*: Laplacian $K$-modes ($\lambda=0.01$, $\sigma=0.1$, ACC: 73.2%, NMI: 83.8%). *Fourth row*: Laplacian $K$-modes with homotopy ($\lambda=0.01$, ACC: 81.5%, NMI: 88.0%) in $\sigma$. *Fourth row*: mean-shift ($\sigma=0.1591$).[]{data-label="f:coil"}](COIL/COIL_centers_kmodes_16.eps "fig:"){width="0.048\linewidth"} ![Centroids found by different algorithms on COIL–20. *First row*: $K$-means ($\lambda=0$, $\sigma\rightarrow\infty$, ACC: 64.8%, NMI: 73.5%). *Second row*: $K$-modes ($\lambda=0$, $\sigma=0.3$, ACC: 65.5%, NMI: 73.0%). *Third row*: Laplacian $K$-modes ($\lambda=0.01$, $\sigma=0.1$, ACC: 73.2%, NMI: 83.8%). *Fourth row*: Laplacian $K$-modes with homotopy ($\lambda=0.01$, ACC: 81.5%, NMI: 88.0%) in $\sigma$. *Fourth row*: mean-shift ($\sigma=0.1591$).[]{data-label="f:coil"}](COIL/COIL_centers_kmodes_17.eps "fig:"){width="0.048\linewidth"} ![Centroids found by different algorithms on COIL–20. *First row*: $K$-means ($\lambda=0$, $\sigma\rightarrow\infty$, ACC: 64.8%, NMI: 73.5%). *Second row*: $K$-modes ($\lambda=0$, $\sigma=0.3$, ACC: 65.5%, NMI: 73.0%). *Third row*: Laplacian $K$-modes ($\lambda=0.01$, $\sigma=0.1$, ACC: 73.2%, NMI: 83.8%). *Fourth row*: Laplacian $K$-modes with homotopy ($\lambda=0.01$, ACC: 81.5%, NMI: 88.0%) in $\sigma$. *Fourth row*: mean-shift ($\sigma=0.1591$).[]{data-label="f:coil"}](COIL/COIL_centers_kmodes_18.eps "fig:"){width="0.048\linewidth"} ![Centroids found by different algorithms on COIL–20. *First row*: $K$-means ($\lambda=0$, $\sigma\rightarrow\infty$, ACC: 64.8%, NMI: 73.5%). *Second row*: $K$-modes ($\lambda=0$, $\sigma=0.3$, ACC: 65.5%, NMI: 73.0%). *Third row*: Laplacian $K$-modes ($\lambda=0.01$, $\sigma=0.1$, ACC: 73.2%, NMI: 83.8%). *Fourth row*: Laplacian $K$-modes with homotopy ($\lambda=0.01$, ACC: 81.5%, NMI: 88.0%) in $\sigma$. *Fourth row*: mean-shift ($\sigma=0.1591$).[]{data-label="f:coil"}](COIL/COIL_centers_kmodes_19.eps "fig:"){width="0.048\linewidth"} ![Centroids found by different algorithms on COIL–20. *First row*: $K$-means ($\lambda=0$, $\sigma\rightarrow\infty$, ACC: 64.8%, NMI: 73.5%). *Second row*: $K$-modes ($\lambda=0$, $\sigma=0.3$, ACC: 65.5%, NMI: 73.0%). *Third row*: Laplacian $K$-modes ($\lambda=0.01$, $\sigma=0.1$, ACC: 73.2%, NMI: 83.8%). *Fourth row*: Laplacian $K$-modes with homotopy ($\lambda=0.01$, ACC: 81.5%, NMI: 88.0%) in $\sigma$. *Fourth row*: mean-shift ($\sigma=0.1591$).[]{data-label="f:coil"}](COIL/COIL_centers_kmodes_20.eps "fig:"){width="0.048\linewidth"}
\[-1ex\] ![Centroids found by different algorithms on COIL–20. *First row*: $K$-means ($\lambda=0$, $\sigma\rightarrow\infty$, ACC: 64.8%, NMI: 73.5%). *Second row*: $K$-modes ($\lambda=0$, $\sigma=0.3$, ACC: 65.5%, NMI: 73.0%). *Third row*: Laplacian $K$-modes ($\lambda=0.01$, $\sigma=0.1$, ACC: 73.2%, NMI: 83.8%). *Fourth row*: Laplacian $K$-modes with homotopy ($\lambda=0.01$, ACC: 81.5%, NMI: 88.0%) in $\sigma$. *Fourth row*: mean-shift ($\sigma=0.1591$).[]{data-label="f:coil"}](COIL/COIL_centers_manikmodes_01.eps "fig:"){width="0.048\linewidth"} ![Centroids found by different algorithms on COIL–20. *First row*: $K$-means ($\lambda=0$, $\sigma\rightarrow\infty$, ACC: 64.8%, NMI: 73.5%). *Second row*: $K$-modes ($\lambda=0$, $\sigma=0.3$, ACC: 65.5%, NMI: 73.0%). *Third row*: Laplacian $K$-modes ($\lambda=0.01$, $\sigma=0.1$, ACC: 73.2%, NMI: 83.8%). *Fourth row*: Laplacian $K$-modes with homotopy ($\lambda=0.01$, ACC: 81.5%, NMI: 88.0%) in $\sigma$. *Fourth row*: mean-shift ($\sigma=0.1591$).[]{data-label="f:coil"}](COIL/COIL_centers_manikmodes_02.eps "fig:"){width="0.048\linewidth"} ![Centroids found by different algorithms on COIL–20. *First row*: $K$-means ($\lambda=0$, $\sigma\rightarrow\infty$, ACC: 64.8%, NMI: 73.5%). *Second row*: $K$-modes ($\lambda=0$, $\sigma=0.3$, ACC: 65.5%, NMI: 73.0%). *Third row*: Laplacian $K$-modes ($\lambda=0.01$, $\sigma=0.1$, ACC: 73.2%, NMI: 83.8%). *Fourth row*: Laplacian $K$-modes with homotopy ($\lambda=0.01$, ACC: 81.5%, NMI: 88.0%) in $\sigma$. *Fourth row*: mean-shift ($\sigma=0.1591$).[]{data-label="f:coil"}](COIL/COIL_centers_manikmodes_03.eps "fig:"){width="0.048\linewidth"} ![Centroids found by different algorithms on COIL–20. *First row*: $K$-means ($\lambda=0$, $\sigma\rightarrow\infty$, ACC: 64.8%, NMI: 73.5%). *Second row*: $K$-modes ($\lambda=0$, $\sigma=0.3$, ACC: 65.5%, NMI: 73.0%). *Third row*: Laplacian $K$-modes ($\lambda=0.01$, $\sigma=0.1$, ACC: 73.2%, NMI: 83.8%). *Fourth row*: Laplacian $K$-modes with homotopy ($\lambda=0.01$, ACC: 81.5%, NMI: 88.0%) in $\sigma$. *Fourth row*: mean-shift ($\sigma=0.1591$).[]{data-label="f:coil"}](COIL/COIL_centers_manikmodes_04.eps "fig:"){width="0.048\linewidth"} ![Centroids found by different algorithms on COIL–20. *First row*: $K$-means ($\lambda=0$, $\sigma\rightarrow\infty$, ACC: 64.8%, NMI: 73.5%). *Second row*: $K$-modes ($\lambda=0$, $\sigma=0.3$, ACC: 65.5%, NMI: 73.0%). *Third row*: Laplacian $K$-modes ($\lambda=0.01$, $\sigma=0.1$, ACC: 73.2%, NMI: 83.8%). *Fourth row*: Laplacian $K$-modes with homotopy ($\lambda=0.01$, ACC: 81.5%, NMI: 88.0%) in $\sigma$. *Fourth row*: mean-shift ($\sigma=0.1591$).[]{data-label="f:coil"}](COIL/COIL_centers_manikmodes_05.eps "fig:"){width="0.048\linewidth"} ![Centroids found by different algorithms on COIL–20. *First row*: $K$-means ($\lambda=0$, $\sigma\rightarrow\infty$, ACC: 64.8%, NMI: 73.5%). *Second row*: $K$-modes ($\lambda=0$, $\sigma=0.3$, ACC: 65.5%, NMI: 73.0%). *Third row*: Laplacian $K$-modes ($\lambda=0.01$, $\sigma=0.1$, ACC: 73.2%, NMI: 83.8%). *Fourth row*: Laplacian $K$-modes with homotopy ($\lambda=0.01$, ACC: 81.5%, NMI: 88.0%) in $\sigma$. *Fourth row*: mean-shift ($\sigma=0.1591$).[]{data-label="f:coil"}](COIL/COIL_centers_manikmodes_06.eps "fig:"){width="0.048\linewidth"} ![Centroids found by different algorithms on COIL–20. *First row*: $K$-means ($\lambda=0$, $\sigma\rightarrow\infty$, ACC: 64.8%, NMI: 73.5%). *Second row*: $K$-modes ($\lambda=0$, $\sigma=0.3$, ACC: 65.5%, NMI: 73.0%). *Third row*: Laplacian $K$-modes ($\lambda=0.01$, $\sigma=0.1$, ACC: 73.2%, NMI: 83.8%). *Fourth row*: Laplacian $K$-modes with homotopy ($\lambda=0.01$, ACC: 81.5%, NMI: 88.0%) in $\sigma$. *Fourth row*: mean-shift ($\sigma=0.1591$).[]{data-label="f:coil"}](COIL/COIL_centers_manikmodes_07.eps "fig:"){width="0.048\linewidth"} ![Centroids found by different algorithms on COIL–20. *First row*: $K$-means ($\lambda=0$, $\sigma\rightarrow\infty$, ACC: 64.8%, NMI: 73.5%). *Second row*: $K$-modes ($\lambda=0$, $\sigma=0.3$, ACC: 65.5%, NMI: 73.0%). *Third row*: Laplacian $K$-modes ($\lambda=0.01$, $\sigma=0.1$, ACC: 73.2%, NMI: 83.8%). *Fourth row*: Laplacian $K$-modes with homotopy ($\lambda=0.01$, ACC: 81.5%, NMI: 88.0%) in $\sigma$. *Fourth row*: mean-shift ($\sigma=0.1591$).[]{data-label="f:coil"}](COIL/COIL_centers_manikmodes_08.eps "fig:"){width="0.048\linewidth"} ![Centroids found by different algorithms on COIL–20. *First row*: $K$-means ($\lambda=0$, $\sigma\rightarrow\infty$, ACC: 64.8%, NMI: 73.5%). *Second row*: $K$-modes ($\lambda=0$, $\sigma=0.3$, ACC: 65.5%, NMI: 73.0%). *Third row*: Laplacian $K$-modes ($\lambda=0.01$, $\sigma=0.1$, ACC: 73.2%, NMI: 83.8%). *Fourth row*: Laplacian $K$-modes with homotopy ($\lambda=0.01$, ACC: 81.5%, NMI: 88.0%) in $\sigma$. *Fourth row*: mean-shift ($\sigma=0.1591$).[]{data-label="f:coil"}](COIL/COIL_centers_manikmodes_09.eps "fig:"){width="0.048\linewidth"} ![Centroids found by different algorithms on COIL–20. *First row*: $K$-means ($\lambda=0$, $\sigma\rightarrow\infty$, ACC: 64.8%, NMI: 73.5%). *Second row*: $K$-modes ($\lambda=0$, $\sigma=0.3$, ACC: 65.5%, NMI: 73.0%). *Third row*: Laplacian $K$-modes ($\lambda=0.01$, $\sigma=0.1$, ACC: 73.2%, NMI: 83.8%). *Fourth row*: Laplacian $K$-modes with homotopy ($\lambda=0.01$, ACC: 81.5%, NMI: 88.0%) in $\sigma$. *Fourth row*: mean-shift ($\sigma=0.1591$).[]{data-label="f:coil"}](COIL/COIL_centers_manikmodes_10.eps "fig:"){width="0.048\linewidth"} ![Centroids found by different algorithms on COIL–20. *First row*: $K$-means ($\lambda=0$, $\sigma\rightarrow\infty$, ACC: 64.8%, NMI: 73.5%). *Second row*: $K$-modes ($\lambda=0$, $\sigma=0.3$, ACC: 65.5%, NMI: 73.0%). *Third row*: Laplacian $K$-modes ($\lambda=0.01$, $\sigma=0.1$, ACC: 73.2%, NMI: 83.8%). *Fourth row*: Laplacian $K$-modes with homotopy ($\lambda=0.01$, ACC: 81.5%, NMI: 88.0%) in $\sigma$. *Fourth row*: mean-shift ($\sigma=0.1591$).[]{data-label="f:coil"}](COIL/COIL_centers_manikmodes_11.eps "fig:"){width="0.048\linewidth"} ![Centroids found by different algorithms on COIL–20. *First row*: $K$-means ($\lambda=0$, $\sigma\rightarrow\infty$, ACC: 64.8%, NMI: 73.5%). *Second row*: $K$-modes ($\lambda=0$, $\sigma=0.3$, ACC: 65.5%, NMI: 73.0%). *Third row*: Laplacian $K$-modes ($\lambda=0.01$, $\sigma=0.1$, ACC: 73.2%, NMI: 83.8%). *Fourth row*: Laplacian $K$-modes with homotopy ($\lambda=0.01$, ACC: 81.5%, NMI: 88.0%) in $\sigma$. *Fourth row*: mean-shift ($\sigma=0.1591$).[]{data-label="f:coil"}](COIL/COIL_centers_manikmodes_12.eps "fig:"){width="0.048\linewidth"} ![Centroids found by different algorithms on COIL–20. *First row*: $K$-means ($\lambda=0$, $\sigma\rightarrow\infty$, ACC: 64.8%, NMI: 73.5%). *Second row*: $K$-modes ($\lambda=0$, $\sigma=0.3$, ACC: 65.5%, NMI: 73.0%). *Third row*: Laplacian $K$-modes ($\lambda=0.01$, $\sigma=0.1$, ACC: 73.2%, NMI: 83.8%). *Fourth row*: Laplacian $K$-modes with homotopy ($\lambda=0.01$, ACC: 81.5%, NMI: 88.0%) in $\sigma$. *Fourth row*: mean-shift ($\sigma=0.1591$).[]{data-label="f:coil"}](COIL/COIL_centers_manikmodes_13.eps "fig:"){width="0.048\linewidth"} ![Centroids found by different algorithms on COIL–20. *First row*: $K$-means ($\lambda=0$, $\sigma\rightarrow\infty$, ACC: 64.8%, NMI: 73.5%). *Second row*: $K$-modes ($\lambda=0$, $\sigma=0.3$, ACC: 65.5%, NMI: 73.0%). *Third row*: Laplacian $K$-modes ($\lambda=0.01$, $\sigma=0.1$, ACC: 73.2%, NMI: 83.8%). *Fourth row*: Laplacian $K$-modes with homotopy ($\lambda=0.01$, ACC: 81.5%, NMI: 88.0%) in $\sigma$. *Fourth row*: mean-shift ($\sigma=0.1591$).[]{data-label="f:coil"}](COIL/COIL_centers_manikmodes_14.eps "fig:"){width="0.048\linewidth"} ![Centroids found by different algorithms on COIL–20. *First row*: $K$-means ($\lambda=0$, $\sigma\rightarrow\infty$, ACC: 64.8%, NMI: 73.5%). *Second row*: $K$-modes ($\lambda=0$, $\sigma=0.3$, ACC: 65.5%, NMI: 73.0%). *Third row*: Laplacian $K$-modes ($\lambda=0.01$, $\sigma=0.1$, ACC: 73.2%, NMI: 83.8%). *Fourth row*: Laplacian $K$-modes with homotopy ($\lambda=0.01$, ACC: 81.5%, NMI: 88.0%) in $\sigma$. *Fourth row*: mean-shift ($\sigma=0.1591$).[]{data-label="f:coil"}](COIL/COIL_centers_manikmodes_15.eps "fig:"){width="0.048\linewidth"} ![Centroids found by different algorithms on COIL–20. *First row*: $K$-means ($\lambda=0$, $\sigma\rightarrow\infty$, ACC: 64.8%, NMI: 73.5%). *Second row*: $K$-modes ($\lambda=0$, $\sigma=0.3$, ACC: 65.5%, NMI: 73.0%). *Third row*: Laplacian $K$-modes ($\lambda=0.01$, $\sigma=0.1$, ACC: 73.2%, NMI: 83.8%). *Fourth row*: Laplacian $K$-modes with homotopy ($\lambda=0.01$, ACC: 81.5%, NMI: 88.0%) in $\sigma$. *Fourth row*: mean-shift ($\sigma=0.1591$).[]{data-label="f:coil"}](COIL/COIL_centers_manikmodes_16.eps "fig:"){width="0.048\linewidth"} ![Centroids found by different algorithms on COIL–20. *First row*: $K$-means ($\lambda=0$, $\sigma\rightarrow\infty$, ACC: 64.8%, NMI: 73.5%). *Second row*: $K$-modes ($\lambda=0$, $\sigma=0.3$, ACC: 65.5%, NMI: 73.0%). *Third row*: Laplacian $K$-modes ($\lambda=0.01$, $\sigma=0.1$, ACC: 73.2%, NMI: 83.8%). *Fourth row*: Laplacian $K$-modes with homotopy ($\lambda=0.01$, ACC: 81.5%, NMI: 88.0%) in $\sigma$. *Fourth row*: mean-shift ($\sigma=0.1591$).[]{data-label="f:coil"}](COIL/COIL_centers_manikmodes_17.eps "fig:"){width="0.048\linewidth"} ![Centroids found by different algorithms on COIL–20. *First row*: $K$-means ($\lambda=0$, $\sigma\rightarrow\infty$, ACC: 64.8%, NMI: 73.5%). *Second row*: $K$-modes ($\lambda=0$, $\sigma=0.3$, ACC: 65.5%, NMI: 73.0%). *Third row*: Laplacian $K$-modes ($\lambda=0.01$, $\sigma=0.1$, ACC: 73.2%, NMI: 83.8%). *Fourth row*: Laplacian $K$-modes with homotopy ($\lambda=0.01$, ACC: 81.5%, NMI: 88.0%) in $\sigma$. *Fourth row*: mean-shift ($\sigma=0.1591$).[]{data-label="f:coil"}](COIL/COIL_centers_manikmodes_18.eps "fig:"){width="0.048\linewidth"} ![Centroids found by different algorithms on COIL–20. *First row*: $K$-means ($\lambda=0$, $\sigma\rightarrow\infty$, ACC: 64.8%, NMI: 73.5%). *Second row*: $K$-modes ($\lambda=0$, $\sigma=0.3$, ACC: 65.5%, NMI: 73.0%). *Third row*: Laplacian $K$-modes ($\lambda=0.01$, $\sigma=0.1$, ACC: 73.2%, NMI: 83.8%). *Fourth row*: Laplacian $K$-modes with homotopy ($\lambda=0.01$, ACC: 81.5%, NMI: 88.0%) in $\sigma$. *Fourth row*: mean-shift ($\sigma=0.1591$).[]{data-label="f:coil"}](COIL/COIL_centers_manikmodes_19.eps "fig:"){width="0.048\linewidth"} ![Centroids found by different algorithms on COIL–20. *First row*: $K$-means ($\lambda=0$, $\sigma\rightarrow\infty$, ACC: 64.8%, NMI: 73.5%). *Second row*: $K$-modes ($\lambda=0$, $\sigma=0.3$, ACC: 65.5%, NMI: 73.0%). *Third row*: Laplacian $K$-modes ($\lambda=0.01$, $\sigma=0.1$, ACC: 73.2%, NMI: 83.8%). *Fourth row*: Laplacian $K$-modes with homotopy ($\lambda=0.01$, ACC: 81.5%, NMI: 88.0%) in $\sigma$. *Fourth row*: mean-shift ($\sigma=0.1591$).[]{data-label="f:coil"}](COIL/COIL_centers_manikmodes_20.eps "fig:"){width="0.048\linewidth"}
\[-1ex\] ![Centroids found by different algorithms on COIL–20. *First row*: $K$-means ($\lambda=0$, $\sigma\rightarrow\infty$, ACC: 64.8%, NMI: 73.5%). *Second row*: $K$-modes ($\lambda=0$, $\sigma=0.3$, ACC: 65.5%, NMI: 73.0%). *Third row*: Laplacian $K$-modes ($\lambda=0.01$, $\sigma=0.1$, ACC: 73.2%, NMI: 83.8%). *Fourth row*: Laplacian $K$-modes with homotopy ($\lambda=0.01$, ACC: 81.5%, NMI: 88.0%) in $\sigma$. *Fourth row*: mean-shift ($\sigma=0.1591$).[]{data-label="f:coil"}](COIL/COIL_centers_manikmodes_homo_01.eps "fig:"){width="0.048\linewidth"} ![Centroids found by different algorithms on COIL–20. *First row*: $K$-means ($\lambda=0$, $\sigma\rightarrow\infty$, ACC: 64.8%, NMI: 73.5%). *Second row*: $K$-modes ($\lambda=0$, $\sigma=0.3$, ACC: 65.5%, NMI: 73.0%). *Third row*: Laplacian $K$-modes ($\lambda=0.01$, $\sigma=0.1$, ACC: 73.2%, NMI: 83.8%). *Fourth row*: Laplacian $K$-modes with homotopy ($\lambda=0.01$, ACC: 81.5%, NMI: 88.0%) in $\sigma$. *Fourth row*: mean-shift ($\sigma=0.1591$).[]{data-label="f:coil"}](COIL/COIL_centers_manikmodes_homo_02.eps "fig:"){width="0.048\linewidth"} ![Centroids found by different algorithms on COIL–20. *First row*: $K$-means ($\lambda=0$, $\sigma\rightarrow\infty$, ACC: 64.8%, NMI: 73.5%). *Second row*: $K$-modes ($\lambda=0$, $\sigma=0.3$, ACC: 65.5%, NMI: 73.0%). *Third row*: Laplacian $K$-modes ($\lambda=0.01$, $\sigma=0.1$, ACC: 73.2%, NMI: 83.8%). *Fourth row*: Laplacian $K$-modes with homotopy ($\lambda=0.01$, ACC: 81.5%, NMI: 88.0%) in $\sigma$. *Fourth row*: mean-shift ($\sigma=0.1591$).[]{data-label="f:coil"}](COIL/COIL_centers_manikmodes_homo_03.eps "fig:"){width="0.048\linewidth"} ![Centroids found by different algorithms on COIL–20. *First row*: $K$-means ($\lambda=0$, $\sigma\rightarrow\infty$, ACC: 64.8%, NMI: 73.5%). *Second row*: $K$-modes ($\lambda=0$, $\sigma=0.3$, ACC: 65.5%, NMI: 73.0%). *Third row*: Laplacian $K$-modes ($\lambda=0.01$, $\sigma=0.1$, ACC: 73.2%, NMI: 83.8%). *Fourth row*: Laplacian $K$-modes with homotopy ($\lambda=0.01$, ACC: 81.5%, NMI: 88.0%) in $\sigma$. *Fourth row*: mean-shift ($\sigma=0.1591$).[]{data-label="f:coil"}](COIL/COIL_centers_manikmodes_homo_04.eps "fig:"){width="0.048\linewidth"} ![Centroids found by different algorithms on COIL–20. *First row*: $K$-means ($\lambda=0$, $\sigma\rightarrow\infty$, ACC: 64.8%, NMI: 73.5%). *Second row*: $K$-modes ($\lambda=0$, $\sigma=0.3$, ACC: 65.5%, NMI: 73.0%). *Third row*: Laplacian $K$-modes ($\lambda=0.01$, $\sigma=0.1$, ACC: 73.2%, NMI: 83.8%). *Fourth row*: Laplacian $K$-modes with homotopy ($\lambda=0.01$, ACC: 81.5%, NMI: 88.0%) in $\sigma$. *Fourth row*: mean-shift ($\sigma=0.1591$).[]{data-label="f:coil"}](COIL/COIL_centers_manikmodes_homo_05.eps "fig:"){width="0.048\linewidth"} ![Centroids found by different algorithms on COIL–20. *First row*: $K$-means ($\lambda=0$, $\sigma\rightarrow\infty$, ACC: 64.8%, NMI: 73.5%). *Second row*: $K$-modes ($\lambda=0$, $\sigma=0.3$, ACC: 65.5%, NMI: 73.0%). *Third row*: Laplacian $K$-modes ($\lambda=0.01$, $\sigma=0.1$, ACC: 73.2%, NMI: 83.8%). *Fourth row*: Laplacian $K$-modes with homotopy ($\lambda=0.01$, ACC: 81.5%, NMI: 88.0%) in $\sigma$. *Fourth row*: mean-shift ($\sigma=0.1591$).[]{data-label="f:coil"}](COIL/COIL_centers_manikmodes_homo_06.eps "fig:"){width="0.048\linewidth"} ![Centroids found by different algorithms on COIL–20. *First row*: $K$-means ($\lambda=0$, $\sigma\rightarrow\infty$, ACC: 64.8%, NMI: 73.5%). *Second row*: $K$-modes ($\lambda=0$, $\sigma=0.3$, ACC: 65.5%, NMI: 73.0%). *Third row*: Laplacian $K$-modes ($\lambda=0.01$, $\sigma=0.1$, ACC: 73.2%, NMI: 83.8%). *Fourth row*: Laplacian $K$-modes with homotopy ($\lambda=0.01$, ACC: 81.5%, NMI: 88.0%) in $\sigma$. *Fourth row*: mean-shift ($\sigma=0.1591$).[]{data-label="f:coil"}](COIL/COIL_centers_manikmodes_homo_07.eps "fig:"){width="0.048\linewidth"} ![Centroids found by different algorithms on COIL–20. *First row*: $K$-means ($\lambda=0$, $\sigma\rightarrow\infty$, ACC: 64.8%, NMI: 73.5%). *Second row*: $K$-modes ($\lambda=0$, $\sigma=0.3$, ACC: 65.5%, NMI: 73.0%). *Third row*: Laplacian $K$-modes ($\lambda=0.01$, $\sigma=0.1$, ACC: 73.2%, NMI: 83.8%). *Fourth row*: Laplacian $K$-modes with homotopy ($\lambda=0.01$, ACC: 81.5%, NMI: 88.0%) in $\sigma$. *Fourth row*: mean-shift ($\sigma=0.1591$).[]{data-label="f:coil"}](COIL/COIL_centers_manikmodes_homo_08.eps "fig:"){width="0.048\linewidth"} ![Centroids found by different algorithms on COIL–20. *First row*: $K$-means ($\lambda=0$, $\sigma\rightarrow\infty$, ACC: 64.8%, NMI: 73.5%). *Second row*: $K$-modes ($\lambda=0$, $\sigma=0.3$, ACC: 65.5%, NMI: 73.0%). *Third row*: Laplacian $K$-modes ($\lambda=0.01$, $\sigma=0.1$, ACC: 73.2%, NMI: 83.8%). *Fourth row*: Laplacian $K$-modes with homotopy ($\lambda=0.01$, ACC: 81.5%, NMI: 88.0%) in $\sigma$. *Fourth row*: mean-shift ($\sigma=0.1591$).[]{data-label="f:coil"}](COIL/COIL_centers_manikmodes_homo_09.eps "fig:"){width="0.048\linewidth"} ![Centroids found by different algorithms on COIL–20. *First row*: $K$-means ($\lambda=0$, $\sigma\rightarrow\infty$, ACC: 64.8%, NMI: 73.5%). *Second row*: $K$-modes ($\lambda=0$, $\sigma=0.3$, ACC: 65.5%, NMI: 73.0%). *Third row*: Laplacian $K$-modes ($\lambda=0.01$, $\sigma=0.1$, ACC: 73.2%, NMI: 83.8%). *Fourth row*: Laplacian $K$-modes with homotopy ($\lambda=0.01$, ACC: 81.5%, NMI: 88.0%) in $\sigma$. *Fourth row*: mean-shift ($\sigma=0.1591$).[]{data-label="f:coil"}](COIL/COIL_centers_manikmodes_homo_10.eps "fig:"){width="0.048\linewidth"} ![Centroids found by different algorithms on COIL–20. *First row*: $K$-means ($\lambda=0$, $\sigma\rightarrow\infty$, ACC: 64.8%, NMI: 73.5%). *Second row*: $K$-modes ($\lambda=0$, $\sigma=0.3$, ACC: 65.5%, NMI: 73.0%). *Third row*: Laplacian $K$-modes ($\lambda=0.01$, $\sigma=0.1$, ACC: 73.2%, NMI: 83.8%). *Fourth row*: Laplacian $K$-modes with homotopy ($\lambda=0.01$, ACC: 81.5%, NMI: 88.0%) in $\sigma$. *Fourth row*: mean-shift ($\sigma=0.1591$).[]{data-label="f:coil"}](COIL/COIL_centers_manikmodes_homo_11.eps "fig:"){width="0.048\linewidth"} ![Centroids found by different algorithms on COIL–20. *First row*: $K$-means ($\lambda=0$, $\sigma\rightarrow\infty$, ACC: 64.8%, NMI: 73.5%). *Second row*: $K$-modes ($\lambda=0$, $\sigma=0.3$, ACC: 65.5%, NMI: 73.0%). *Third row*: Laplacian $K$-modes ($\lambda=0.01$, $\sigma=0.1$, ACC: 73.2%, NMI: 83.8%). *Fourth row*: Laplacian $K$-modes with homotopy ($\lambda=0.01$, ACC: 81.5%, NMI: 88.0%) in $\sigma$. *Fourth row*: mean-shift ($\sigma=0.1591$).[]{data-label="f:coil"}](COIL/COIL_centers_manikmodes_homo_12.eps "fig:"){width="0.048\linewidth"} ![Centroids found by different algorithms on COIL–20. *First row*: $K$-means ($\lambda=0$, $\sigma\rightarrow\infty$, ACC: 64.8%, NMI: 73.5%). *Second row*: $K$-modes ($\lambda=0$, $\sigma=0.3$, ACC: 65.5%, NMI: 73.0%). *Third row*: Laplacian $K$-modes ($\lambda=0.01$, $\sigma=0.1$, ACC: 73.2%, NMI: 83.8%). *Fourth row*: Laplacian $K$-modes with homotopy ($\lambda=0.01$, ACC: 81.5%, NMI: 88.0%) in $\sigma$. *Fourth row*: mean-shift ($\sigma=0.1591$).[]{data-label="f:coil"}](COIL/COIL_centers_manikmodes_homo_13.eps "fig:"){width="0.048\linewidth"} ![Centroids found by different algorithms on COIL–20. *First row*: $K$-means ($\lambda=0$, $\sigma\rightarrow\infty$, ACC: 64.8%, NMI: 73.5%). *Second row*: $K$-modes ($\lambda=0$, $\sigma=0.3$, ACC: 65.5%, NMI: 73.0%). *Third row*: Laplacian $K$-modes ($\lambda=0.01$, $\sigma=0.1$, ACC: 73.2%, NMI: 83.8%). *Fourth row*: Laplacian $K$-modes with homotopy ($\lambda=0.01$, ACC: 81.5%, NMI: 88.0%) in $\sigma$. *Fourth row*: mean-shift ($\sigma=0.1591$).[]{data-label="f:coil"}](COIL/COIL_centers_manikmodes_homo_14.eps "fig:"){width="0.048\linewidth"} ![Centroids found by different algorithms on COIL–20. *First row*: $K$-means ($\lambda=0$, $\sigma\rightarrow\infty$, ACC: 64.8%, NMI: 73.5%). *Second row*: $K$-modes ($\lambda=0$, $\sigma=0.3$, ACC: 65.5%, NMI: 73.0%). *Third row*: Laplacian $K$-modes ($\lambda=0.01$, $\sigma=0.1$, ACC: 73.2%, NMI: 83.8%). *Fourth row*: Laplacian $K$-modes with homotopy ($\lambda=0.01$, ACC: 81.5%, NMI: 88.0%) in $\sigma$. *Fourth row*: mean-shift ($\sigma=0.1591$).[]{data-label="f:coil"}](COIL/COIL_centers_manikmodes_homo_15.eps "fig:"){width="0.048\linewidth"} ![Centroids found by different algorithms on COIL–20. *First row*: $K$-means ($\lambda=0$, $\sigma\rightarrow\infty$, ACC: 64.8%, NMI: 73.5%). *Second row*: $K$-modes ($\lambda=0$, $\sigma=0.3$, ACC: 65.5%, NMI: 73.0%). *Third row*: Laplacian $K$-modes ($\lambda=0.01$, $\sigma=0.1$, ACC: 73.2%, NMI: 83.8%). *Fourth row*: Laplacian $K$-modes with homotopy ($\lambda=0.01$, ACC: 81.5%, NMI: 88.0%) in $\sigma$. *Fourth row*: mean-shift ($\sigma=0.1591$).[]{data-label="f:coil"}](COIL/COIL_centers_manikmodes_homo_16.eps "fig:"){width="0.048\linewidth"} ![Centroids found by different algorithms on COIL–20. *First row*: $K$-means ($\lambda=0$, $\sigma\rightarrow\infty$, ACC: 64.8%, NMI: 73.5%). *Second row*: $K$-modes ($\lambda=0$, $\sigma=0.3$, ACC: 65.5%, NMI: 73.0%). *Third row*: Laplacian $K$-modes ($\lambda=0.01$, $\sigma=0.1$, ACC: 73.2%, NMI: 83.8%). *Fourth row*: Laplacian $K$-modes with homotopy ($\lambda=0.01$, ACC: 81.5%, NMI: 88.0%) in $\sigma$. *Fourth row*: mean-shift ($\sigma=0.1591$).[]{data-label="f:coil"}](COIL/COIL_centers_manikmodes_homo_17.eps "fig:"){width="0.048\linewidth"} ![Centroids found by different algorithms on COIL–20. *First row*: $K$-means ($\lambda=0$, $\sigma\rightarrow\infty$, ACC: 64.8%, NMI: 73.5%). *Second row*: $K$-modes ($\lambda=0$, $\sigma=0.3$, ACC: 65.5%, NMI: 73.0%). *Third row*: Laplacian $K$-modes ($\lambda=0.01$, $\sigma=0.1$, ACC: 73.2%, NMI: 83.8%). *Fourth row*: Laplacian $K$-modes with homotopy ($\lambda=0.01$, ACC: 81.5%, NMI: 88.0%) in $\sigma$. *Fourth row*: mean-shift ($\sigma=0.1591$).[]{data-label="f:coil"}](COIL/COIL_centers_manikmodes_homo_18.eps "fig:"){width="0.048\linewidth"} ![Centroids found by different algorithms on COIL–20. *First row*: $K$-means ($\lambda=0$, $\sigma\rightarrow\infty$, ACC: 64.8%, NMI: 73.5%). *Second row*: $K$-modes ($\lambda=0$, $\sigma=0.3$, ACC: 65.5%, NMI: 73.0%). *Third row*: Laplacian $K$-modes ($\lambda=0.01$, $\sigma=0.1$, ACC: 73.2%, NMI: 83.8%). *Fourth row*: Laplacian $K$-modes with homotopy ($\lambda=0.01$, ACC: 81.5%, NMI: 88.0%) in $\sigma$. *Fourth row*: mean-shift ($\sigma=0.1591$).[]{data-label="f:coil"}](COIL/COIL_centers_manikmodes_homo_19.eps "fig:"){width="0.048\linewidth"} ![Centroids found by different algorithms on COIL–20. *First row*: $K$-means ($\lambda=0$, $\sigma\rightarrow\infty$, ACC: 64.8%, NMI: 73.5%). *Second row*: $K$-modes ($\lambda=0$, $\sigma=0.3$, ACC: 65.5%, NMI: 73.0%). *Third row*: Laplacian $K$-modes ($\lambda=0.01$, $\sigma=0.1$, ACC: 73.2%, NMI: 83.8%). *Fourth row*: Laplacian $K$-modes with homotopy ($\lambda=0.01$, ACC: 81.5%, NMI: 88.0%) in $\sigma$. *Fourth row*: mean-shift ($\sigma=0.1591$).[]{data-label="f:coil"}](COIL/COIL_centers_manikmodes_homo_20.eps "fig:"){width="0.048\linewidth"}
\[-1ex\] ![Centroids found by different algorithms on COIL–20. *First row*: $K$-means ($\lambda=0$, $\sigma\rightarrow\infty$, ACC: 64.8%, NMI: 73.5%). *Second row*: $K$-modes ($\lambda=0$, $\sigma=0.3$, ACC: 65.5%, NMI: 73.0%). *Third row*: Laplacian $K$-modes ($\lambda=0.01$, $\sigma=0.1$, ACC: 73.2%, NMI: 83.8%). *Fourth row*: Laplacian $K$-modes with homotopy ($\lambda=0.01$, ACC: 81.5%, NMI: 88.0%) in $\sigma$. *Fourth row*: mean-shift ($\sigma=0.1591$).[]{data-label="f:coil"}](COIL/COIL_centers_ms_01.eps "fig:"){width="0.048\linewidth"} ![Centroids found by different algorithms on COIL–20. *First row*: $K$-means ($\lambda=0$, $\sigma\rightarrow\infty$, ACC: 64.8%, NMI: 73.5%). *Second row*: $K$-modes ($\lambda=0$, $\sigma=0.3$, ACC: 65.5%, NMI: 73.0%). *Third row*: Laplacian $K$-modes ($\lambda=0.01$, $\sigma=0.1$, ACC: 73.2%, NMI: 83.8%). *Fourth row*: Laplacian $K$-modes with homotopy ($\lambda=0.01$, ACC: 81.5%, NMI: 88.0%) in $\sigma$. *Fourth row*: mean-shift ($\sigma=0.1591$).[]{data-label="f:coil"}](COIL/COIL_centers_ms_02.eps "fig:"){width="0.048\linewidth"} ![Centroids found by different algorithms on COIL–20. *First row*: $K$-means ($\lambda=0$, $\sigma\rightarrow\infty$, ACC: 64.8%, NMI: 73.5%). *Second row*: $K$-modes ($\lambda=0$, $\sigma=0.3$, ACC: 65.5%, NMI: 73.0%). *Third row*: Laplacian $K$-modes ($\lambda=0.01$, $\sigma=0.1$, ACC: 73.2%, NMI: 83.8%). *Fourth row*: Laplacian $K$-modes with homotopy ($\lambda=0.01$, ACC: 81.5%, NMI: 88.0%) in $\sigma$. *Fourth row*: mean-shift ($\sigma=0.1591$).[]{data-label="f:coil"}](COIL/COIL_centers_ms_03.eps "fig:"){width="0.048\linewidth"} ![Centroids found by different algorithms on COIL–20. *First row*: $K$-means ($\lambda=0$, $\sigma\rightarrow\infty$, ACC: 64.8%, NMI: 73.5%). *Second row*: $K$-modes ($\lambda=0$, $\sigma=0.3$, ACC: 65.5%, NMI: 73.0%). *Third row*: Laplacian $K$-modes ($\lambda=0.01$, $\sigma=0.1$, ACC: 73.2%, NMI: 83.8%). *Fourth row*: Laplacian $K$-modes with homotopy ($\lambda=0.01$, ACC: 81.5%, NMI: 88.0%) in $\sigma$. *Fourth row*: mean-shift ($\sigma=0.1591$).[]{data-label="f:coil"}](COIL/COIL_centers_ms_04.eps "fig:"){width="0.048\linewidth"} ![Centroids found by different algorithms on COIL–20. *First row*: $K$-means ($\lambda=0$, $\sigma\rightarrow\infty$, ACC: 64.8%, NMI: 73.5%). *Second row*: $K$-modes ($\lambda=0$, $\sigma=0.3$, ACC: 65.5%, NMI: 73.0%). *Third row*: Laplacian $K$-modes ($\lambda=0.01$, $\sigma=0.1$, ACC: 73.2%, NMI: 83.8%). *Fourth row*: Laplacian $K$-modes with homotopy ($\lambda=0.01$, ACC: 81.5%, NMI: 88.0%) in $\sigma$. *Fourth row*: mean-shift ($\sigma=0.1591$).[]{data-label="f:coil"}](COIL/COIL_centers_ms_05.eps "fig:"){width="0.048\linewidth"} ![Centroids found by different algorithms on COIL–20. *First row*: $K$-means ($\lambda=0$, $\sigma\rightarrow\infty$, ACC: 64.8%, NMI: 73.5%). *Second row*: $K$-modes ($\lambda=0$, $\sigma=0.3$, ACC: 65.5%, NMI: 73.0%). *Third row*: Laplacian $K$-modes ($\lambda=0.01$, $\sigma=0.1$, ACC: 73.2%, NMI: 83.8%). *Fourth row*: Laplacian $K$-modes with homotopy ($\lambda=0.01$, ACC: 81.5%, NMI: 88.0%) in $\sigma$. *Fourth row*: mean-shift ($\sigma=0.1591$).[]{data-label="f:coil"}](COIL/COIL_centers_ms_06.eps "fig:"){width="0.048\linewidth"} ![Centroids found by different algorithms on COIL–20. *First row*: $K$-means ($\lambda=0$, $\sigma\rightarrow\infty$, ACC: 64.8%, NMI: 73.5%). *Second row*: $K$-modes ($\lambda=0$, $\sigma=0.3$, ACC: 65.5%, NMI: 73.0%). *Third row*: Laplacian $K$-modes ($\lambda=0.01$, $\sigma=0.1$, ACC: 73.2%, NMI: 83.8%). *Fourth row*: Laplacian $K$-modes with homotopy ($\lambda=0.01$, ACC: 81.5%, NMI: 88.0%) in $\sigma$. *Fourth row*: mean-shift ($\sigma=0.1591$).[]{data-label="f:coil"}](COIL/COIL_centers_ms_07.eps "fig:"){width="0.048\linewidth"} ![Centroids found by different algorithms on COIL–20. *First row*: $K$-means ($\lambda=0$, $\sigma\rightarrow\infty$, ACC: 64.8%, NMI: 73.5%). *Second row*: $K$-modes ($\lambda=0$, $\sigma=0.3$, ACC: 65.5%, NMI: 73.0%). *Third row*: Laplacian $K$-modes ($\lambda=0.01$, $\sigma=0.1$, ACC: 73.2%, NMI: 83.8%). *Fourth row*: Laplacian $K$-modes with homotopy ($\lambda=0.01$, ACC: 81.5%, NMI: 88.0%) in $\sigma$. *Fourth row*: mean-shift ($\sigma=0.1591$).[]{data-label="f:coil"}](COIL/COIL_centers_ms_08.eps "fig:"){width="0.048\linewidth"} ![Centroids found by different algorithms on COIL–20. *First row*: $K$-means ($\lambda=0$, $\sigma\rightarrow\infty$, ACC: 64.8%, NMI: 73.5%). *Second row*: $K$-modes ($\lambda=0$, $\sigma=0.3$, ACC: 65.5%, NMI: 73.0%). *Third row*: Laplacian $K$-modes ($\lambda=0.01$, $\sigma=0.1$, ACC: 73.2%, NMI: 83.8%). *Fourth row*: Laplacian $K$-modes with homotopy ($\lambda=0.01$, ACC: 81.5%, NMI: 88.0%) in $\sigma$. *Fourth row*: mean-shift ($\sigma=0.1591$).[]{data-label="f:coil"}](COIL/COIL_centers_ms_09.eps "fig:"){width="0.048\linewidth"} ![Centroids found by different algorithms on COIL–20. *First row*: $K$-means ($\lambda=0$, $\sigma\rightarrow\infty$, ACC: 64.8%, NMI: 73.5%). *Second row*: $K$-modes ($\lambda=0$, $\sigma=0.3$, ACC: 65.5%, NMI: 73.0%). *Third row*: Laplacian $K$-modes ($\lambda=0.01$, $\sigma=0.1$, ACC: 73.2%, NMI: 83.8%). *Fourth row*: Laplacian $K$-modes with homotopy ($\lambda=0.01$, ACC: 81.5%, NMI: 88.0%) in $\sigma$. *Fourth row*: mean-shift ($\sigma=0.1591$).[]{data-label="f:coil"}](COIL/COIL_centers_ms_10.eps "fig:"){width="0.048\linewidth"} ![Centroids found by different algorithms on COIL–20. *First row*: $K$-means ($\lambda=0$, $\sigma\rightarrow\infty$, ACC: 64.8%, NMI: 73.5%). *Second row*: $K$-modes ($\lambda=0$, $\sigma=0.3$, ACC: 65.5%, NMI: 73.0%). *Third row*: Laplacian $K$-modes ($\lambda=0.01$, $\sigma=0.1$, ACC: 73.2%, NMI: 83.8%). *Fourth row*: Laplacian $K$-modes with homotopy ($\lambda=0.01$, ACC: 81.5%, NMI: 88.0%) in $\sigma$. *Fourth row*: mean-shift ($\sigma=0.1591$).[]{data-label="f:coil"}](COIL/COIL_centers_ms_11.eps "fig:"){width="0.048\linewidth"} ![Centroids found by different algorithms on COIL–20. *First row*: $K$-means ($\lambda=0$, $\sigma\rightarrow\infty$, ACC: 64.8%, NMI: 73.5%). *Second row*: $K$-modes ($\lambda=0$, $\sigma=0.3$, ACC: 65.5%, NMI: 73.0%). *Third row*: Laplacian $K$-modes ($\lambda=0.01$, $\sigma=0.1$, ACC: 73.2%, NMI: 83.8%). *Fourth row*: Laplacian $K$-modes with homotopy ($\lambda=0.01$, ACC: 81.5%, NMI: 88.0%) in $\sigma$. *Fourth row*: mean-shift ($\sigma=0.1591$).[]{data-label="f:coil"}](COIL/COIL_centers_ms_12.eps "fig:"){width="0.048\linewidth"} ![Centroids found by different algorithms on COIL–20. *First row*: $K$-means ($\lambda=0$, $\sigma\rightarrow\infty$, ACC: 64.8%, NMI: 73.5%). *Second row*: $K$-modes ($\lambda=0$, $\sigma=0.3$, ACC: 65.5%, NMI: 73.0%). *Third row*: Laplacian $K$-modes ($\lambda=0.01$, $\sigma=0.1$, ACC: 73.2%, NMI: 83.8%). *Fourth row*: Laplacian $K$-modes with homotopy ($\lambda=0.01$, ACC: 81.5%, NMI: 88.0%) in $\sigma$. *Fourth row*: mean-shift ($\sigma=0.1591$).[]{data-label="f:coil"}](COIL/COIL_centers_ms_13.eps "fig:"){width="0.048\linewidth"} ![Centroids found by different algorithms on COIL–20. *First row*: $K$-means ($\lambda=0$, $\sigma\rightarrow\infty$, ACC: 64.8%, NMI: 73.5%). *Second row*: $K$-modes ($\lambda=0$, $\sigma=0.3$, ACC: 65.5%, NMI: 73.0%). *Third row*: Laplacian $K$-modes ($\lambda=0.01$, $\sigma=0.1$, ACC: 73.2%, NMI: 83.8%). *Fourth row*: Laplacian $K$-modes with homotopy ($\lambda=0.01$, ACC: 81.5%, NMI: 88.0%) in $\sigma$. *Fourth row*: mean-shift ($\sigma=0.1591$).[]{data-label="f:coil"}](COIL/COIL_centers_ms_14.eps "fig:"){width="0.048\linewidth"} ![Centroids found by different algorithms on COIL–20. *First row*: $K$-means ($\lambda=0$, $\sigma\rightarrow\infty$, ACC: 64.8%, NMI: 73.5%). *Second row*: $K$-modes ($\lambda=0$, $\sigma=0.3$, ACC: 65.5%, NMI: 73.0%). *Third row*: Laplacian $K$-modes ($\lambda=0.01$, $\sigma=0.1$, ACC: 73.2%, NMI: 83.8%). *Fourth row*: Laplacian $K$-modes with homotopy ($\lambda=0.01$, ACC: 81.5%, NMI: 88.0%) in $\sigma$. *Fourth row*: mean-shift ($\sigma=0.1591$).[]{data-label="f:coil"}](COIL/COIL_centers_ms_15.eps "fig:"){width="0.048\linewidth"} ![Centroids found by different algorithms on COIL–20. *First row*: $K$-means ($\lambda=0$, $\sigma\rightarrow\infty$, ACC: 64.8%, NMI: 73.5%). *Second row*: $K$-modes ($\lambda=0$, $\sigma=0.3$, ACC: 65.5%, NMI: 73.0%). *Third row*: Laplacian $K$-modes ($\lambda=0.01$, $\sigma=0.1$, ACC: 73.2%, NMI: 83.8%). *Fourth row*: Laplacian $K$-modes with homotopy ($\lambda=0.01$, ACC: 81.5%, NMI: 88.0%) in $\sigma$. *Fourth row*: mean-shift ($\sigma=0.1591$).[]{data-label="f:coil"}](COIL/COIL_centers_ms_16.eps "fig:"){width="0.048\linewidth"} ![Centroids found by different algorithms on COIL–20. *First row*: $K$-means ($\lambda=0$, $\sigma\rightarrow\infty$, ACC: 64.8%, NMI: 73.5%). *Second row*: $K$-modes ($\lambda=0$, $\sigma=0.3$, ACC: 65.5%, NMI: 73.0%). *Third row*: Laplacian $K$-modes ($\lambda=0.01$, $\sigma=0.1$, ACC: 73.2%, NMI: 83.8%). *Fourth row*: Laplacian $K$-modes with homotopy ($\lambda=0.01$, ACC: 81.5%, NMI: 88.0%) in $\sigma$. *Fourth row*: mean-shift ($\sigma=0.1591$).[]{data-label="f:coil"}](COIL/COIL_centers_ms_17.eps "fig:"){width="0.048\linewidth"} ![Centroids found by different algorithms on COIL–20. *First row*: $K$-means ($\lambda=0$, $\sigma\rightarrow\infty$, ACC: 64.8%, NMI: 73.5%). *Second row*: $K$-modes ($\lambda=0$, $\sigma=0.3$, ACC: 65.5%, NMI: 73.0%). *Third row*: Laplacian $K$-modes ($\lambda=0.01$, $\sigma=0.1$, ACC: 73.2%, NMI: 83.8%). *Fourth row*: Laplacian $K$-modes with homotopy ($\lambda=0.01$, ACC: 81.5%, NMI: 88.0%) in $\sigma$. *Fourth row*: mean-shift ($\sigma=0.1591$).[]{data-label="f:coil"}](COIL/COIL_centers_ms_18.eps "fig:"){width="0.048\linewidth"} ![Centroids found by different algorithms on COIL–20. *First row*: $K$-means ($\lambda=0$, $\sigma\rightarrow\infty$, ACC: 64.8%, NMI: 73.5%). *Second row*: $K$-modes ($\lambda=0$, $\sigma=0.3$, ACC: 65.5%, NMI: 73.0%). *Third row*: Laplacian $K$-modes ($\lambda=0.01$, $\sigma=0.1$, ACC: 73.2%, NMI: 83.8%). *Fourth row*: Laplacian $K$-modes with homotopy ($\lambda=0.01$, ACC: 81.5%, NMI: 88.0%) in $\sigma$. *Fourth row*: mean-shift ($\sigma=0.1591$).[]{data-label="f:coil"}](COIL/COIL_centers_ms_19.eps "fig:"){width="0.048\linewidth"} ![Centroids found by different algorithms on COIL–20. *First row*: $K$-means ($\lambda=0$, $\sigma\rightarrow\infty$, ACC: 64.8%, NMI: 73.5%). *Second row*: $K$-modes ($\lambda=0$, $\sigma=0.3$, ACC: 65.5%, NMI: 73.0%). *Third row*: Laplacian $K$-modes ($\lambda=0.01$, $\sigma=0.1$, ACC: 73.2%, NMI: 83.8%). *Fourth row*: Laplacian $K$-modes with homotopy ($\lambda=0.01$, ACC: 81.5%, NMI: 88.0%) in $\sigma$. *Fourth row*: mean-shift ($\sigma=0.1591$).[]{data-label="f:coil"}](COIL/COIL_centers_ms_20.eps "fig:"){width="0.048\linewidth"}
---------- -------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- -------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- -------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- -------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- -------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- -------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- -------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- -------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- -------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- -------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- -------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- -------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- -------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- -------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- -------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- -------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- -------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- -------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- -------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- --------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
Another key advantage of Laplacian $K$-modes is that the centroids are interpretable patterns of the dataset. We show the centroids (each as an image) found by centroid-based algorithms (using optimal hyperparameters) on MNIST in Fig. \[f:mnist\] and COIL–20 in Fig. \[f:coil\], all using the $K$-means initialization. For such high-dimensional problems with small $K$, GMS tends to have a majority of centroids associated with very few points that are outliers with unusual patterns, and we do not show them here. Not surprisingly, some $K$-means centroids are blurry images consisting of an average of digits/objects of different identity and style. This implies some centroids lie between different branches of data manifolds, thus in low-density areas and not prototypical. The $K$-modes centroids have cleaner shapes, but the identities of the different centroids somewhat overlap, so that some classes are represented by multiple centroids, at the expense of other classes, which may be represented by no centroids. This is because $K$-modes only takes into account the kde’s, and classes with higher sample density will receive more centroids. However, this will result in a class being abruptly partitioned, and the smoothness term in Laplacian $K$-modes prevents this, so its centroids not only have prototypical shapes, but also cover more digit/object identities. On COIL-20, starting from the $K$-means solution, and by slowly varying the hyperparameter $\sigma$, the centroids gradually move to high-density areas under the homotopy algorithm, being more prototypical, while the assignments separate different objects better.
Applying our algorithm at an intermediate $\sigma$ achieves just the right amount of smoothing. This is clearly seen from the centroids obtained on MNIST. It allows the centroids to look like valid digit images, but at the same time to average out noise, unusual strokes or other idiosyncrasies of the dataset images (while not averaging digits of different identities or different styles, as $K$-means does). This yields centroids that are more representative even than individual images of the dataset.
Conclusion
==========
Our Laplacian $K$-modes algorithm enjoys some of the best properties of a range of clustering algorithms. It is nonparametric and allows the user to work with a kernel density estimate that produces exactly $K$ clusters (as in $K$-means and $K$-modes), even in high dimension (unlike in mean-shift), and which can be nonconvex (as in mean-shift and spectral clustering). It also finds centroids that are valid patterns and lie in high-density areas, are representative of their cluster and neighborhood, yet they average out noise or idiosyncrasies that exist in individual data points. Computationally, our current alternating optimization scheme is simple, efficient and scales well. Experiments demonstrate the superior performance of Laplacian $K$-modes compared to well-known algorithms.
[33]{} \[1\][\#1]{} \[1\][`#1`]{} urlstyle \[1\][doi: \#1]{}
R. Arora, M. Gupta, A. Kapila, and M. Fazel. Clustering by left-stochastic matrix factorization. In L. Getoor and T. Scheffer, editors, *Proc. of the 28th Int. Conf. Machine Learning (ICML 2011)*, pages 761–768, Bellevue, WA, June 28 – July 2 2011.
D. Arthur and S. Vassilvitskii. `k-means++`: The advantages of careful seeding. In *Proc. of the 18th ACM-SIAM Symposium on Discrete Algorithms (SODA 2007)*, pages 1027–1035, New Orleans, LA, Jan. 7–9 2007.
A. Beck and M. Teboulle. A fast iterative shrinkage-thresholding algorithm for linear inverse problems. *SIAM J. Imaging Sciences*, 20 (1):0 183–202, 2009.
M. Belkin, P. Niyogi, and V. Sindhwani. Manifold regularization: A geometric framework for learning from labeled and unlabeled examples. *J. Machine Learning Research*, 7:0 2399–2434, Nov. 2006.
C. M. Bishop. *Pattern Recognition and Machine Learning*. Springer Series in Information Science and Statistics. Springer-Verlag, Berlin, 2006.
D. Cai, X. He, J. Han, and T. S. Huang. Graph regularized nonnegative matrix factorization for data representation. *IEEE Trans. Pattern Analysis and Machine Intelligence*, 330 (8):0 1548–1560, Aug. 2011.
M. [Á]{}. Carreira-Perpi[ñ]{}[á]{}n. Mode-finding for mixtures of [Gaussian]{} distributions. *IEEE Trans. Pattern Analysis and Machine Intelligence*, 220 (11):0 1318–1323, Nov. 2000.
M. [Á]{}. Carreira-Perpi[ñ]{}[á]{}n. Acceleration strategies for [Gaussian]{} mean-shift image segmentation. In C. Schmid, S. Soatto, and C. Tomasi, editors, *Proc. of the 2006 IEEE Computer Society Conf. Computer Vision and Pattern Recognition (CVPR’06)*, pages 1160–1167, New York, NY, June 17–22 2006.
M. [Á]{}. Carreira-Perpi[ñ]{}[á]{}n. Gaussian mean shift is an [EM]{} algorithm. *IEEE Trans. Pattern Analysis and Machine Intelligence*, 290 (5):0 767–776, May 2007.
M. [Á]{}. Carreira-Perpi[ñ]{}[á]{}n and W. Wang. The [$K$]{}-modes algorithm for clustering. Unpublished manuscript, arXiv:1304.6478, Apr. 23 2013.
M. [Á]{}. Carreira-Perpi[ñ]{}[á]{}n and R. S. Zemel. Proximity graphs for clustering and manifold learning. In L. K. Saul, Y. Weiss, and L. Bottou, editors, *Advances in Neural Information Processing Systems (NIPS)*, volume 17, pages 225–232. MIT Press, Cambridge, MA, 2005.
Y. Cheng. Mean shift, mode seeking, and clustering. *IEEE Trans. Pattern Analysis and Machine Intelligence*, 170 (8):0 790–799, Aug. 1995.
C. Chennubhotla and A. Jepson. : Half-lives of [EigenFlows]{} for spectral clustering. In S. Becker, S. Thrun, and K. Obermayer, editors, *Advances in Neural Information Processing Systems (NIPS)*, volume 15, pages 705–712. MIT Press, Cambridge, MA, 2003.
D. Comaniciu and P. Meer. Mean shift: A robust approach toward feature space analysis. *IEEE Trans. Pattern Analysis and Machine Intelligence*, 240 (5):0 603–619, May 2002.
J. Duchi, S. Shalev-Shwartz, Y. Singer, and T. Chandra. Efficient projections onto the $\ell_1$-ball for learning in high dimensions. In A. [McCallum]{} and S. Roweis, editors, *Proc. of the 25th Int. Conf. Machine Learning (ICML’08)*, pages 272–279, Helsinki, Finland, July 5–9 2008.
K. Fukunaga and L. D. Hostetler. The estimation of the gradient of a density function, with application in pattern recognition. *IEEE Trans. Information Theory*, IT–210 (1):0 32–40, Jan. 1975.
T. J. Hastie, R. J. Tibshirani, and J. H. Friedman. *The Elements of Statistical Learning—Data Mining, Inference and Prediction*. Springer Series in Statistics. Springer-Verlag, second edition, 2009.
G. Hinton and S. T. Roweis. Stochastic neighbor embedding. In S. Becker, S. Thrun, and K. Obermayer, editors, *Advances in Neural Information Processing Systems (NIPS)*, volume 15, pages 857–864. MIT Press, Cambridge, MA, 2003.
L. Kaufman and P. J. Rousseeuw. *Finding Groups in Data: An Introduction to Cluster Analysis*. Wiley Series in Probability and Mathematical Statistics. John Wiley & Sons, 1990.
Y. [LeCun]{}, L. Bottou, Y. Bengio, and P. Haffner. Gradient-based learning applied to document recognition. *Proc. IEEE*, 860 (11):0 2278–2324, Nov. 1998.
D. D. Lee and H. S. Seung. Learning the parts of objects by non-negative matrix factorization. *Nature*, 4010 (6755):0 788–791, Oct. 21 1999.
W. Liu, J. He, and S.-F. Chang. Large graph construction for scalable semi-supervised learning. In J. F[ü]{}rnkranz and T. Joachims, editors, *Proc. of the 27th Int. Conf. Machine Learning (ICML 2010)*, Haifa, Israel, June 21–25 2010.
S. A. Nene, S. K. Nayar, and H. Murase. Columbia object image library [(COIL-20)]{}. Technical Report , Dept. of Computer Science, Columbia University, Feb. 1996.
J. Shi and J. Malik. Normalized cuts and image segmentation. *IEEE Trans. Pattern Analysis and Machine Intelligence*, 220 (8):0 888–905, Aug. 2000.
M. Vladymyrov and M. [Á]{}. Carreira-Perpi[ñ]{}[á]{}n. Entropic affinities: Properties and efficient numerical computation. In S. Dasgupta and D. [McAllester]{}, editors, *Proc. of the 30th Int. Conf. Machine Learning (ICML 2013)*, pages 477–485, Atlanta, GA, June 16–21 2013.
U. von Luxburg. A tutorial on spectral clustering. *Statistics and Computing*, 170 (4):0 395–416, Dec. 2007.
M. P. Wand and M. C. Jones. *Kernel Smoothing*. Number 60 in Monographs on Statistics and Applied Probability. Chapman & Hall, London, New York, 1994.
W. Wang and M. [Á]{}. Carreira-Perpi[ñ]{}[á]{}n. Projection onto the probability simplex: An efficient algorithm with a simple proof, and an application. Unpublished manuscript, arXiv:1309.1541, Sept. 3 2013.
Z. Yang and E. Oja. Clustering by low-rank doubly stochastic matrix decomposition. In J. Langford and J. Pineau, editors, *Proc. of the 29th Int. Conf. Machine Learning (ICML 2012)*, pages 831–838, Edinburgh, Scotland, June 26 – July 1 2012.
S. X. Yu and J. Shi. Multiclass spectral clustering. In *Proc. 9th Int. Conf. Computer Vision (ICCV’03)*, pages 313–319, Nice, France, Oct. 14–17 2003.
X. Yuan, B.-G. Hu, and R. He. Agglomerative mean-shift clustering. *IEEE Trans. Knowledge and Data Engineering*, 240 (2):0 209–219, Feb. 2010.
L. Zelnik-Manor and P. Perona. Self-tuning spectral clustering. In L. K. Saul, Y. Weiss, and L. Bottou, editors, *Advances in Neural Information Processing Systems (NIPS)*, volume 17, pages 1601–1608. MIT Press, Cambridge, MA, 2005.
X. Zhu, Z. Ghahramani, and J. Lafferty. Semi-supervised learning using [Gaussian]{} fields and harmonic functions. In T. Fawcett and N. Mishra, editors, *Proc. of the 20th Int. Conf. Machine Learning (ICML’03)*, pages 912–919, Washington, DC, Aug. 21–24 2003.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'The current understanding of radiation tolerance of Silicon Photomultipliers (SiPMs) is reviewed. Radiation damage in silicon sensors is briefly introduced, surface and bulk effects are separately addressed. Results on the operation of irradiated SiPMs with X-ray, gamma, electron, proton and neutron sources are presented. The most critical effect of radiation on SiPMs is the increase of dark count rate, which makes it impossible to resolve signals generated by a single photon from the noise. Methods to characterize irradiated SiPMs after their single photo-electron resolution is lost are discussed. Due to the important similarity in the operation below the breakdown voltage, also studies on radiation damage of avalanche photo-diodes (APD) are reviewed. Finally, ideas are presented on how to approach the development of radiation hard SiPMs in the future.'
address:
- 'University of Hamburg, Hamburg, Germany'
- 'University of Notre Dame, Notre Dame, IN 46556, USA'
- 'Instutute for Nuclear Research RAS, pr. 60-letiya Oktyabrya 7a, 117312 Moscow, Russia'
author:
- 'E. Garutti'
- 'Yu. Musienko'
bibliography:
- 'bibliography.bib'
title: Radiation damage of SiPMs
---
Silicon photomultiplier ,radiation damage ,annealing
Introduction {#sec:intro}
============
This paper deals with the effects of radiation damage on SiPMs. As SiPMs detect single charge carriers, radiation damage is a major concern when operating these devices in harsh radiation environments (i.e. CMS and LHCb detectors at LHC, detectors at the proposed International Linear Collider (ILC), detectors for space experiments, etc.). Most of the experiments at lepton colliders or at lower energy machines as well as detectors for space and medical applications will receive fluences below 10$^{12}$ particles/cm$^2$ throughout their lifetime. New detectors for the upgrade of the LHC experiments demand to operate SiPMs up to fluences of $\sim$10$^{14}$ particles/cm$^2$. To understand how radiation can affect the operation of a SiPM it is necessary to first inspect its structure. A SiPM is a matrix of avalanche photo-diodes connected in parallel and operated above the breakdown voltage, in Geiger mode. For a detailed description of the working principle of a SiPM we refer the reader to [@Piemonte:thisVolume]. The single photo-diode will be referred to as pixel of the SiPM. Fig. \[Fig:SiPM-crosss\] schematically shows the cross section of two possible implementations of SiPM pixels. The left design is used for instance in the Hamamatsu MPPC, with a depth of the p-epitaxial layer of about 2 $\mu$m, corresponding approximately to the thickness of the multiplication region. In the right structure the p-n junction can be less deep and the multiplication region can be as thin as 1 $\mu$m or even less for special UV-sensitive designs. In both cases most of the bulk of the silicon material (denoted as “substrate”) is not depleted and only a fraction of the electric charges generated in this volume will reach the multiplication region by diffusion. In the following we first give a short summary of the effects of radiation on the silicon crystal, including the effect on the SiO$_2$. Then we review radiation damage caused by electromagnetically interacting particles (photons, electrons and positrons) and by hadrons (protons, neutrons) on SiPMs in Sec. \[sec:SiPM\]. In Sec. \[sec:apd\] we discuss radiation damage of Avalanche Photo-Diodes (APD), which have been more extensively studied than SiPMs in the past and may offer important insights for future SiPM studies. Finally, we present in Sec. \[sec:outlook\] a summary of the factors limiting SiPM operation in high radiation environment and a list of possible precautions and design considerations when developing a SiPM for these environments.
![Schematic cross-section of two possible single pixels of blue sensitive SiPMs. The poly-silicon quenching resistor $R_q$ and the Al-contact line are isolated from the silicon via a SiO$_2$ layer. A typical anti-reflecting coating material used for the entrance window is Si$_{3}$N$_{4}$.[]{data-label="Fig:SiPM-crosss"}](FIGURES/HPKsketch "fig:"){width="0.48\linewidth"} ![Schematic cross-section of two possible single pixels of blue sensitive SiPMs. The poly-silicon quenching resistor $R_q$ and the Al-contact line are isolated from the silicon via a SiO$_2$ layer. A typical anti-reflecting coating material used for the entrance window is Si$_{3}$N$_{4}$.[]{data-label="Fig:SiPM-crosss"}](FIGURES/Keteksketch "fig:"){width="0.44\linewidth"}
Radiation damage of silicon - a short summary {#sec:sirad}
==============================================
The effect of radiation in silicon detectors below the voltage at which avalanche multiplication becomes significant is very well studied and documented in dozens of books and papers, see for reference [@LUTZ1999; @LINDSTROM200330]. Here we present only a short review aimed to introduce the effects relevant to SiPMs, mainly below the breakdown voltage, $V_{bd}$. When discussing the effect of radiation to silicon detectors one has to distinguish between two types of damage: Bulk damage due to Non Ionizing Energy Loss (NIEL), and surface damage due to Ionizing Energy Loss (IEL).
Radiation damage due to NIEL {#sec:sirad-niel}
----------------------------
Bulk damage is primarily produced by high energy particles (protons, pions, electron and photons) and by neutrons, which can displace atoms out of their lattice site generating crystal defects. The minimum energy transfer required for one silicon atom displacement is $\sim$25 eV. This Primary Knock-on Atom (PKA) leaves a vacancy in the crystal lattice and relocates itself in a new position as interstitial among other atoms. If the kinetic energy of a PKA is sufficiently high ($>1$ keV) it can displace additional atoms leading to the creation of a cluster defect. For PKA energies $>12$ keV also multiple-cluster defects can form. Table \[tab1\] from Ref. [@LUTZ1999] presents the kinematic collision properties of 1 MeV particles in Si, including Si atoms as PKA. $T_{max}$ is the maximum- and $T_{av}$ is the average transferred energy. $E_{min}$ is the minimal energy needed to create a point or cluster defect. Looking at this table, one realizes that 1 MeV electrons will produce point defects but almost no clusters, whereas protons and neutrons can produce both type of defects.
Radiation e– p n Si+
------------------- ------- ------------- -------- -------
Interaction EM EM + strong strong EM
$T_{max}$ 0.155 133.7 133.9 1000
$T_{av}$ 0.046 0.210 50 0.265
$E_{min}$ point 260 0.190 0.190 0.025
$E_{min}$ cluster 4600 15 15 2
: Kinematic collision properties of 1 MeV particles in Si, from [@LUTZ1999]. Energies are in keV.[]{data-label="tab1"}
Interstitials and vacancies move inside the lattice and are very mobile above 150 K. Interstitials may annihilate with vacancies at a regular lattice position curing these defects or may diffuse out of the surface. This effect can be enhanced or sped up increasing the sensor temperature for a given time. This procedure is called annealing. Alternatively, dislocated atoms may combine with other defects and form stable secondary defects. These can be combinations of interstitials (I), vacancies (V) with C, O, P atoms, leading to permanent formation or removal of donors and acceptors, e.g.: VP, VO, Divacancy (V2), Trivacancy (V3).
According to the NIEL hypothesis the radiation damage is proportional to the non-ionizing energy loss of the penetrating particles (radiation) and this energy loss is again proportional to the energy used to dislocate lattice atoms (displacement energy). The NIEL hypothesis does not consider atom transformations nor annealing effects and is therefore not exact. Nevertheless, it is common to scale the damage effects of different particles using the NIEL hypothesis. However, different effects (leakage current, doping concentration, charge collection efficiency) require different NIEL coefficients to describe their scaling as a function of fluence. The damage functions for various particle types and energies are discussed in Ref. [@LINDSTROM200330]. It provides a hardness factor $\kappa$ for each particle type, allowing to compare the damage efficiency of radiation sources with different particles and energy spectra $\Phi(E)$. Macroscopically, the defects generated by radiation in the silicon crystal lead to changes of the detector performance, related to the newly introduced energy levels in the energy gap between valence and conduction band, $E_g \sim$1.12 eV[^1]. Depending on the position of the energy level and the capture cross sections, different effects can occur:
- Increase of leakage current:\
Defects with energy levels close to the middle of the band gap ($\sim$0.56 eV) facilitate the thermal excitation of electrons and holes, increasing the dark current generated by generation-recombination.
- Decrease of signal:\
Charged defects act as trapping centers. The deep-level ones, far from the band edges have long de-trapping times and release the trapped carriers too late to contribute to the signal formation.
- Change of effective doping density:\
Depending on their occupation, defect states contribute to the effective doping, and thus to the electric field in the amplification region. The occupation depends on the density of free charge carriers, on the dark current, and on the distance of the defect state from the band gap. In addition, the radiation removes dopant atoms by nuclear interactions. These effects have an impact on the depletion- and the breakdown voltage.
Fig. \[Fig:DefectLevels\] sketches the various positions of defects in the band gap of silicon and their macroscopic effects. Additionally, for devices operated in avalanche or Geiger mode, the change of the multiplication coefficient (or ionization coefficient) as function of fluence should also be considered, which is not properly discussed in the literature.
![Schematic of possible defects in the band gap of silicon and their macroscopic effect on the operation of a silicon sensor. []{data-label="Fig:DefectLevels"}](FIGURES/SRH_eff){width="0.7\linewidth"}
The change of leakage current with particle fluence has been extensively studied, see for instance [@Moll:1999kv] and references therein. A proportional dependence of the increase of leakage current with fluence has been reported by many authors and it is often described by the damage parameter $\alpha = \frac{\Delta I}{A\cdot w \cdot \Phi_{eq}} \sim 4 \cdot 10^{-17}$ A/cm, with $\Phi_{eq}$ being the fluence equivalent to 1 MeV neutrons, and $A\cdot w$ the depleted volume. The damage parameter $\alpha$ is often quoted at 20 $^{\circ}$C and after an annealing of 60 minutes at 80 $^{\circ}$C, for better comparison. The unwanted increase of leakage current can be mitigated by cooling the detector, or partially recovered by annealing procedures.
To understand the temperature dependence of the leakage current it is necessary to first outline the main contributing mechanisms: the diffusion of minority charge carriers from quasi neutral regions into the depleted region [@Haitz2006; @Pagano2012]; the generation of electron-hole pairs due to defects in the depletion region [@LUTZ1999], which are enhanced by a high electric field leading to the mechanism of trap-assisted tunneling or the Poole-Frenkel effect. For higher electric field strengths, direct band-to-band tunneling contributes to the generation of electron-hole pairs, as reported in [@Vincent1979; @Hurkx1992]. Which of the named mechanisms is dominating, depends on the electric field strength and operation temperature. The diffusion and generation currents have the following temperature dependence [@Grove1967]: $$I_{diff} \propto T^3 e^{-\frac{E_g}{kT}},$$ $$I_{gen} \propto T^2 e^{-\frac{E_a}{kT}}.
\label{Eq:Igen}$$ For the activation energy, $E_a$ a value of 0.605 eV is found by Chilingarov Ref. [@Chilin:2013], which using the SRH model corresponds to a trap energy $E_t= E_a + \frac{E_g}{2}$ = 45 meV from mid gap.
A practical approximation for the temperature dependence of the leakage current of a diode, in the absence of strong fields, i.e. amplification gain equal to one, is a decrease by a factor 2 for 8 degrees temperature reduction at room temperature.
For electric fields of the order of $10^5$ V/cm or higher Eq. \[Eq:Igen\] needs to be corrected by a trap-assisted tunneling term, $I_{gen+tat}$. The correction depends on the effective field strength, $F_{eff}$ and modifies Eq. \[Eq:Igen\], as $$I_{gen+tat} \propto \left(1+ \Gamma\right) T^2 e^{-\frac{E_{a}}{kT}},
\label{eq:gentat}$$ where $\Gamma \approx \frac{F_{eff}}{(kT)^{3/2}} e^{\left(\frac{F_{eff}}{(kT)^{3/2}}\right)^2} $ is the term defined by Hurkx in Ref. [@Hurkx1992:field], which accounts for the effects of tunneling.
The band-to-band tunneling current $I_{bbt}$ is independent of T, if the temperature dependence of the band-gap energy is neglected.
For a device operating at gain larger than unity, namely in linear amplification mode (like an APD) or in Geiger mode (like a SiPM), the bulk leakage current is also linearly proportional to the amplification gain.
How temperature affects the operation of SiPMs is exemplary demonstrated in Fig. \[Fig:DCRvsT\]. Here, the temperature dependence of the DCR of FBK SiPMs is shown using an Arrhenius-like plot. Two regions are clearly separated: at high temperatures (low $1/T$) thermal generation is the dominating mechanism responsible for the dark rate generation. At low temperature DCR saturates and becomes T-independent. The main contribution to DCR originates from tunneling. Two SiPM types are presented in the plot, which differ in the design of the electric field of the multiplication region. For lower electric field value, the magnitude of the tunneling component is highly suppressed.
![DCR for NUV-HD FBK SiPMs with standard field (triangular markers) and low field (circular markers) as function of function of 1/T and over-voltage. From Ref. [@Acerbi:2016ikf].[]{data-label="Fig:DCRvsT"}](FIGURES/DCRvsT){width="0.9\linewidth"}
Radiation damage due to IEL {#sec:sirad-iel}
---------------------------
Surface damage is primarily produced by photons and charged particles, generating charges in the oxide (SiO$_2$) and at the Si-SiO$_2$ interface, and interface traps at the Si-SiO$_2$ interface.
Photons with energies below 300 keV, which is the threshold energy for the formation of defects in the silicon bulk, generate only surface defects. The effects of X-ray radiation damage are reported by Barnaby in Ref [@Barnaby2006] and by Oldham in Ref [@Oldham1999].
In SiO$_2$, X-rays produce on average one electron-hole (e-h) pair every 18 eV of deposited energy. Depending on ionization density and electric field, a fraction of the e-h pairs recombine. The remaining charge carriers move in the SiO$_2$ by diffusion and, if an electric field is present, by drift. Most electrons, due to their high mobility and relatively low trapping probability, leave the SiO$_2$. However holes, which move via polaron hopping, are typically captured by deep traps in the SiO$_2$ or at the Si-SiO$_2$ interface, where interface traps are mainly formed by the depassivation of dangling bonds, resulting in fixed positive charge states and interface traps. We denote the surface density of oxide charges by $N_{ox}$, and the density of the Si-SiO$_2$ interface traps by $N_{it}$. The interface traps, if exposed to an electric field, act as generation centers for a surface current with density $J_{surf}$.
Results on $N_{ox}$ and $J_{surf}$ from MOS-Capacitors and Gate-Controlled-Diodes produced by different vendors and for different crystal orientations for X-ray doses between 10 kGy and 1 GGy are reported by Klanner [@Klanner2013] and Zhang [@Zhang:2013]. For a dose of 10 kGy the values for $N_{ox}$ are between 0.4$\cdot$10$^{12}$ and 1.2$\cdot$10$^{12}$ cm$^{-2}$, and for $J_{surf}$ between 0.1 and 1 $\mu$A cm$^{-2}$ at room temperature. Depending on technology and crystal orientation for doses of the order of 1 MGy the values of $N_{ox}$ and $J_{surf}$ saturate at 1.5-3.5$\cdot$10$^{12}$ cm$^{-2}$ and 2-6 $\mu$A cm$^{-2}$, respectively. Before irradiation typical values are a few 10$^{10}$ cm$^{-2}$ and a few nA/cm$^2$, respectively. We note that in addition to differences due to technology, the values of $N_{ox}$ and of $J_{surf}$ at a given dose depend on the value and the orientation of the electric field in the oxide, and that there are significant annealing effects as reported by Fretwurst, Lindstroem and Moll in Refs. [@Lindstrom:1999mw; @Moll:1999nh]. The depleted Si-SiO$_2$-interface areas generate surface currents, and therefore a significant increase in dark current below the breakdown voltage. If a fraction of the surface current reaches the amplification region it gets amplified and it increases also the dark-count rate. This however depends on the electric field distribution of the device.\
In conclusion, the total dark current $I_{dark}$ of a silicon device is the sum of the surface and the bulk currents. Relevant for SiPMs is to distinguish between the part of current (bulk and part of surface current) that results in a Geiger discharge and the part that does not.
Defect characterization {#sec:defects}
-----------------------
Various techniques are available for the microscopic investigation of defects in silicon. Often used ones are deep-level transient spectroscopy (DLTS), thermally stimulated current (TSC) techniques and IR-absorption spectroscopy.
![Left) TSC spectra for 200 $\mu$m thick p-type silicon pad diodes, after 23 MeV proton irradiation ($\Phi_{eq}=0.5\cdot 10^{14}$ cm$^{−2}$) and annealed for 8 minutes at T=80 $^{\circ}$C. From Ref. [@RADECSDonegani]. Right) TSC spectra, after irradiation by electrons and normalized to $\Phi_{eq}=10^{14}$ cm$^{−2}$, and annealed for 30 min at at T=80 $^{\circ}$C. The x-axis of the plots can be converted in energy from the conduction or valence band depending on the type of defect. High temperatures correspond to energies with large distance from the bands, i.e. close to mid gap energy. From Ref. [@DONEGANI201815].[]{data-label="Fig:RADECS"}](FIGURES/RADECS-Donegani.png "fig:"){width="0.49\linewidth"} ![Left) TSC spectra for 200 $\mu$m thick p-type silicon pad diodes, after 23 MeV proton irradiation ($\Phi_{eq}=0.5\cdot 10^{14}$ cm$^{−2}$) and annealed for 8 minutes at T=80 $^{\circ}$C. From Ref. [@RADECSDonegani]. Right) TSC spectra, after irradiation by electrons and normalized to $\Phi_{eq}=10^{14}$ cm$^{−2}$, and annealed for 30 min at at T=80 $^{\circ}$C. The x-axis of the plots can be converted in energy from the conduction or valence band depending on the type of defect. High temperatures correspond to energies with large distance from the bands, i.e. close to mid gap energy. From Ref. [@DONEGANI201815].[]{data-label="Fig:RADECS"}](FIGURES/TSC-electrons.png "fig:"){width="0.5\linewidth"}
Some typical results of defect spectroscopy on irradiated silicon diodes are shown in Fig. \[Fig:RADECS\]. The peaks in the spectrum correspond to specific electrically charged defects in the silicon band gap. The left plot compares the defects generated by proton irradiation in silicon materials (Magnetic Czochralski and Float-Zone). The right plot shows defects generated in a silicon diode irradiated with electrons of various energies. In Ref. [@RADECSDonegani] a correlation between the leakage current and the concentrations of three deep defects close to mid-gap (V$_2$, V$_3$ and H(220K)) is reported; and changes in the space charge in p-type sensors are correlated to the concentration of the donor E(30K), the acceptor BiOi, and of the three main deep acceptors (H(116K), H(140K) and H(152K)).
These techniques have been tested also on SiPMs. Due to large capacitance and high dopant concentration DLTS and TSC are not applicable for SiPMs. The DLTS sensitivity limit is about $N_{trap}=10^{-4} \cdot N_d$ with $N_d$ the doping concentration. For non-irradiated SiPMs the doping concentration is $N_{d} \leq 10^{16}$ cm$^{-3}$, so one would be sensitive to a trap concentration of $N_{trap}>10^{12}$ cm$^{-3}$, which is too high with respect to the expected trap concentration from radiation damage[^2]. Single pixel structures yielding a smaller dark current may be better suited for this characterization, but still the high doping concentration required for the multiplication region may be an obstacle. IR-spectroscopy measurements on SiPMs are not reported in the literature, but could be an alternative way to characterize generation of defects by radiation damage.
An interesting study of the position dependence of radiation damage in SiPM pixels has been presented by Barnyakov in Ref. [@Barnyakov:2016bwt]. The authors have investigated the radiation damage of digital SiPMs exposed to 800 MeV protons. In a digital SiPM, the DCR of every individual cell can be monitored separately, such that it is possible to generate plots like Fig. \[Fig:DigiSIPM\] where the DCR of single cells as function of the number of beam protons is shown. The step-like increase of the DCR indicates that a single interaction of a proton with a Si atom may result in a drastic DCR increase and that the increase may differ by orders of magnitude for each proton interaction. Most likely this effect is linked to the formation of cluster-like defects in one pixel.
![The dark count rate of individual cells of a Philips digital photon counter as a function of total accumulated fluence. From Ref. [@Barnyakov:2016bwt]. The x-axis is the total number of protons traversing the SiPM. The fluence for the highest irradiation was $\Phi_{protons} \sim 4\cdot 10^{11}$ cm$^{-2}$, and the hardness factor for this proton energy is $\kappa \sim$ 1.2.[]{data-label="Fig:DigiSIPM"}](FIGURES/DigitalSiPM.png){width="0.65\linewidth"}
An elegant way to visualize the effect of radiation damage is to observe the light emission in the dark from a SiPM biased above breakdown. This is based on the phenomenon that every Geiger discharge emits a certain number of optical and IR photons produced in the high field region[^3]. For randomly distributed DCR on the SiPM volume, the light emission is expected to be homogeneous. In the case of local defects in silicon, hotspots can form, which are more likely to generate Geiger avalanches in the dark.
![Light intensity images for a non-irradiated (a) and a neutron irradiated SiPM test structure, operated at $\Delta$V = 4 V. The irradiated SiPM was exposed to $\Phi_{eq} = 10^{10}$ cm$^{-2}$ and no annealing was applied. The effect of radiation in increasing the number of hot-spots is evident in these images. From [@Engelmann:phd].[]{data-label="Fig:Engelmann"}](FIGURES/Engelmann1){width="0.65\linewidth"}
Fig. \[Fig:Engelmann\] from Engelmann PhD thesis, [@Engelmann:phd], demonstrates very clearly the increase in the number of hot-spots due to the effects of radiation. A SiPM exposed to $\Phi_{eq} = 10^{10}$ cm$^{-2}$ is compared to a non-irradiated device of the same type. The color scale indicates the intensity of the light emitted. It can be seen that the light intensity emitted[^4] by the neutron induced defects (Fig. \[Fig:Engelmann\]b) is about two orders of magnitude larger than for ion implantation induced defects visible in the non-irradiated sample (Fig. \[Fig:Engelmann\]a). From this result the author concludes that the dark count rate generated by neutron induced crystal defects is two orders of magnitude larger than the dark count rate generated by implantation defects. However, one has to consider that the implantation defects are evaluated after an annealing process, whereas the neutron defects are evaluated without a significant annealing.
Annealing of radiation damage (defect kinetics) {#sec:sirad-annealing}
-----------------------------------------------
As discussed the reverse current of a diode without gain increases proportionally to the fluence of non-ionizing radiation. To compare data, the damage parameter $\alpha$ is normalized to a given temperature (usually 20 $^{\circ}$C). It is observed in Refs. [@Moll:1999kv; @Moll:1999nh] that $\alpha$ can be reduced by annealing, an effect which is referred to as beneficial annealing. Fig. \[Fig:annealing\] demonstrates that a factor 2-3 reduction of $\alpha$, and therefore of the leakage current of an irradiated diode, is attainable after 830 hours at 60 $^{\circ}$C. Permanent damage will not be cured by this procedure such that the leakage current level of the diode before irradiation can not be fully recovered.
![Annealing function of current-related damage parameter $\alpha$ for standard (solid line) and DOFZ (dots) silicon detectors, taken from [@LINDSTROM200330].[]{data-label="Fig:annealing"}](FIGURES/alpha-annealing.png){width="0.8\linewidth"}
A recommendation from Lindstroem [@LINDSTROM200330] for the comparison of radiation damage effects of various devices is to perform an annealing of 80 min at 60 $^{\circ}$C (or equivalently 10 min at 80 $^{\circ}$C) to have results with little dependence on the detailed temperature history after the irradiation.
Radiation damage of Silicon Photomultipliers {#sec:SiPM}
============================================
Effects of radiation damage in silicon have been introduced in Sec. \[sec:sirad\]. The main macroscopic effect in SiPMs reported in several studies is the significant increase of dark current ($I_{dark}$) below and above breakdown. For SiPM operation this translates in an increase of dark count rate (DCR). The scaling of $I_{dark}$ and DCR, with particle type, energy and irradiation fluence is not yet well documented. Some examples from literature are reported in the following. Fig. \[Fig:IV\] shows an example of current-voltage curves for a SiPM irradiated with neutrons, and operated at -30$^{\circ}$C. In the region of unit gain (V $\sim$ 5 V) the dark current increases by about three orders of magnitude after $\Phi_{eq} = 5 \cdot 10^{14}$ cm$^{-2}$, whereas above breakdown voltage the increase is more than six orders of magnitude. For too high over-voltages the measurements at the highest fluences are affected by saturation due to the current limit of the power supply.
![Exemplar current-voltage curves for a KETEK SiPM (15 $\mu$m pixel size) irradiated with neutrons up to $\Phi_{eq} = 5 \cdot 10^{14}$ cm$^{-2}$ and operated at -30 $^{\circ}$C. From Ref. [@Garutti:2017ipx].[]{data-label="Fig:IV"}](FIGURES/IV.png){width="0.90\linewidth"}
Another major effect of radiation on SiPM is the loss of single photon counting resolution due to the noise increase. This effect is reported in all studies dealing with bulk damage in SiPMs, and occurs at different fluences depending on the SiPM design, operating temperature, type of irradiation, and other parameters. It is safe to state that between $\Phi_{eq} = 10^{9}-10^{10}$ cm$^{-2}$, all SiPMs lose single photon counting resolution when operated at room temperature. Fig. \[Fig:SPE\] shows an example of a pulse-height spectrum for a SiPM before irradiation in which the photo-electron peaks are resolved. In the spectrum after $\Phi_{eq} = 10^{9}$ cm$^{-2}$, the photo-electron peaks are no longer resolved at room temperature. Tsang and co-authors demonstrate in Ref. [@Tsang2016] that the same SiPM cooled at 84 K still resolves single photons.
![Single photoelectron charge signal pulses at $V-V_{bd}$=3 V (b) before irradiation, (c) after neutron irradiation to $\Phi_{eq} = 10^{9}$ cm$^{-2}$. From Tsang Ref. [@Tsang2016].[]{data-label="Fig:SPE"}](FIGURES/SPE-Tsang){width="0.65\linewidth"}
Another example is given in Fig. \[Fig:waveforms\]. The top plot compares waveforms acquired in the dark with a KETEK SiPM operated at -30$^{\circ}$C. The single photo-electron peak visible before irradiation (blue curve) cannot be resolved in the noisy baseline after neutron irradiation to $\Phi_{eq} = 10^{13}$ cm$^{-2}$. The bottom plot demonstrates that the device is still a functional photo-detector for larger light intensity.
![Waveforms of a KETEK SiPM operated at -30$^{\circ}$C and 2.5 V overvoltage. The blue lines are before and the orange lines after irradiation with neutrons to $\Phi_{eq} = 10^{13}$ cm$^{-2}$. The waveforms in the top plot are taken in the dark, those in the bottom plot are taken with the SiPM under illumination with an LED light intensity of approximately 100 photoelectrons. From S. Cerioli doctoral thesis in progress.[]{data-label="Fig:waveforms"}](FIGURES/Waveforms-dark "fig:"){width="0.9\linewidth"} ![Waveforms of a KETEK SiPM operated at -30$^{\circ}$C and 2.5 V overvoltage. The blue lines are before and the orange lines after irradiation with neutrons to $\Phi_{eq} = 10^{13}$ cm$^{-2}$. The waveforms in the top plot are taken in the dark, those in the bottom plot are taken with the SiPM under illumination with an LED light intensity of approximately 100 photoelectrons. From S. Cerioli doctoral thesis in progress.[]{data-label="Fig:waveforms"}](FIGURES/Waveforms-light "fig:"){width="0.9\linewidth"}
For the characterization of SiPMs this implies that when the single photon resolution is lost, new methods need to be applied to extract the SiPM characteristic parameters, e.g. gain, DCR, correlated noise, PDE, etc...[^5]. A straight forward method allowing to investigate possible changes in the product of the factors contributing to the SiPM response is to plot the normalized ratio of photo-currents[^6] before ($\Phi =0$) and after ($\Phi$) irradiation [@Garutti2016:IEEE].
The ratio of normalized photo-currents is defined as: $$R=\frac{I_{photo}^{norm}(\Phi)}{I_{photo}^{norm}(\Phi =0)} = \frac{( G_{\Phi} ~ECF_{\Phi}~ PDE_{\Phi}) / (M_{\Phi}~ QE_{\Phi})}{\underbrace{(G_0 ~ECF_0 ~PDE_0 )}_\text{$V>V_{bd}$} / \underbrace{(M_0 ~QE_0)}_\text{$V \ll V_{bd}$}},$$ with the Excess Charge Factor (ECF) defined as the ratio between the mean values of measured $\langle N_{pe} \rangle$ and primary produced Geiger avalanches, $\langle N_{pG} \rangle$: $$ECF=\frac{\langle N_{pe} \rangle}{\langle N_{pG} \rangle}.
\label{eq:ecf}$$ where it is assumed that the number of primary produced avalanches follows the same distribution as the impinging light, and that the light source emission follows a Poisson distribution.
The photo-current is normalized to its value, $I_{photo} = q_0 ~N_{\gamma}~M ~QE $, at $V<<V_{bd}$ where the amplification gain can be considered $M \approx 1$. A value of $R=1$ for $V\ll V_{bd}$ confirms that the product of amplification factor and quantum efficiency ($M~QE$) of the SiPM is unchanged after irradiation. For $V>V_{bd}$ a value $R=1$ indicates that the product of gain, correlated noise, and PDE is not changed after irradiation. Any deviation from $R=1$ indicates that one or more of these parameters are affected (an example of such result is discussed in Sec. \[sec:had\]).
![The normalized LED photo-current of a non-irradiated SiPM and a SiPM irradiated to $\Phi_{eq} = 10^{12}$ cm$^{-2}$, taken from Ref. [@Garutti2016:IEEE]. Although the normalized photo-current spans five orders of magnitude, the ratio of the measurements before and after irradiation shows a variation of less than 7%. []{data-label="Fig:photocurrent"}](FIGURES/eta.png){width="0.90\linewidth"}
Fig. \[Fig:photocurrent\] shows $R$ for a KETEK SiPM irradiated with neutrons to $\Phi_{eq} = 10^{12}$ cm$^{-2}$, and operated at -30 $^{\circ}$C. $R$ agrees well with unity below breakdown and above breakdown is at most 7% higher than unity. This suggests that after irradiation the product $G_{\Phi}~ECF_{\Phi}~PDE_{\Phi}$ has not changed by more than 7% with respect to the original value. This and other observations discussed in Sec. \[sec:had\] suggested to split the discussion of bulk damaged SiPM in two fluence regions: Medium-fluence irradiations, for which the values of $G$, $ECF$, $PDE$ can be assumed not to change, and high-fluence irradiation for which the changes in these SiPM parameters must be taken into consideration in the interpretation of results.
In the following discussion we present studies of radiation damage on SiPMs separated in: Surface radiation damage, induced by X-ray irradiations (Sec. \[sec:Xray\]), and bulk radiation damage, induced by high-energetic electrons and gammas (Sec. \[sec:em\]), or by protons and neutrons (Sec. \[sec:had\]). It should be noted that aside of the latter all other particle types induce both surface and bulk damage in SiPMs, but the separation of these effects is not further discussed.
Surface radiation damage - X-ray photons {#sec:Xray}
----------------------------------------
Electrons and photon with energies below the threshold for bulk defects ($\sim$300 keV) generate only defects in the dielectrics, at the Si-SiO$_2$ interface and at the interface between dielectrics. About $\sim$18 eV are required to generate an e-h pair in SiO$_2$. This is the typical case for X-ray irradiation where photon energies are usually in the range $<$100 keV. Details of the expected surface damage effects are discussed in Sec. \[sec:sirad-iel\].
The SiPM electric parameters ($R_q, C_{pix}, \tau$) are extracted from measurements of current-voltage, and capacitance/conductance-voltage for frequencies between 100 Hz and 2 MHz just below $V_{bd}$. Breakdown voltage, gain, dark-count rate, cross-talk probability and pulse shape above the breakdown voltage are also investigated. The values of pixel capacitance and quenching resistor do not alter after irradiation. Their product $\tau_{RC} = C_{pix}R_q$ defines the pixel recovery time. This quantity is compared to $\tau_{exp}$ extracted as exponential slope of the signal from current transients. For this device the two definitions of $\tau$ are found to agree within experimental errors.
After 20 MGy the current below $V_{bd}$ increases by three orders of magnitude compared to the value before irradiation. This effect is ascribed to an increase of the surface-generation current from the Si-SiO$_2$ interface. The same increase is also visible in the current above breakdown, whereas the dark-count rate increases only by an order of magnitude at the same dose. The difference between dark-current and dark-count rate increase indicates that the large fraction of the current is not amplified in the multiplication region. These studies indicate that when developing SiPMs for applications in high dose X-ray environment, one must pay attention to the surface design, to minimize the increase of surface current due to radiation and to prevent that the surface current reaches the multiplication region.
Bulk radiation damage - electrons and gammas {#sec:em}
--------------------------------------------
SiPMs exposed to $^{60}$Co gamma irradiation ($E_{\gamma}$ = 1.33, 1.17 MeV) will experience a combination of surface and bulk defects, and are expected to have similar effects as SiPMs irradiated with electrons or positrons.
Matsubara and co-authors in Ref. [@Matsubara:2006zz] have irradiated a prototype SiPM from Hamamatsu (Type No. T2K-11-100C) under bias up to 240 Gy of $^{60}$Co $\gamma$-rays and measured the dark current, dark-count rate, gain, and cross talk. Whereas gain and cross talk did not significantly change with dose, large dark-count pulses and localized spots with leakage current along the outer edge of the active region and the bias lines were observed for about half an hour after irradiation for doses above 200 Gy. Renker commented in Ref. [@RenkerP04004] that very likely this effect was caused by accumulated and stationary charges at the Si-SiO$_2$ interface generated by the breakup of SiO$_2$ molecules. The authors also observed that immediately after powering the SiPM had large dark counts with an amplitude corresponding to a signal of more than 10 photo-electrons. After a couple of minutes, all signals with large amplitudes and most of the dark current disappeared. This phenomenon reappears after power cycling the device to zero and back to operating voltage.
The study was extended to higher doses by Lombardo et al. in Ref. [@Pagano:2014bua], where SiPMs from ST Microelectronics (400 pixels, 0.64 mm$^2$ active area, 0.47 fill factor) were irradiated with a $^{60}$Co source to 0, 10, 136 Gy, and 1.3, 9.4 kGy. Current-voltage and transient characteristics were recorded. Single photo-electron peaks are visible in the integrated spectrum up to 136 Gy. Above 1 kGy the noise increase is such that no structure can be distinguished in the pulse-height spectra, so the gain of the device can no longer be measured. The increase of dark current, dark count rate and cross talk as function of the dose is shown in Fig. \[Fig:Pagano\]. The authors explain that the method used to determine cross talk, using the fraction $CT(\%) \sim \frac{DC_{1.5}}{DC_{0.5}}\times 100$ may not be suitable for doses larger than 1.3 kGy. The measurement of dark count rate at 9.4 kGy may also be affected by saturation due to the method used, namely to count pulses above a certain threshold in a one second long transient. For a recovery time $\tau = R_qC_{pix} \sim$ 100 ns, the counting method is supposed to saturate at about 10$^7$ counts/sec, and this is the order of magnitude reached after 9.4 kGy. The increase of dark counts for the highest dose measurement of Matsubara et al. agrees qualitatively with the trend presented in Ref. [@Pagano:2014bua] (orange star in Fig. \[Fig:Pagano\]).
![Ratio of measured quantities vs. irradiation dose. Modified from [@Pagano:2014bua]. The red squared indicate the Dark Counts (DC) from [@Pagano:2014bua] taken at $\Delta V$ = 3 V ($\Delta V/V_{bd}$ = 11%). The two additional star points indicate measurements from [@Matsubara:2006zz] (orange) and [@Sudo2009] (blue).[]{data-label="Fig:Pagano"}](FIGURES/Pagano-remake.png){width="0.9\linewidth"}
In Ref. [@RenkerP04004] it is reported, that several SiPMs have been irradiated up to 500 Gy by a $^{60}$Co-source without applying a bias voltage during irradiation. No evidence for large pulses has been found after the irradiation. In Ref.[@QIANG2013234], in which the radiation hardness of Hamamatsu SiPMs was investigated, footnote 1 states: “An early irradiation test on SiPMs using a series of high activity $^{137}$Cs-sources in Jefferson Lab showed that SiPMs are insensitive to electromagnetic radiation and there was no significant change in performance of SiPMs up to 2 krad of gamma irradiation.”.
Achenbach et al. have irradiated green-sensitive SiPMs (SSPM-0701BG-TO18 from Photonique S A, Geneva) with 14 MeV electrons to fluences between 3.1$\cdot$10$^{11}$ cm$^{-2}$ and 3.8$\cdot$10$^{12}$ cm$^{-2}$ and observed a large increase in dark-count rate and a decrease in effective gain, see Ref. [@SANCHEZMAJOS2009506]. After irradiation the single photo-electron peaks are still visible as demonstrated in Fig. \[Fig:SPEelec\]. Surface effects were deemed responsible for the observed shift in the pedestal position, as consequence of the increase in leakage current.
![Single photoelectron spectra for low light intensity illumination of a Photonique green-sensitive SiPM before and after irradiation with 14 MeV electrons, to $\Phi = 3.1 \cdot 10^{11}$ cm$^{-2}$. From Achenbach Ref. [@SANCHEZMAJOS2009506].[]{data-label="Fig:SPEelec"}](FIGURES/SPE-electron.png){width="0.7\linewidth"}
The SiPM response to a medium-intensity laser light was investigated. A progressive reduction of signal charge is observed as a function of the irradiation dose. The paper does not attempt to disentangle the effects of decrease in gain, PDE in the single pixel or increase of pixel occupancy. A similar study was conducted in Ref. [@MUSIENKO2007433], using 28 MeV positrons with fluences up to 8$\cdot$10$^{10}$ cm$^{-2}$. PDE, gain, dark current and dark count rate were measured at room temperature as a function of bias voltage before and 2 days after irradiation. No change of the PDE and gain were found. As in the other publication the dark current and counts are shown to increase as function of dose. The authors introduce a new variable to quantify the effect of radiation damage on the dark count rate ($DCR$):\
$$\Delta DCR_{norm} = \frac{DCR(\Phi)-DCR(\Phi=0)}{ A\cdot PDE_{515nm}}
\label{Eq:DN}$$ where $A$ is the active area of the device, and the PDE is calculated at 515 nm.
![Increase of dark count rate after irradiation, normalized to active area and PDE as expressed in Eq. \[Eq:DN\] and plotted vs the PDE of several SiPM irradiated with 28 MeV positrons to $\Phi = 8\cdot 10^{10}$ cm$^{-2}$. The variable labeled $\Delta N_{norm}$ on the y-axis is the asme as defined in Eq. \[Eq:DN\]. Plot taken from Ref. [@MUSIENKO2007433].[]{data-label="Fig:positron"}](FIGURES/Musienko-positron.png){width="0.9\linewidth"}
By plotting $\Delta DCR_{norm}$ vs PDE, the authors argue that one can recognize devices with similar depletion layer thickness, and that good quality SiPM should have small PDE dependence of $\Delta DCR_{norm}$ (labeled $\Delta N_{norm}$ in Fig. \[Fig:positron\]). Unfortunately, due to lack of information it is not possible to extract the same quantity from other electromagnetic irradiation results so that the comparison between positron, electron and gamma radiation damage effects remains purely qualitative.
Bulk radiation damage - hadrons {#sec:had}
-------------------------------
SiPMs have been irradiated by several groups with neutrons [@Nakamura09; @QIANG2013234; @Andreotti:2013nra; @Musienko:2015lia; @Heering:2016lmu; @Garutti:2017ipx; @CentisVignali:2017zpz; @Musienko:2017znn; @Cattaneo17] and protons [@Matsumura:2009he; @DANILOV2009183; @BOHN2009722; @LI201663; @MUSIENKO200987; @SANCHEZMAJOS2009506], and their properties investigated. In this review we combine results on hadron irradiation without distinguishing particle type, since no dedicated studies have been performed so far on SiPMs on this topic. Fluences are quoted in 1 MeV neutron equivalent, according to the NIEL scaling explained in Sec. \[sec:sirad-niel\]. We group the studies in two categories: Medium-fluence irradiations, up to $\Phi_{eq} \approx 10^{12}$ cm$^{-2}$, and high-fluence irradiation to higher fluences. For medium fluences the majority of the studies report no significant change in SiPM parameters $V_{bd}$, $R_q$, $C_{pix}$, PDE, gain. Whereas, dark current (DC) and dark count rate (DCR) significantly increase proportionally to the fluence. For high fluences, $\Phi_{eq} > 10^{12}$ cm$^{-2}$, also other parameters are affected; but most importantly, the methods to determine the parameters are strongly affected by the high dark current and their limitations need to be discussed carefully.
### Effects on $R_q$, $C_{pix}$, $V_{bd}$ {#sec:rq}
The values of $R_q$, $C_{pix}$ can be extracted from impedance-frequency scans below breakdown, both before and after irradiation. In Fig. \[Fig:cpix\] results are presented exemplary for KETEK SiPMs irradiated with neutrons. These plots are an extension to higher fluences of the analysis already published in Ref. [@Garutti2016:IEEE]. From the C-V measurements below the breakdown voltage, which were taken at 25 frequencies between 100 Hz and 2 MHz, the SiPM electrical parameters have been determined using a simple R-C model. It is found that the value of $C_{pix}$ neither depends on temperature nor on neutron fluence, whereas the value of $R_q$ increases for $\Phi_{eq} > 10^{12}$ cm$^{-2}$. As expected for a poly-Si resistor, $R_q$ increases with increasing temperature. Ref. [@MUSIENKO200987] shows similar results for proton irradiation up to $\Phi_{eq} = 2\cdot 10^{10}$ cm$^{-2}$.
![SiPM electrical parameters of a KETEK SiPM (15 $\mu$m pixel size) as the function of neutron fluence measured at +20$^{\circ}$C and -30$^{\circ}$C. (Top) Pixel capacitance, $C_{pix}$, and (bottom) quenching resistance, $R_q$.[]{data-label="Fig:cpix"}](FIGURES/Cpix.png "fig:"){width="0.9\linewidth"} ![SiPM electrical parameters of a KETEK SiPM (15 $\mu$m pixel size) as the function of neutron fluence measured at +20$^{\circ}$C and -30$^{\circ}$C. (Top) Pixel capacitance, $C_{pix}$, and (bottom) quenching resistance, $R_q$.[]{data-label="Fig:cpix"}](FIGURES/Rq.png "fig:"){width="0.9\linewidth"}
In the absence of single-photon counting capability the turn-off voltage cannot be extracted from the linear relation between the excess bias voltage and the gain, $G=C_{pix}(V_{bias}-V_{bd})/q_0$[^7]. I-V measurements are use to determine $V_{bd}$ in this case. It is recommended to use I-V measurements taken under illumination of the SiPM with a source of light, either pulsed or continuous, and to apply a simple quadratic interpolation to the minimum of the inverse logarithmic derivative of the current, or analogously to the maximum of the logarithmic derivative. This procedure is stable against observed changes in the I-V dependence of dark noise below breakdown before and after irradiation possibly due to surface current, and also against the large uncertainty in the current measurement, which may affect I-V below breakdown at low temperatures or for low fluences.
One should keep in mind that the observable obtained with the I-V method, is not the same as that obtained from gain vs voltage linear regression [@Chmill:2016msk]. In the first case the turn-on voltage for the Geiger avalanche is measured (the breakdown voltage), whereas the former gives access to the turn-off voltage, which is the voltage relevant for the gain calculation in SiPMs. The two values can be significantly different, and it is not known if they both react to radiation damage in the same way.
![$V_{bd}$ of a KETEK SiPM (15 $\mu$m pixel size) as the function of neutron fluence measured at -30$^{\circ}$C.[]{data-label="Fig:vbd"}](FIGURES/Vbd.png){width="0.9\linewidth"}
Ref. [@Heering:2016lmu] presents the first evidence of a $V_{bd}$ shift due to radiation damage. Hamamatsu SiPMs exposed to high-fluence proton irradiation, $\Phi_{eq} = 6\cdot 10^{12} - 2\cdot 10^{14}$ cm$^{-2}$, show a linear increase of $V_{bd}$ with fluence, up to 4 V. Fig. \[Fig:vbd\] presented in Ref. [@Klanner:ICASIPM] shows $V_{bd}$ constant within uncertainties up to $\Phi_{eq} = 5\cdot 10^{13}$ cm$^{-2}$ and a possible increase of 300 mV after $\Phi_{eq} = 5\cdot 10^{14}$ cm$^{-2}$ for KETEK SiPMs. The difference in the magnitude of the $V_{bd}$ shift could be explained by the difference in the multiplication layer width. For Hamamatsu SiPM the multiplication layer is wider ($d\sim$ 2 $\mu m$) than for KETEK SiPMs ($d\sim$ 1.2 $\mu m$). It appears that thinner multiplication regions are less sensitive to breakdown voltage shifts by radiation damage. This could be linked to electric charges creation inside the depleted volume. If one assumes uniform acceptor generation during irradiation, the $V_{bd}$ shift should be proportional to the square of the depletion depth, as proposed in Ref. [@Musienko:2000im] for breakdown voltage shifts in APDs.
### Effects on $I_{dark}$ and DCR {#sec:DCR}
Up to $\Phi_{eq} \approx 10^{10}$ cm$^{-2}$ the typical increase of dark current is by a factor 30-100, confirmed by various studies on different devices. Cibinetto et al. [@Andreotti:2013nra] have irradiated samples from various producers with neutron fluences up to $\Phi_{eq} = 6.2\cdot 10^{9}$ cm$^{-2}$, and show that the relative increase in current is independent of the pixel size despite the significant higher current values for larger pixel size or capacitance (see Fig. \[Fig:pixsize\]).
![Dark current vs neutron dose for 1 mm$^2$ HAMAMATSU SiPMs with different pixel size. The blue line is for the 25 $\mu$m pixel, black 50 $\mu$m and red 100 $\mu$m. From Ref. [@Andreotti:2013nra]. In the paper similar measurements for AdvanSiD and SensL are also reported.[]{data-label="Fig:pixsize"}](FIGURES/IV-pixelsize.png){width="0.9\linewidth"}
Quiang et al. [@QIANG2013234] present a linear correlation between dark current and dark count rate, expected if one assumes that all the dark current is generated by carriers that traverse the multiplication region initiating a Geiger discharge: $$I_{dark} = q_0 ~G~ECF~DCR.
\label{eq:idark}$$ Note that in the absence of correlated noise (cross-talk and afterpulse) $ECF=1$. For SiPMs typically $ECF<1.2$, with values increasing as a function of the excess bias voltage. Fig. \[Fig:photocurrent\] demonstrates that the assumption that $ECF$ is independent of fluence is justified for medium-fluences. For high-dose irradiation the SiPM current can easily exceed hundreds of micro-ampere[^8] also when cooling the device below zero degree. These high currents can cause a local warm up of the pixel. The pixel temperature as function of $I_{dark}$ is hard to estimate, but a current-induced temperature increase could possibly be responsible for a change in $V_{bd}$ as function of fluence.
For non-irradiated SiPMs the DCR is often determined from the assumption of Poisson distributed dark count signals in a time interval $t_{gate}$ (e.g. the width of the gate), as: $$DCR=-\frac{\ln(P_0)}{t_{gate}}$$ with $P_0$ the fraction of events with pulse areas corresponding to $N_{pe} < 0.5$. In irradiated SiPM this method can no longer be applied due to the loss of single photon resolution. Other methods have been tested and should be investigated more thoroughly. In Ref. [@Garutti:2017ipx] the DCR of irradiated SiPM is scaled from the DCR before irradiation using the ratio of the dark current: $$DCR_{IV}(\Phi) = \frac{I_{dark}(\Phi)}{I_{dark}(\Phi = 0)}~ DCR(\Phi = 0).
\label{eq:DCR1}$$ However, this relation holds only under the assumption that gain and ECF are unchanged after irradiation, which has been confirmed in Fig. \[Fig:photocurrent\] for $\Phi_{eq} \leq 10^{12}$ cm$^{-2}$.
Another method is to define the DCR from the rms of the charge spectrum obtained with no illumination, $\sigma_0$. An often used definition is: $$DCR^* \propto \frac{\sigma^2_{0}}{\Delta t}.$$ If $\sigma_{0}$ has the units of charge it should be divided by $q_0 ~G$ to obtain the DCR in units of counts/second. This approximation does not take into account the non-Poisson distribution of the SiPM signal. The time interval $\Delta t$ is sometimes taken to be the integration gate, with no further assumption on its relation to the signal shape, which may not be correct in case of short gates. Following the derivation presented in Ref. [@Klanner:thisVolume] the relation between DCR and $\sigma^2_{0}$ for the simple case of a SiPM signal with one recovery time component $\tau = R_q C_{pix}$ is: $$DCR_{rms} = \frac{\sigma^2_{0}}{q_0^2~G^2 ~ENF~ECF^2 \cdot \left(t_{gate}+\tau \left( e^{-\frac{t_{gate}}{\tau}} -1 \right)\right)}.
\label{eq:DCR2}$$ where the expression in brackets comes from the variance of the distribution of $N_{DC}$ dark photo-electron pulses in a time interval $\Delta t_1$, defined such that $N_{DC} = DCR \cdot \Delta t_1 $. In the case of a signal much shorter than the integration gate, $\tau << t_{gate}$, the expression in brackets simplifies to $t_{gate}$. For more complex signal shapes another expression needs to be derived.
The Excess Charge Factor (ECF) is defined in Eq. \[eq:ecf\] and the Excess Noise Factor (ENF) is defined as: $$ENF=\left(\frac{\sigma_Q}{\langle Q \rangle}\right)^2 \left( \frac{\langle Q_P \rangle }{(\sigma_Q)_P} \right)^2
= \left(\frac{\sigma_Q}{\langle Q \rangle} \right)^2 \langle N_{pG} \rangle,
%=\frac{\sigma^2}{\sigma^2_P}\cdot \frac{1}{ECF^2}.
\label{eq:enf}$$
To extract DCR from Eq. \[eq:DCR2\] for an irradiated SiPM requires the assumption that gain and correlated noise are not affected by radiation damage. In addition, it is assumed that the signal shape does not change significantly after irradiation, otherwise different gate lengths may be needed for measurements before and after irradiation. Fig. \[Fig:DCR\] shows a very preliminary study comparing DCR values obtained using Eq. \[eq:idark\] and Eq. \[eq:DCR2\] on the same SiPM. For fluences between $10^{11}$ cm$^{-2}$ and $10^{13}$ cm$^{-2}$ the agreement is reasonably good. For the highest fluence of $\Phi_{eq} = 5 \cdot 10^{13}$ cm$^{-2}$ a saturation effect is observed in the transient RMS method, due to high pixel occupancy. This effect is not implemented in the analysis yet. The DCR data obtained from transient measurements are affected by about 30% correlated systematic uncertainties so that at this point it is only possible to say that the shapes are well in agreement, but the absolute values are subject to further investigations. Several of the assumptions still have to be verified, an example is the possible effect of an increase in after-pulsing and late cross-talk, which may mean that the ENF and ECF are different between a 100 ns gate and the current measurement (infinite gate length).
![Comparison of DCR as extracted from Eq. \[eq:idark\] (lines) and Eq. \[eq:DCR2\] (points) for KETEK SiPMs irradiated with neutrons and operated at -30$^{\circ}$C. No annealing is performed. Adapted from [@Klanner:ICASIPM]. []{data-label="Fig:DCR"}](FIGURES/DCRcomp){width="0.90\linewidth"}
### Effects on signal, PDE, gain {#sec:gain}
As already argued, generally in irradiated devices the PDE and gain cannot be directly measured due to the noise increase. With the analysis of the normalized photo-current as presented in Fig. \[Fig:photocurrent\], one can determine the product $PDE(\Phi) ~G(\Phi) ~ECF(\Phi)$. As long as $R=I_{photo}^{norm}(\Phi)/I_{photo}^{norm}(\Phi=0) =1 $ it can be argued that the three parameters have not changed as a consequence of irradiation. In Fig. \[Fig:photocurrent2\] from Ref. [@CentisVignali:2017zpz], $R$ is presented for a KETEK SiPM operated at -30$^{\circ}$C after neutron irradiation to $\Phi_{eq} = 5\cdot 10^{13}$ cm$^{-2}$.
![The normalized LED photo-current of a non-irradiated SiPM and a SiPM irradiated to $\Phi_{eq} = 5\cdot 10^{13}$ cm$^{-2}$. The KETEK SiPM is operated at -30$^{\circ}$C. From Ref. [@CentisVignali:2017zpz][]{data-label="Fig:photocurrent2"}](FIGURES/photocurrent2.png){width="0.9\linewidth"}
Whereas, below breakdown $R$ is still close to unity, it drops rapidly from 1 to 0 above $V_{bd}$. This indicates that one or more parameters are affected at this fluence. A possible explanation is that the PDE decreases due to an increase of pixel occupancy. The same argument was also offered by Angelone et al. in Ref. [@Angelone:2010mg] to explain a signal reduction of SiPM irradiated with neutrons to $7\cdot 10^{10}$ cm$^{-2}$.
The pixel occupancy due to dark counts is the probability of a Geiger discharge inside a pixel in a time interval $\Delta t$. Using $\Delta t = \tau = R_q (C_{pix}+C_q) \approx \tau = R_q C_{pix}$ is justified for SiPMs without a large fast component (a large $C_q$ parallel to $R_q$). Then for a device with $N_{pix}$ pixels one obtains [@Garutti:2017ipx]: $$\eta_{DC} = \frac{I_{dark}}{\Delta V} \frac{R_q}{N_{pix}} = \frac{\tau ~ECF}{N_{pix}}DCR.
\label{Eq:eta}$$ The first equation shows that for $\eta_{DC} =1$ the current over the quenching resistor is equal to the excess bias voltage, so the voltage is always at $V_{bd}$ and the SiPM never recovers. For very high DCR, pulses may often occur in a pixel in which the voltage is not fully recovered from a previous discharge, i.e. at an effective lower overvoltage resulting in a smaller gain. This effect may lead to a signal reduction after irradiation. In the second part of Eq. \[Eq:eta\], $I_{dark}$ is substituted using Eq. \[eq:idark\] and the proportionality to the DCR is made explicit. The values of $\eta_{DC}$ measured on SiPMs irradiated with neutrons to $\Phi_{eq} = 1\cdot 10^{11} - 5\cdot 10^{14}$ cm$^{-2}$ are shown in Fig. \[Fig:eta\]. The value of $\eta_{DC}$ increases linearly with fluence, and exceeds 1% at $\Phi_{eq} > 10^{12}$ cm$^{-2}$, for $\Delta V >$ 2 V and $T = -30^{\circ}$C. For a high pixel occupancy a large number of pixels does not recover within the time of the signal and is not available to detect photons, consequently the effective PDE is reduced. It is expected that the signal of the SiPM in response to low intensity of light will decrease at least by the same amount, if no additional effects contribute. One should note that the value obtained with this method is an average pixel occupancy and that the deviations between pixels can be very large as already discussed when quoting the work by Barnyakov on digital photon counters, [@Barnyakov:2016bwt].
![The values of $\eta_{DC}$ for SiPM irradiated with neutrons to $\Phi_{eq} = 1\cdot 10^{11} - 5\cdot 10^{14}$ cm$^{-2}$. The measurements are performed at -30 $^{\circ}$C. For fluences larger than $\Phi_{eq} = 10^{12}$ cm$^{-2}$ the pixel occupancy is larger than 1% at the operating voltage, and therefore a reduction of the effective PDE is expected. In Ref. [@Garutti:2017ipx], $V_{bd}$ is determined using the inverse logarithmic derivative of the current, and $V_{ex} = V_{bias} -V_{bd}$. []{data-label="Fig:eta"}](FIGURES/eta-new.png){width="0.90\linewidth"}
Alternatively, the parameters of irradiated SiPM can be measured from the response to a pulsed light signal (LED or laser) of sufficiently high intensity to be separable from the increased noise. For a pulse of fixed light intensity the integrated SiPM charge follows a distribution with mean $Q$ and variance $\sigma_{Q}^2$ (see Appendix \[appendix\]). The $PDE$ is the ratio between mean number of photons photons initiating primary Geiger discharges, $\langle N_{pG} \rangle$, and the average number of photons impinging the photo-detector surface $\langle N_{\gamma} \rangle$. The value of $\langle N_{pG} \rangle$ can be determined in various ways (see Eq.s \[eq:mu1\], \[eq:mu2\], \[eq:mu3\]) each depending on a different assumption. Often used is Eq. \[eq:mu3\],
$$\label{Eq:PDE}
PDE = \frac{\langle N_{pG} \rangle}{\langle N_{\gamma} \rangle}= \frac{ENF}{\langle N_{\gamma} \rangle}\frac{Q^2}{\sigma_{Q}^2} = ENF~PDE^*,$$
where the variance measured in dark conditions is already subtracted from the variance of the light signal: $\sigma_{Q}^2 = \sigma_{Q_{light}}^2 -\sigma_{Q_0}^2$.
In Ref. [@MUSIENKO200987], results for PDE$^* $ for SiPMs from various producers are compared before and after irradiation with protons to $\Phi_{eq} = 2 \cdot 10^{10}$ cm$^{-2}$. Relative changes in PDE$^*$ are below the 10% level despite the significant increase of $I_{dark}$ and DCR.
The effect of high-fluence proton irradiation is reported in Ref. [@Heering:2016lmu]. In Eq. \[Eq:PDE\] the mean number of photo-electrons is directly derived from the amplitude and width of the signal response to a known $N_{\gamma}$. Note that this formula only holds under the assumption of Poisson distributed light intensity, which for instance for laser light of low intensity could become an issue, and for LED pulsed light requires the stability of the pulse generator to be of order 0.1%. In addition, the light intensity should be low to avoid effects of SiPM non-linearity, and the readout system (preamplifier, ADC, ...) must be linear and have no saturation effects. In Ref. [@Heering:2016lmu] it is argued that the observed decrease of this quantity indicates a PDE$^* $ decrease by 25% after $\Phi_{eq} = 2.2 \cdot 10^{14}$ cm$^{-2}$. For the same fluence the signal response to LED light, $Q$ decreases by 50%. The PDE$^* $ reduction can be explained with the pixel occupancy mechanism, but the additional signal loss must have another origin. Like the photo-current, also the signal is proportional to the product $PDE(\Phi) ~G(\Phi) ~ECF(\Phi)$. To explain the large observed signal reduction either the gain or the $ECF$ must reduce as well.
A possible additional effect not affecting the pixel occupancy is the decrease of gain. As argued in Ref. [@Heering:2016lmu], self-heating effects at the pixel level could cause a local breakdown voltage change for the hot pixels. The exact temperature inside the discharging pixel is not easily measurable, but possibly local temperature increase could change the pixel breakdown voltage with respect to the $V_{bd}$ value obtained from I-V measurements. A possible way to further investigate this effect is to compare the three definitions of $\langle N_{pG} \rangle$ given in Appendix \[appendix\], which have different dependencies on the gain (none, first order and quadratic) and therefore may be more sensitive to gain changes. New methods to quantify the gain for irradiated SiPMs would also be beneficial.
Annealing effects {#sec:annealing}
-----------------
Several groups have measured the effect of thermal annealing on irradiated SiPMs. Nakamura et al. [@Nakamura09] investigate SiPMs irradiated with neutrons up to $\Phi_{eq} \approx 10^{12}$ cm$^{-2}$, and demonstrate that the single photon detection capability is lost immediately after irradiation, but up to $\Phi_{eq} \approx 10^{10}$ cm$^{-2}$ it recovers after about two months of room temperature annealing. This self-annealing effect is also reported in Ref. [@Andreotti:2013nra; @BOHN2009722; @SANCHEZMAJOS2009506].
Quiang et al. [@QIANG2013234] further investigated the temperature dependent annealing on SiPMs irradiated to $\Phi_{eq} = 3.7\cdot 10^{9}$ cm$^{-2}$. They confirm a self-annealing time constant of about 10 days at room temperature, and show an exponential decrease of the annealing time constant with increasing temperature in the range of -5 / +60 $^{\circ}$C.
![Effect of high temperature and forward bias annealing on two types of SiPMs irradiated to $\Phi_{eq} = 10^{9}$ cm$^{-2}$ and $\Phi_{eq} = 10^{10}$ cm$^{-2}$. Left SiPMs from Hamamatsu (HPK) and right from FBK. From Ref. [@Tsang2016].[]{data-label="Fig:tsang"}](FIGURES/annealing_Tsang.png){width="0.90\linewidth"}
Tsang has presented results of annealing at +250 $^{\circ}$C, using forward bias with the SiPM current reaching 10 mA, [@Tsang2016]. A remarkable effect of this high temperature annealing is demonstrated in Fig. \[Fig:tsang\]. Single photo-electron resolution was recovered after this procedure for devices irradiated up to $\Phi_{eq} = 10^{12}$ cm$^{-2}$ with cooling them to about -50 $^{\circ}$C. Still the dark current before irradiation cannot be completely recovered even after days at the annealing conditions.
Gotti at al. have recently published a study [@Gotti2018:arxiv] of SiPMs irradiated with neutrons up to $\Phi_{eq} = 10^{14}$ cm$^{-2}$ and demonstrated that single photo-detection with a DCR below 1 kHz is possible after this fluence if the devices are annealed for several days at 175 $^{\circ}$C, and subsequently operated at cryogenic temperatures (-77 $^{\circ}$C).
Radiation damage of avalanche photodiodes {#sec:apd}
=========================================
Avalanche photodiodes (APD) have been established as photo-detectors few years before SiPMs were developed and have been investigated in great detail. Most of the studies of radiation damage in APDs were performed for their application in the electromagnetic calorimeter of the Compact Muon Solenoid (CMS) experiment at the Large Hadron Collider (LHC) [@Chatrchyan:2008aa]. Analogously to a SiPM, the principle of operation of an APD is based on the conversion of the energy of the absorbed photons into electron-holes pairs in silicon diodes and their multiplication through an avalanche process [@Sze2006]. The diode has a region with electric field of 1.5-2.5$\cdot 10^5$ V/cm (so lower than the field required for a SiPM to reach Geiger mode) where electrons (or holes) are accelerated and acquire a kinetic energy large enough to create additional electron-hole pairs via impact ionization. The breakdown voltage is (in contrast to SiPMs) the maximum voltage that can be applied to an APD. The effect of radiation in silicon APDs is very similar to that in silicon diodes operated without gain [@ANTUNOVIC2005379]. However, there are several important differences related to their operation as photo-detectors with internal gain. It is relevant to give a short review on APDs radiation hard studies, which may prove valuable to interpret results on SiPMs. In the following we report separately on studies performed irradiating APDs with photons and hadrons.
Radiation damage by gammas {#sec:apd-gamma}
--------------------------
The effect of the gamma radiation has been studied in [@Kirn:1997]. There it has been established that APDs, which had a front window made of SiO$_2$, show a decrease of the quantum efficiency in the short wavelength region after gamma irradiation, while the photosensors, with a Si$_3$N$_4$ window, show no change in the quantum efficiency after exposure to a dose of 55 kGy. More than 3000 CMS APDs were irradiated with $^{60}$Co source up to 2.5 kGy gamma dose [@ANTUNOVIC2005379]. A dark current increase of $\sim$150-500 nA was measured after irradiation at the operating voltage (corresponding to $G = 50$). The origin of this current was associated to damage in the APD dielectric layer, as it had only a weak dependence on the APD gain. The maximum operating voltage, corresponding to the breakdown voltage was in the range of 430-440 V before irradiation, and was reduced by 4-30 V for approximately 0.7% of the irradiated APDs. The other APDs presented no effect of breakdown voltage decrease.
Radiation damage by hadrons
---------------------------
### Dark current, QE and gain {#sec:apd-DC}
An extensive series of studies of the APD resistance to hadron radiation has been carried out by different groups participating in the CMS collaboration. These studies have been performed utilizing reactor neutrons, a spallation neutron source, a $^{252}$Cf source, and a proton beam (see Ref. [@ANTUNOVIC2005379] and Refs. therein).
Fig. \[Fig:APD2\] shows the increase in the bulk current as a function of the neutron fluence [@BACCARO1999206]. The parameter “bulk current” or “generation bulk current” was introduced in this paper because it was found that the ratio of dark current and gain is almost constant in a wide range of gains ($G >$ 10). The bulk current is defined as the dark current divided by the APD gain ($G =$ 50 in this case). It can be noticed that the bulk current grows linearly with the fluence, as for diodes.
![Radiation-induced bulk current versus fluence for one of the Hamamatsu APDs (25 mm$^2$ in area). The measurements are taken at 18$^{\circ}$C and 2 days after the irradiation [@BACCARO1999206]. []{data-label="Fig:APD2"}](FIGURES/APD2){width="0.7\linewidth"}
When trying to compare the damage parameter $\alpha$ for APDs and diodes (see solid line in Fig. \[Fig:APD2\]), it was observed that the volume of the APD contributing to the dark current generation is smaller than the geometrical volume obtained multiplying the active area by the depletion depth (for Hamamatsu APDs: A = 25 mm$^2$, $L_{dep}$ = 50 $\mu m$). This is because the carriers generated throughout the depletion region in a device with multiplication do not all contribute the same to the total dark current. The correct formalism to describe the noise generation in a device with multiplication gain was developed by McIntyre in Ref. [@McIntyre:1966]. In Ref. [@Musienko:2000jj] the concept of an effective thickness of the APD, $L_{eff}$ was introduced for simplicity. It is found that the contribution to the bulk current is generated by a region with $L_{eff} \approx$ 5 $\mu m$, compared to the depletion thickness of $L_{dep}$ = 50 $\mu m$. The same effect is also relevant when discussing the region contributing to the dark current of a SiPMs.
In Ref. [@Musienko:2000jj] two APDs from EG&G with different depletion region thicknesses, 196 and 243 $\mu$m, and one APD from Hamamatsu, with depletion region thickness, $\sim$50 $\mu$m, were irradiated to $\Phi_{eq}=$ 2.33$\cdot 10^{13}$ cm$^{-2}$. The quantum efficiencies of the investigated EG&G APDs remained unchanged over a wide range of wavelengths. The Hamamatsu APD showed a significant decrease of quantum efficiency (by 10-40 %), especially for short wavelengths, as shown in Fig. \[Fig:APD5\].
![Quantum efficiency losses vs. wavelength after irradiation (G = 19, T = 25 $^{\circ}$C), after irradiation to $\Phi_{eq}=5\cdot 10^{14}$ cm$^{-2}$ and several years of annealing at room temperature. Plot from Ref. [@Musienko:2015fiz].[]{data-label="Fig:APD5"}](FIGURES/APD5){width="0.7\linewidth"}
In Ref. [@Musienko:2000jj] it was observed that the curves of gain as a function of bias voltage for the EG&G APDs after irradiation are shifted compared to those before irradiation (see Fig. \[Fig:APD3\]). This effect is related to a change of the doping concentration inside the depletion region of the APD due to the creation of acceptor-like states [@Musienko:2000im], that caused the APD breakdown voltage increase. The gain as a function of bias voltage for the Hamamatsu APD remained unchanged after irradiation, probably due to a thin depletion region ($\sim$50 $\mu$m).
![Gain as a function of bias voltage for the 196 um EG&G APD, measured before and after irradiation.[]{data-label="Fig:APD3"}](FIGURES/APD3){width="0.7\linewidth"}
The CMS ECAL APDs (produced by Hamamatsu) were irradiated up to $\Phi_{eq}=5\cdot 10^{14}$ cm$^{-2}$ [@Musienko:2015fiz]. Dark current and gain vs. bias dependences were measured at T = 25 $^{\circ}$C, 15 $^{\circ}$C and 5 $^{\circ}$C after several years of annealing at room temperature. The gain vs. bias and QE vs. wavelength dependences were compared with that of non-irradiated APDs. The main results of the study are:
- APDs are still operational as light detectors with $G>15$ also without cooling;
- The dark current increases linearly with fluence. After $\Phi_{eq}=5\cdot 10^{14}$ cm$^{-2}$, and for $G=10$ and T = 25 $^{\circ}$C, it is $I_{dark} \approx$ 80 $\mu$A;
- The dark current can be significantly decreased by reduction of temperature ($\sim$2.2 times per 10 $^{\circ}$C, consistent with the temperature dependence of leakage current in diodes, in the absence of strong fields) $-$ For SiPMs the reduction of dark current with temperature is weaker.
- QE losses of 20-50% were found for the 350-500 nm wavelength range (see Fig. \[Fig:APD5\]) $-$ Similar measurements are not available for SiPMs;
- The gain at a fixed voltage was reduced due to doping profile change, caused by creation of acceptor-like states in the APD depleted region. The problem can be solved by $\sim$30 V bias voltage increase $-$ This effect is qualitatively consistent with the shift in $V_{bd}$ observed in SiPMs;
### Annealing studies {#sec:apd-annealing}
The recovery of the dark current can be described [@BACCARO1999206] as a sum of exponentials, each exponential being attributed to the recovery of one defect with its proper recovery-time $\tau_i$ and its weight $g_i$: $$I_{dark}^{irr}(t) = I_{dark}^{irr}(0) \sum{g_i e^{-\frac{t}{\tau_i}}}
\label{Eq:APD3}$$
![Room temperature annealing of the dark current of one of the Hamamatsu APDs irradiated with 2.7$\cdot 10^{12}$ (E=72 MeV) protons/cm$^{2}$ at PSI [@CMSECAL:1997].[]{data-label="Fig:APD4"}](FIGURES/APD4){width="0.65\linewidth"}
Fig. \[Fig:APD4\] shows the recovery of the dark current of an APD irradiated to $\Phi=$2.7$\cdot 10^{12}$ (E=72 MeV) protons/cm$^{2}$. The measurements were done at $G=50$ and at room temperature. The measurements started 10 min after irradiation. Table \[Tab:tau\] shows the decay-times and the weights of the various components fitted from these data points.
\[Tab:tau\]
$g_i$ $\tau_i$
------- ----------
0.21 576 min
0.36 13 days
0.17 600 days
0.26 infinite
: Dark current annealing decay times and their weights for Hamamatsu APDs irradiated with 2.7$\cdot 10^{12}$ (E=72 MeV) protons/cm$^{2}$ at PSI [@CMSECAL:1997]. The values are extracted by a fit to Fig. \[Fig:APD4\] using Eq. \[Eq:APD3\].
Several other APDs have been irradiated with neutrons. They show similar annealing behaviors (see Ref. [@BACCARO1999206]).
Approaches to develop radiation harder SiPMs {#sec:outlook}
============================================
There are several factors limiting the operation of SiPMs in high radiation environment:
- dark noise increase due to deep level defects produced by radiation in silicon bulk;
- high cell occupancy due to high generation-recombination rate and limited cell recovery time causes a reduction of SiPM gain and PDE;
- large leakage current results in increase of the p-n junction temperature and high power consumption;
- increase of the SiPM breakdown voltage due to induced changes in doping concentration;
- reduction of the SiPM photon detection efficiency due to damage in dielectric-silicon interface and charge trapping in non-depleted entrance layer;
Next we discuss approaches to overcome these limitations.\
**Dark noise reduction**\
The main problem for the optimization of radiation tolerance of the SiPMs is presented by the increase of DCR. The main sources of the DCR increase are:
- surface current reaching the multiplication region;
- diffusion from the non-depleted bulk;
- defect-induced generation in the depletion region.
Dark noise generation in SiPMs can be reduced by
- avoiding that surface current reaches the multiplication region;
- reducing diffusion from the non-depleted bulk;
- optimizing the field in the depletion region, “field shaping”.
This last point can be achieved by:
- reducing thickness of the depleted region;
- reducing trap-assisted tunnelling (reducing peak electric field in the SiPM p-n-junction).
Unfortunately these two approaches contradict each other: a reduction of electric field requires wider avalanche and depletion regions. A special R&D has to be performed to find an optimum (compromise) between these two approaches.\
**Cell occupancy reduction**\
The cell occupancy reduction can be achieved by reducing the cell active volume (smaller cell size) and cell recovery time. However SiPMs with high cell density have large non-sensitive zones (cell separation regions, area occupied by polysilicon quenching resistors which are almost non-transparent). The solution of this problem can be very thin trenches ($<$0.5 $\mu$m) separating cells and use of metal film resistors with high transparency to visible light. Smaller cell size results also in smaller cell capacitance and reduction of the recovery time. Reduction of quenching resistor value is limited. Few hundreds k$\Omega$ resistors are required to quench the cell discharge and allow stable SiPM operation in a reasonable range of over-voltages (2-3 V). The reduction of the SiPM cell size and capacitance has another positive outcome, namely the reduction of the SiPM dark noise, or leakage current (mainly because of gain reduction). This is also linked to a reduction of power consumption. Related to the need of short recovery time is another important aspect, namely that of including the signal shaping (fast recovery time, one vs two time component, ...) in the optimization process. So far this aspect has not been covered in literature.\
**Breakdown voltage increase minimization**\
The SiPM $V_{bd}$ increase with irradiation strongly depends on the thickness of the depletion region. It can be reduced by reducing this thickness. SiPMs with low $V_{bd}$ value have significantly smaller $V_{bd}$ increase after hadron irradiation [@Andreotti:2013nra; @Heering:2016lmu].\
**Reduction of the damage in SiPM entrance window**\
The SiPM entrance window material has to be chosen appropriately for radiation hard applications, i.e. avoiding materials with hydrogen or boron content. The thickness of the SiO$_2$ layer has to be properly adjusted taking into account surface effects. The thickness of non-depleted region near the SiPM entrance window has to be minimized to reduce nuclear effects in its volume.\
Detailed and possibly realistic simulations of SiPMs should be performed in order to optimize the SiPM designs for specific applications. For this purpose information is still missing:
- Fluence dependence of relevant parameters: mobility and multiplication coefficient as function of doping; field enhanced generation of traps,
- more systematic measurements of SiPM parameters,
- microscopic measurements on defects (can defect engineering help?).
The outcome of the optimization will depend on the targeted wavelength and the expected operation temperature for the SiPM.
Conclusions
===========
This review is an attempt to summarize the current knowledge of radiation damage of SiPMs, a complex multi-parameter problem that recently has attracted large attention in the SiPMs users community. The main issue with irradiated SiPMs is the increase of dark count rate, which at some point also affects other SiPM parameters. Due to the increased DCR, the single photo-electron separation from noise is lost already at relatively low fluences $\Phi_{eq} \sim 10^{10}$ cm$^{-2}$. This limit depends on many factors related to the SiPM design and the operation conditions, so it should be tested for each specific application. The DCR also affects the pixel occupancy, which if no longer negligible yields to $PDE$ reduction.
Once single photo-electrons can no longer be resolved, many characterization methods used for the investigation of non-irradiated SiPM parameters fail. Some recently developed characterization methods are presented in this review. We tried to spell our the assumptions made by each method. They all still need further testing and validation on larger sets of data. It is reported that SiPM parameters like $R_q$, $C_{pix}$, $V_{bd}$, $G$, $ECF$, $ENF$, $PDE$ for $\Phi_{eq} < 10^{12}$ cm$^{-2}$, whereas above this fluence changes are observed. Again this limit is SiPM design and temperature dependent. A proper quantification of the fluence dependence of the individual parameters requires further dedicated experiments. Also the annealing state of the SiPM is important to define limits on the operation range. Annealing studies indicate the possibility to recover single photo-electron resolution at least for operation at low/cryogenic temperatures.
An optimization of the SiPM design for operation in high radiation environment is not yet available and lacks still dedicated experimental input. Further effort in this field is strongy encouraged.
Statistical treatment of SiPM signal {#appendix}
====================================
The statistical treatment of SiPM signals is presented in several papers, notable are among other Ref. [@Vinogradov:2011vr; @Arosio:2015sra; @Chmill:2016ghf].
Given a Poisson distribution of impinging photons on the SiPM, its response does not follow a Poisson distribution due to its correlated noise. A generalized Poisson formalism is used often to account for in-time correlated signal generation (prompt cross-talk). But SiPMs are also affected by late correlated noise such as delayed cross-talk and after-pulses. We report here some useful definitions related to the treatment of SiPM signals and relevant for the formalism in this paper.
The charge distribution of a SiPM integrated in a time $t_{gate}$, is characterized by the first moment $\langle Q \rangle$ and the second central moment or variance, $\sigma_Q^2 = Var_Q = E[(Q-\langle Q \rangle)^2]$. In the following the variables $\langle Q \rangle$, and $\sigma_Q^2$ will be used for the mean and variance of the charge distribution. Another often used distribution to characterize SiPMs is that of the number of measured Geiger discharges with mean $\langle N_{pe} \rangle$ and the variance $\sigma_{pe}^2$. The distributions of charge and number of measured Geiger discharges, or photo-electrons, are linked by the relations: $$\langle Q \rangle = q_0 ~G ~\langle N_{pe} \rangle,
\label{eq:multimean0}$$ $$\sigma_Q^2 = q_0^2 ~G^2 ~\sigma_{pe}^2.
\label{eq:multivar0}$$
The definition of the Excess Charge Factor (ECF) (Eq. \[eq:ecf\]) relates the mean charge to that of the primary Geiger discharges (Poisson distributed): $$ECF =\frac{\langle Q \rangle}{\langle Q_P \rangle} =\frac{\langle N_{pe} \rangle}{\langle N_{pG} \rangle},
\label{eq:eqf}$$ with $\langle N_{pG} \rangle = \mu$ being the mean number of primary Geiger discharges with $\sigma_{pG} = \sqrt{\langle N_{pG} \rangle}$ . The Excess Noise Factor (ENF) relates the variances of the two distributions (Eq. \[eq:enf\]): $$ENF=\frac{\sigma_Q^2}{\langle Q \rangle ^2} \frac{\langle Q^2_P \rangle}{(\sigma_Q^2)_P}
=\frac{1}{ECF^2} \frac{\sigma_Q^2}{(\sigma_Q^2)_P}
=\frac{1}{ECF^2} \frac{\sigma_{pe}^2}{\langle N_{pG} \rangle}.
\label{eq:enf1}$$ We note that for the Generalized Poisson distribution $GP(k;\mu, \lambda) = (\mu \cdot (\mu + k \cdot \lambda)^{k-1} \cdot e^{-(\mu + k\cdot\lambda)}/k!$, $ECF = ENF = 1/(1 -\lambda)$ [@Vinogradov:2011vr].
Substituting Eq. \[eq:eqf\] in Eq. \[eq:multimean0\], and Eq. \[eq:enf1\] in Eq. \[eq:multivar0\] one gets:
$$\langle Q \rangle= q_0 ~G ~ECF ~\langle N_{pG} \rangle,
\label{eq:multimean}$$
$$\sigma_Q^2 = q_0^2 ~G^2 ~ENF ~ECF^2 ~\langle N_{pG} \rangle.
\label{eq:multivar}$$
From Eq. \[eq:multimean\] and Eq. \[eq:multivar\] the gain factor (G) is obtained as: $$G = \frac{1}{q_0~ ENF~ECF}\frac{\sigma_Q^2}{\langle Q \rangle}.
\label{eq:gain2}$$ Eq. \[eq:gain2\] can be used when a SiPM does not provide single photo-electron resolution, as for instance after radiation damage.\
For the analysis of light signals, the mean number of primary Geiger avalanches, $\langle N_{pG} \rangle$ can be determined in two different ways using either Eq. \[eq:multimean\] or Eq. \[eq:multivar\]: $$\langle N_{pG} \rangle_1 = \frac{\langle Q \rangle}{q_0 ~G ~ECF}
\label{eq:mu1}$$ $$\langle N_{pG} \rangle_2= \frac{\sigma_Q^2} {q_0^2 ~G^2 ~ENF ~ECF^2 },
\label{eq:mu2}$$ Combining Eq. \[eq:eqf\] in Eq. \[eq:enf1\] a third independent definition is: $$\langle N_{pG} \rangle_3= ENF \frac{\langle Q \rangle ^2}{\sigma_Q^2} = ENF \frac{\langle N_{pe} \rangle ^2}{\sigma_{pe}^2},
\label{eq:mu3}$$ where mean and variance can be taken consistently from the distribution of charge, number of photo-electrons or even ADC channels with no need of further calibration, since additional conversion factors cancel in the ratio.\
[^1]: This value refers to the band gap energy of silicon at room temperature.
[^2]: This holds true for low and medium fluences up to $\Phi_{eq} \sim 10^{12}$ cm$^{-2}$. For higher fluences the trap concentration may reach a measurable level. Still the very high current is a limit for DLTS application, while the very small depleted volume may be a limit for TSC applications. This still has to be verified experimentally.
[^3]: The phenomenon of the emission of light with energies larger than 1.12 eV (the band gap energy in silicon) from a p-n junction operated above breakdown voltage was first observed by Newman [@Newman], then quantified by Lacaita, Zappa and Bigliardi, [@Lacaita] to produce in average 3$\cdot$10$^{-5}$ photons per charge carrier crossing the junction. The dominant production mechanism is bremsstrahlung, but other effects also contribute.
[^4]: It should be remarked that the light intensity emitted by Geiger discharges is at a given gain is relatively constant, such that the measured intensity on the CCD can be used to calculate the number of discharges of the SiPM in the measurement time interval.
[^5]: For a detailed overview of standard SiPM characterization methods we refer to Ref. [@Klanner:thisVolume] within this review volume.
[^6]: The photo-current is the current in the presence of light subtracted from the dark current: $I_{photo}=I_{light}-I_{dark}$, and it is given by $I_{photo}= q_0 N_{\gamma}~G ~ECF ~PDE $.
[^7]: More correctly this expression should read $G=(C_{pix}+C_q)(V_{bias}-V_{off})/q_0$, taking into consideration the difference between the turn-on voltage for the Geiger avalanche (the breakdown voltage), and the turn-off voltage demonstrated in Ref. [@Chmill:2016msk].
[^8]: Note that as a consequence of the significant current increase the bias voltage applied to the SiPM, $V_{bias}$, must be calculated correcting the power supply voltage $V_{PS}$ by the voltage drop over the series resistance of the biasing circuit, $V_{bias}=V_{PS}-I_{leak}\cdot R$. Where $R$ is a value specific for a given set-up. This can cause voltage drops of several volts depending on the circuit for current values in the range of mA.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We describe the realization of a magnetically guided beam of cold rubidium atoms, with a flux of $7\times 10^9$ atoms/s, a temperature of 400 $\mu$K and a mean velocity of 1 m/s. The rate of elastic collisions within the beam is sufficient to ensure thermalization. We show that the evaporation induced by a radio-frequency wave leads to appreciable cooling and increase in phase space density. We discuss the perspectives to reach the quantum degenerate regime using evaporative cooling.'
author:
- 'T. Lahaye, J. M. Vogels, K. Günter, Z. Wang, J. Dalibard, and D. Guéry-Odelin'
title: Realization of a magnetically guided atomic beam in the collisional regime
---
The recent realization of slow atom sources using laser cooling methods has constituted a major advance for many applications in metrology and matter wave interferometry [@Interferometry]. The possibility to complement laser cooling by forced evaporation has led to even colder and denser atomic gases, and has culminated with the observation of Bose-Einstein condensation (BEC) [@BEC] and pulsed “atom lasers" extracted from the condensate [@atomlaser]. These advances in atom cooling allow in principle to build a continuous and coherent source of matter waves which would be the equivalent of a monochromatic laser. It should lead to unprecedented performances in terms of focalization or collimation, and provide a unique tool for future developments in atom manipulation and precision measurements.
Two paths have been considered to reach the goal of a continuous and coherent source of atoms (“cw atom laser”). The first possibility [@ScienceK02] consists in using the “standard" condensation procedure to periodically replenish with new condensates a gas sample held in an optical dipole trap. The second possibility, that we are currently investigating, consists in transposing the evaporative cooling method to a guided atomic beam [@Mandonnet00]. As the beam progresses in a magnetic guide [@mguide; @cren], the most energetic atoms are removed and the remaining particles thermalize at a lower temperature and a larger density, so that the beam can eventually reach the degenerate regime. This method should in principle allow for a large flux, since it consists in a parallel implementation of the various steps to BEC. Instead of performing sequentially the optical cooling and the evaporation sequences at the same place, these occur simultaneously in different locations along the guided beam. This gain in flux can constitute a major advantage for applications.
To implement this method, the initial atomic beam needs to be deeply in the collisional regime. The number $N_{\rm col }$ of elastic collisions undergone by each atom as it travels along the guide must be large compared to unity, to provide the thermalization required for evaporative cooling. Such beams were so far not available. In this Letter we report on the realization of a rubidium atomic beam in the collisional regime, with $N_{\rm
col }\sim 8$. It propagates in a 4.5 m long magnetic guide with a mean velocity of 1 m/s. We also demonstrate a first step of evaporative cooling, and show that the measured increase in phase space density is in good agreement with expectations. Finally we discuss possible improvements of the current setup which should allow to reach the degenerate regime.
Fig. 1 shows the layout of our experimental setup, composed of three main parts lying in the $xy$ horizontal plane: the Zeeman slower along the $y$ axis, the magneto-optical trap (MOT) and the magnetic guide along the $x$ axis. The Zeeman slower [@zeemanslower] consists in a 1.1 m long tapered solenoid in which atoms emerging from an oven are decelerated by resonant radiation pressure. It delivers a flux of $2\times
10^{11}$s$^{-1}$ $^{87}$Rb atoms, with a velocity of 20 m/s.
The slow atoms are captured in a 3-dimensional MOT formed by 3 pairs of counter-propagating beams. One pair of beam is aligned with the vertical ($z$) axis, and the two other pairs are along the directions ${\boldsymbol}e_x \pm {\boldsymbol}e_y$. Each trapping beam has a circular profile, with a $1/e^2$ radius of 17 mm and a power of 15 mW. It is spatially filtered using a single mode optical fiber and intensity stabilized to better than $1\%$. The MOT magnetic gradients are $b'_x=0.6$ G/cm, $b'_y=5.4$ G/cm and $b'_z=-6$ G/cm. The MOT is thus cigar-shaped, and its long axis coincides with the magnetic guide direction. The capture rate of the MOT is $2\times
10^{10}$ atoms/s.
![Top view of the setup. The atoms emerging from the oven are slowed down by radiation pressure, captured in the magneto-optical trap, and launched into the magnetic guide.[]{data-label="fig1"}](fig1.eps){width="80mm"}
The entrance of the magnetic guide is located 4.6 cm away from the MOT center. The guide, which is placed inside the vacuum system, consists of four parallel, water-cooled copper tubes, in which we run a current $I$ up to $400$ A. Unless otherwise stated, parameters and results given below correspond to $I=240$ A. The entrance of the guide consists in a ring shaped metal part, that is used for the recombination of the electric and cooling water circuits [@cren]. The guide is held in place and shaped by means of ceramic pieces. At the entrance of the guide the distance between the centers of adjacent tubes is 14 mm, resulting in a gradient of $b'_0=200$ G/cm. The first 40 cm of the magnetic guide form a tapered section, obtained by reducing progressively the distance between the tubes to 8 mm. This results in an increase of the trapping gradient to $b'_1=$600 G/cm. The guide provides a confinement of the atoms with a linear potential $\mu b'
(y^2+z^2)^{1/2}$, where $\mu$ is the magnetic moment of the atoms [@Majorana].
To optimize the transfer of the atoms from the MOT into the magnetic guide, we operate in pulsed mode [@vogels], with a duration of $t_{\rm seq}=283$ ms per sequence. In the first 140 ms we capture $3\times10^{9}$ atoms in the MOT. All beams are set at the same frequency $\omega_L$, corresponding to a detuning $\Delta=\omega_L-\omega_A=-3~\Gamma$, where $\Gamma/2\pi=5.9$ MHz is the natural width of the $5P_{3/2}$ excited state and $\omega_A$ the resonance frequency of a rubidium atom. The trapped cloud is cigar-shaped, with a $1/\sqrt{e}$ radius of 1.4 mm and a length of 35 mm.
The MOT magnetic field is then turned off and a 3 ms launching phase occurs. The detunings of the beams in the $xy$ plane are shifted by $\pm k_L v/\sqrt{2}$ with respect to the detuning $\Delta$ of the vertical beams, where $k_L=\omega_L/c$. This forms a moving molasses which sets the atoms in motion at velocity $v=1.1$ m/s in the direction of the magnetic guide [@launch]. The detuning $\Delta$ during the launching phase is ramped to $-10~\Gamma$, which reduces the temperature of the gas to $T_{0}=40 \,\mu$K in the frame moving at velocity $v$.
The final phase of the sequence consists in (i) optical pumping of the moving atoms into the weak-field-seeking ground state $|F=1,m_F=-1\rangle$, and (ii) magnetic pre-guiding of the atom cloud. The pre-guiding is provided by a two-dimensional ($yz$) quadrupole magnetic field $b'_2=65\,$G/cm generated by four elongated coils (not shown in Fig. \[fig1\]), located around the MOT chamber. It prevents the atom cloud from expanding during its flight towards the entrance of the magnetic guide. During the pre-guiding phase we also apply at the MOT location a $56$ G bias field along the $x$ axis to match the oscillation frequency in the pre-guide with the cloud parameters. The overall efficiency of the pumping+pre-guiding process is $60\%$ so that each packet loaded into the guide contains $N_{\rm load}\sim 2\times 10^9$ atoms. We measure the atom number and mean velocity by monitoring the absorption of a resonant probe beam at the exit of the 4.5 m magnetic guide. Losses due to collisions with residual gas are negligible.
We have estimated numerically the influence of the rapid variation of the magnetic gradient from $b'_2$ to $b'_0$ at the entrance in the main guide by analyzing single particle trajectories. The cloud is compressed to a final transverse radius $r_0=400\, \mu$m and an energy per atom of $k_B\times 500\,\mu$K in the moving frame. In the subsequent propagation into the tapered section of the guide, the gradient increases to $b'_1$ and the bias field decreases to 0 [@Majorana]. It leads to a further, quasi-adiabatic, compression of the cloud and the energy per atom raises to $k_B\times 1400$ $\mu$K. Assuming that thermalization occurs via elastic collisions, this corresponds to a temperature of $T_1=400\,\mu$K shared as $3k_BT_1/2$ for the kinetic energy in the moving frame and $2k_BT_1$ for the potential energy of the transverse linear confinement. This estimate for $T_1$ is in good agreement with our measurement as we shall see below.
![a) Absorption signal measured at the end of the guide for 1, 2, 4 and 24 launches. b) Number of atoms (area of the absorption signal) as a function of the number of launches. The time origin corresponds to the launching of the first packet.[]{data-label="fig2"}](fig2.eps){width="80mm"}
The next capture+launch sequence starts 140 ms after the beginning of the pre-guiding phase, so that the atom packet has moved $\sim
14$ cm away from the MOT center and is now well inside the main guide. A remarkable feature of this protocol is that the light scattered from the MOT during the preparation of a packet does not produce detectable losses on the previous packet. Fig. \[fig2\] shows the signal detected for different numbers of launches. One can check that the area of the signal scales linearly with this number.
To generate a continuous beam we repeat the sequence with a cycling rate of $t_{\rm seq}^{-1}=3.5$ Hz. For $T=400\,\mu$K the packets overlap after a propagation of 60 cm. The flux is $\phi=N_{\rm load} t_{\rm seq}^{-1}= 7\,(\pm 2) \times 10^{9}$ atoms/s.
We now turn to the characterization of the beam and the demonstration that the collisional regime has indeed been reached. The main tools for this study are radio-frequency (r.f.) antennas located along the guide, which generate an oscillating magnetic field at frequency $\nu$. An antenna flips the magnetic moment of the atoms passing at a distance $r_{\rm e}=h\nu/(\mu b')$ from the $x$ axis (r.f. evaporation [@rfevaporation]). For an atom with transverse energy $E$ and angular momentum $L$ along the $x$ axis, the evaporation criterion $r=r_{\rm e}$ is fulfilled at some points of the atom trajectory if $$E\; \geq \; \frac{L^2}{2mr_{\rm e}^2}+\mu b' r_{\rm e}\ ,
\label{evapcond}$$ where $m$ is the atom mass. The range of an antenna is $\sim
20$ cm, so that it constitutes a local probe of the $(E,L)$ distribution at the scale of the whole guide.
![ Experiment with two antennas separated by a distance $d$: remaining fraction after the second antenna, with the first antenna set at a fixed value $\nu_1$. a) Antenna 1 off, flux $\phi$, $b'_1=600$ G/cm and $d=3$ m. Using (\[fitfunction\]) as a fit function (solid line), we find $T_1=384\, (\pm\, 13)\,\mu$K. b) $\nu_1=24$ MHz with the same conditions as for a). The fit gives $T_2=281 \,(\pm8)\;\mu$K indicating a significant cooling. c) $\nu_1=8$ MHz and reduced number of collisions between antennas: flux $\phi/3$, $b'_1=400$ G/cm, $d=0.85$ m. The dotted line is the prediction for the collisionless regime. The solid line is the result expected if thermalization had occurred.[]{data-label="fig3"}](fig3.eps){width="85mm"}
The temperature in the guide is deduced from the variation of the fraction $f(\nu)$ of remaining atoms after the antenna. For $r_{\rm e}$ small compared to the thermal size $r_{\rm
th}=2k_BT/(\mu b')$, only a very small fraction of atoms has a sufficiently low angular momentum to be evaporated, as a consequence $f(\nu\rightarrow 0)=1$. In the opposite limit $r_{\rm
e}\gg r_{\rm th}$, the evaporated fraction is negligible because of the exponential decay of the energy distribution, thus $f(\infty)=1$. We have used a Monte-Carlo determination of $f(\nu)$, that is accurately fitted by the analytical form: $$f(\nu)=1-A\,\eta^{B}\,e^{-C\eta} \quad \mbox{ with }\quad
\eta=h\nu/(k_BT)
\label{fitfunction}$$ with $A=1.65$, $B=1.13$ and $C=0.92$. A minimum value $f^*\simeq
0.33$ is expected for $ h\nu^*/(k_BT) \simeq 1.25\,$.
A typical measurement of $f(\nu)$ is given in Fig. 3a, together with a fit using (\[fitfunction\]). The transverse temperature corresponding to the best fit is $384\, (\pm\, 13)\,\mu$K, in good agreement with the expected temperature given above. For the fit we allow the coefficient $A$ in (\[fitfunction\]) to vary, to account for a possible uncomplete efficiency of the evaporation process. For the data of Fig. \[fig3\]a, the efficiency is better than 90%.
The thermal equilibrium assumption used above is valid only if a sufficient number of elastic collisions (at least a few per atom) has occurred in the guide. The good agreement between the shapes of the experimental and predicted functions $f(\nu)$ is a first indication that the collisional regime has indeed been reached. To investigate this point further, we now turn to a two-antenna experiment (Fig. \[fig3\]b). These antennas are located respectively $d_1=1$ m and $d_2=4$ m after the entrance of the guide. The frequency $\nu_1$ of the first antenna is fixed at $\nu_1=24$ MHz, which corresponds to $\eta_1\equiv
h\nu_1/(k_BT_1)=3$, where $T_1$ is the temperature deduced from the analysis of Fig. \[fig3\]a. This antenna evaporates the atoms fulfilling (\[evapcond\]), thus producing a non thermal ($E,L)$ distribution. We measure the function $f(\nu_2)$ with the second antenna, and check whether a thermal distribution is recovered during the propagation from $d_1$ to $d_2$. The result for $f(\nu_2)$ in Fig. \[fig3\]b shows clearly that this is indeed the case. We infer from this set of data a temperature $T_2=281 \,(\pm8)\;\mu$K. The fact that $T_2<T_1$ is a clear demonstration of evaporative cooling of the beam by the first antenna.
![Ratio $T_2/T_1$ between the temperature $T_2$ of the beam when the first antenna is set at $\nu_1$ and the temperature $T_1$ of the beam in absence of evaporation, as a function of the dimensionless parameter $\eta_1=h\nu_1/(k_BT_1)$. The solid line is the theoretical prediction for a linear potential (no adjustable parameter). The open square corresponds to the data of Fig. \[fig3\]b. The squares (resp. circles) have been obtained with $I=240$ A (resp. $400$ A).[]{data-label="fig4"}](fig4.eps){width="80mm"}
We have investigated the thermalization process as a function of the evaporation frequency $\nu_1$ (Fig. \[fig4\]). For $\eta_1
\lesssim 1.25$ the evaporated atoms have less energy than average, leading to heating ($T_2>T_1$). Cooling is observed in the opposite case $\eta_1 \gtrsim 1.25$. The continuous line in Fig. \[fig4\] is the theoretical prediction assuming full thermalization between $d_1$ and $d_2$. It is in excellent agreement with the experimental results. The phase space density at the center of the beam scales as $\phi/T^{7/2}$. When we choose $\eta_1=3$ (Fig. \[fig3\]b), we measure a decrease of the flux by $36\%$ and a decrease of temperature of $27\%$. This corresponds to an *increase* in phase space by a factor $1.92\,(\pm \,0.15)$, to be compared with the expected gain of 1.90. It shows that this transverse evaporation mechanism is indeed well suited for our final goal, i.e. bringing the beam in the degenerate regime.
We have also investigated the transition between the collisional and collisionless regimes. For this study we have reduced the collision rate by decreasing the flux to $\phi/3$ and by operating the magnetic guide at a lower current (160 A). The measured temperature is in this case $T_1=230\,(\pm\,6)\;\mu$K. To minimize the thermalization process between the antennas, we have also reduced the distance $d_2-d_1$ to 0.85 m. The result for $f(\nu_2)$ is shown in Fig. \[fig3\]c for the particular choice $\nu_1=8\,$MHz. The dashed line gives the result expected in the collisionless regime. In this case the second antenna does not induce extra loss when $\nu_2=\nu_1$ since no modification of the $(E,L)$ distribution occurs between the two antennas. The measured $f(\nu_2)$ is in good agreement with this prediction. For comparison we also indicate in Fig. \[fig3\]c the expected signal in case of a complete thermalization between the antennas. It clearly does not agree with our data, contrary to the large flux case (Fig. \[fig3\]b).
To summarize, our guided beam is deeply in the collisional regime when operated at the maximum flux. The collision rate in the linear guide is $\gamma_c=n_0\, \sigma\, \Delta
v/\sqrt{\pi}=2\,$s$^{-1}$, where $n_0=2.8\times
10^{10}\,$cm$^{-3}$ is the central density, $\sigma=6.8\times
10^{-12}\,$cm$^2$ is the s-wave elastic cross-section at low temperature for spin polarized $^{87}$Rb atoms and $\Delta
v=\sqrt{k_BT/ m}=19\,$cm/s. This corresponds to $N_{\rm col}=8$ collisions per atom along the whole guide. For the thermalization experiment of Fig. \[fig3\]b and taking into account the reduction of flux by the first antenna, we estimate that each atom undergoes 3 collisions between the two antennas. This is known to be the minimum value to ensure thermalization (see e.g. [@timecoll]).
Finally we briefly discuss the possibility to use this device for producing a coherent continuous atomic wave. The current phase space density at the center of the beam is $2\times 10^{-8}$. To reach quantum degeneracy, we plan to use several antennas with decreasing frequencies. Operating all antennas at the same $\eta$, the cooling distance is minimized for $\eta\sim\,$4. A total of 40 antennas (each providing a phase space increase of 1.63) is needed. The expected output flux is $2\times 10^{-4}\phi$ for a cooling distance of $\sim 100\, v/\gamma_c$. To match this distance with the length of our guide, we need to increase the initial collision rate $\gamma_c$ by one order of magnitude. A fraction of this gain can be obtained by slowing down the beam by gravity, using a bent or tilted guide [@ldea]. In addition we plan to transpose to our elongated MOT the “dark spot" technique, which is known to lead to a significant gain in terms of spatial density of the laser cooled atomic source [@Darkspot]. A continuous coherent atomic beam will be a very useful tool not only for the interferometric and metrology applications mentioned above, but also for studies of fundamental issues, such as the existence of superfluidity in quasi one-dimensional systems. Another exciting challenge will be the miniaturization of the setup, along the line of what has been achieved recently with “atom chips” for the production of standard BEC [@chips].
We acknowledge stimulating discussion with the ENS laser cooling group. We thank A. Senger for his participation to the data acquisition. K. Günter acknowledges support from the Marie Curie Host Fellowship “QPAF” HPMT-CT-2000-00102. J. M. Vogels acknowledges support from the Research Training Network “Cold Quantum Gases” HPRN-CT-2000-00125. This work was partially supported by the CNRS, the Région Ile-de-France, the Collège de France, the Ecole Normale Supérieure, the Bureau National de la Métrologie, the University of Paris VI and the Délégation Générale de l’Armement.
[99]{}
Unité de Recherche de l’Ecole Normale Supérieure et de l’Université Pierre et Marie Curie, associée au CNRS.
P. R. Berman, [*Atom interferometry*]{}, (Academic Press, London, 1997).
C. J. Pethick and H. Smith, [*Bose-Einstein Condensation in Dilute Gases*]{}, (Cambridge University Press, Cambridge, 2002); L. P. Pitaevskii and S. Stringari, [*Bose-Einstein condensation*]{}, (Clarendon Press, Oxford, 2003).
K. Helmerson, D. Hutchinson and W. Phillips, Physics World [**12**]{}, 31 (1999) and refs. therein.
A. P. Chikkatur [*et al.*]{}, Science **296**, 2193 (2002).
E. Mandonnet [*et al.*]{}, Eur. Phys J. D **10**, 9 (2000).
M. Key [*et al.*]{}, Phys. Rev. Lett. **84**, 1371 (2000); N. H. Dekker [*et al.*]{}, Phys. Rev. Lett. **84**, 1124 (2000); B. K. Teo and G. Raithel, Phys. Rev. A **63**, 031402 (2001); J. A. Sauer, M. D. Barrett, and M. S. Chapman, Phys. Rev. Lett. **87**, 270401 (2001).
P. Cren [*et al.*]{}, Eur. Phys J. D **20**, 107 (2002).
H. J. Metcalf and P. van der Straten, [*Laser Cooling and Trapping*]{}, (Springer-Verlag, New York, 1999).
Strictly speaking, the magnetic field at the center of the guide does not vanish thanks to the $x$ component of the Earth magnetic field, which prevents Majorana spin flips.
J. M. Vogels [*et al.*]{}, cond-mat/0404560.
A. Clairon, C. Salomon, S. Guellati. and W.D. Phillips, Europhys. Lett., [**16**]{}, 165 (1991).
O. Luiten, M. Reynolds, J. Walraven, Phys. Rev. A **53**, 381 (1996); W. Ketterle and N. van Druten, Adv. At. Mol. Opt. Phys. **37**, 181 (1996).
C. R. Monroe [*et al.*]{}, Phys. Rev. Lett. **70**, 414, (1993).
T. Lahaye, P. Cren, C. Roos, D. Guéry-Odelin, Comm. Nonlin. Sci. Num. Sim. [**8**]{}, 315 (2003).
W. Ketterle [*et al.*]{}, Phys. Rev. Lett. **70**, 2253, (1993).
W. Hänsel, P. Hommelhoff, T. W. Hänsch, and J. Reichel, Nature (London) **413**, 498 (2001); H. Ott, J. Fortägh, G. Schlotterbeck, A. Grossmann, and C. Zimmermann, Phys. Rev. Lett. **87**, 230401 (2001); A. E. Leanhardt [*et al.*]{}, ibid **89**, 040401 (2002); S. Schneider [*et al.*]{} Phys. Rev. A **67**, 023612 (2003).
| {
"pile_set_name": "ArXiv"
} |
-30pt plus 2mm minus 2mm 22.0cm
****
Final-State Interaction in Electron Scattering\
by Polarized Nuclei[^1]
[J.E. Amaro$^{a}$ and T.W. Donnelly$^b$]{}
$^a$ [*Departamento de Física Moderna, Universidad de Granada, Granada 18071, Spain*]{}\
$^b$ [*Center for Theoretical Physics, Laboratory for Nuclear Science and Dept. of Physics,* ]{}\
[*Massachusetts Institute of Technology, Cambridge, MA 02139, U.S.A.*]{}\
1.5 cm
------------------------------------------------------------------------
[**Abstract**]{}
The cross section for the reaction $^{39}\vec{{\rm K}}(e,e'p)^{38}$Ar is computed, and the dependence of the FSI on the initial polarization angles of the nucleus $^{39}$K is analyzed. The results are explained in a semi-classical picture of the reaction. This procedure allows us to find the best initial kinematical conditions for minimizing the FSI.
------------------------------------------------------------------------
Introduction
============
In this talk we explore the effects of the FSI in $(e,e'p)$ reactions using polarized nuclei. All of the measurements to date involving medium and heavy nuclei have been performed with unpolarized targets [@Bof96] and hence only the global effects of FSI averaged over all polarization directions have been addressed experimentally. Using instead polarized nuclei as targets, new possibilities to extract the full tri-dimensional momentum distribution of nuclei will become available [@Cab93; @Cab94; @Gar95; @Cab95].
The few theoretical studies of $(e,e'p)$ reactions involving polarized, medium and heavy nuclei in DWIA [@Bof88; @Gar95; @Ama98; @Ama99] report a dependence of FSI effects — or nuclear transparency — on the choice of polarization angles. In the present work we show that these variations of the transparency can be understood in terms of the orientation of the initial-state nucleon’s orbit and of the attenuation of the ejected nucleon’s flux.
We shall show that one is able to predict the orientations of the target polarization that are optimal for minimizing the FSI effects, providing the ideal situations for nuclear structure studies. This situation occurs when the nucleon is ejected directly away from the nuclear surface. On the other hand, when the nucleon is ejected from the nuclear surface but in the opposite direction — into the nucleus — it has to cross the entire nucleus to exit on the opposite side, and the FSI effects are then found to be maximal. This second situation is ideal for detailed studies of the absorptive part of the FSI. All of these situations can be selected simply by changing the direction of the nuclear polarization.
Coincidence cross section of polarized nuclei. Results
======================================================
Here we present the results of a DWIA calculation of the $^{39}\vec{\rm K}(e,e'p)^{38}{\rm Ar}_{\rm g.s.}$ cross section in the extreme shell model. The present choice is prototypical and the results can be generalized for any polarized nucleus and can be addressed using more sophisticated nuclear models.
We describe the ground state of $^{39}\vec{\rm K}$ as a hole in the $d_{3/2}$ shell of $^{40}$Ca. The initial nuclear state is 100% polarized in the direction $\Omega^*=(\theta^*,\phi^*)$. $$|A(\Omega^*) \rangle = R(\Omega^*)
|d_{3/2}^{-1},m=\textstyle\frac32\rangle ,$$ where $R(\Omega^*)$ is a rotation operator. In this simple model the nuclear polarization is carried by the hole in the $d_{3/2}$ shell. The polarization angles $(\theta^*,\phi^*)$ are the spherical coordinates of the polarization vector ${\mbox{\boldmath$\Omega$}}^*$ with respect to the ${{\bf q}}$-direction ($z$-axis) and with the $x$-axis in the scattering plane.
The final hadronic state is given by a proton in the continuum with energy $\epsilon'$ and momentum ${{\bf p}}'$, plus a daughter $A-1$ nucleus ($^{38}$Ar) in the ground state. This is described in the shell model as two holes in the $d_{3/2}$ shell coupled to total spin $J=0$.
The hole wave function is obtained by solving the Schrödinger equation with a Woods-Saxon potential. The wave function of the ejected proton is obtained by solving the Schrödinger equation with an optical potential for positive energies.
We compute the cross section as $$\Sigma \equiv \frac{d\sigma}{dE'_ed\Omega'_ed\Omega'}
= \sigma_M\left( v_L{{\cal R}}^L + v_T{{\cal R}}^T +
v_{TL}{{\cal R}}^{TL} + v_{TT}{{\cal R}}^{TT}
\right),$$ where $\sigma_M$ is the Mott cross section, $v_K$ are the electron kinematical factors given in [@Ras89], and ${{\cal R}}^K$ are the nuclear response functions. See Refs. [@Ama98; @Ama99] for more details of the model.
Next we show results of a calculation of the $(e,e'p)$ cross section for different nuclear polarizations $(\theta^*,\phi^*)$. The kinematics correspond to the quasi-elastic peak and in-plane emission $$q=500 \,{\rm Mev}/c,
\kern 5mm
\omega=133.5\, {\rm MeV},
\medskip
\kern 5mm
\phi=0,
\medskip
\kern 5mm
\theta_e = 30^{\rm o}$$
[ Cross section computed for two values of the nuclear polarization angles. Left: $\theta^*=90^{\rm o}$, $\phi^*=-90^{\rm o}$; right: $\theta^*=90^{\rm o}$, $\phi^*=90^{\rm o}$. The meaning of the curves is the following: solid: DWIA; dashed: PWIA; dotted: DWIA but with just the imaginary part of the central optical potential; dash-dotted: DWIA without spin-orbit contributions. ]{}
In fig. 1 we show the cross section for $^{39}\vec{\rm K}$ polarized in the $-y$ direction, (left) and in the $y$ direction (right). The solid lines are the full DWIA calculation, using the optical potential of Schandt et al. The dashed lines are the cross sections computed in PWIA, i.e., without FSI. The dotted lines correspond to the DWIA, but including in the FSI just the central imaginary part of the optical potential, while the dash-dotted lines include in addition the central real part of the potential.
Comparing the solid and dashed lines, we see that the effect of the FSI (solid lines relative to dashed lines) is quite dependent on the polarization of the nucleus. This fact suggest that the “transparency” of the nucleus to proton propagation can be maximized or minimized by selecting a particular polarization of the nucleus and that if one is able to understand physically the different behavior seen for the FSI effects in fig. 1, then it could be possible to make specific predictions about the reaction for future experiments.
A semi-classical picture of the reaction
========================================
In order to understand physically the above results we will consider a semi-classical model of the reaction by assuming it to take place in two or more steps as follows: first a proton with (missing) momentum ${{\bf p}}$ and energy $\epsilon$ is knocked-out by the virtual photon and it acquires momentum ${{\bf p}}'$ and energy $\epsilon'=\epsilon+\omega$. Second, as this high-energy nucleon traverses the nucleus it undergoes elastic and inelastic scattering which, in our model, are produced by the real and imaginary parts of the optical potential.
The important point here is that the nucleus is polarized in a specific direction. Accordingly, the initial-state nucleon can be localized in an oriented (quantum) orbit. From the knowledge of this orbit and of the missing momentum one can predict the most probable location of the struck proton, and therefore one can specify the quantity of nuclear matter that the proton must cross before exiting from the nucleus with momentum ${{\bf p}}'$.
We illustrate the case of a particle in a $d_{3/2}$ wave. which is the relevant state for our calculation. First consider that the particle is polarized in the $z$-direction (${\mbox{\boldmath$\Omega$}}^* = {{\bf e}}_3$). The corresponding wave function can be written as $$\textstyle
|\frac32\frac32\rangle
= \psi_1 |\uparrow\rangle +\psi_2 |\downarrow\rangle .
\label{e11}$$ where the up and down components are given by $$\begin{aligned}
\psi_1 &=& -\sqrt{\frac{3}{8\pi}}\sin\theta\cos\theta\;
{\rm e}^{i\phi}R(r)
\label{e12}\\
\psi_2 &=& -\sqrt{\frac{3}{8\pi}}\sin^2\theta\;
{\rm e}^{2i\phi}R(r).
\label{e13}\end{aligned}$$ Here the angles $(\theta,\phi)$ are the spherical coordinates of the particle’s position ${{\bf r}}$ and $R(r)$ is its radial wave function. The spatial distribution is then given by the single-particle probability density $$\rho({{\bf r}}) = |\psi_1|^2+|\psi_2|^2 =
\frac{3}{8\pi}\sin^2\theta|R(r)|^2.
\label{e17}$$
[ Pictorial representation of the spatial distribution of a proton in the $d_{3/2}$ shell, shown as a torus-like distribution for two different polarizations. ]{}
Taking into account the form of the radial wave function for the $d_{3/2}$ wave, we can see that the particle is distributed around the center of the nucleus in a toroidal-like (quantum) orbit as shown schematically in fig. 2 (upper part). In a semi-classical picture of the bound state, we can imagine the particle performing a rotatory orbit within the torus in a counter-clock sense, as shown in the figure. The shape of the distribution for arbitrary polarization ${\mbox{\boldmath$\Omega$}}^*$ is just a rotation of the above distribution, as also shown in fig. 2 (bottom).
The next step is to localize the particle within the orbit for a given value of the missing momentum ${{\bf p}}$. From elementary quantum mechanics we employ the Fourier transform $\tilde{\psi}({{\bf p}})$ of the wave function and the position operator in momentum space ${{\bf \hat{r}}}=i\nabla_p$ to define the local position of the nucleon in the orbit for momentum ${{\bf p}}$ in the following way: $${{\bf r}}({{\bf p}}) = \frac{{\rm Re}\; \tilde\psi^{\dagger}({{\bf p}})(i\nabla_p)
\tilde\psi({{\bf p}})} {\tilde\psi^{\dagger}({{\bf p}})\tilde\psi({{\bf p}})} .$$ This is a well-defined vector which represents the most probable location of a particle with momentum ${{\bf p}}$ when it is described by a wave function $\psi$. Henceforth ${{\bf r}}({{\bf p}})$ represents the position of the particle in the orbit in the present semi-classical model.
For the case of interest here of the $d_{3/2}$ orbit polarized in the $z$-direction, we compute the position ${{\bf r}}({{\bf p}})$ by using the wave function given in eqs. (\[e11\]–\[e13\]) in momentum space: $${\rm Re}\, \tilde\psi^{\dagger}({{\bf p}})i\nabla_p\tilde\psi({{\bf p}})=
-\frac{3}{8\pi}\frac{|\tilde{R}(p)|^2}{p}\sin\theta(1+\sin^2\theta){\hat{\mbox{\boldmath$\phi$}}},$$ where now $(\theta,\phi)$ are the spherical coordinates of the missing momentum ${{\bf p}}$, $\tilde{R}(p)$ is the radial wave function in momentum space, and ${\hat{\mbox{\boldmath$\phi$}}}$ is the unit vector in the azimuthal direction. As we see, upon dividing by the momentum distribution (given by eq. (\[e17\]), but in momentum space) $${\tilde\psi^{\dagger}({{\bf p}})\tilde\psi({{\bf p}})}= \frac{3}{8\pi}\sin^2\theta
|\tilde{R}(p)|^2 ,$$ the radial dependence in the numerator and denominator goes away, and we obtain an expectation value of position which is independent of the radial wave function $${{\bf r}}({{\bf p}}) = -\frac{1+\sin^2\theta}{p\sin\theta}{\hat{\mbox{\boldmath$\phi$}}}.$$
This expression has been obtained for the polarization in the $z$-direction. For a general polarization vector ${\mbox{\boldmath$\Omega$}}^*$ we just perform a rotation of the vector ${{\bf r}}({{\bf p}})$. Introducing the angle $\theta_p^*$ between ${{\bf p}}$ and ${\mbox{\boldmath$\Omega$}}^*$, we can write the nucleon position in a general way $${{\bf r}}({{\bf p}})=-\frac{1+\sin^2\theta_p^*}{p^2\sin^2\theta_p^*}
\,{\mbox{\boldmath$\Omega$}}^*\times{{\bf p}}.
\label{e23}$$
Using the above definitions we can give a physical interpretation of the results shown in fig. 1. The kinematics for the case of the $^{39}$K nucleus polarized in the $-y$ direction are illustrated in fig. 3(a). Therein, the momentum transfer points in the $z$-direction and we show the missing-momentum vector ${{\bf p}}$ corresponding to the maximum of the momentum distribution, $p\sim 140$ MeV/c. The momentum of the ejected proton ${{\bf p}}'$ is also shown in the picture. For ${\mbox{\boldmath$\Omega$}}^*$ pointing in the $-y$ direction, the semi-classical orbit lies in the $xz$-plane and follows a counter-clockwise direction of rotation. For these conditions, the most probable position of the proton before the interaction is indicated with a black dot near the bottom of the orbit. As the particle is going up with momentum ${{\bf p}}'$ after the interaction with the virtual photon, it has to cross all of the nucleus and exit it by the opposite side; thus one expects that the FSI will be large in this situation, as shown in the left panel of fig. 1.
[ Semi-classical orbit and location of the proton for the same kinematics and nuclear polarizations as in Fig. 1. ]{}
In fig. 3(b) we show the picture for the opposite polarization in the $y$-direction. In this case the nucleon distribution in the orbit is the same as in (a), but the rotation direction is the opposite, Hence now it is more probable for the nucleon to be located near the upper part of the orbit. As the nucleon is still going up with the same momentum ${{\bf p}}'$, the distance that it has to travel through the nucleus is much smaller than in case (a), and hence one expects small FSI effects, namely, what is seen in the right panel of fig. 1.
We have arrived at a very intuitive physical picture of why the FSI effects differ for different orientations of the nuclear spin: the polarization direction fixes the orientation of the nucleon distribution (quantum orbit). For a given value of the missing momentum one can locate the particle in a definite position within the orbit, and therefore within the nucleus. Assuming that the particle leaves the nucleus with the known momentum ${{\bf p}}'$, one can determine the quantity of nuclear matter that it has to cross before exiting.
In order to check the above picture for any nuclear polarization, we have computed the cross section for a set of 26 different nuclear polarization angles expanding the $(\theta^*,\phi^*)$ plane. Using equation (\[e23\]) we have computed the distance $s$ of the nucleon trajectory within the nucleus, by choosing some appropriate value for the nuclear radius $R$. A model of exponential attenuation of the cross section due to nuclear absorption can be crafted in the following way: $$\Sigma_{DWIA}\simeq \Sigma_{PWIA}\;{\rm e}^{-s/\lambda} ,$$ where $\lambda$ is a free parameter to be interpreted as the mean free path (MFP). Within this approximation, the nuclear transparency, defined as the ratio between the DWIA and PWIA results, can be written as $$\label{e34}
T \equiv \frac{\Sigma_{DWIA}}{\Sigma_{PWIA}} \simeq {\rm e}^{-s/\lambda}.$$
[ Nuclear transparency as a function of the nucleon path $s$ within the nucleus for different nuclear polarizations. The FSI only include the imaginary part of the central optical potential. ]{}
In fig. 4 we show the nuclear transparency as a function of the distance $s$ to the nuclear surface, computed for different polarizations, at the maximum of the cross section. For the FSI we have used just the central, imaginary part of the optical potential. In this figure we see that the dependence of $\log T$ can in fact be approximated by a straight line. By performing a linear regression we obtain a MFP of $\lambda=8.4$ fm. This value is quite independent of the radius $R$ in the region between $r_{1/2}$ and $r_{1/10}$, where the nuclear density $\rho(r)$ takes the values $\rho(0)/2$ and $\rho(0)/10$, respectively.
Applications to two-particle emission reactions
===============================================
Finally we give a possible application of the above model to two-hadron emission reactions. Consider $(e,e'N\pi)$ reactions from polarized nuclei in the $\Delta$-region. By selecting the appropriate nuclear polarization, one could reduce or enhance the FSI of the final $\Delta$ in the nuclear medium. In fact, using the above model, one have control on the point of the nucleus where the $\Delta$ is created. Making a crude estimation of the length that the $\Delta$ travels before decaying $$x \sim \frac{\hbar c}{\Gamma_{\Delta}} \sim \frac{200}{120}\,\rm fm
\sim 1.7\, fm$$ we see that it could be possible to produce the $\Delta$ in the two situations shown in figure 5.
[ Electroproduction of a $\Delta$ from a polarized nucleus for two different nuclear polarization directions. ]{}
In the case (a) the $\Delta$ is created near the nuclear surface and propagates into the nucleus. It has large FSI and decays inside the nucleus into a pair $N+\pi$ which also interacts with the nucleus. In the second case (b) the $\Delta$ propagates out of the nucleus. The FSI of the $\Delta$ is expected to be smaller. The $\Delta$ decays outside of the nucleus. This situation is cleaner to study the $\Delta$ electroproduction amplitude in nuclei without too much distortion by FSI. Case (a) is ideal to study the $\Delta$ properties in the nuclear medium.
Summary and conclusions
=======================
We have studied the reaction $^{39}$K$(e,e'p)^{38}\rm Ar_{gs}$ for polarized $^{39}$K in DWIA. We have studied the dependence of the FSI as a function of the nuclear polarization direction and introduced a physical picture of the results in order to understand the different effects seen in the cross section.
The argument to explain the FSI effects is based on the PWIA and it has been illustrated by introducing the semi-classical concept of a nucleon orbit within the nucleus. For given kinematics we can fix the expectation value of the position of the nucleon within the nucleus before the interaction. From this information we have computed the length of the path that the nucleon travels across the nucleus for each polarization.
Our results show that when the FSI effects are large the computed nucleon path through the nucleus is also large, whereas the opposite happens when the FSI effects are small. Thus, by selecting the appropriate nuclear polarization, one can reduce or enhance the FSI effects. Such control should prove to be very useful in analyzing the results from future experiments with polarized nuclei.
Finally, our model can also be applied to the $(e,e'N\pi)$ reaction in the $\Delta$ peak. Since by flipping the nuclear polarization one can go from big to small FSI effects of the $\Delta$ , this opens the possibility of using this kind of reaction to distinguish the FSI effects from other issues of interest, such as the $\Delta$ electroproduction amplitudes in the medium.
Acknowledgments {#acknowledgments .unnumbered}
===============
This work is supported in part by funds provided by the U.S. Department of Energy (D.O.E.) under cooperative agreement \#DE-FC01-94ER40818, in part by DGICYT (Spain) under Contract No. PB92-0927 and the Junta de Andalucía (Spain) and in part by NATO Collaborative Research Grant \#940183.
[Expo92]{}
S. Boffi, C. Giusti, F. D. Pacati and M Radici, Electromagnetic Response of Atomic Nuclei, Oxford University Press, 1996
J.A. Caballero, T.W. Donnelly and G.I. Poulis, Nucl. Phys. [**A555**]{} (1993) 709.
J.A. Caballero, T.W. Donnelly, G.I. Poulis, E. Garrido and E. Moya de Guerra, Nucl. Phys. [**A577**]{} (1994) 528.
E. Garrido, J.A. Caballero, E. Moya de Guerra, P. Sarriguren and J.M. Udías, Nucl. Phys. [**A584**]{} (1995) 256.
J.A. Caballero, E. Garrido, E. Moya de Guerra, P. Sarriguren, J.M. Udias, Ann. Phys. [**239**]{} (1995) 351.
S. Boffi, C. Giusti and F.D. Pacati, Nucl. Phys [**A476**]{} (1988) 617.
J.E. Amaro and T.W. Donnelly Ann. Phys. (N.Y.) [**263**]{} (1998) 56.
J.E. Amaro and T.W. Donnelly Nucl. Phys. [**A 646**]{} (1999) 187.
A.S. Raskin and T.W. Donnelly, Ann. Phys. (NY) [**191**]{} (1989) 78.
[^1]: Talk presented by J.E. Amaro at the IVth Workshop on Electromagnetically induced Two-Hadron Emission, Granada, Spain, 1999
| {
"pile_set_name": "ArXiv"
} |
---
address: 'Max-Planck-Institut für Mathematik in den Naturwissenschaften, Inselstrasse 22, 04103 Leipzig, Germany'
author:
- Marc Josien and Felix Otto
title: 'The annealed Calderón-Zygmund estimate as convenient tool in quantitative stochastic homogenization'
---
Abstract {#abstract .unnumbered}
========
This article is about the quantitative homogenization theory of linear elliptic equations in divergence form with random coefficients. We derive gradient estimates on the homogenization error, i. e. on the difference between the actual solution and the two-scale expansion of the homogenized solution, both in terms of strong norms (*oscillation*) and weak norms (*fluctuation*). These estimates are optimal in terms of scaling in the ratio between the microscopic and the macroscopic scale.
The purpose of this article is to highlight the usage of the recently introduced *annealed* Calderón-Zygmund (CZ) estimates in obtaining the above, previously known, error estimates. Moreover, the article provides a novel proof of these annealed CZ estimate that completely avoids quenched regularity theory, but rather relies on functional analysis. It is based on the observation that even on the level of operator norms, the Helmholtz projection is close to the one for the homogenized coefficient (for which annealed CZ estimates are easily obtained).
In this article, we strive for simple proofs, and thus restrict ourselves to ensembles of coefficient fields that are local transformations of Gaussian random fields with integrable correlations and Hölder continuous realizations. As in earlier work, we use the natural objects from the general theory of homogenization, like the (potential and flux) correctors and the homogenization commutator. Both oscillation and fluctuation estimates rely on a sensitivity calculus, *i.e.* on estimating how sensitively the quantity of interest does depend on an infinitesimal change in the coefficient field, which is fed into the Spectral Gap inequality. In this article, the annealed CZ estimate is the only form in which elliptic regularity theory enters.
Stochastic homogenization, Calderón-Zygmund estimates, corrector estimates, oscillations, fluctuations
35B27, 35R60, 35J15, 35B45, 42B15
[ ]{}
Introduction
============
The role of annealed CZ estimates$^{\ref{1}}$
---------------------------------------------
We start with a preamble on language. Homogenization capitalizes on a separation between two length scales: a micro-scale set by the coefficient field $a$ and a macro-scale set by the other data (*e.g.* domain, right-hand side). In the case of stochastic homogenization, the micro-scale is set by a correlation length of the ensemble $\langle\cdot\rangle$; in this article, it is of the same order as the (potentially much smaller) scale up to which the coefficient field $a$ varies little. In this article, the domain is always the whole $d$-dimensional space $\mathbb{R}^d$, so the macro-scale is just set by the scale intrinsic to the right-hand side. Throughout the entire article, the micro-scale is set to be unity – the reader will (almost) nowhere find the $\epsilon$ dear to the mathematical homogenization literature!
There is an intimate connection between elliptic regularity theory and quantitative stochastic homogenization. This became already apparent in the first work on error estimates [@Yurinskii], where Nash’s heat kernel decay for general uniformly elliptic coefficient fields $a$ is used. Like for much of the last decade’s contributions, the present article follows the strategy introduced in [@NaddafSpencer_1998] of using a Spectral Gap inequality, which is assumed to hold for the ensemble $\langle\cdot\rangle$ under consideration and quantifies its ergodicity, to capture stochastic cancellations. It is in the same article [@NaddafSpencer_1998] where CZ estimates for $-\nabla\cdot a\nabla$ in form of the $\LL^4_{\mathbb{R}^d}$-boundedness of the $a$-Helmholtz projection $-\nabla(-\nabla\cdot a\nabla)^{-1}\nabla\cdot$ entered the stage as a part of the sensitivity estimate. In that work, this boundedness was guaranteed by assuming a small ellipticity contrast, appealing to Meyers’ perturbative argument. In [@GloriaOtto_2011], the assumption of small ellipticity was removed by using de Giorgi’s theory in order to obtain optimal decay estimates on the Green function $G(a;x,y)$, upgraded to optimal decay of its mixed derivative with help of Caccioppoli’s estimate.
Those were *quenched* estimates on the *quenched* Green function $G(a;x,y)$, that is, estimates valid for almost every realization $a$ of the coefficient field, but they required some macroscopic spatial averaging in case of the derivatives. Optimal [*pointwise*]{} decay estimates on the mixed derivative of the Green function were first obtained in [@Marahrens_Otto]; these estimates can only hold with high probability, and were formulated as *annealed* estimates in the sense that all algebraic moments with respect to the ensemble $\langle\cdot\rangle$ have the desired behavior. Note that the mixed derivative of the Green function is the kernel of the $a$-Helmholtz projection. Hence, loosely speaking, the estimates in [@Marahrens_Otto] yield “up to a logarithm” the boundedness of the $a$-Helmholtz projection from $\LL^p_{\mathbb{R}^d}(\LL^r_{\langle\cdot\rangle})$ to $\LL^p_{\mathbb{R}^d}(\LL^{r'}_{\langle\cdot\rangle})$, with an arbitrarily small loss $r'<r$ in the stochastic integrability. We call this family of estimates *annealed CZ estimates*.
The result in [@Marahrens_Otto] showed that with overwhelming probability, the random Green function has no worse decay than the constant-coefficient one. This emergence of large-scale regularity is well-known in periodic homogenization since the work [@AvellanedaLin] (see the recent monograph [@Shen]), where like in classical regularity theory, a Campanato iteration propagates Hölder regularity over dyadic scales, but unlike classical theory, the regularity of the homogenized operator is propagated [*downward*]{} in scales, appealing to qualitative homogenization itself. In [@Armstrong], this approach was adapted to the random case, by substituting the compactness argument by a more quantitative one, establishing the pivotal large-scale Lipschitz estimate. This was refined to large-scale quenched $\CC^{\alpha}$ (Schauder) estimates for the $a$-Helmholtz projection in [@Gloria_Neukamm_Otto_2019]. Here, by a large scale quenched regularity theory we mean that there exists a stationary random radius $\rstar\sim 1$ starting from which the theory holds almost surely (which is reminiscent of the random time scale starting from which Euclidean heat kernel decay holds in percolation theory). In [@Armstrong_CZ_2016], the large-scale Lipschitz estimate was used to establish large-scale quenched CZ estimates, which however had a small loss in terms of spatial integrability. The optimal large-scale quenched CZ theory, that is, boundedness of the $a$-Helmholtz projection in $\LL^p_{\mathbb{R}_x^d}(\LL^2_{\Boule_{\rstar(x)}(x)})$ (loosely speaking, $\LL^2$ up to scales $\rstar$, $\LL^p$-from scales $\rstar$ onwards) was first obtained in [@Gloria_Neukamm_Otto_2019]. A simpler proof, based on modern kernel-free approaches to classical CZ theory, may be found in [@Armstrong_book_2018] and [@DuerinckxOtto_2019 Proposition 6.4].
While in many error estimates of stochastic homogenization, weighted $\LL^{2+\epsilon}$-estimates are an alternative to $\LL^p$-estimates, the usage of CZ estimates seems crucial for the theory of fluctuations as conceived in [@DuerinckxGO_2016], where again $\LL^4_{\mathbb{R}^d}$ comes up naturally. As worked out in [@DuerinckxOtto_2019], in particular in the elaborated case of a higher-order theory of fluctuations considered there, the annealed version of CZ estimates is very convenient. In that paper, see [@DuerinckxOtto_2019 Section 6], the annealed estimates were derived from the quenched ones, appealing to the good stochastic integrability of $\rstar$ established in [@Gloria_Neukamm_Otto_2019], and appealing to the same tool that was used to derive the quenched estimates from the Lipschitz estimates.
In this paper, we completely avoid the large-scale quenched regularity theory, be it Lipschitz, Schauder, or CZ. We derive the annealed CZ theory in a single step by, loosely speaking, establishing that the $a$-Helmholtz projection is close in the corresponding [*operator norm*]{} to the Helmholtz projection for the homogenized coefficient $\abar$. This amounts to a particularly strong version of an estimate of the homogenization error. Indeed, on the level of the Helmholtz projection, which itself has no regularizing effect, the homogenization error has a chance of being small only if the r. h. s. lives on scales $\gg 1$. Estimating the homogenization error in an operator norm thus requires splitting the r. h. s. into a large-scale part and its remainder. This is best done *intrinsically* and here is carried out by considering the family of elliptic operators with massive term $\frac{1}{T}-\nabla\cdot a\nabla$, which provides the flexibility of choosing a length scale $\sqrt{T}$. The main challenge arises from the fact that the estimate on the homogenization error relies on an estimate of the corrector which because of its stochastic nature reduces the exponent $r$ in $\LL^p_{\mathbb{R}^d}(\LL^r_{\langle\cdot\rangle})$ and thus prevents a straightforward buckling. The key idea is to trade-in smallness in terms of $\sqrt{T}$ to compensate the loss of integrability in terms of $r$ in form of a (real) interpolation argument. We hope that this functional-analytic argument finds applications in situations where the more PDE-based argument does not apply.
The developments in quantitative stochastic homogenization
----------------------------------------------------------
Qualitative stochastic homogenization has been studied for forty years [@Varadhan_1979; @JKO]. Nevertheless, as opposed to the periodic framework, where optimal estimates of the homogenization error were obtained quite early (see *e.g.* [@AvellanedaLin; @Tartar] and the monograph [@Shen]), quantitative stochastic homogenization remained long out of reach. On a qualitative level, for instance when it comes to the construction of the corrector, the periodic case and the random ergodic case can be treated in parallel, by lifting the corrector equation to the torus and the probability space, respectively. For a quantitative treatment, it is crucial that the torus admits a Poincaré estimate, while there is no analogue in the general ergodic case, so that the qualitative sublinearity cannot be easily improved upon.
There has been substantial progress in quantitative stochastic homogenization in the last decade: A suitable substitute for the missing Poincaré estimate on probability space has been identified in the mathematical physics literature [@ConlonNaddaf; @NaddafSpencer_1998] in form of a spectral gap estimate, which quantifies ergodicity. This was taken up and combined with more elliptic regularity theory in [@GloriaOtto_2011] to treat, for the first time, the standard non-perturbative situation; [@Gloria_Otto; @Gloria_Neukamm_Otto_2019] are key papers that pushed this approach further. An alternative approach based on quantified ergodicity in terms of a finite range assumption has been introduced in [@Armstrong], extending the regularity theory [@AvellanedaLin] from periodic to random, and culminating in the monograph [@Armstrong_book_2018].
One merit of homogenization is the reduction in numerical complexity it entails, reducing the need to resolve the micro-scale to a couple of model calculations. Numerics is indeed a major motivation behind a quantitative theory for stochastic homogenization, see the review [@Anantharaman_2012]. In particular, devising efficient methods for computing the homogenized coefficient from the knowledge of the heterogeneous media has attracted much attention. See in particular [@Gloria_Otto; @Mourrat_2019] for general methods, and the attempts [@Minvielle_2015; @Fischer_ARMA_2019; @CloJoOtXu_2020] to reduce different sources of error (namely the random and the systematic one) in specific settings.
Organization of the paper
-------------------------
Next to providing a novel, and self-contained, proof for the annealed CZ estimates, see Proposition \[PropCZ\], the second purpose of this paper is to display their benefit in error estimates. To this aim, we display how the annealed CZ estimate is crucial in obtaining two natural estimates of the homogenization error on the level of the gradient, namely in strong norms (*oscillations*), *cf.* Corollary \[Cor1\], and in weak norms (*fluctuations*), *cf.* Proposition \[Propfluct\]. These results are not new, see [@Gloria_Neukamm_Otto_2019] for oscillations, and [@DuerinckxGO_2016] for fluctuations. The usage of the annealed CZ estimates to derive fluctuation estimates was recently introduced in [@DuerinckxOtto_2019].
However, the statements given here are more transparent and the proofs are simpler because we allow ourselves to work in the simplest class of ensembles $\langle\cdot\rangle$, namely those that can be obtained by a pointwise Lipschitz transformation of stationary Gaussian ensembles with integrable correlations (which is the crucial large-scale assumption) and Hölder continuous realizations (which is a convenient small-scale assumption). Also, presenting the corrector estimates, *cf.* Proposition \[Propcorr\], alongside with the oscillation and fluctuation estimates allows us to highlight the many similarities: the use of an annealed CZ estimate in the perturbative regime in Proposition \[Propcorr\], the explicit representation of the Malliavin derivatives in Proposition \[Propcorr\] and \[Propfluct\] via auxiliary dual problems on which the annealed CZ estimates are applied, and the use of the $\LL^p$-version of the Spectral Gap estimate.
This presentation, together with the new proof of Proposition \[PropCZ\], stresses the role of functional analysis techniques for quantitative stochastic homogenization, and minimizes the use of finer PDE-ingredients, which hopefully resonates well with some audiences. In this sense, the article has also the nature of a review article – in fact, it grew out of a mini-course the second author gave in Toulouse in the spring of 2019. We now address the content of the individual sections.
In Section \[SecHcv\], we introduce the deterministic part of the setting: Next to the uniformly elliptic coefficient fields $a$ on $\mathbb{R}^d$, we recall the notions of the (extended) correctors $(\phi_i,\sigma_i)_{i=1,\cdots,d}$, of the homogenization commutator, and of the two-scale expansion. In order to stress the importance of a few key identities involving these quantities, which will be crucially used in later sections, and for the convenience of the reader not familiar with homogenization, we derive two elementary statements in the spirit of Tartar’s $H$-convergence in Proposition \[PropHcon\].
In Section \[SecCorrEstim\], we get to the stochastic part of the setting by introducing the class of ensembles $\langle\cdot\rangle$ of coefficient fields $a$ considered here, and derive the two main features that we actually need: 1) An $\LL^p_{\langle\cdot\rangle}$-version of the Spectral Gap estimate with the carré-du-champs $\LL^2_{\mathbb{R}_x^d}(\LL^1_{\Boule_1(x)})$ (i.e. $\LL^1$ on scales up to $1$ and $\LL^2$ on scales larger than $1$) and 2) classical Schauder and CZ regularity theory for the elliptic operator on scales up to $1$ with overwhelming probability, see Lemma \[LemSG\].
In Section \[SecCorrEstim2\], we state and prove the corrector estimates, which take a particularly simple form thanks to Lemma \[LemSG\], see Proposition \[Propcorr\]. In particular, this allows for a simple passage from the estimate of spatial averages of the gradient $\nabla \phi_i$ to the estimates of increments of $\phi_i$, and for a simple passage from the flux $q_i:=a(e_i+\nabla\phi_i)$ to $\nabla\sigma_i$. Modulo this easy post-processing, Proposition \[Propcorr\] relies on the estimate of spatial averages of the field/flux pair $(\nabla\phi_i,q_i)$, which requires the buckling of a stochastic and a deterministic estimate. The stochastic estimate relies on a representation of the Malliavin derivative of the spatial average with help of the solution $\nabla v$ of a dual problem, the estimate of $\nabla v$ via the perturbative version of the annealed CZ estimate, and the $\LL^p$-version of the Spectral Gap estimate. The deterministic estimate relies on an elementary but subtle regularity property of $a$-harmonic functions, here applied to the harmonic coordinate $x_i+\phi_i$, which is reminiscent of [@Bella_Giunti_Otto_2017 Lem. 4]. While none of the elements is novel, the combination here is slightly more efficient than in the existing literature.
In Section \[Sec:Osc\] , as a simple application of Proposition \[Propcorr\], we derive the oscillation estimates, *cf.* Corollary \[Cor1\]. In the case of $\LL^p_{\mathbb{R}^d}$ with $p\not=2$, we use Proposition \[PropCZ\] in a straightforward way.
In Section \[SecFluctu\], we state and prove the fluctuation estimates, *cf.* Proposition \[Propfluct\]. In the proof, we stress the similarity with the proof of Proposition \[Propcorr\]. Proposition \[PropCZ\], in its non-perturbative part, now enters in a substantial way.
In Section \[SecAnnealed\], we finally establish the annealed CZ estimate, *cf.* Proposition \[PropCZ\]. In the perturbative regime, this is done via Meyers’ argument and does not involve a loss of stochastic integrability. As discussed in Section \[SS\_AnnealedCZ\], we derive this non-perturbative part of Proposition \[PropCZ\] from Proposition \[Propcorr\] – this is not circular, since in the latter proposition, only the perturbative part of Proposition \[PropCZ\] enters. We will need a version of Proposition \[PropCZ\] for $-\nabla\cdot a \nabla$ replaced by the massive operator $\frac{1}{T}-\nabla\cdot a\nabla$ which will be established along the same lines (see Proposition \[PropmassCZ\]). The strategy for the proof of Proposition \[PropCZ\] is discussed in all detail there. In Section \[SecLocal\], we briefly sketch the alternative strategy relying on quenched CZ estimates.
The appendix contains several auxiliary statements, in particular those that have to do with the usage of local regularity on scales less than $1$.
Notations
---------
Throughout this article, a few conventions will be used without further notice.
We denote by $e_i$ the canonical basis of $\R^d$. We use Einstein’s convention for summation. The symbol $\cdot$ denotes the scalar product. In particular, $\nabla \cdot$ stands for the divergence operator. The divergence of a matrix is taken with respect to the second coordinate, that is, $$\begin{aligned}
e_j \cdot (\nabla \cdot \sigma_i)=\partial_k \sigma_{ijk}.
\end{aligned}$$ For simplicity, we omit the parentheses when considering operators of the type $\nabla \cdot a\nabla$: We write $\nabla \cdot a \nabla u$ for $\nabla \cdot ( a \nabla u)$. If $f$ and $g$ are vector fields with $m$ and $n$ coordinates in the canonical basis, respectively, we denote by $(f,g)$ the $(m+n)$-dimensional vector field obtained by concatenating the coordinates of $f$ and $g$. For simplicity, we omit to specify explicitly in the functional space the dimension of the vector field it refers to. For example, we may abusively write $\LL^2(\R^d)$ instead of $\LL^2(\R^d,\R^d)$.
The symbol “$\lesssim_{\alpha_1,\cdots,\alpha_n}$” reads “$\leq C$ for a constant $C$ depending only on the tuple $(\alpha_1,\cdots,\alpha_n)$ of previously defined parameters”. Similarly, the symbol “$\ll_{\alpha_1,\cdots,\alpha_n} 1$” (“$\gg_{\alpha_1,\cdots,\alpha_n} 1$”) in an assumption reads “sufficiently small” (“sufficiently large”, respectively) in the sense of being below a threshold that depends only on the parameters $\alpha_1,\cdots,\alpha_n$. For simplicity, in the course of the proofs, the subscript might be omitted.
We denote by $\Boule_R(x)$ the ball of center $x$ and radius $R$. When it is clear from the context, the variables $x$ and $R$ may be omitted (by default, we have $\Boule_R:=\Boule_R(0)$). Also, if $\Boule$ is the ball $\Boule_R(x)$ we abusively denote by $\frac{1}{2}\Boule$ the ball $\Boule_{R/2}(x)$. The function $\mathds{1}_{\Omega}$ is the characteristic function of a domain $\Omega$. When the set of integration is omitted, we use the convention that the integral is taken on the whole space $\R^d$; namely $\int_{\R^d} f:=\int f$. Also, unless otherwise stated, the equations we consider are satisfied on $\R^d$. We denote the homogeneous Hölder semi-norm and the Hölder norm: $$\begin{aligned}
[ h ]_{\CC^{\beta}(\Omega)}:= \sup_{x,x' \in \Omega} \frac{|h(x)-h(x')|}{|x-x'|^\beta} \et \|h\|_{\CC^{\beta}(\Omega)} = [ h ]_{\CC^{\beta}(\Omega)} + \|h\|_{\LL^\infty(\Omega)}.
\end{aligned}$$
If $p \in [1,\infty]$ is an exponent, its conjugated exponent is denoted $p^\star$ and is defined by $1/p+1/p^\star=1$.
General notions in homogenization {#SecHcv}
=================================
Correctors, homogenization commutator, and two-scale expansion
--------------------------------------------------------------
This article is concerned with the following elliptic equation in divergence form $$\begin{aligned}
\nabla \cdot ( a \nabla u + f )=0 \dans \R^d,
\end{aligned}$$ where $a$ is a $\lambda$-uniformly elliptic coefficient field, *i.e.* for some $\lambda>0$, there holds $$\begin{aligned}
\label{Ellipticite}
\xi \cdot a(x) \xi \geq \lambda |\xi|^2 \quad\et \quad \xi \cdot a^{-1}(x) \xi \geq |\xi|^2.
\end{aligned}$$ While we adopt scalar notation and language, everything remains unchanged if $u$ has values in some finite-dimensional vector space.
In most general terms, homogenization means relating the heterogeneous coefficient field $a$ to a homogeneous coefficient $\abar$. In homogenization, this does not rely on the smallness of $a-\abar$ in a strong or (usual) weak sense. Almost tautologically, the suitable topology is provided by the notion of $H$-convergence [@Allaire Def. 1.2.15, p. 25]. On a purely *algebraic* level, this topology is captured on the level of an $a$-Helmholtz decomposition, giving rise to a scalar potential (inducing the definition of the corrector $\phi_i$) and a vector potential (inducing the definition of the flux corrector $\sigma_i$); for $d\not=3$, the vector potential is replaced by a skew symmetric tensor field (i. e. an alternating $(d-2)$-form).
We decompose $(a-\abar)e_i$ into two parts as follows: $$\begin{aligned}
\label{Id02}
(a-\abar) e_i&=- a \nabla \phi_i + \nabla \cdot \sigma_i,
\end{aligned}$$ in the sense of $a_{ji} - \abar_{ji} = - a_{jk} \partial_k \phi_i + \partial_k \sigma_{ijk}$. Here, $\phi_i$ is a scalar field and $\sigma_i$ is a skew-symmetric tensor field: $$\begin{aligned}
\label{Id01}
\sigma_{ijk}&=-\sigma_{ikj}.
\end{aligned}$$ Equation is an $a$-Helmholtz decomposition of $(a-\abar) e_i$ in the following sense: The first r. h. s. term of is the product of $a$ and a curl-free vector field (which, in case of the trivial topology of $\R^d$, amounts to a gradient field), whereas the second one is a divergence-free vector field. This is an $a$-Helmholtz decomposition since the identification of the $1$-form $\nabla \phi_i$ with a $(d-1)$-form is done via the “metric” $a$. We introduce the analogous objects $(\phi_i^\star,\sigma_i^\star)$ on the level of the transposed coefficient field $a^\star$: $$\begin{aligned}
\label{Id02prime}
(a^\star-\abar^\star) e_i&=- a^\star \nabla \phi^\star_i + \nabla \cdot \sigma^\star_i.
\end{aligned}$$
Decomposition immediately implies the classical equation for $\phi_i$: $$\begin{aligned}
\label{Id04}
-\nabla \cdot a \nabla (x_i+ \phi_i)&=
\nabla \cdot( \abar e_i + \nabla \cdot \sigma_i )\overset{\eqref{Id01}}{=}0.
\end{aligned}$$ Thus, the functions $\phi_i$ are called *correctors*, for they correct the $\abar$-harmonic coordinates $x_i$ to $a$-harmonic functions $x_i+\phi_i$. (By *$a$-harmonic* functions, we mean functions $u$ such that $\nabla \cdot a \nabla u=0$.) It is clear that this notion is only useful when the correctors grow sublinearly. The corrector $\phi_i$ induces a natural quantity, namely the flux (or current) $q_i$ defined by $$\begin{aligned}
\label{Defqi}
q_i:= a (e_i+\nabla \phi_i ).
\end{aligned}$$
We call the functions $\sigma_i$ *flux correctors*, for they correct the flux as follows: $$\begin{aligned}
\label{Num:1}
q_i=a(e_i+\nabla \phi_i)=\abar e_i + \nabla \cdot \sigma_i.
\end{aligned}$$ Note that does not determine $\sigma$ (not even up to the addition of a constant as in the case of $\phi$). In fact, $\sigma_i$, which should be considered as an alternating $(d-2)$-form only determined up to a $(d-3)$-form. A natural choice of gauge is given by $-\Delta \sigma_i = \nabla \times q_i$, where we use the $(d=3)$-notation as an abbreviation for $$\begin{aligned}
\label{Id03}
-\Delta \sigma_{ijk}&=\partial_j q_{ik} - \partial_k q_{ij}.
\end{aligned}$$ Note that is indeed consistent with in the form of because in view of $\nabla \cdot q_i=0$ it yields $-\Delta ( \nabla \cdot \sigma_i - q_i )=0$. We will henceforth call the pair $(\phi,\sigma)$ *extended correctors*.
The extended correctors appear in two fundamental identities, which are proved at the end of the section. The first one is at the core of homogenization in form of $H$-convergence: The effective tensor $\abar$ is supposed to relate macroscopic averages of fields to macroscopic averages of fluxes. It is easiest to make this precise on the level of $a$-harmonic functions $u$, for which $\nabla u$ is (up to the sign) the field and $a\nabla u$ the flux, so that the *homogenization commutator* $a\nabla u-\abar\nabla u$ should have small macroscopic averages. This is embodied by the following formula $$\begin{aligned}
\label{1a}
e_j\cdot(a\nabla u-\abar\nabla u)=-\nabla\cdot((\phi_j^\star a+\sigma_j^\star)\nabla u)
\quad\mbox{provided}\;\nabla\cdot a\nabla u=0.
\end{aligned}$$ Indeed, due to its divergence form, macroscopic averages of the r. h. s. are smaller than those of each term on the l. h. s., provided of course that the extended correctors grow sublinearly.
The second fundamental identity intertwines the variable-coefficient operator $-\nabla\cdot a\nabla$ with the constant-coefficient operator $-\nabla\cdot \abar\nabla$ via what is called the two-scale expansion, namely $$\begin{aligned}
\label{2scale}
\tilde u:=(1+\phi_i\partial_i)\bar u,
\end{aligned}$$ where here and in the sequel we use Einstein’s convention of implicit summation over repeated indexes. This intertwining is embodied by the following formula $$\begin{aligned}
\label{1b}
\nabla\cdot a\nabla(1+\phi_i\partial_i)\bar u=\nabla\cdot\abar\nabla \bar u
+\nabla\cdot((\phi_ia-\sigma_i)\nabla\partial_i\bar u).
\end{aligned}$$ Note that for affine $\bar u$, turns into the familiar . For general macroscopically varying $\bar u$, the two-scale expansion provides a microscopic modulation adapted to the variable-coefficient operator. Indeed, the second r. h. s. term of will be much smaller than the first r. h. s. term because of the additional derivative on the macroscopically varying $\bar u$, again provided the extended correctors grow sublinearly.
Typically, (\[1b\]) is used in the following way: Given a square integrable $f$, let the square integrable $\nabla u$ and $\nabla\bar u$ solve $$\begin{aligned}
\nabla\cdot(a\nabla u+f)&=0,\label{Complexe}\\
\nabla\cdot(\abar\nabla \bar u+f)&=0,\label{Simple}
\end{aligned}$$ where here and in the sequel we think of these equations as being posed on the entire $\mathbb{R}^d$ if not stated otherwise. Then we have $$\begin{aligned}
\label{12}
-\nabla\cdot a\nabla\big(u-(1+\phi_i\partial_i)\bar u\big)=
\nabla\cdot((\phi_ia-\sigma_i)\nabla\partial_i\bar u).
\end{aligned}$$ Summing up, we learn from that $a\nabla u-\abar\nabla u$ is small in the weak topology, while we learn from that $\nabla u-\nabla(1+\phi_i\partial_i)\bar u$ is small in the strong topology. We will make this precise in Section \[SecHcv\_2\].
Let us close with a numerical consideration. Homogenization in general and the two-scale expansion in particular are useful from a numerical point of view. Indeed, the problem defining $\ubar$ is far simpler than , since it involves a homogeneous coefficient instead of a heterogeneous one. Also, the homogenized coefficient $\abar$ and the correctors $\phi_i$ are independent of $f$ and may be recovered from . If $a$ is periodic, solving the latter equation is simple, since it reduces to a problem set on a periodic cell. In the stochastic context, recovering $\abar$ and the correctors $\phi_i$ is more difficult, since is posed on the entire $\R^d$ [@Anantharaman_2012]. However, it is definitely an efficient strategy when one needs to solve for a large number of right-hand sides $\nabla \cdot f$, as may be the case inside an optimization procedure. In this case, $\bar{a}$ and $\phi_i$ can be computed –at least approximately– and stored during an “offline” stage while is solved in an “online stage”.
First note that for any function $v$ there holds $$\begin{aligned}
\label{Id1}
\nabla \cdot (v \nabla \cdot \sigma_i)=(\nabla \cdot \sigma_i) \nabla v=-\nabla \cdot ( \sigma_i \nabla v).
\end{aligned}$$ Indeed, written componentwise, takes the form $$\begin{aligned}
\partial_j \lt(v \partial_k \sigma_{ijk} \rt)\overset{\eqref{Id01}}{=}
\partial_k\sigma_{ijk} \partial_j v
=\partial_k (\sigma_{ijk} \partial_j v ) - \sigma_{ijk} \partial_k\partial_j v
\overset{\eqref{Id01}}{=}
-\partial_k (\sigma_{ikj} \partial_j v ).
\end{aligned}$$
We now prove by a straightforward computation: $$\begin{aligned}
e_i \cdot ( a \nabla u - \abar \nabla u)
=~&((a^\star-\abar^\star)e_i ) \cdot \nabla u
\\
\overset{\eqref{Id02prime}}{=}~&(-a^\star \nabla \phi_i^\star + \nabla \cdot \sigma_i^\star )\cdot \nabla u
\\
=~&-\nabla \phi_i^\star \cdot a \nabla u +(\nabla \cdot \sigma_i^\star) \nabla u
\\
\overset{\eqref{Id1}}{=}&\phi_i^\star \nabla \cdot a\nabla u -\nabla \cdot ( \phi_i^\star a \nabla u) -\nabla \cdot (\sigma^\star_i \nabla u ).
\end{aligned}$$
Here comes the proof of . We note $$\begin{aligned}
a \nabla ( (1+\phi_i\partial_i) \ubar )
=~&\partial_i \ubar a ( e_i + \nabla \phi_i) + \phi_i a \nabla \partial_i \ubar
\\
\overset{\eqref{Id02}}{=}\;&\partial_i \ubar ( \abar e_i + \nabla \cdot\sigma_i ) + \phi_i a \nabla \partial_i \ubar
\\
=~&\abar \nabla \ubar + \partial_i \ubar\nabla \cdot\sigma_i + \phi_i a \nabla \partial_i \ubar.
\end{aligned}$$ Applying the divergence, we obtain from .
$H$-convergence {#SecHcv_2}
---------------
The main goal of this section is to familiarize the reader with the notions introduced in the previous section; later parts of the paper do not rely on it. Proposition \[PropHcon\] below (see [@Gloria_Neukamm_Otto_2019 Prop. 1]) relates the relative error in the two-scale expansion to the sublinearity of the extended corrector. The qualitative arguments establishing Proposition 2.1 below are at the core of the notion of $H$-convergence.
\[PropHcon\] Assume that the coefficient field $a:\R^d \rightarrow \R^{d\times d}$ satisfies and admits $\abar \in \R^{d \times d}$ as a constant homogenized coefficient and $(\phi,\sigma)$ and $(\phi^\star,\sigma^\star)$ as extended correctors in the sense of and . Let $\epsilon>0$ be given. Then, provided $\delta\ll_{d,\lambda,\epsilon} 1$, the following properties hold:
\(i) For any ball $\Boule \subset \R^d$ of radius $R>0$, if the extended correctors satisfy the following estimate: $$\begin{aligned}
\label{CorrSub}
\frac{1}{R} \Big(\fint_{\Boule} |(\phi,\phi^\star,\sigma,\sigma^\star)|^2\Big)^{\frac{1}{2}} \leq \delta,
\end{aligned}$$ then, for any $a$-harmonic function $u$ in $\Boule$ and for any $\abar$-harmonic function $\ubar$ in $\Boule$ so that $u=\ubar$ on the boundary $\partial B$, there holds: $$\begin{aligned}
\label{Prop1Estim1}
\frac{1}{R^2} \fint_{\Boule} (u-\ubar)^2 \leq \epsilon \fint_{\Boule} |\nabla u|^2.
\end{aligned}$$
\(ii) Moreover, we have the following interior bound on the error of the two-scale expansion: $$\begin{aligned}
\label{Prop1Estim2}
\fint_{\frac{1}{4}\Boule} |\nabla u - \partial_i \ubar (e_i + \nabla \phi_i) |^2 \leq \epsilon \fint_{\Boule}|\nabla u|^2.
\end{aligned}$$
The proof of and rely on the two fundamental algebraical formulas and , respectively. The proof is divided into two steps. First, we make use of in order to obtain that $a\nabla u - a \nabla \ubar$ weakly vanishes when $\delta \downarrow 0$. Then, by a compactness argument that is reminiscent of the $H$-convergence [@Allaire Def. 1.2.15], we deduce from this fact that holds. In Step 2, thanks to , we upgrade the latter estimate to . In this proof, the symbols $\lesssim$ will only depend on $d$ and $\lambda$ (but *not* on $\epsilon$).
We start with a couple of simplifications. W. l. o. g. by scaling, we may restrict to $R=1$ and by translation invariance to $x=0$, so that $\Boule=\Boule_1(0)$. Also, we assume that $\fint_{\Boule} u=0$, so that, by the Poincaré-Wirtinger inequality, $$\begin{aligned}
\fint_{\Boule} u^2 \lesssim \fint_{\Boule} |\nabla u|^2.
\end{aligned}$$ By the triangle inequality combined with the Poincaré inequality, once more the triangle inequality, and the energy estimate, we deduce that $$\begin{aligned}
\fint_{\Boule} \ubar^2 \lesssim~& \fint_{\Boule} (\ubar-u)^2 + \fint_{\Boule} u^2
\lesssim \fint_{\Boule} |\nabla (\ubar-u)|^2 + \fint_{\Boule} |\nabla u|^2
\\
\lesssim~& \fint_{\Boule} |\nabla \ubar|^2 + \fint_{\Boule} |\nabla u|^2 \lesssim \fint_{\Boule} |\nabla u|^2.
\end{aligned}$$ As a consequence, by homogeneity, we may w. l. o. g. suppose that $$\begin{aligned}
\label{Assum1}
\int_{\Boule} ( u^2 + |\nabla u|^2 + \ubar^2 + |\nabla \ubar |^2 ) \leq 1.
\end{aligned}$$ Also, up to choosing $\delta$ sufficiently small, we may assume that: $$\begin{aligned}
\label{Ellipticiteabar}
\xi \cdot \abar \xi \geq \frac{1}{2}\lambda |\xi|^2 \quad \et \quad \xi \cdot \abar^{-1} \xi \geq \frac{1}{2}|\xi|^2.
\end{aligned}$$ (The easy argument of how this follows from and via is skipped since automatically holds, even without the factor $\frac{1}{2}$, in stochastic homogenization [@Gloria_Neukamm_Otto_2019 (4)].)
We now give an indirect argument for . We consider a sequence $a_n$ with associated homogenized matrices $\abar_n$, extended correctors $(\phi_n,\sigma_n)$ and $(\phi^\star_n,\sigma^\star_n)$, as well as $a_n$-harmonic functions $u_n$ and $\abar_n$-harmonic functions $\ubar_n$, satisfying $$\begin{aligned}
\lim_{n \uparrow \infty} \fint_{\Boule} |(\phi_n,\phi^\star_n,\sigma_n,\sigma^\star_n)|^2 =0 \qquad \text{and} \qquad \liminf_{n \uparrow \infty} \fint_{\Boule} (u_n-\ubar_n)^2 >0
\label{Contradict}
\end{aligned}$$ next to . We argue that such a sequence cannot exist, thus establishing . Up to a subsequence, the following convergences hold: $\abar_n \rightarrow \abar$ and, by the Rellich theorem, $$\begin{aligned}
(u_n,\ubar_n) \rightarrow (u,\ubar) \quad\et \quad\quad (\nabla u_n,\nabla \ubar_n) \rightharpoonup (\nabla u,\nabla \ubar) \dans \LL^2(\Boule).
\end{aligned}$$ We will obtain a contradiction to by arguing that $u=\ubar$.
By assumption, $u=\ubar$ on $\partial \Boule$ (in the sense of $u-\ubar \in \HH^1_0(\Boule)$), so that it remains to argue that $$\begin{aligned}
\label{ToEstab}
-\nabla \cdot \abar \nabla u=0 \dans \HH^{-1}(\Boule).
\end{aligned}$$ Indeed, by weak convergence of $\nabla \ubar_n$ and convergence of $\abar_n$, we know that $\ubar$ is $\abar$-harmonic. Identity will follows from $$\begin{aligned}
a_n \nabla u_n \rightharpoonup \abar \nabla u \dans \LL^2(\Boule),
\end{aligned}$$ which is equivalent to $$\begin{aligned}
\label{ToEstab2}
a_n \nabla u_n- \abar_n \nabla u_n \rightharpoonup 0 \dans \LL^2(\Boule).
\end{aligned}$$ The latter is a consequence of . Indeed, for any $v \in \CC_0^\infty(\Boule)$ and $j\in \{1,\cdots,d\}$, $$\begin{aligned}
\int_{\Boule} v e_j \cdot ( a_n \nabla u_n - \abar_n \nabla u_n) \overset{\eqref{1a}}{=}\int_{\Boule} \nabla v \cdot (\phi^{\star}_{n,j} a_n + \sigma^{\star}_{n,j} ) \nabla u_n \rightarrow 0
\end{aligned}$$ since $\nabla v \in \LL^\infty(\Boule)$, $\phi^{\star}_{n,j} a + \sigma^{\star}_{n,j} \rightarrow 0$ in $\LL^2(\Boule)$ and $\nabla u_n$ is bounded in $\LL^2(\Boule)$. This establishes and therefore .
The proof of relies on the previous step and on two ingredients from elliptic regularity. The first one is the well-known Caccioppoli estimate: If $w$ solves $\nabla \cdot ( a \nabla w + f )=0$ in $\frac{1}{2}\Boule$, then $$\begin{aligned}
\label{Caccioppoli}
\fint_{\frac{1}{4}\Boule} |\nabla w |^2 \lesssim \fint_{\frac{1}{2}\Boule} (w^2 + |f|^2 ).
\end{aligned}$$ (Recall that $\Boule=\Boule_1(0)$.) We reproduce in Appendix \[ProofCaccioppoli\] a proof of this estimate in a generalized context. The second ingredient is an interior regularity result for elliptic equations with constant coefficients. More precisely, if $\ubar$ is $\abar$-harmonic in $\Boule$, then there holds $$\begin{aligned}
\label{Regubar}
\sup_{\frac{1}{2} \Boule} ( |\nabla^2 \ubar |^2 + |\nabla \ubar|^2 ) \lesssim \fint_{\Boule} |\nabla \ubar |^2.
\end{aligned}$$ (It is an easy consequence of the Sobolev embedding and an iterated application of the Caccioppoli estimate to derivatives of the constant-coefficient elliptic equation.)
Equipped with and , we proceed with the proof of . Appealing to , we apply to $w:=u-(1+\phi_i\partial_i)\ubar$ and $f:=(\phi_i a-\sigma_i)\nabla \partial_i \ubar$: $$\begin{aligned}
\fint_{\frac{1}{4}\Boule} | \nabla u - \partial_i \ubar (e_i+\nabla \phi_i)|^2
&~\lesssim\fint_{\frac{1}{4}\Boule} | \nabla w |^2
+\fint_{\frac{1}{4}\Boule} | \phi_i \nabla \partial_i \ubar|^2
\\
&\overset{\eqref{Caccioppoli}}{\lesssim}
\fint_{\frac{1}{2}\Boule} w^2 + \fint_{\frac{1}{2}\Boule}
|(\phi,\sigma)|^2 |\nabla^2 \ubar |^2
\\
&~\lesssim
\fint_{\frac{1}{2}\Boule} ( u -\ubar )^2 + \fint_{\frac{1}{2}\Boule} |(\phi,\sigma)|^2 ( |\nabla^2 \ubar |^2 + |\nabla \ubar|^2 ).
\end{aligned}$$ Making use of , this yields $$\begin{aligned}
&\fint_{\frac{1}{4}\Boule} | \nabla u + \partial_i \ubar (e_i+\nabla \phi_i)|^2
\lesssim \fint_{\frac{1}{2}\Boule} ( u -\ubar )^2 + \fint_{\Boule} |(\phi,\sigma)|^2 \fint_{\Boule}|\nabla u|^2.
\end{aligned}$$ Appealing to , this finally shows for $\delta \ll_{d,\lambda,\epsilon} 1$.
Our framework for stochastic homogenization {#SecCorrEstim}
===========================================
General assumptions {#SecAssumpGen}
-------------------
In stochastic homogenization, one is given a probability measure on the space of coefficient fields $a$ that are $\lambda$-uniformly elliptic in the sense of (with the topology induced by the $H$-convergence, which makes it into a compact space). Using physics jargon, we call this probability measure an *ensemble* and denote the expectation by $\langl\cdot\rangl$. This ensemble is assumed to be:
- [stationary, that is, for all shift vectors $z\in \R^d$, $a(\cdot+z)$ and $a(\cdot)$ have the same law under $\langl \cdot \rangl$;]{}
- [ergodic (see [@JKO Chap. 7 pp. 222–225]). Since this assumption will be strengthened below in Section \[SecAssumpGauss\], we do not state it precisely. Let us just mention that this property is required to use the Birkhoff ergodic theorem [@JKO Th. 7.2, p. 225], which relates the expectation to the spatial average.]{}
Under these assumptions, by [@Gloria_Neukamm_Otto_2019 Lem. 1], there exists a random scalar field $\phi_i$ (the corrector), a random tensor field $\sigma_i$ (the flux corrector) for $i \in \{1,\cdots,d\}$, and a deterministic coefficient $\abar$ (the homogenized coefficient) that are related to $a$ in the sense of , , and , respectively. Moreover, the following properties hold:
- [$\nabla \phi_i$ and $\nabla \sigma_i$ are stationary. We say that a random field $f(a,x)$ is stationary if it is shift-covariant in the sense of $$\begin{aligned}
&&f(a(\cdot + z),x)=f(a(\cdot),x+z) \qquad \pourtout x, z \in \R^d \text{ and for } \langl\cdot\rangle\text{-a.e. } a.
\end{aligned}$$ It implies that $f(a,\cdot)$ and $f(a,\cdot+z)$ have the same law under $\langl \cdot \rangl$. ]{}
- [$\nabla \phi_i$ and $\nabla \sigma_i$ have a finite second moment and vanishing expectation, that is, $$\begin{aligned}
\label{NablaCorrBdd}
&\langl |\nabla \phi_i(x)|^2 \rangl +\langl |\nabla \sigma_i(x)|^2 \rangl \lesssim 1,
\\
\label{Vanish_expectation}
&\langl \nabla \phi_i(x)\rangl =0 \et \langl \nabla \sigma_i(x) \rangl=0.
\end{aligned}$$ (By stationarity, this estimate and these identities are independent of $x \in \R^d$). ]{}
- [The coefficient $\abar$ can be expressed as $\abar e_i = \langl q_i \rangl$ for $q_i$ defined by , and it satisfies . ]{}
These results are only qualitative. In order to obtain a quantitative theory of stochastic homogenization, we additionally assume that the ensemble $\langl \cdot \rangl$ satisfies a spectral gap [@Gloria_Otto]. This condition amounts to a quantification of ergodicity. It reads: For any random variable $F$, which is a (measurable) functional of $a$, there holds $$\begin{aligned}
\label{SG}
\langl (F-\langl F \rangl )^2 \rangl \leq \Big\langl \int \big| \frac{\partial F}{\partial a}\big|^2 \Big\rangl,
\end{aligned}$$ where the random tensor field $\frac{\partial F}{\partial a}$ (depending on $(a,x)$) is the functional (or vertical, or Malliavin) derivative of $F$ with respect to $a$ defined by $$\label{Def:Malliavin}
\lim_{\epsilon \rightarrow 0} \frac{F(a+\epsilon \delta a) - F(a)}{\epsilon} = \int \frac{\partial F(a)}{\partial a_{ij}(x)} (\delta a)_{ij}(x) \dd x.$$ The spectral gap replaces the Poincaré inequality used in the periodic case. It leads to a sensitivity analysis through the study of the functional derivative.
In the next section, we consider a more specific class of ensembles $\langl \cdot \rangl$ satisfying the above assumptions.
A class of Gaussian ensembles {#SecAssumpGauss}
-----------------------------
In this section, we introduce the class of ensembles we will be working with. Let $\langl\cdot\rangl$ denote an ensemble of stationary and centered Gaussian fields $g$ (with values in some finite dimensional vector space, but we adopt scalar language and notation). It is thus characterized by its covariance function $c$, the (non-negative) Fourier transform $\FF c$ of which is assumed to satisfy $$\begin{aligned}
\label{wg01}
\FF c(k)\;\le (1+|k|)^{-d-2\alpha}\pourtout k\in\R^d,
\end{aligned}$$ for some fixed exponent $\alpha \in (0,1)$. Loosely speaking, encodes that the covariance function is integrable at infinity (in a somewhat weakened way) and $2\alpha$-Hölder continuous at the origin. These assumptions contain the popular Matérn kernel [@Stein_1999 p. 31 & (32), p. 49], the Fourier transform of which reads $\FF c(k):= C (1+|k|^2)^{-\nu-d/2}$ for $C>0$ and $\nu>0$.
In this framework, we set $a:=A(g)$, where $A$ is a Lipschitz map from $\R$ to the set of $d\times d$ tensors satisfying (see Figure \[Fig\_Coeff\] for an example). For conciseness, we henceforth suppress the dependence in the estimates on the $4$-tuple of fixed constants $$\label{Def_Gamma}
\gamma:=\big(d,\lambda,\alpha,\|A'\|_{\LL^\infty}\big).$$
![A realization of a random (scalar) coefficient field $a$ generated by a Gaussian field.[]{data-label="Fig_Coeff"}](Gaussian_Field.pdf){width="\textwidth"}
Assumption (\[wg01\]) on the ensemble $\langl\cdot\rangl$ has two beneficial consequences: On the one hand, we have a spectral gap estimate with a carré-du-champs that behaves like the $\LL^1$-norm on scales up to order one and like the $\LL^2$-norm on scales larger than order $1$ (and thus is smaller than the $\LL^2$-norm, thereby strengthening the spectral gap property w. r. t. ). On the other hand, the realizations of $g$ and thus $a$ are Hölder continuous. More precisely, we show the following result (the proof of is relegated to Appendix \[SecProofLemloc\]):
\[LemSG\] Let $g$ and $a$ satisfy the above assumptions. Then, for any functional $F=F(a)$ and $r \in [1,\infty)$, there holds $$\begin{aligned}
\label{Prop21a}
\langl (F-\langl F \rangl)^{2r}\rangl^{\frac{1}{r}}
\lesssim_{\gamma, r}
\bigg\langl \Big( \int \Big( \fint_{\Boule_1(x)} \big| \frac{\partial F}{\partial a(y)}\big| \dd y \Big)^2 \dd x \Big)^{r} \bigg\rangl^{\frac{1}{r}},
\end{aligned}$$ and, for any $0<\alpha'<\alpha$ and $r \in [1,\infty)$, there holds $$\begin{aligned}
\label{wg03}
\big\langl\| a \|_{\CC^{\alpha'}(\Boule_1)}^r\big\rangl^\frac{1}{r}
\lesssim_{\gamma, r, \alpha'} 1.
\end{aligned}$$
*En passant*, notice that estimate allows to appeal to classical Schauder and CZ estimates for scales up to order $1$ since the dependence of these estimates on $\|a\|_{\CC^{\alpha'}(\R^d)}$ scales polynomially, see Lemma \[LemAppendHold\] with $T=\infty$. In particular, if $u$ is $a$-harmonic in $\Boule_1$, then $$\label{Lipschitz}
\langl | \nabla u(0) |^{r'}\rangl^\frac{1}{r'}
\lesssim_{\gamma, r',r} \langl \| \nabla u \|_{\LL^2(\Boule_1)}^{r} \rangl^{\frac{1}{r}} \pourtout 1\leq r'<r< \infty.$$
W. l. o. g. we assume that $\langl F \rangl=0$. We derive by a two-step argument. First, we show that we have the following strengthening of : $$\begin{aligned}
\label{wg02}
\langl(F-\langl F\rangl)^2\rangl\lesssim
\Big\langl \int
\Big(\fint_{\Boule_1(x)}\big|\frac{\partial F}{\partial g}\big|\Big)^2\dd x\Big\rangl,
\end{aligned}$$ from which we deduce in a second step.
On the Fourier side, the Helffer-Sjöstrand representation [@Helffer_1998 (2.2)] implies[^1] $$\begin{aligned}
\langl F^2\rangl
\leq~& \Big\langl \int \int \frac{\partial F}{\partial g(x)} c(x-y) \frac{\partial F}{\partial g(y)} \dd y \dd x \Big\rangl
= \Big\langl\frac{1}{(2\pi)^{d}}\int \FF c\big|\FF \frac{\partial F}{\partial g}\big|^2 \dd k\Big\rangl.
\end{aligned}$$ Therefore, in view of (\[wg01\]), it is enough to estimate the carré-du-champs for an arbitrary function $\zeta=\zeta(x)$ as follows: $$\begin{aligned}
\int (1+|k|)^{-d-2\alpha}|\FF \zeta|^2\dd k
\lesssim
\int \Big(\fint_{\Boule_1(x)}|\zeta|\Big)^2\dd x,
\end{aligned}$$ which by duality is equivalent to $$\begin{aligned}
\label{wg04}
\int \sup_{\Boule_1(x)} \zeta^2\dd x
\lesssim
\int (1+|k|)^{d+2\alpha}|\FF \zeta|^2\dd k.
\end{aligned}$$ Since the Sobolev inequality is unrelated to homogenization, we relegate its proof to Appendix \[ProofE\].
We now improve to . By the chain rule, $$\begin{aligned}
\frac{\partial}{\partial g(y)} = A'(g(y)) \frac{\partial }{\partial a(y)},
\end{aligned}$$ and since $A'$ is bounded (recall the definition of $\gamma$), yields $$\begin{aligned}
\label{SGImproved}
\langl (F-\langl F\rangl)^2\rangl\lesssim
\Big\langl \int \Big(\fint_{\Boule_1(x)}\big|\frac{\partial F}{\partial a}\big|\Big)^2\dd x\Big\rangl.
\end{aligned}$$
We apply to $F$ replaced by $|F|^r$ for $r \geq 1$: $$\begin{aligned}
\label{1a0}
\langl F^{2r} \rangl \lesssim \langl |F|^r \rangl^2
+ \Big\langl \int \Big( \fint_{\Boule_1(x)} \big| \frac{\partial}{\partial a} |F|^r \big| \Big)^2 \Big\rangl.
\end{aligned}$$ By the chain rule $$\begin{aligned}
\frac{\partial}{\partial a} |F|^r = r |F|^{r-2} F \frac{\partial F}{\partial a},
\end{aligned}$$ and using the Hölder inequality in probability, we obtain $$\begin{aligned}
\Big\langl \int \Big( \fint_{\Boule_1(x)} | \frac{\partial}{\partial a} |F|^r | \Big)^2 \dd x \Big\rangl
&= r^2
\Big\langl F^{2(r-1)}\int \Big( \fint_{\Boule_1(x)}\big| \frac{\partial F}{\partial a}\big|\Big)^2 \dd x \Big\rangl
\\
&\leq r^2
\langl F^{2r} \rangl^{\frac{r-1}{r}} \Big\langl \Big(\int \Big( \fint_{\Boule_1(x)}\big| \frac{\partial F}{\partial a}\big|\Big)^2 \dd x \Big)^r \Big\rangl^{\frac{1}{r}}.
\end{aligned}
\label{1a1}$$ Moreover, the Hölder inequality followed by the spectral gap yields $$\begin{aligned}
\langl |F|^r \rangl^2
\leq~&
\langl F^{2r} \rangl^{\frac{r-2}{r-1}} \langl F^2 \rangl^{\frac{r}{r-1}}
\lesssim
\langl F^{2r} \rangl^{\frac{r-2}{r-1}} \Big\langl \int \Big( \fint_{\Boule_1(x)}\big| \frac{\partial F}{\partial a}\big|\Big)^2 \dd x \Big\rangl^{\frac{r}{r-1}}.
\end{aligned}$$ Applying Jensen’s inequality to the last factor, we deduce $$\begin{aligned}
\langl |F|^r \rangl^2
\lesssim
\langl F^{2r} \rangl^{\frac{r-2}{r-1}}
\Big\langl \Big(\int \Big( \fint_{\Boule_1(x)}\big| \frac{\partial F}{\partial a}\big|\Big)^2 \dd x \Big)^r \Big\rangl^{\frac{1}{r-1}}.
\label{1a2}
\end{aligned}$$ Now, follows from inserting and into and appealing to Young’s inequality.
Stochastic corrector estimates {#SecCorrEstim2}
==============================
Estimates on the extended corrector are at the core of quantitative stochastic homogenization. The following proposition collects all the estimates needed in this paper:
\[Propcorr\] Under the assumptions of Section \[SecAssumpGauss\], for any $r \in [1, \infty)$, there holds: $$\begin{aligned}
\label{i}
\langl | \nabla (\phi,\sigma)|^{2r} \rangl^{\frac{1}{r}} \lesssim_{\gamma, r} 1.
\end{aligned}$$ Moreover, the spatial averages of the gradient of the extended correctors display cancellations: $$\begin{aligned}
\label{ii}
\Big\langl \Big| \int g \cdot \nabla (\phi,\sigma) \Big|^{2r} \Big\rangl^{\frac{1}{r}} \lesssim_{\gamma, r} \int |g|^2 \quad \text{ for all deterministic vector fields }g.
\end{aligned}$$ Last, the increments of the extended correctors are controlled as follows: $$\begin{aligned}
\label{iii}
\langl |(\phi,\sigma)(x) - (\phi,\sigma)(0)|^{2r} \rangl^{\frac{1}{r}} \lesssim_{\gamma, r} \mu_d^2(|x|) \pourtout x \in \R^d,
\end{aligned}$$ where the function $\mu_d$ is defined by $$\begin{aligned}
\label{Defmu}
\mu_d(r):=
\begin{cases}
\sqrt{r+1} &\si d=1,\\
\ln^{\frac{1}{2}}(r+2) &\si d=2,\\
1 & \si d>2.
\end{cases}
\end{aligned}$$
Estimate is identical to what one would obtain if $\nabla(\phi,\sigma)$ were replaced by white noise. It is thus not surprising that the large-scale behavior of increments of the extended corrector, *cf.* , is that of the Gaussian free field (on small scales, $(\phi,\sigma)$ is more regular).
For abbreviation, we denote $$\begin{aligned}
\label{Def_chiu}
\chiu_i(x):=x_i + \phi_i(x),
\end{aligned}$$ and henceforth omit the index $i$.
The proof is divided into two main parts. Part 1 deals with the corrector $\phi$ and Part 2 deals with the flux corrector $\sigma$.
We describe the first part: In Step 1, we derive a representation formula for the Malliavin derivative $\frac{\partial}{\partial a} \int g \cdot \nabla \phi$. In Step 2, the $\LL^r$ version of the spectral gap is used to estimate the $r$-th moments of the spatial averages of $\nabla \phi$ as $$\begin{aligned}
\label{Num:809}
\Big\langl \Big| \int g \cdot \nabla \phi \Big|^{2r} \Big\rangl^{\frac{1}{r}}
\lesssim
\Big\langl \Big( \fint_{\Boule_1} |\nabla \chiu|^2 \Big)^{r} \Big\rangl^{\frac{1}{r}} \int |g|^2 \pour 1 \ll r < \infty.
\end{aligned}$$ There, the representation formula for Malliavin derivative derived in Step 1 plays a central role, as well as an annealed CZ estimate of Proposition \[PropCZ\](i). Step 3 involves a PDE ingredient, namely the Caccioppoli estimate, to control the $r$-th moments of $\fint_{\Boule_1} |\nabla \chiu|^2$ by spatial averages of $\nabla \chiu$ $$\begin{aligned}
\label{Claim01}
\Big\langl \Big(\fint_{\Boule_1} |\nabla \chiu|^2 \Big)^r \Big\rangl^{\frac{1}{r}}
\lesssim
R^{d(1-\frac{1}{r})}\Big\langl \Big| \fint_{\Boule_{R}} \nabla \chiu \Big|^{2r} \Big\rangl^{\frac{1}{r}} \pour R \gg 1.
\end{aligned}$$ In Step 4, the two above inequalities and are combined in order to buckle, yielding $$\begin{aligned}
\label{Step3}
\Big\langl \Big(\fint_{\Boule_1} |\nabla \chiu|^2 \Big)^r \Big\rangl^{\frac{1}{r}} \lesssim 1.
\end{aligned}$$ By the local regularity estimate , we deduce for $\nabla\phi$. Inserting this into yields for $\nabla\phi$. Finally, in Step 5, potential theory shows that is a consequence of . The second part is divided into three steps: We show the analogue of for the flux $q_i-\abar e_i$, which, using the estimate for $\nabla \phi$, entails that $\nabla\sigma$ satisfies , from which we deduce that $\sigma$ also satisfies .
We establish the following representation formula for the Malliavin derivative: $$\begin{aligned}
\label{Prop21b}
\frac{\partial}{\partial a} \int g \cdot \nabla \phi= \nabla v \otimes \nabla \chiu,
\end{aligned}$$ where $\nabla v$ denotes the square-integrable solution of $$\label{Prop21b_prime}
\nabla \cdot ( a^\star \nabla v +g )=0.$$
Appealing to , we give ourselves a smooth and compactly supported infinitesimal variation $\delta a$ of $a$ and denote by $\nabla \delta \phi$ the corresponding infinitesimal variation of $\nabla \phi$, which is the square-integrable solution of[^2] $$\begin{aligned}
\label{DefDeltaphi}
-\nabla \cdot a \nabla \delta \phi=\nabla \cdot \delta a (e + \nabla \phi ).
\end{aligned}$$ Since $g$ is deterministic, this implies $$\int \Big(\frac{\partial}{\partial a} \int g \cdot \nabla \phi\Big) : \delta a
= \int g \cdot \nabla \delta \phi.$$ Inserting the variational formulations of and into this, we get $$\int \Big(\frac{\partial}{\partial a} \int g \cdot \nabla \phi\Big) : \delta a
=-\int \nabla v \cdot a \nabla \delta \phi
=\int \nabla v \cdot \delta a \nabla \chiu.$$ This yields .
The proof relies on three ingredients. The first one is the $\LL^r$ version of spectral gap. The second one is the representation formula for the Malliavin derivative . The third one is the annealed CZ estimate given by Proposition \[PropCZ\](i) below.
Equipped with these ingredients, we establish . From , , and , followed by the Cauchy-Schwarz inequality in $\LL^2(\Boule_1(x))$, we deduce that $$\begin{aligned}
\Big\langl \Big| \int g \cdot \nabla \phi \Big|^{2r} \Big\rangl^{\frac{1}{r}}
\lesssim~& \Big\langl \Big( \int \Big( \fint_{\Boule_1(x)} | \nabla v \otimes \nabla \chiu |\Big)^2 \dd x\Big)^{r} \Big\rangl^{\frac{1}{r}}
\\
\leq~& \Big\langl \Big( \int \Big( \fint_{\Boule_1(x)} |\nabla v|^2 \Big) \Big( \fint_{\Boule_1(x)}|\nabla \chiu|^2\Big) \dd x \Big)^r \Big\rangl^{\frac{1}{r}}.
\end{aligned}$$ Thus by duality in probability, with $1/r + 1/r^\star=1$, $$\label{Dual1}
\begin{aligned}
&\Big\langl \Big| \int g \cdot \nabla \phi \Big|^{2r} \Big\rangl^{\frac{1}{r}}
\lesssim \sup_{\langl F^{2r^\star} \rangl=1} \Big\langl F^2 \int \Big( \fint_{\Boule_1(x)} |\nabla v|^2 \Big) \Big( \fint_{\Boule_1(x)}|\nabla \chiu|^2\Big) \dd x \Big\rangl.
\end{aligned}$$ We use the Hölder inequality and the stationarity of $\nabla \chiu$ (*cf.* ) to the effect of $$\label{Dual2}
\begin{aligned}
&\int \Big\langl F^2 \Big( \fint_{\Boule_1(x)} |\nabla v|^2 \Big) \Big( \fint_{\Boule_1(x)}|\nabla \chiu|^2\Big) \Big\rangl\dd x
\\
&\qquad \qquad \leq \int \Big\langl \Big( \fint_{\Boule_1(x)}|F\nabla v|^2\Big)^{r^\star} \Big\rangl^{\frac{1}{r^\star}} \dd x
\Big\langl \Big( \fint_{\Boule_1} |\nabla \chiu|^2 \Big)^r \Big\rangl^{\frac{1}{r}}.
\end{aligned}$$ Since $u \mapsto \langl |u|^{r^\star} \rangl^{\frac{1}{r^\star}}$ is a norm, by Jensen’s inequality, $$\label{Num:713}
\begin{aligned}
\int \Big\langl \Big( \fint_{\Boule_1(x)}|F\nabla v|^2\Big)^{r^\star} \Big\rangl^{\frac{1}{r^\star}} \dd x
&\leq
\int \fint_{\Boule_1(x)}\langl (|F\nabla v|^2)^{r^\star} \rangl^{\frac{1}{r^\star}} \dd x
\\
&= \int \langl |F\nabla v|^{2r^\star} \rangl^{\frac{1}{r^\star}},
\end{aligned}$$ where we used in the last equality the identity $$\begin{aligned}
\label{Num:556}
\int h(x) \dd x
=\int \fint_{\Boule_1(x)} h(y) \dd y \dd x.
\end{aligned}$$ Then, since may be written as $\nabla \cdot ( a^\star \nabla (Fv) + Fg)=0$, we learn from Proposition \[PropCZ\](i) below that $$\label{Num:714}
\int \big\langl|F\nabla v|^{2r^\star} \big\rangl^{\frac{1}{r^\star}}
\lesssim
\int \langl| F g|^{2r^\star} \rangl^{\frac{1}{r^\star}}
=
\langl F^{2r^\star} \rangl^{\frac{1}{r^\star}} \int |g|^2.$$ In the last lines, we have extensively used that $F$ is a random variable independent of $x$, whereas $g$ is deterministic. The application of Proposition \[PropCZ\](i) requires $|2r^\star-2| \ll 1$, which amounts to $r \gg 1$. Inserting and into , which we use in , entails .
By stationarity we have $$\begin{aligned}
\Big\langl \Big(\fint_{\Boule_1} |\nabla \chiu|^2 \Big)^r \Big\rangl^{\frac{1}{r}}
=\Big\langl \fint_{\Boule_R} \Big(\fint_{\Boule_1(x)} |\nabla \chiu|^2 \Big)^r \dd x\Big\rangl^{\frac{1}{r}}.
\end{aligned}$$ We now appeal to the elementary string of inequalities: $$\begin{aligned}
\fint_{\Boule_R} \Big(\fint_{\Boule_1(x)} |\nabla \chiu|^2 \Big)^r \dd x
&= \fint_{\Boule_R} \Big(\fint_{\Boule_1(x)} |\nabla \chiu|^2 \Big)^{r-1} \Big(\fint_{\Boule_1(x)} |\nabla \chiu|^2 \Big) \dd x
\\
&\leq (2R)^{d(r-1)}\Big(\fint_{\Boule_{2R}} |\nabla \chiu|^2 \Big)^{r-1} \fint_{\Boule_R}\fint_{\Boule_1(x)} |\nabla \chiu|^2 \dd x
\\
&\leq 2^d(2R)^{d(r-1)}\Big(\fint_{\Boule_{2R}} |\nabla \chiu|^2 \Big)^{r}.
\end{aligned}$$ Therefore, we obtain $$\begin{aligned}
\label{2b}
\Big\langl \Big(\fint_{\Boule_1} |\nabla \chiu|^2 \Big)^r \Big\rangl^{\frac{1}{r}}
\lesssim~&R^{d(1-\frac{1}{r})} \Big\langl \Big(\fint_{\Boule_{2R}} |\nabla \chiu|^2 \Big)^{r}\Big\rangl^{\frac{1}{r}}.
\end{aligned}$$ Since $u$ is $a$-harmonic, *c.f.* and , the above right-hand term may be bounded thanks to the Caccioppoli estimate (with $T=\infty$) $$\begin{aligned}
\label{Cacc_1}
\fint_{\Boule_{2R}} |\nabla \chiu|^2
\lesssim
\frac{1}{R^2}\fint_{\Boule_{4R}} (\chiu-\Cmu)^2 \qquad \text{for all constants } \Cmu.
\end{aligned}$$ We denote by $u_{R'}$ the mollification of $u$ on scale $R' \leq R$ by convolution: $$\label{Num:719}
u_{R'}(x):=\fint_{\Boule_{R'}(x)}u.$$ Choosing $\Cmu:=\fint_{\Boule_{4R}} u_{R'}(x)$, using the convolution estimate and the Poincaré-Wirtinger estimate, that is, $$\fint_{\Boule_{4R}}(u-u_{R'})^2\lesssim {R'}^2\fint_{\Boule_{8R}}|\nabla u|^2
\et \quad
\fint_{\Boule_{4R}}(u_{R'}-c)^2
\lesssim
R^2\fint_{\Boule_{4R}}|\nabla u_{R'}|^2,$$ we obtain by the triangle inequality $$\label{Num:739}
\begin{aligned}
\fint_{\Boule_{4R}} (\chiu-\Cmu)^2
\lesssim~& \fint_{\Boule_{4R}} \big( \chiu-\chiu_{R'}\big)^2
+\fint_{\Boule_{4R}} \big(\chiu_{R'}-\Cmu\big)^2
\\
\lesssim~& {R'}^2 \fint_{\Boule_{8R}} |\nabla \chiu|^2 + R^2 \fint_{\Boule_{4R}} |\nabla\chiu_{R'}|^2.
\end{aligned}$$ Inserting this into entails $$\begin{aligned}
\fint_{\Boule_{2R}} |\nabla \chiu|^2
\lesssim~& \big(\frac{R'}{R}\big)^2 \fint_{\Boule_{8R}} |\nabla \chiu|^2
+ \fint_{\Boule_{4R}} |\nabla\chiu_{R'}|^2.
\label{2c}
\end{aligned}$$ Since we may cover $\Boule_{8R}$ by balls $\Boule_{2R}(x_n)$ for $n \in \{1,\cdots,N\}$ and $N\lesssim 1$, $$\begin{aligned}
\fint_{\Boule_{8R}} |\nabla \chiu|^2
\lesssim \sum_{n=1}^N \fint_{\Boule_{2R}(x_n)} |\nabla \chiu|^2.
\end{aligned}$$ Therefore, by the stationarity of $\nabla \chiu$ and the triangle inequality, $$\begin{aligned}
\label{Num:771}
\Big\langl \Big(\fint_{\Boule_{8R}} |\nabla \chiu|^2\Big)^r\Big\rangl^{\frac{1}{r}}
\lesssim
\Big\langl \Big(\fint_{\Boule_{2R}} |\nabla \chiu|^2\Big)^r\Big\rangl^{\frac{1}{r}}.
\end{aligned}$$ As a consequence, it follows from that $$\begin{aligned}
\Big\langl \Big(\fint_{\Boule_{2R}} |\nabla \chiu|^2\Big)^r \Big\rangl^{\frac{1}{r}}
\lesssim~&
\big(\frac{R'}{R}\big)^2 \Big\langl \Big(\fint_{\Boule_{2R}} |\nabla \chiu|^2\Big)^r\Big\rangl^{\frac{1}{r}}
+\Big\langl \Big( \fint_{\Boule_{4R}} |\nabla\chiu_{R'}|^2\Big)^r \Big\rangl^{\frac{1}{r}}.
\end{aligned}$$ Hence, if $R'=\theta R$ with $\theta \ll 1$, the first right-hand term may be absorbed in the left-hand side. Inserting this into yields: $$\begin{aligned}
\Big\langl \Big(\fint_{\Boule_1} |\nabla \chiu|^2 \Big)^r \Big\rangl^{\frac{1}{r}}
\lesssim
R^{d(1-\frac{1}{r})}\Big\langl \Big( \fint_{\Boule_{4R}} |\nabla\chiu_{\theta R}|^2\Big)^r \Big\rangl^{\frac{1}{r}}.
\end{aligned}$$ By Jensen’s inequality and the stationarity of $\nabla \chiu$, we finally obtain $$\Big\langl \Big(\fint_{\Boule_1} |\nabla \chiu|^2 \Big)^r \Big\rangl^{\frac{1}{r}}
\lesssim
R^{d(1-\frac{1}{r})}\big\langl |\nabla\chiu_{\theta R}|^{2r} \big\rangl^{\frac{1}{r}},$$ which may be rewritten as (by replacing $\theta R \rightsquigarrow R$).
Note that, in light of , reduces to $$\begin{aligned}
\label{1prime}
\Big\langl \Big| \int g \cdot \nabla \phi \Big|^{2r} \Big\rangl^{\frac{1}{r}}
\lesssim
\int |g|^2
\pourtout 1 \ll r < \infty.
\end{aligned}$$ Also, taking advantage of the local regularity estimate , estimate may be upgraded to $$\begin{aligned}
\label{EstimNablaLp}
\langl |\nabla \chiu |^{2r} \rangl^{\frac{1}{r}} \lesssim 1.
\end{aligned}$$ Therefore, we obtain and for $\nabla\phi$.
Here comes the argument for . W. l. o. g. we may assume that $r\gg 1$ (Jensen’s inequality will provide the full range $[1,\infty)$). With the abbreviation $g_R=\frac{1}{|\Boule_{R}|}\mathds{1}_{\Boule_{R}}$, we obtain $$\begin{aligned}
\Big\langl \Big| \fint_{\Boule_{R}} \nabla \chiu\Big|^{2r} \Big\rangl^{\frac{1}{r}}
\overset{\eqref{Def_chiu}}{\leq}&1 + \Big\langl \Big| \int g_R \nabla \phi\Big|^{2r} \Big\rangl^{\frac{1}{r}}
\\
\overset{\eqref{Num:809}}{\lesssim}& 1 + \Big\langl\Big(\fint_{\Boule_1} |\nabla \chiu|^{2}\Big)^{r} \Big\rangl^{\frac{1}{r}} \int g_R^2
\\
\lesssim~& 1 + R^{-d}\Big\langl\Big(\fint_{\Boule_1} |\nabla \chiu|^{2}\Big)^{r} \Big\rangl^{\frac{1}{r}}.
\end{aligned}$$ Inserting this estimate into yields $$\begin{aligned}
\Big\langl \Big(\fint_{\Boule_1} |\nabla \chiu|^2 \Big)^r \Big\rangl^{\frac{1}{r}}
\lesssim
R^{d(1-\frac{1}{r})} +R^{-\frac{d}{r}}\Big\langl\Big(\fint_{\Boule_1} |\nabla \chiu|^{2}\Big)^{r} \Big\rangl^{\frac{1}{r}}.
\end{aligned}$$ Choosing $R\gg 1$, the above r. h. s. may be absorbed into the above left-hand side. This establishes .
This step is dedicated to proving $$\begin{aligned}
\label{iiia}
\langl |\phi(x) - \phi(0)|^{2r} \rangl^{\frac{1}{r}} \lesssim \mu_d^2(|x|).
\end{aligned}$$ By the triangle inequality and the stationarity of $x \mapsto \phi(x)-\fint_{\Boule_1(x)} \phi$, it is sufficient to show the following estimates: $$\begin{aligned}
\Big\langl \Big| \fint_{\Boule_1(x)} \phi - \fint_{\Boule_1(0)} \phi\Big|^{2r} \Big\rangl^{\frac{1}{r}} &\lesssim \mu_d^2(|x|),
\label{4a}
\\
\Big\langl \Big| \phi(0) - \fint_{\Boule_1(0)} \phi\Big|^{2r} \Big\rangl^{\frac{1}{r}} &\lesssim 1.
\label{4b}
\end{aligned}$$
Here comes the argument for . We have the representation formula $$\begin{aligned}
\label{RepFormula}
\fint_{\Boule_1(x)} \phi - \fint_{\Boule_1(0)} \phi
&=\int \nabla \ubar \cdot \nabla \phi,
\end{aligned}$$ where $\ubar$ denotes the decaying solution of $$\begin{aligned}
-\Delta \ubar =\frac{1}{|\Boule_1|} ( \mathds{1}_{\Boule_1(x)} - \mathds{1}_{\Boule_1(0)} ).
\end{aligned}$$ By classical potential theory[^3], $$\begin{aligned}
\label{IneqPot}
\int |\nabla \ubar|^2 \lesssim \mu^2_d(|x|).
\end{aligned}$$ Since $\ubar$ is deterministic, as a consequence of and , we obtain $$\begin{aligned}
\Big\langl \Big| \fint_{\Boule_1(x)} \phi - \fint_{\Boule_1(0)} \phi\Big|^{2r} \Big\rangl^{\frac{1}{r}}
&\lesssim \int |\nabla \ubar|^2,
\end{aligned}$$ which entails by .
We now argue for . By a Sobolev embedding, provided $2r >d$, there holds: $$\begin{aligned}
\Big| \phi(0) - \fint_{\Boule_1(0)} \phi\Big|^{2r}
\lesssim \fint_{\Boule_1(0)} |\nabla \phi|^{2r}.
\end{aligned}$$ Hence, taking the expectation, and recalling that $\nabla \phi$ is stationary and satisfies , we deduce that $$\begin{aligned}
\Big\langl \Big| \phi(0) - \fint_{\Boule_1(0)} \phi\Big|^{2r}\Big\rangl
\lesssim \Big\langl \fint_{\Boule_1(0)} |\nabla \phi|^{2r} \Big\rangl \lesssim 1.
\end{aligned}$$ This shows and concludes the proof of .
We first show the analogue of for the flux $q_i$ (*cf.* ); that is, if $g$ is a deterministic vector field, then $$\begin{aligned}
\label{1flux}
\Big\langl \Big| \int g \cdot ( q_i-\abar e_i) \Big|^{2r} \Big\rangl^{\frac{1}{r}}
\lesssim
\int |g|^2,
\end{aligned}$$ from which we shall deduce for $\nabla\sigma$. The only change is in the representation formula: $$\begin{aligned}
\label{1fluxprime}
\frac{\partial}{\partial a} \int g \cdot ( q_i-\abar e_i) =( \nabla v + g ) \otimes \nabla \chiu_i,
\end{aligned}$$ where the square-integrable vector field $\nabla v$ satisfies $\nabla \cdot a^\star ( \nabla v + g)=0$. Indeed, we have $$\begin{aligned}
\int \Big(\frac{\partial}{\partial a}\int g \cdot ( q_i-\abar e_i) \Big): \delta a
\overset{\eqref{Defqi}}=~&
\int g \cdot ( \delta a ( e_i + \nabla \phi_i)+ a \nabla \delta \phi_i),
\end{aligned}$$ where $\nabla \delta\phi_i$ satisfies . The rightmost term in the integrand is dealt with as in , so that is proved.
Next, we show that $\nabla\sigma$ satisfies . Denoting by $\nabla \vbar$ the square-integrable solution of $\Delta \vbar+\nabla \cdot g=0$, we obtain $$\begin{aligned}
\int g \cdot \nabla \sigma_{ijk}
=~& -\int ( \vbar\partial_j q_{ik} - \vbar \partial_k q_{ij})
\overset{\eqref{Id02}}{=}\int
\big( \partial_j \vbar(q_{ik}- \abar_{ki}) - \partial_k\vbar (q_{ij}-\abar_{ji})\big),
\end{aligned}$$ where justifies this formal integration by parts. Therefore, we deduce from that $$\begin{aligned}
\Big\langl \Big|\int g \cdot \nabla \sigma_{ijk}\Big|^{2r} \Big\rangl^{\frac{1}{r}}
\lesssim
\int |\nabla \vbar |^2
\leq
\int |g |^2.
\end{aligned}$$
By the CZ estimates applied to the constant-coefficient equation (by a post-processing of [@GT Th. 9.11, p. 235]), there holds $$\begin{aligned}
\fint_{\Boule_R} |\nabla \sigma |^{2r}
\lesssim
\Big(\fint_{\Boule_{2R}} |\nabla \sigma |^{2} \Big)^{r}
+
\fint_{\Boule_{2R}} |q |^{2r}.
\end{aligned}$$ By ergodicity and stationarity, when $R \uparrow \infty$, each of these spatial averages converges almost surely to the associated expectation (this is a consequence of the Birkhoff theorem). Therefore, we get $$\begin{aligned}
\langl |\nabla \sigma|^{2r} \rangl^{\frac{1}{r}}
\lesssim
\langl|\nabla \sigma |^{2} \rangl
+\langl |q |^{2r}\rangl^{\frac{1}{r}},
\end{aligned}$$ from which we deduce that $\nabla\sigma$ satisfies , by and .
By arguments identical to Part 1, Step 5, we deduce that $\sigma$ satisfies .
Oscillations: Estimate of homogenization error in strong norms {#Sec:Osc}
==============================================================
Equipped with the stochastic corrector estimates, we now may tackle the homogenization error. In this section, we address the error in the two-scale expansion on the level of the gradient in strong $\LL^p$-norms. In fact, in view of , the error estimate is a corollary of the corrector estimates in Proposition \[Propcorr\] and the annealed CZ estimate of Proposition \[PropCZ\](ii).
\[Cor1\] Suppose that $f$ a deterministic, smooth, and compactly supported vector field. Under the assumptions of Section \[SecAssumpGauss\], let the square-integrable vector fields $\nabla u$, $\nabla \ubar$ be related to $f$ by $$\begin{aligned}
\nabla \cdot ( a \nabla u + f ) =0 =\nabla \cdot( \abar \nabla \ubar + f).
\end{aligned}$$ Then, for any $p \in (1,\infty)$ and $r\in[1,\infty)$, there holds: $$\begin{aligned}
\begin{aligned}
\Big\langl \Big( \int | \nabla ( u - (1+\phi_i\partial_i) \ubar ) |^p \Big)^{r} \Big\rangl^{\frac{1}{pr}}
\lesssim_{\gamma,p, r} \Big( \int \big|\mu_d(|\cdot| ) \nabla f\big|^p \Big)^{\frac{1}{p}}.
\end{aligned}
\label{Estim02}
\end{aligned}$$
It is customary to rescale the space variable according to macroscopic coordinates $\hat{x}=\epsilon x$, where $\epsilon \ll 1$ corresponds to the ratio of the microscale coming from the coefficient field (implicitly contained in our assumption ) and the macroscale coming from the r. h. s. $f$. By scaling, we have $$\dd \hat{x}=\epsilon^d \dd x,\quad \hat{\nabla} = \epsilon^{-1}\nabla, \quad \hat{u}(\hat{x})=\epsilon u(x) \quad \et \quad \hat{\phi}(\hat{x})=\epsilon\phi(x).$$ In this new notation, Corollary \[Cor1\] reads:
\[Cor2\] Suppose that $\hat{f}$ a deterministic, smooth, and compactly supported vector field. Under the assumptions of Section \[SecAssumpGauss\], let the square-integrable vector fields $\nabla \hat{u}$, $\nabla \ubar$ be related to $\hat{f}$ by $$\begin{aligned}
\hat{\nabla} \cdot \big( a\big(\frac{\cdot}{\epsilon}\big) \hat{\nabla} \hat{u} + \hat{f} \big)
=0
=\hat{\nabla} \cdot( \abar \hat{\nabla} \ubar + \hat{f})
\end{aligned}$$ for some $\epsilon\leq 1$. Then, for any $p \in (1,\infty)$ and $r\in [1,\infty)$, there holds: $$\begin{aligned}
\begin{aligned}
\Big\langl \Big( \int \big| \hat{\nabla} \big(\hat{u} - \big(1+\hat{\phi}_i \hat{\partial}_i\big) \ubar \big) \big|^p \dd \hat{x} \Big)^{r} \Big\rangl^{\frac{1}{pr}}
&\lesssim_{\gamma,p, r, \hat{f}} \epsilon \mu_d(1/\epsilon)
\\
&=
\begin{cases}
\epsilon^{\frac{1}{2}}
& \si \; d=1,\\
\epsilon \ln^{\frac{1}{2}}(1/\epsilon +2)
& \si \; d=2,\\
\epsilon
& \si \; d>2.
\end{cases}
\end{aligned}
\label{Estim03}
\end{aligned}$$
Hence, for dimensions $d>2$, stochastic homogenization has linear order of convergence in $\epsilon$, as in periodic homogenization. In fact, higher-order correctors provide the dimension-dependent order of convergence $d/2$ (with a logarithmic correction in even dimensions), *c.f.* [@DuerinckxOtto_2019 Prop. 2.7].
We first note that is it sufficient to establish $$\begin{aligned}
\label{Estim01}
\begin{aligned}
\Big\langl \Big( \int | \nabla ( u - (1+\phi_i\partial_i) \ubar ) |^p \Big)^{r} \Big\rangl^{\frac{1}{r}}
\lesssim_{\gamma,p, r} \int \big|\mu_d(|\cdot|) \nabla^2 \ubar\big|^p.
\end{aligned}
\end{aligned}$$ Indeed, observe that $\mu_d^p(|\cdot|)$ is a Muckenhoupt weight of class $A_p$ [@Grafakos_book Def. 7.1.3, p. 503] if $d\geq 2$. Thus, by weighted CZ estimates [@Grafakos_book Th. 7.4.6, p. 540] for the operator $\nabla (-\nabla \cdot \abar \nabla)^{-1} \nabla \cdot$, there holds $$\int \big|\mu_d (|\cdot|) \nabla^2 \ubar\big|^p \lesssim \int \big|\mu_d (|\cdot|) \nabla f\big|^p.$$ (This estimate is trivially true if $d=1$.) As a consequence, yields .
Here comes the argument for . We multiply by an arbitrary random variable $F=F(a) \geq 0$, to the effect of $$\begin{aligned}
-\nabla\cdot a \nabla(F ( u - (1+\phi_i\partial_i) \ubar )) = \nabla \cdot F (\phi_i a - \sigma_i ) \nabla \partial_i \ubar.
\end{aligned}$$ By Proposition \[PropCZ\] (ii), this yields $$\begin{aligned}
\int \langl |F \nabla ( u - (1+\phi_i\partial_i) \ubar)|^p \rangl
\lesssim
\int \langl F^{s}|(\phi_i a - \sigma_i ) \nabla \partial_i \ubar|^{s} \rangl^{p/s}
\end{aligned}$$ for some $s >p$ to be fixed later. The integrand of the above right-hand side may be estimated by means of the Hölder inequality with $t, t^\star>1$ satisfying $1/t+1/t^\star=1$ (to be fixed later) as $$\begin{aligned}
\langl F^{s}|(\phi_i a - \sigma_i ) \nabla \partial_i \ubar|^{s} \rangl^{p/s}
=~&\langl F^{s}|\phi_i a - \sigma_i|^{{s}} \rangl^{p/{s}} |\nabla \partial_i \ubar|^p
\\
\leq~& \langl F^{{s}t^\star}\rangl^{\frac{p}{{s}t^\star}}
\langl|\phi_i a - \sigma_i|^{{s}t} \rangl^{\frac{p}{{s}t}} |\nabla \partial_i \ubar|^p.
\end{aligned}$$ W. l. o. g. we may assume the anchoring $(\phi,\sigma)(0)=0$ so that by Proposition \[Propcorr\] we have $$\begin{aligned}
\langl|(\phi_i a - \sigma_i )(x)|^{{s}t} \rangl^{\frac{p}{{s}t}} \lesssim \mu^p_d(|x|).
\end{aligned}$$ Therefore, we deduce from the above estimates that $$\begin{aligned}
&\Big\langl F^p\int |\nabla ( u - (1+\phi_i\partial_i) \ubar)|^p \Big\rangl
\lesssim
\langl (F^p)^{\frac{{s}t^\star}{p}}\rangl^{\frac{p}{{s}t^\star}}
\int \big| \mu_d(|\cdot|) \nabla^2 \ubar \big|^p.
\label{Abo1}
\end{aligned}$$ We now pick an exponent $t^\star < r^\star$ and specify $$\begin{aligned}
s:=\frac{pr^\star}{t^\star} >p.
\end{aligned}$$ Hence, reads $$\begin{aligned}
\Big\langl F^p\int |\nabla ( u - (1+\phi_i\partial_i) \ubar)|^p \Big\rangl
\lesssim
\langl F^{pr^\star}\rangl^{\frac{1}{r^\star}}
\int \big| \mu_d(|\cdot|) \nabla^2 \ubar\big|^p,
\end{aligned}$$ so that a duality argument yields .
Rescaling , we obtain: $$\begin{aligned}
\Big\langl \Big( \int \big| \hat{\nabla} \big(\hat{u} - \big(1+\hat{\phi}_i \hat{\partial}_i\big) \ubar \big) \big|^p \dd \hat{x} \Big)^{r} \Big\rangl^{\frac{1}{pr}}
\lesssim
\Big( \int \big|\epsilon \mu_d\big(\frac{|\hat{x}|}{\epsilon}\big) \hat{\nabla} \hat{f}\big|^p \dd \hat{x}\Big)^{\frac{1}{p}}.
\end{aligned}$$ Recalling the expression of $\mu_d$ directly yields for $d\neq 2$. If $d=2$, we additionally make use of the inequality $\ln(|x|/\epsilon+2) \lesssim \ln(|x|+2) \ln( 1/\epsilon+2)$ to get .
Fluctuations: Estimate of homogenization error in weak norms {#SecFluctu}
============================================================
This section is devoted to studying the fluctuations of an observable $$\begin{aligned}
\label{MacroF}
G:=\int g \cdot \nabla u,
\end{aligned}$$ where the square-integrable vector fields $\nabla u$ and $f$ are related through for deterministic vector fields $g$ and $f$. More precisely, we are interested in macroscopic observables, that is $$\begin{aligned}
\label{Macrohf}
g(x):=\epsilon^{d} \hat{g}(\epsilon x) \et f(x):=\hat{f}(\epsilon x),
\end{aligned}$$ where $\hat{g}$ and $\hat{f}$ are fixed vector fields of scale $1$ and $\epsilon \ll 1$. In [@Marahrens_Otto], it has been observed that the variance of $G$ has a central-limit theorem scaling in $\epsilon^{-1}$. Moreover, in [@GuMourrat_2016], it is shown that the leading order (in $\epsilon \ll 1$) of this variance may be explicitly characterized in terms of a quartic tensor $\QU$ introduced in [@Mourrat_Otto_2016].
Surprisingly enough, the naive (but natural) idea of replacing $\nabla u$ by its two-scale expansion in gives the wrong leading order, as was discovered by [@GuMourrat_2016]. However, the two-scale expansion may be used in a more subtle way, as was discovered in [@DuerinckxGO_2016]: We define $\nabla\vbar $ as the square-integrable vector field related to $g$ through $$\begin{aligned}
\label{Defh}
\nabla \cdot ( \abar^\star \nabla \vbar + g)=0.
\end{aligned}$$ Then, $G$ may be written in terms of the homogenization commutator (see ) as $$\begin{aligned}
\label{Gread}
G=~&-\int \nabla \vbar \cdot \abar \nabla u=\int \nabla \vbar \cdot (a-\abar) \nabla u + F,
\end{aligned}$$ where $F=-\int \nabla \vbar \cdot a \nabla u=\int \nabla \vbar \cdot f$ is deterministic. It turns out that in the homogenization commutator $(a-\abar)\nabla u$, it is legitimate to approximate $\nabla u$ by its two-scale expansion. This leads to the *standard homogenization commutator*, which is the stationary random field $\Xi$ defined by $$\begin{aligned}
\label{Def_Xi}
\Xi e_i := ( a - \abar ) ( e_i+\nabla \phi_i)
\end{aligned}$$ (see [@DuerinckxGO_2016], and the discussion of the literature there). This observation is made rigorous in the result below. This result is identical to [@DuerinckxOtto_2019 Prop. 3.2], which itself is a continuum version of [@DuerinckxGO_2016 Th. 2]. We reproduce the streamlined proof in order to highlight the similarity to the one of Proposition \[Propcorr\] and the use of the annealed CZ estimates, namely Proposition \[PropCZ\](ii).
\[Propfluct\] Let the ensemble $\langl \cdot \rangl$ be defined as in Section \[SecAssumpGauss\]. Assume that $h$ and $f$ are square-integrable deterministic vector fields. Let the square-integrable vector fields $\nabla u$ and $\nabla \ubar$ be related to $f$ through and . Then, the random variable $$\begin{aligned}
\label{DefH}
H:=\int h \cdot ( a - \abar ) \big(\nabla u -\partial_i \ubar (e_i+\nabla \phi_i )\big)
\end{aligned}$$ satisfies, for $\mu_d$ defined by , $$\begin{aligned}
\label{Num:004}
\langl ( H - \langl H \rangl)^{2 r} \rangl^{\frac{1}{r}}
\lesssim_{\gamma, r} \Big( \int | h |^4 \int \big|\mu_d(|\cdot|)\nabla f\big|^4
+ \int |f|^4\int \big|\mu_d(|\cdot|) \nabla h \big|^4 \Big)^{\frac{1}{2}}.
\end{aligned}$$
Let us briefly compare Proposition \[Propfluct\] and Corollary \[Cor1\]. Both give an estimate on the error in the two-scale expansion on the level of the gradient $\nabla(u-(1+\phi_i\partial_i)\ubar) \simeq \nabla u - \partial_i \ubar (e_i+\nabla \phi_i)$. While Corollary \[Cor1\] gives an estimate in a strong norm (*oscillations*), Proposition \[Propfluct\], modulo the homogenization commutator, gives an estimate in a weak norm (*fluctuations*), *i.e.* when tested against a macroscopic $h$.
Proposition \[Propfluct\] allows to approximate the random variable $G$ defined by by its *two-scale expansion* $\tilde G$ defined by $$\label{Def_tildeG}
\tilde{G}:=\int \nabla \vbar \cdot \Xi \nabla \ubar.$$ This is valuable because $\tilde G$ only relies on solving the constant-coefficient problem , the dual constant-coefficient problem , and involves randomness only in form of the standard homogenization commutator , which is independent on the r. h. s. $f$ and the averaging function $g$. In fact, on large scales, $\Xi-\langle \Xi\rangle$ behaves like (tensorial) white noise and is characterized by the above-mentioned quartic tensor $\QU$. As for $\abar$, $\QU$ may be approximated by the representative volume element method [@DuerinckxGO_2016 Th. 2].
\[CoroFluctu\] Let the ensemble $\langl \cdot \rangl$ be defined as in Section \[SecAssumpGauss\]. Assume that $\hat{f}$ and $\hat{g}$ are smooth and compactly supported, and let $f$ and $g$ be defined by . Let $\nabla u, \nabla \ubar$, and $\nabla \vbar $ be square-integrable functions satisfying , , and . Let $G$ and $\tilde{G}$ be defined by and . Then, for $\mu_d$ defined by and $r \in [1,\infty)$, there holds $$\label{Ineq4}
\begin{aligned}
&\epsilon^{-d/2}\langl ( G -\tilde{G} - \langl G-\tilde{G} \rangl )^{2r} \rangl^{\frac{1}{2r}}
&\lesssim_{\gamma, r,\hat{f},\hat{g}} \epsilon \mu_d(1/\epsilon).
\end{aligned}$$
By and as a consequence of Proposition \[Propfluct\], there holds $$\begin{aligned}
\begin{aligned}
&\langl ( G -\tilde{G} - \langl G-\tilde{G} \rangl )^{2r} \rangl^{\frac{1}{r}}
\\
&\qquad \lesssim \Big( \int | h |^4 \int \big|\mu_d(|\cdot|)\nabla f\big|^4 +
\int |f|^4\int \big|\mu_d(|\cdot|) \nabla h \big|^4 \Big)^{\frac{1}{2}}
\end{aligned}
\label{Ineq3}
\end{aligned}$$ for $h:=\nabla \vbar $. We recall that, in dimensions $d \geq 2$, $\mu_d^4(|\cdot|)$ are Muckenhoupt weights of class $A_{4}$. Therefore, by weighted CZ estimates [@Grafakos_book Th. 7.4.6, p. 540], we also have $$\label{Num:718}
\int | h |^4 \lesssim \int |g|^4 \et \int \big|\mu_d(|\cdot|)\nabla h \big|^4 \lesssim \int \big|\mu_d(|\cdot|) \nabla g\big|^4.$$ (A similar estimate also holds in dimension $d=1$.) Hence, using the scaling induced by , estimate yields $$\begin{aligned}
\langl (G -\tilde{G} - \langl G-\tilde{G}\rangl )^{2r} \rangl^{\frac{1}{r}}
&\lesssim \Big( \int | \epsilon^{d} \hat{g}(\epsilon x) |^4 \dd x \int \big|\mu_d(|x|)\epsilon\hat{\nabla} \hat{f}(\epsilon x)\big|^4 \dd x\Big)^{\frac{1}{2}}
\\
&\quad + \Big( \int |\hat{f}(\epsilon x)|^4 \dd x \int \big|\mu_d(|x|) \epsilon^{d+1} \hat{\nabla} \hat{g}(\epsilon x) \big|^4 \dd x\Big)^{\frac{1}{2}},
\end{aligned}$$ which implies by the change of variables $\hat{x}=\epsilon x$.
The proof, which has the same ingredients as the proof of Proposition \[Propcorr\], is divided into three steps. In Step 1, which is reminiscent of Part 1, Step 1 of the proof of Proposition \[Propcorr\], we derive the following representation of the Malliavin derivative of $H$: $$\label{DeriveH}
\begin{aligned}
\frac{\partial H}{\partial a}
=~&h_j ( e_j + \nabla \phi^\star_j) \otimes ( \nabla w + \phi_i \nabla \partial_i \ubar)
\\
&
+ ( \nabla w^\star + \phi_j^\star \nabla h_j) \otimes \nabla u
\\
&- ( \nabla w_i^\star + \phi^\star_j \nabla ( h_j \partial_i \ubar)) \otimes ( e_i + \nabla \phi_i ),
\end{aligned}$$ where $\nabla w$, $\nabla w^\star$, and $\nabla w^\star_j$ are the square-integrable solutions of $$\begin{aligned}
\label{w1}
&\nabla \cdot ( a \nabla w + ( \phi_i a - \sigma_i) \nabla \partial_i \ubar )=0,\\
\label{w2}
&\nabla \cdot ( a^\star \nabla w^\star + ( \phi^\star_j a^\star - \sigma_j^\star ) \nabla h_j )=0,\\
\label{w3}
&\nabla \cdot ( a^\star \nabla w^\star_j + ( \phi^\star_i a^\star - \sigma^\star_i ) \nabla (h_i \partial_j \ubar ) )=0.
\end{aligned}$$ (Note that and define $d+1$ functions $w^\star, w_1^\star, \cdots, w_d^\star$.) Step 2 relies on the spectral gap and makes use of the annealed estimate in Proposition \[PropCZ\](ii) below. Finally, in Step 3, we appeal to the correctors estimates to establish .
Here comes the argument for . Given a smooth and compactly supported infinitesimal variation $\delta a$ of $a$, since $h$ is deterministic, we have for the generated infinitesimal variation of $H$ defined by $$\begin{aligned}
\delta H
=~& \int h \cdot \delta a ( \nabla u - \partial_i \ubar ( e_i + \nabla \phi_i))
+ \int h \cdot ( a - \abar) ( \nabla \delta u -\partial_i \ubar \nabla \delta \phi_i).
\end{aligned}
\label{Diff0}$$ Since $f$ is deterministic, differentiating the equation satisfied by $u$ we obtain $$\begin{aligned}
\label{Diff1}
&\nabla \cdot ( a \nabla \delta u + \delta a \nabla u )=0.
\end{aligned}$$ Moreover, we recall that $\nabla \delta \phi_i$ satisfies . Next, we note, and will prove below, that $w^\star$ and $w^\star_j$ defined by and satisfy $$\begin{aligned}
\label{w21}
&\nabla \cdot \big( a^\star \nabla ( w^\star + \phi_j^\star h_j )
+ (a^\star- \abar^\star) h \big)=0,
\\
\label{w31}
&\nabla\cdot \big( a^\star \nabla ( w^\star_j + \phi_i^\star h_i \partial_j \ubar ) + \partial_j \ubar ( a^\star- \abar^\star) h \big)=0.
\end{aligned}$$ This allows us to rewrite the r. h. s. of as $$\begin{aligned}
\int h \cdot ( a - \abar) \nabla \delta u
&\overset{\eqref{w21}}{=}-\int \nabla ( w^\star + \phi_j^\star h_j )\cdot a \nabla \delta u
\\
&\overset{\eqref{Diff1}}{=} \int \nabla ( w^\star + \phi_j^\star h_j ) \cdot \delta a \nabla u,
\end{aligned}
\label{Diff4}$$ and, similarly, $$\begin{aligned}
-\int \partial_i \ubar h \cdot ( a - \abar) \nabla \delta \phi_i
&\overset{\eqref{w31}}{=}\int \nabla ( w^\star_i + \phi_j^\star h_j \partial_i \ubar ) \cdot a \nabla \delta \phi_i
\\
&\overset{\eqref{DefDeltaphi}}{=} -\int \nabla (w_i^\star + \phi_j^\star h_j \partial_i \ubar )\cdot \delta a ( e_i + \nabla \phi_i).
\end{aligned}
\label{Diff5}$$ Inserting and into , using Leibniz’ rule, and reordering the terms yields $$\label{Diff6}
\begin{aligned}
\delta H
=~&
\int h_j (e_j + \nabla \phi_j^\star )\cdot \delta a ( \nabla u - \partial_i \ubar ( e_i + \nabla \phi_i))
\\
&+\int ( \nabla w^\star + \phi_j^\star \nabla h_j ) \cdot \delta a \nabla u
\\
&-\int \big(\nabla w_i^\star + \phi_j^\star \nabla (h_j \partial_i \ubar) \big) \cdot \delta a ( e_i + \nabla \phi_i).
\end{aligned}$$ Comparing with yields by uniqueness (and Leibniz’ rule) $$\nabla w + \phi_i \nabla \partial_i \ubar=
\nabla u - \partial_i \ubar ( e_i + \nabla \phi_i).$$ Inserting this information into the first right-hand term of yields the desired identity .
We finally argue that $w^\star$ and $w^\star_j$ satisfy and , respectively. We consider the l. h. s. of , in which we replace the first term by means of , and then use (for $\sigma^\star_j$): $$\begin{aligned}
&\nabla \cdot \big( a^\star \nabla ( w^\star + \phi_j^\star h_j ) + (a^\star- \abar^\star) h \big)
\\
&\qquad\overset{\eqref{w2}}{=}
\nabla \cdot \big( \sigma^\star_j \nabla h_j + h_j a^\star \nabla \phi_j^\star + ( a^\star - \abar^\star) h \big)
\\
&\qquad\overset{\eqref{Id1}}{=} \nabla \cdot \big( h_j ( a^\star ( e_j + \nabla \phi^\star_j)- \abar^\star e_j -\nabla \cdot \sigma^\star_j ) \big)
\overset{\eqref{Id02prime}}{=}0.
\end{aligned}$$ Similarly, we obtain by replacing $w^\star$ by $w^\star_j$ and $h$ by $\partial_j \ubar h$ in the above manipulations.
We establish the following estimate $$\begin{aligned}
\langl (H- \langl H \rangl)^{2r}\rangl^{\frac{1}{r}}
\lesssim~&
\Big(\int h_j^4 \langl | e_j + \nabla \phi^\star_j |^{4r} \rangl^{\frac{1}{r}}
\int \langl|(\phi, \sigma)|^{8 r}\rangl^{\frac{1}{2r}} |\nabla^2 \ubar|^4\Big)^{\frac{1}{2}}
\\
&+\Big(\int \langl|(\phi^\star, \sigma^\star) |^{8r}\rangl^{\frac{1}{2r}} |\nabla h|^4 \int |f|^{4}\Big)^{\frac{1}{2}}
\\
&+\int \langl|(\phi^\star,\sigma^\star)|^{8r}\rangl^{\frac{1}{4r}} | \nabla(\partial_i \ubar h) |^2\langl |e_i + \nabla \phi_i|^{4r} \rangl^{\frac{1}{2r}}.
\end{aligned}
\label{Num:006}$$ Indeed, by an application of the spectral gap into which we insert , and making use of Jensen’s inequality, we obtain $$\begin{aligned}
\langl (H- \langl H \rangl)^{2r}\rangl^{\frac{1}{r}}
\lesssim~& \int \Big\langl\big| \frac{\partial H}{\partial a(x)}\big|^{2r}\Big\rangl^{\frac{1}{r}} \dd x
\\
\lesssim~&
\int \big\langl | h_j ( e_j + \nabla \phi^\star_j) \otimes ( \nabla w + \phi_i \nabla \partial_i \ubar)|^{2r} \big\rangl^{\frac{1}{r}}
\\
&+\int \big\langl| ( \nabla w^\star + \phi_j^\star \nabla h_j) \otimes \nabla u|^{2r} \big\rangl^{\frac{1}{r}}
\\
&+\int \big\langl| ( \nabla w_i^\star + \phi^\star_j \nabla ( h_j \partial_i \ubar)) \otimes ( e_i + \nabla \phi_i )|^{2r}\big\rangl^{\frac{1}{r}}.
\end{aligned}$$ Let us focus on the first above right-hand term. By the Cauchy-Schwarz inequality in probability and in space, this yields $$\begin{aligned}
&\int \langl | h_j ( e_j + \nabla \phi^\star_j) \otimes ( \nabla w + \phi_i \nabla \partial_i \ubar)|^{2r} \rangl^{\frac{1}{r}}
\\
&\qquad \leq \Big(\int \langl | h_j ( e_j + \nabla \phi^\star_j)|^{4r}\rangl^{\frac{1}{r}} \Big)^{\frac{1}{2}}
\Big( \int\langl | \nabla w + \phi_i \nabla \partial_i \ubar|^{4r}\rangl^{\frac{1}{r}}\Big)^{\frac{1}{2}}.
\end{aligned}$$ By similar manipulations and the stationarity of $\nabla \phi_i$, we get $$\begin{aligned}
\langl (H- \langl H \rangl)^{2r}\rangl^{\frac{1}{r}}
& \lesssim
\Big(\int \langl | h_j ( e_j + \nabla \phi^\star_j) |^{4r} \rangl^{\frac{1}{r}}\Big)^{\frac{1}{2}}
\Big(\int \langl | \nabla w + \phi_i \nabla \partial_i \ubar|^{4r} \rangl^{\frac{1}{r}}\Big)^{\frac{1}{2}}
\\
&~~~~+\Big(\int \langl | \nabla w^\star + \phi_j^\star \nabla h_j |^{4r} \rangl^{\frac{1}{r}}\Big)^{\frac{1}{2}}
\Big(\int \langl | \nabla u|^{4r} \rangl^{\frac{1}{r}}\Big)^{\frac{1}{2}}
\\
&~~~~+\int \langl |\nabla w_i^\star + \phi^\star_j \nabla ( h_j \partial_i \ubar)|^{4r} \rangl^{\frac{1}{2r}}\langl |e_i + \nabla \phi_i|^{4r} \rangl^{\frac{1}{2r}}.
\end{aligned}
\label{Num:005}$$ We now invoke the annealed estimate from Proposition \[PropCZ\](ii) below with $r' \rightsquigarrow 4r$, $r\rightsquigarrow 8r$, and $p \rightsquigarrow 4$ on $\nabla \cdot ( a \nabla u + f )=0$, obtaining, since $f$ is deterministic, $$\begin{aligned}
\int \langl | \nabla u|^{4r} \rangl^{\frac{1}{r}}
\lesssim \int \langl |f|^{8r} \rangl^{\frac{1}{2r}} =\int |f|^{4}.
\end{aligned}$$ Similarly, based on , , and , we have that $$\begin{aligned}
\int \langl | \nabla w |^{4r} \rangl^{\frac{1}{r}}
\lesssim~&
\int \langl|(\phi,\sigma)|^{8 r}\rangl^{\frac{1}{2r}} |\nabla^2 \ubar|^4,
\\
\int \langl|\nabla w^\star|^{4r}\rangl^{\frac{1}{r}}
\lesssim~& \int \langl|(\phi^\star, \sigma^\star) |^{8r}\rangl^{\frac{1}{2r}} |\nabla h|^4,
\\
\int \langl |\nabla w^\star_i|^{4r} \rangl^{\frac{1}{2r}}
\lesssim~& \int \langl|(\phi^\star,\sigma^\star)|^{8r}\rangl^{\frac{1}{4r}} | \nabla(\partial_i \ubar h) |^2.
\end{aligned}$$ Inserting these four estimates into (and using Jensen’s inequality in probability) entails .
By the corrector estimates and , and Leibniz’ rule followed by the Cauchy-Schwarz inequality in the last term, turns into $$\begin{aligned}
&\langl |H- \langl H \rangl|^{2r}\rangl^{\frac{1}{r}}
\\
&\qquad \lesssim
\Big(\int |h|^4 \int \big|\mu_d(|\cdot|) \nabla^2 \ubar\big|^4\Big)^{\frac{1}{2}}
+\Big(\int \big|\mu_d(|\cdot|) \nabla h\big|^4 \int |(f,\nabla \ubar)|^{4}\Big)^{\frac{1}{2}}.
\end{aligned}
\label{Num:007}$$ As for , we have $$\begin{aligned}
\int |\nabla \ubar|^4 \lesssim \int |f|^4 \et \int \big|\mu_d(|\cdot|) \nabla^2 \ubar\big|^4 \lesssim \int \big|\mu_d(|\cdot|) \nabla f\big|^4.
\end{aligned}$$ Inserting these estimates into yields .
Our tool: Annealed CZ estimates {#SecAnnealed}
===============================
General statement, and proof of the perturbative CZ estimates
-------------------------------------------------------------
The main contribution of this paper is a novel proof of the following result from [@DuerinckxOtto_2019 Th. 6.1]:
\[PropCZ\] Let the random fields $\nabla u$ and $f$ be square-integrable and related by $$\begin{aligned}
\label{Prop3Defv}
\nabla \cdot ( a \nabla u + f ) = 0.
\end{aligned}$$ Then:
\(i) [For any ensemble $\langl \cdot \rangl$ of elliptic coefficient fields $a$ satisfying , there holds $$\begin{aligned}
\label{Prop3E1}
\int \langl |\nabla u|^r \rangl^{\frac{2}{r}}
\lesssim_{d,\lambda}
\int \langl |f|^r \rangl^{\frac{2}{r}}, \qquad \text{ provided } |r-2| \ll_{d,\lambda} 1.
\end{aligned}$$ ]{}
\(ii) [For any ensemble $\langl \cdot\rangl$ as in Section \[SecAssumpGauss\], there holds $$\begin{aligned}
\label{Estim13_ter}
\int \langl|\nabla u|^{r'}\rangl^\frac{p}{r'}
\lesssim_{\gamma, p, r',r}
\int \langl|f|^{r}\rangl^\frac{p}{r} \pourtout 1\leq r'<r\leq \infty \text{ and } 1 < p<\infty.
\end{aligned}$$ ]{}
Note that the ensemble occurs in the *inner* part of the norms in Proposition \[PropCZ\]. In other words, the estimates are not about stochastic moments of the constant in a quenched CZ estimate, *i.e.* in norms $\LL^r_{\langl\cdot\rangl}(\LL_{\R^d}^p)$, but rather about *vector-valued* CZ estimates, *i.e.* in norms $\LL_{\R^d}^p(\LL^r_{\langl\cdot\rangl})$ (in the terminology of [@McConnell_1984], where the vector is represented by the values with respect to the ensemble). In particular, even in the case when $a$ is a deterministic constant coefficient, estimate is not trivial and requires refined tools (which however are well-established, see the proof of Proposition \[PropCZ\](i) below).
Proposition \[PropCZ\](i) holds under very general assumptions and rests only on the perturbative Meyers’ approach to lift the above-mentioned corresponding statement for constant coefficients. On the contrary, Proposition \[PropCZ\](ii) requires homogenization techniques. Namely, we follow the basic idea of lifting the constant-coefficient regularity theory for the Helmholtz projection (see Section \[SS\_AnnealedCZ\] for a discussion).
Proposition \[PropCZ\](ii) holds at the price of a loss of stochastic integrability. This loss is unavoidable for $p$ or $r$ far from $2$. This is not only due to the randomness of the local regularity of $a$, but more importantly, it is also a consequence of the randomness on large scales. To explain this, we fix $|p-2| \gg_{d,\lambda} 1$ and consider the simpler framework of $\R^d$ replaced by the lattice $\Z^d$ (see [@Gloria_Otto]). The random coefficient field $a$ now lives on the edges of the lattice $\Z^d$ and takes two values according to i. i. d. Bernoulli variables. Then, it is easily seen that any given configuration on a domain of finite size may be found with positive probability. Choosing a configuration with an arbitrarily large constant $C$ in the CZ estimate [^4], that is $$\int |\nabla u|^p \geq C \int |f|^p$$ for some suitable $f$, we may build a random r. h. s. $f$ such that $$\begin{aligned}
\int \langle |\nabla u|^p \rangle \geq C \quad \text{ but }\int \langle |f|^p \rangle = 1,
\end{aligned}$$ so that cannot hold for $r'=r=p$.
The main ingredient is the following: If $a=\Id$, then Proposition \[PropCZ\](i) is satisfied (here, we denote by $\Id$ the identity matrix). A perturbation argument *à la Meyers* yields the desired result for general elliptic coefficient field $a$ (see *e.g.* [@Meyers_Estim] and a more recent presentation in [@MuellerI Chap. 2, Th. 2.6.2, p. 122] for the proof of the Meyers estimate).
By [@McConnell_1984 Th. 1.1.], if $a=\Id$, then holds for all $r \in (1,\infty)$ since the space $\LL^{r}_{\langl \cdot\rangl}$ is UMD by [@Weis_Book Prop. 4.2.15, p. 291]. In this special case, we denote by $C_{r}$ the best constant in $$\begin{aligned}
\Big( \int \langl |\nabla u|^r \rangl^{\frac{2}{r}} \Big)^{\frac{1}{2}}
\leq C_{r} \Big( \int \langl |f|^r \rangl^{\frac{2}{r}} \Big)^{\frac{1}{2}}.
\end{aligned}$$ We now rewrite as $$\begin{aligned}
\label{Truc}
\nabla u = \nabla \Delta^{-1} \nabla \cdot ( \Id - a ) \nabla u - \nabla \Delta^{-1} \nabla \cdot f.
\end{aligned}$$ By , we have in the sense of operator norm on $\R^d$ $$\begin{aligned}
|\Id-a| \leq \sqrt{1-\lambda}.
\end{aligned}$$ Indeed, for all $\xi \in \R^d$, there holds $$|(\Id-a)\xi|^2 = \xi^2 + |a\xi|^2 -2 \xi \cdot a \xi \overset{\eqref{Ellipticite}}{\leq} \xi^2 - \xi \cdot a \xi \overset{\eqref{Ellipticite}}{\leq} (1-\lambda) |\xi|^2.$$ Therefore, the operator $\nabla \Delta^{-1} \nabla \cdot ( \Id - a )$ appearing in is a contraction in $\LL_{\R^d}^{2}\LL^r_{\langl \cdot \rangl}$ provided $$\begin{aligned}
\label{Condition}
C_{r} \sqrt{1-\lambda} <1.
\end{aligned}$$ By the energy estimate we already know that $C_{2}=1$. Fix an exponent $\bar{r} \in (2,\infty)$; by complex interpolation between spaces $\LL_{\R^d}^{2}\LL^r_{\langl \cdot \rangl}$, we obtain that for $1/r=(1-\theta)/2 + \theta/\bar{r}$ there holds: $$\begin{aligned}
C_{r} \leq C_{2}^{1-\theta} C_{\bar{r}}^{\theta} = C_{\bar{r}}^{\theta}, \quad \text{so that}\; \limsup_{r \downarrow 2} C_{r} \leq 1.
\end{aligned}$$ Thus, if $|r-2| \ll_{d,\lambda} 1$, then is satisfied, so that the operator $\nabla \Delta^{-1} \nabla \cdot ( \Id - a )$ appearing in is indeed a contraction on $\LL_{\R^d}^{2}\LL^r_{\langl \cdot \rangl}$. This shows that the operator $\nabla ( \nabla \cdot a \nabla )^{-1} \nabla \cdot$ is bounded in $\LL_{\R^d}^{2}\LL^r_{\langl \cdot \rangl}$, which amounts to .
Strategy for proving the non-perturbative CZ estimates: massive equation and correctors {#SS_AnnealedCZ}
---------------------------------------------------------------------------------------
The proof of Proposition 7.1 (ii) follows the philosophy of [@AvellanedaLin] in the sense that we appeal to homogenization to deduce the boundedness of the $a$-Helmholtz projection from that of the $\abar$-Helmholtz projection. Smallness of the homogenization error necessarily requires some amount of regularity of the r. h. s. $f$, see for instance Corollary \[Cor1\]. Since for the operator norm, we have to consider arbitrary $f$, this suggests to decompose $f$ into a fairly smooth or large-scale part $f_<$ (we use the language of *low-pass*) and an oscillatory small scale part $f_>$ (*high-pass*); the corresponding $\nabla u_>$ then requires an independent argument. The latter argument is based on locality, meaning that $\nabla u_>(x)$ depends on $f$ only in a (large) ball around $x$.
Given a length scale, such a decomposition into a smooth part and a local part is provided by the semi-group, which here we substitute by the resolvent ${\frac{1}{\tau}-\nabla\cdot a\nabla}$, which is called the *massive* version of the elliptic operator. The notation is motivated by the interpretation of $\frac{1}{\tau}-\nabla\cdot a\nabla$ as the generator of a diffusion coupled to desorption at exponential rate $\frac{1}{\tau}$, and the language comes from quantum field theory where this term is well-known to provide an infra-red cut-off beyond length scales $\sqrt{\tau}$. More precisely, we consider $u_>:=(\frac{1}{\tau}-\nabla\cdot a\nabla)^{-1}\nabla\cdot f$, which up to exponentially small tails has the desired locality properties on scale $\sqrt{\tau}$ (as can be guessed from the form of the fundamental solution of $\frac{1}{\tau}-\triangle$).
As mentioned, we shall apply homogenization to $u_<:=u-u_>$. We depart from the general strategy in [@AvellanedaLin] by using a whole-space argument (as opposed to arguing on dyadically increasing balls typical for a Campanato iteration). Since we will appeal to the divergence-form representation of the residuum in the error of the two-scale expansion, *cf.* , this whole-space argument would require the correctors $(\phi,\sigma)$ to be stationary, which would impose the restriction $d>2$. (This easier argument will be carried out in [@JosienOtto_2020].) Here, we avoid this restriction by equivalently characterizing the low-pass part as the solution of the *massive* equation $u_<=(\frac{1}{\tau}-\nabla\cdot a\nabla)^{-1}\frac{1}{\tau}u$, so that when comparing $u_<$ to its homogenized counterpart $\bar u$ $:=(\frac{1}{\tau}-\nabla\cdot \abar_\tau\nabla)^{-1}\frac{1}{\tau}u$, the analogue of will involve the *massive* correctors (and the corresponding homogenized coefficient $\abar_\tau$ defined by below), see Section \[Sec\_MassCorr\]. The massive correctors are trivially stationary and satisfy estimates analogous to the standard correctors, see Lemma \[LemMassCorr\].
A “collateral damage” of carrying out the homogenization on the level of $u_<=(\frac{1}{\tau}-\nabla\cdot a\nabla)^{-1}\frac{1}{\tau}u$ (instead of $u_<=(-\nabla\cdot a\nabla)^{-1}\frac{1}{\tau}u_>$), leading to $\bar u$ $=(\frac{1}{\tau}-\nabla\cdot \abar_\tau\nabla)^{-1}\frac{1}{\tau}u$ is the following: We need to split also $\bar u$ into $\bar u_>$ $:=(-\nabla\cdot \abar_\tau\nabla)^{-1}\frac{1}{\tau}u_>$ coming from the local part, and the remainder $\bar u_<:=\bar u-\bar u_>$, which again is expected to be small by homogenization, *cf.* and .
In view of this crucial role of the massive operator, it is convenient to unfold our task by establishing the boundedness of the massive $a$-Helmholtz projection $\nabla(\frac{1}{\tau}-\nabla\cdot a\nabla)^{-1}\nabla\cdot$, and to include the zero-order term on the l. h. s. into the estimate, as well as to allow for a non-divergence form r. h. s., *cf.* . This comes with the notational disadvantage that we need to monitor *two* cut-off parameters $T\ge\tau$. We recover the massless CZ estimate as the limit of its massive counterpart (along with ).
### The massive equation
We consider a massive version of $$\begin{aligned}
\label{Ann_ae1}
\frac{1}{T}u-\nabla\cdot a\nabla u=\frac{1}{T}g + \nabla\cdot f
\end{aligned}$$ for a parameter $T\geq 1$, which we think of as being large. As mentioned above, our strategy is to derive annealed estimates for (see Proposition \[PropmassCZ\]). As explained above, we split $u$ as follows: $$\begin{aligned}
\label{Ann_ae3}
u=u_>+u_<,
\end{aligned}$$ where, for some $\tau \leq T$, $u_>$ is the high-pass part defined by $$\label{Def_u_tau}
\frac{1}{\tau}u_>-\nabla\cdot a\nabla u_>=\frac{1}{T}g + \nabla\cdot f,$$ and where the low-pass part $u_<$ satisfies $$\begin{aligned}
\label{Ann_ae4}
\frac{1}{\tau} u_< - \nabla \cdot a\nabla u_<= \big(\frac{1}{\tau}-\frac{1}{T}\big)u.
\end{aligned}$$ We approximate the low-pass part $u_<$ by the solution of the homogenized equation $$\begin{aligned}
\label{Ann_ae5}
\frac{1}{\tau} \ubar - \nabla \cdot \abar_\tau \nabla \ubar = \big(\frac{1}{\tau}-\frac{1}{T}\big)u
\end{aligned}$$ for $\abar_\tau$ defined by below. However, as mentioned above, on the homogenized level of $\ubar$, it turns out that we need to split once more into the high-pass and the low-pass parts: $$\label{Ann_ae3_bis}
\ubar
=\ubar_>+\ubar_<,$$ where $\ubar_>$ is the high-pass part defined by $$\label{Ann_ae10_bis}
\frac{1}{T} \ubar_> - \nabla \cdot \abar_\tau \nabla \ubar_>
=\big(\frac{1}{\tau}-\frac{1}{T}\big)u_>,$$ and where the low-pass part $\ubar_<$ satisfies $$\label{Ann_ae11_bis}
\frac{1}{T} \ubar_< - \nabla \cdot \abar_\tau \nabla \ubar_<
=\big(\frac{1}{\tau}-\frac{1}{T}\big)(u_<-\ubar ).$$
### The massive correctors {#Sec_MassCorr}
Expressing the homogenization error when passing from to requires the use of the *massive* extended correctors $(\phi_\tau,\sigma_\tau,\psi_\tau)$ (see [@GloriaOtto_2015 (48), (50) & (51)] and also [@GloriaOtto_2017_Corr]). These objects are stationary solutions to the upcoming equations: Generalizing and , the massive correctors $\phi_{\tau,i}$ are defined through $$\begin{aligned}
\label{Ann_as12}
\frac{1}{\tau}\phi_{\tau,i}-\nabla\cdot a(\nabla\phi_{\tau,i}+e_i)=0,
\end{aligned}$$ and the massive flux correctors $\sigma_{\tau,i}=\{\sigma_{\tau,ijk}\}_{j, k=1,\cdots,d}$ are skew symmetric tensor fields given by $$\label{Ann_as12s}
\big(\frac{1}{\tau}-\Delta\big)\sigma_{\tau,ijk}=\partial_j q_{\tau,ik}-\partial_k q_{\tau,ij},$$ where $$\label{Def_qT}
q_{\tau,ij}:=e_j\cdot a(e_i+\nabla\phi_{\tau,i}) \et \quad \abar_\tau e_i:=\langl q_{\tau,i}\rangl.$$ The vector field $\psi_{\tau,i}$, which has no analogue in the case of $\tau=\infty$, is defined through $$\begin{aligned}
\label{Ann_as11}
\big(\frac{1}{\tau}-\Delta\big)\psi_{\tau,i}=q_{\tau,i}-\abar_\tau e_i-\nabla\phi_{\tau,i}.
\end{aligned}$$ The merit of $\psi_{\tau,i}$ is that the massive extended correctors $(\phi_\tau,\sigma_\tau,\psi_\tau)$ together satisfy the following generalization of : $$\begin{aligned}
\label{Ann_as10}
a(e_i+\nabla\phi_{\tau,i})=q_{\tau,i}=\abar_\tau e_i+\nabla\cdot\sigma_{\tau,i}+\frac{1}{\tau}\psi_{\tau,i}.
\end{aligned}$$ Note that each of the equations , and has a unique solution in the class of bounded fields, which is thus stationary (see *e.g.* [@GloriaOtto_2017_Corr Lem. 2.7]).
The massive correctors enjoy properties similar to their massless counterparts (see Proposition \[Propcorr\]):
\[LemMassCorr\] Let $T \geq 1$. Under the assumptions of Section \[SecAssumpGauss\], for any $r \in [1, \infty)$, there holds: $$\begin{aligned}
\label{Ann_ap56}
\Big\langl\big|\big(\nabla\phi_T,\nabla\sigma_T,\frac{\nabla\psi_T}{\sqrt{T}}\big)\big|^{2r}\Big\rangl^\frac{1}{r}
&\lesssim_{\gamma,r} 1,
\\
\label{Ann_ap57}
\Big\langl\big|\big(\phi_T,\sigma_T,\frac{\psi_T}{\sqrt{T}}\big)\big|^{2r}\Big\rangl^\frac{1}{r}
&\lesssim_{\gamma,r}
\mu^2_d(\sqrt{T}),
\end{aligned}$$ where $\mu_d$ is defined by .
We postpone the proof of Lemma \[LemMassCorr\] until Section \[Sec\_LemMassCorr\].
By the uniqueness result [@GloriaOtto_2017_Corr Lem. 2.7], it is enough to check (\[Ann\_as10\]) after applying the operator $\big(\frac{1}{\tau}-\Delta\big)$ to it, that is, in form of $$\begin{aligned}
\big(\frac{1}{\tau}-\Delta\big)q_{\tau,i}=\frac{1}{\tau}\abar_\tau e_i+\nabla\cdot\big(\frac{1}{\tau}-\Delta\big)\sigma_{\tau,i} +\frac{1}{\tau}\big(\frac{1}{\tau}-\Delta\big)\psi_{\tau,i}.
\end{aligned}$$ We first eliminate $\psi_\tau$ via its definition (\[Ann\_as11\]), leading to $$\begin{aligned}
-\Delta q_\tau=\nabla\cdot\big(\frac{1}{\tau}-\Delta\big)\sigma_\tau-\frac{1}{\tau}\nabla\phi_\tau.
\end{aligned}$$ We then eliminate $\sigma_\tau$ by inserting $\nabla\cdot\big(\frac{1}{\tau}-\Delta\big)\sigma=\nabla\nabla\cdot q_\tau-
\Delta q_\tau$, which is obtained by applying $\partial_k$ to equation (\[Ann\_as12s\]), to the effect of $$\begin{aligned}
0=\nabla\nabla\cdot q_\tau-\frac{1}{\tau}\nabla\phi_\tau.
\end{aligned}$$ This is nothing else than the operator $-\nabla$ applied to equation , appealing to definition .
### Splitting {#Sec:0_Splitting}
As announced, the massive extended correctors allow us to characterize the error in the massive two-scale expansion $$\begin{aligned}
\label{Ann_ae6}
w:=u_<-(1+\phi_{\tau,i} \partial_i)\ubar
\end{aligned}$$ via $$\begin{aligned}
\frac{1}{\tau} w - \nabla \cdot a \nabla w
=~& \nabla \cdot \big( ( \phi_{\tau,i} a -\sigma_{\tau,i}) \nabla \partial_i \ubar + \frac{1}{\tau} \partial_i \ubar \psi_{\tau,i} \big)
- \frac{1}{\tau}\phi_{\tau,i} \partial_i \ubar .
\end{aligned}
\label{Ann_ae7}$$ This induces the following splitting $(u,\nabla u)=(v_>,h_>)+(v_<,h_<)$ into high-pass and low-pass parts via $$\begin{aligned}
\left\{
\begin{aligned}
v_>&:=u_> + \ubar_>,
\\
v_<&:=w+\ubar_< + \phi_{\tau,i} \partial_i \ubar ,
\\
h_>&:=\nabla u_> + \partial_i \ubar_>(e_i + \nabla \phi_{\tau,i}),
\\
h_<&:=\nabla w + \partial_i \ubar_< (e_i + \nabla \phi_{\tau,i}) + \phi_{\tau,i} \nabla \partial_i \ubar,
\end{aligned}
\right.
\label{Ann_ae12}
\end{aligned}$$ *cf.* , , and , which will be used in Parts 1 and 2 of the proof of Proposition \[PropCZ\].
Equation is derived by taking the difference between and , and by appealing to the intertwining relation (which generalizes ) $$\begin{aligned}
\nabla\cdot a\nabla(1+\phi_{\tau,i}\partial_i)\ubar =\nabla\cdot\abar_\tau\nabla \ubar
+\nabla\cdot\big((\phi_{\tau,i} a-\sigma_{\tau,i}) \nabla \partial_i\ubar +\frac{1}{\tau} \partial_i\ubar \psi_{\tau,i} \big).
\end{aligned}$$ This intertwining relation itself follows via $\nabla(1+\phi_{\tau,i}\partial_i)\ubar $ $=\partial_i\ubar (e_i+\nabla\phi_{\tau,i})+\phi_{\tau,i}\nabla \partial_i\ubar $ from multiplying (\[Ann\_as10\]) by $\partial_i\ubar $ and using the identity in form of $\nabla\cdot(\partial_i\ubar \nabla\cdot\sigma_{\tau,i})=-\nabla\cdot(\sigma_{\tau,i}\nabla \partial_i\ubar )$, which relies on the skew symmetry of $\sigma_{\tau,i}$.
### Result for the massive equation
For further discussion, we denote the norm appearing in by $$\label{Num:701}
\|h\|_{p,r}:=\Big(\int \big\langl|h|^r\big\rangl^\frac{p}{r}\Big)^\frac{1}{p}.$$ As announced, we generalize Proposition \[PropCZ\](ii) to:
\[PropmassCZ\] Fix the exponents $p\in (1,\infty)$ and $1 \leq r'<r < \infty$. Let $\barC_{p,r',r}(T) \geq 1$ denote the smallest constant such that for all square-integrable random fields $u$, $g$, and $f$ related through the massive equation , we have $$\begin{aligned}
\label{Ann_ee5}
\big\|\big(\frac{u}{\sqrt{T}},\nabla u\big)\big\|_{p,r'} \leq
\barC_{p,r',r}(T) \big\|\big(\frac{g}{\sqrt{T}},f\big)\big\|_{p,r}.
\end{aligned}$$ Under the assumptions of Section \[SecAssumpGauss\], this constant satisfies $$\label{Borne_C}
\barC_{p,r',r}(T) \lesssim_{\gamma,p,r',r} 1 \quad \pourtout T \geq 1.$$
The proof of Proposition \[PropmassCZ\] is done in Section \[Sec:ProofPropMassCZ\] and relies on only two ingredients:
- [the homogenized (constant-coefficient) operator $\frac{1}{T}-\nabla\cdot\abar_\tau\nabla$ satisfies , with a constant $\bar{C}_{p,r',r}$ independent of $T$, and even for $r'=r$ (see Section \[Ann\_ssconst\]);]{}
- [the massive correctors are strictly sublinear in $\sqrt{T}$ (see Lemma \[LemMassCorr\]).]{}
Philosophically speaking, we substitute any regularity theory for the operator $\frac{1}{T}-\nabla\cdot a\nabla$ by regularity theory for the operator $\frac{1}{T}-\nabla\cdot\abar_\tau\nabla$. This substitution relies on functional analysis, mostly interpolation. Literally, the only large-scale regularity theory ingredient for the massive operator $\frac{1}{T}-\nabla\cdot a\nabla$ is the whole-space energy estimate (see Section \[Sec:Lem\_SsOpti\]). This approach quite different from the one based on quenched CZ estimates (briefly described in Section \[SecLocal\]).
### Strategy for dealing with the loss in the stochastic integrability {#Sec:Strat}
The strategy is to start from the standard annealed CZ estimates for the solution $\ubar$ of the constant-coefficient equation (*cf.* Section \[Ann\_ssconst\]) and to buckle on the level of the optimal constant $\barC_{p,r',r}(T)$ in by applying CZ estimates also to the equation for $w$. The main challenge in this strategy is the loss in stochastic integrability coming from the need to feed in the estimates on the massive correctors (\[Ann\_ap57\]). To overcome this challenge, we use a real interpolation argument. The loss in stochastic integrability of the low-pass part $u_<$ is compensated by its smallness when $\tau$ is large and by the gain in stochastic integrability of the high-pass part (see Parts 1 and 2 of the proof of Proposition \[PropmassCZ\]).
To carry out this real interpolation argument, we need an independent estimate of the high-pass part $u_>$. It is provided in Section \[Sec:Lem\_SsOpti\] by Lemma \[Lem\_SsOpti\], which is based on the locality of $u_>$. The estimate therein is highly suboptimal in its scaling in $\sqrt{T}$ (which plays the role of $\sqrt{\tau}$); this however is balanced by the smallness of $u_<$ in our real interpolation argument. Finally, by complex interpolation starting from the pivotal energy estimate , we iteratively enlarge the zone in which we have $\barC_{p,r',r}(T) \lesssim 1$, reaching any admissible $3$-tuple of exponents $(p,r',r)$ in a finite number of steps (see Parts 3 and 4 of the proof of Proposition \[PropmassCZ\]).
### The constant-coefficient estimates {#Ann_ssconst}
We only need one result for the constant-coefficient equation, namely:
\[LemCstMassCZ\] Assume that $\abar$ is a constant coefficient satisfying . Then, for any $p, r \in (1,\infty)$, the solution $\ubar$ to the massive equation $$\begin{aligned}
\label{Ann_masseq}
\frac{1}{T} \ubar - \nabla \cdot \abar \nabla \ubar =\frac{1}{T} g + \nabla \cdot f
\end{aligned}$$ satisfies $$\begin{aligned}
\label{Ann_eee23}
\big\|\big( \frac{\ubar}{\sqrt{T}}, \nabla \ubar \big) \big\|_{p,r} \lesssim_{d,\lambda,p,r} \big\|\big( \frac{g}{\sqrt{T}}, f \big)\big\|_{p,r}.
\end{aligned}$$
The proof is based on a version of the Mikhlin theorem, which involves Fourier multipliers [@McConnell_1984 Th. 1.1]. By scaling, we may assume $T=1$. The functional space $\LL^{r}_{\langl\cdot\rangl}$ is a UMD space (see [@Weis_Book Def. 4.2.1, p. 281] for a definition). Moreover, the Fourier multiplier $m$ corresponding to the solution operator $\mathfrak{M}:(g,f) \mapsto (\ubar,\nabla \ubar )$ of has the block structure $$\begin{aligned}
m(k)=\frac{1}{1 + k \cdot \abar k} \left( \begin{array}{c|c} 1 & \ii k^\star \\ \hline \ii k & - k \otimes k \end{array} \right),
\end{aligned}$$ where $k^\star$ is the vector $k$ transposed. Obviously, the symbol $m$ belongs to $\CC^{d+1}(\R^d \backslash \{0\})$, and satisfies $$\begin{aligned}
\sup_{k \in \R^d\backslash\{0\}}\sum_{j=0}^{d+1} |k|^j |\nabla^j m(k)| \lesssim_{d,\lambda} 1.
\end{aligned}$$ Therefore, by [@McConnell_1984 Th. 1.1], the operator $\mathfrak{M}$ extends from $\LL^p(\R^d)$ to $\LL^{p}(\R^d,\LL^r_{\langl\cdot\rangl})$. This establishes .
Suboptimal CZ estimates for the massive equation {#Sec:Lem_SsOpti}
------------------------------------------------
As mentioned above, we need a robust but suboptimal estimate of $\barC_{p,r',r}(T)$:
\[Lem\_SsOpti\] Let $T \geq 1$. Under the assumptions of Section \[SecAssumpGauss\], the operator norm denoted by $\barC_{p,r',r}(T)$ in satisfies $$\begin{aligned}
\label{Ann_ap25}
\barC_{2,2,2}(T)&\lesssim_{d,\lambda} 1,
\\
\label{Ann_ee2*}
\barC_{p,r',r}(T) &\lesssim_{\gamma,p,r',r}
\sqrt{T}^d \qquad \text{provided } p \in (1,\infty) \text{ and } 1\leq r'<r<\infty.
\end{aligned}$$
Lemma \[Lem\_SsOpti\] relies on the locality on scale $\sqrt{T}$ of solutions to the massive equation, and on the regularity on scale $1$ of the coefficient field (see Lemma \[Lem\_AnnealedCZ\_2\]). These two properties allow for estimates between different spatial $\LL^p$-norms, where locality provides the large-scale cut-off, and regularity the small-scale cut-off. In particular, we will jump between the $\LL^2_{\R^d}$-norm, on which scale we have energy estimates, and the $\LL^\infty_{\R^d}$-norm, where we may handle the $\LL^r_{\langle\cdot\rangle}$-norm.
Throughout the proof, the square-integrable random fields $u$, $g$, and $f$ are related by . By a duality argument and Jensen’s inequality, we may restrict to the case of $2\le r'<r$. As mentioned above, the local regularity of $a$ allows us to use CZ estimates on scales $\le 1$. We capitalize on this in form of Lemma \[Lem\_AnnealedCZ\_2\](i), via a family of norms that treat scales $\le 1$ separately, namely $$\begin{aligned}
\label{Ann_ap20}
\|h\|_{p,r,q}:=\bigg(\int \Big\langl\Big(\fint_{\Boule_1(x)}|h|^q\Big)^\frac{r}{q}\Big\rangl^\frac{p}{r}\dd x\bigg)^\frac{1}{p},
\end{aligned}$$ where we think of the exponent $q \in(1,\infty)$ of the innermost spatial norm as being close to $1$ (of course, the precise value $1$ of the radius in is not important). In Step 1, we will derive the crucial property of these norms, namely that they decrease with increasing spatial exponent $p$, see . As mentioned above, the massive term provides an approximate locality on scale $\sqrt{T}$, which in Step 2 we capture through the energy estimate with exponential weight on that scale, see . In Step 3, we use local regularity to express this weighted energy estimate on the level of the norms . In Step 4, we derive the statement of this lemma on the scale of the norms , see . In Step 5, finally, we use once more local regularity to return to the original norms .
We claim the following properties of the norms : $$\begin{aligned}
\label{Ann_ap39}
&\|h\|_{\bar{p},r,q}\lesssim\|h\|_{p,r,q} &&\mbox{provided}\;p\le \bar{p},
\\
\label{Num:03_bis}
&\|h\|_{p,r'}\leq\|h\|_{p,r} \quad \et \quad \|h\|_{p,r',q}\leq\|h\|_{p,r,q} &&\mbox{provided}\;r'\le r,
\\
\label{Num:706}
&\|h\|_{p,r,q}\leq \|h\|_{p,r} &&\mbox{provided}\; q \leq \min\{p,r\}.
\end{aligned}$$
We start with the argument for (\[Ann\_ap39\]). The core is the following discrete $\ell_{\bar{p}}-\ell_p$ estimate, where we introduce the abbreviation $\bar{h}_R(y):=\langl(\fint_{\Boule_R(y)}|h|^q)^\frac{r}{q}\rangl^\frac{1}{r}$: $$\begin{aligned}
\big(\sum_{z\in\mathbb{Z}^d}|\bar h_1(x+z)|^{\bar{p}}\big)^\frac{1}{{\bar{p}}}
\le
\big(\sum_{z\in\mathbb{Z}^d}|\bar h_1(x+z)|^p\big)^\frac{1}{p}.
\end{aligned}$$ Using that $0\le\bar {h}_1(y)\lesssim\bar {h}_2(y')$ for $|y'-y|<1$ and thus $|\bar h_1(y)|^p\lesssim\fint_{\Boule_1(y)}|\bar h_2|^p$, the above inequality may be upgraded to $$\begin{aligned}
\big(\sum_{z\in\mathbb{Z}^d}|\bar h_1(x+z)|^{\bar{p}}\big)^\frac{1}{{\bar{p}}}
\lesssim
\Big(\int |\bar h_2|^p\Big)^\frac{1}{p}.
\end{aligned}$$ Taking the $\LL^{\bar{p}}$-norm in $x\in[0,1)^d$ of this estimate, we obtain $$\begin{aligned}
\label{Ann_ee21}
\Big(\int |\bar h_1|^{\bar{p}}\Big)^\frac{1}{{\bar{p}}}
\lesssim\Big(\int |\bar h_2|^p\Big)^\frac{1}{p}.
\end{aligned}$$ Using now the elementary geometric fact that there exist $N\lesssim 1$ shift vectors $z_1,\cdots,z_N\in\R^d$ such that $\Boule_2(x)$ $\subset\bigcup_{n=1}^N \Boule_1(x-z_n)$ and therefore $(\fint_{\Boule_2(x)}|h|^q)^\frac{1}{q}$ ${\leq\sum_{n=1}^N(\fint_{\Boule_1(x-z_n)}|h|^q)^\frac{1}{q}}$, we obtain by the triangle inequality and the shift invariance of the norm $(\int \langl|h|^r\rangl^\frac{{\bar{p}}}{r})^\frac{1}{{\bar{p}}}$ that $$\begin{aligned}
\label{Ann_ap23}
\bigg(\int \Big\langl\Big(\fint_{\Boule_2(x)}|h|^q\Big)^\frac{r}{q}\Big\rangl^\frac{p}{r}\dd x\bigg)^\frac{1}{p}
\le N\|h\|_{p,r,q}\lesssim \|h\|_{p,r,q}.
\end{aligned}$$ By definitions of $\bar h_1$ and of $\bar h_2$, the latter in conjunction with , turns into (\[Ann\_ap39\]).
Inequalities are obvious from Jensen’s inequality in probability.
We finally turn to , which is a consequence of Jensen’s inequality: Rewriting definition in form of $$\begin{aligned}
\|h\|_{p,r,q}&=\Big(\int\big\langle\big(\fint_{\Boule_1}|h(x+z)|^qdz\big)^\frac{r}{q}
\big\rangle^\frac{p}{r}dx\Big)^\frac{1}{p}
=\Big\|\big\|\fint_{\Boule_1}|h(\cdot+z)|^qdz\big\|_{\LL^\frac{r}{q}_{\langle\cdot\rangle}}
\Big\|_{\LL^\frac{p}{q}_{\mathbb{R}^d}}^\frac{1}{q},
\end{aligned}$$ we learn from the convexity and translation invariance of the involved norms (here we use $q\le\min\{p,r\}$) that $$\begin{aligned}
\|h\|_{p,r,q}&\le
\Big(\fint_{\Boule_1}\big\|\||h(\cdot+z)|^q\|_{\LL^\frac{r}{q}_{\langle\cdot\rangle}}
\big\|_{\LL^\frac{p}{q}_{\mathbb{R}^d}}dz\Big)^\frac{1}{q}
\stackrel{\eqref{Num:701}}{=}\|h\|_{p,r}.
\end{aligned}$$
In fact, in the next step, we need its local version $$\label{Ann_ap34}
\big\|\omega_T \big(\frac{u}{\sqrt{T}},\nabla u\big)\big\|_{2,2}
\lesssim_{d,\lambda} \big\|\omega_T \big(\frac{g}{\sqrt{T}},f\big)\big\|_{2,2},$$ where the weight is of exponential form $$\label{Ann_ap21}
\omega_T(x):=\exp(-\frac{|x|}{C\sqrt{T}})$$ for a constant $C=C(d,\lambda)$ fixed below.
Starting with , we test with $u$, use the uniform $\lambda$-ellipticity of $a$ and the Cauchy-Schwarz inequality, which gives $$\label{Energ}
\int \big|\big(\frac{u}{\sqrt{T}},\nabla u\big)\big|^2
\lesssim
\int \big|\big(\frac{g}{\sqrt{T}},f\big)\big|^2.$$ We now take the expectation, which we exchange with the spatial integral, yielding in the form of $$\begin{aligned}
\label{Ann_ap97}
\big\|\big(\frac{u}{\sqrt{T}},\nabla u\big)\big\|_{2,2}
\lesssim\big\|\big(\frac{g}{\sqrt{T}},f\big)\big\|_{2,2}.
\end{aligned}$$
We now upgrade to : Multiplying (\[Ann\_ae1\]) with a cut-off function $\omega$ we obtain by Leibniz’ rule $$\begin{aligned}
\frac{1}{T}\omega u-\nabla\cdot a\nabla (\omega u)=\frac{1}{T}\omega g-\nabla\omega\cdot(a\nabla u+f)+\nabla\cdot(\omega f-au\nabla\omega).
\end{aligned}$$ Appealing to yields $$\begin{aligned}
\big\|\omega\big(\frac{u}{\sqrt{T}},\nabla u\big)\big\|_{2,2}
\lesssim\big\|\omega\big(\frac{g}{\sqrt{T}},f\big)\big\|_{2,2}
+\sqrt{T}\|\nabla\omega\cdot(a\nabla u+f)\|_{2,2}
+\|u\nabla\omega\|_{2,2}
.
\end{aligned}$$ Specifying $\omega$ to be of the form of , for which $|\nabla\omega_T|\le\frac{1}{C\sqrt{T}}\omega_T$, this entails $$\begin{aligned}
&\big\|\omega_T\big(\frac{u}{\sqrt{T}},\nabla u\big)\big\|_{2,2}
\\
&\qquad \lesssim \frac{1+C}{C}\|\omega_T f\|_{2,2} +\frac{1}{\sqrt{T}}\|\omega_T g\|_{2,2}
+\frac{1}{C\sqrt{T}}\|\omega_T u\|_{2,2}+\frac{1}{C}\|\omega_T\nabla u\|_{2,2}.
\end{aligned}$$ For $C \gg_{d,\lambda} 1$, the two last r. h. s. terms may be absorbed, giving rise to .
We claim that, for any $s'>2$ and $q \in (1,2]$, there exists a constant $C \gg_{d,\lambda,s',q} 1$ such that, defining $\omega_T$ by , there holds $$\label{Num:702}
\big\|\omega_T \big(\frac{u}{\sqrt{T}},\nabla u\big)\big\|_{2,2,q}
\lesssim_{\gamma,s',q} \big\|\omega_T \big(\frac{g}{\sqrt{T}},f\big)\big\|_{2,s',q}.$$
The argument relies on complex interpolation. We choose a $q' \in (1,q)$ and then define $\theta \in (0,1)$ and $s \in (s',\infty)$ through $$\frac{1}{q}= \frac{\theta}{2}+\frac{1-\theta}{q'} \quad \et \quad \frac{1}{s'}=\frac{\theta}{2}+\frac{1-\theta}{s}.$$ We make use of the estimate of Lemma \[Lem\_AnnealedCZ\_2\](ii) (replacing $(r,q) \rightsquigarrow (s,q')$), which we copy here: $$\label{Num:116_bis}
\big\|\big(\frac{u}{\sqrt{T}},\nabla u\big)\big\|_{2,2,q'} \lesssim \big\|\big(\frac{u}{\sqrt{T}},\nabla u\big)\big\|_{2,s,q'}.$$ Appealing to , we may identify the norms $\|\cdot\|_{2,2}$ and $\|\cdot\|_{2,2,2}$ so that reads $$\label{Ann_ap34_bis}
\big\|\omega_T \big(\frac{u}{\sqrt{T}},\nabla u\big)\big\|_{2,2,2}
\lesssim \big\|\omega_T \big(\frac{g}{\sqrt{T}},f\big)\big\|_{2,2,2},$$ By complex interpolation between and (using the Stein-Weiss theorem [@BerghLofstrom Th. 5.4.1 p. 115]), we get $$\big\|\omega_T^\theta \big(\frac{u}{\sqrt{T}},\nabla u\big)\big\|_{2,2,q}
\lesssim_{\gamma,s',q} \big\|\omega_T^\theta \big(\frac{g}{\sqrt{T}},f\big)\big\|_{2,s',q}.$$ Since $\omega_T^\theta(x)=\exp(-\frac{\theta|x|}{C\sqrt{T}})$ by , we obtain by adapting the definition of $C$.
We now are given $p \in (1,\infty)$, $2< r'<r$, and $q \leq \min\{p,2\} \in (1,2]$, and establish $$\label{Num:703}
\big\|\big(\frac{u}{\sqrt{T}},\nabla u\big)\big\|_{p,r',q}
\lesssim_{\gamma,p,r',r,q} \sqrt{T}^d \big\|\big(\frac{g}{\sqrt{T}},f\big)\big\|_{p,r,q}.$$
The two ingredients for are the following norm relations, to be established below: $$\begin{aligned}
&\sup_{x} \Big\langl \Big(\fint_{\Boule_1(x)} |h|^q \Big)^{\frac{2}{q}} \Big\rangl^{\frac{1}{2}} =: \|h\|_{\infty,2,q}
\lesssim\|h\|_{2,2,q} \quad\mbox{and}
\label{Ann_ap35}
\\
&
\left\{
\begin{aligned}
&\|\omega_T h\|_{2,r,q}\lesssim\sqrt{T}^{\max\{0,d(\frac{1}{2}-\frac{1}{p})\}}\|\omega_{4T} h\|_{p,r,q},
\\
&\|\omega_{\frac{T}{4}}h\|_{p,r',q}\lesssim\sqrt{T}^\frac{d}{p}\|\omega_{T} h\|_{\infty,r',q}.
\end{aligned}
\right.
\label{Ann_ap38}
\end{aligned}$$ The merit of passing to the spatial exponent $p=\infty$ is that by duality, $$\begin{aligned}
\|h\|_{\infty,r',q}
=~&\sup_{\langl|F|^\frac{2r'}{r'-2}\rangl\le 1}\|Fh\|_{\infty,2,q}.
\label{Ann_ap36}
\end{aligned}$$ Estimate is immediate from . In the case of $p\ge 2$, the first estimate in follows from appealing to the relation $\omega_T=\omega_{4T}^2$, *cf.* , which allows to use the Hölder inequality, so that it reduces to the obvious $\|\omega_{4T}\|_{\frac{2p}{p-2},\infty,\infty}$ ${\lesssim\sqrt{T}^{d(\frac{1}{2}-\frac{1}{p})}}$ (recall $T \geq 1$). In the case of $p\le 2$, we appeal to and the obvious $\omega_T$ $\le\omega_{4T}$. For the second estimate in , we start from $\omega_{T/4}=\omega_T^2$, use the Hölder inequality, and $\|\omega_{T}\|_{p,\infty,\infty} \lesssim\sqrt{T}^{\frac{d}{p}}$.
Turning now to , we define the exponent $s'>2$ by $$\label{Num:751}
\frac{1}{s'}:=\frac{1}{2}-\frac{1}{r'}+\frac{1}{r}=\frac{r'-2}{2r'}+\frac{1}{r}.$$ We specify the constant $C$ in to be the one of Step 3 belonging to $q$ and $s'$. Let $F$ be an auxiliary random variable. Since it does not depend on space, is preserved by multiplication with $F$. Hence by followed by the Hölder inequality in probability (based on the definition of $s'$) we obtain $$\begin{aligned}
\big\|\omega_{T}F \big(\frac{u}{T},\nabla u\big)\big\|_{2,2,q}
&\lesssim
\|\omega_T F \big(\frac{g}{\sqrt{T}},f\big)\|_{2,s',q}
\le
\big\|\omega_T \big(\frac{g}{\sqrt{T}},f\big)\big\|_{2,r,q}
\langl|F|^{\frac{2r'}{r'-2}}\rangl^\frac{r'-2}{2r'}.
\end{aligned}$$ In combination with , this yields $$\begin{aligned}
\big\|\omega_{T}F \big(\frac{u}{\sqrt{T}},\nabla u\big)\big\|_{\infty,2,q}
\lesssim
\big\|\omega_T \big(\frac{g}{\sqrt{T}},f\big)\big\|_{2,r,q}
\langl|F|^{\frac{2r'}{r'-2}}\rangl^\frac{r'-2}{2r'},
\end{aligned}$$ so that by $$\begin{aligned}
\big\|\omega_{T}\big(\frac{u}{\sqrt{T}},\nabla u\big)\big\|_{\infty,r',q}\lesssim
\big\|\omega_T \big(\frac{g}{\sqrt{T}},f\big)\big\|_{2,r,q},
\end{aligned}$$ which by implies, using $\max\{0,d(\frac{1}{2}-\frac{1}{p})\}+\frac{d}{p} \leq d$, $$\begin{aligned}
\|\omega_{\frac{T}{4}}\big(\frac{u}{\sqrt{T}},\nabla u\big)\|_{p,r',q}
\lesssim\sqrt{T}^{d}
\|\omega_{4T} \big(\frac{g}{\sqrt{T}},f\big)\|_{p,r,q}.
\end{aligned}$$ Since by the definition and the properties of $\omega_T$ we have $$\begin{aligned}
\|\omega_T(\cdot-z)h\|_{p,r,q}\sim
\Big(\int\Big(\omega_T(x-z)\big\langle\big(\fint_{\Boule_1(x)}|h|^q\big)^\frac{r}{q}\big\rangle^\frac{1}{r}\Big)^pdx
\Big)^\frac{1}{p},
\end{aligned}$$ the desired follows from taking the $L^p$-norm in the shift $z$.
Let $p \in (1,\infty)$. By duality and Jensen’s inequality, it suffices to establish for $2 \leq r' < r < \infty$. We set $q:=\min\{p,r\}$ and select an $r'<s<r$. Appealing to Lemma \[Lem\_AnnealedCZ\_2\](i) we have $$\begin{aligned}
\|(\frac{u}{\sqrt{T}},\nabla u)\|_{p,r'}\lesssim\|(\frac{u}{\sqrt{T}},\nabla u)\|_{p,s,q}
+\|(\frac{g}{\sqrt{T}},f)\|_{p,s}.
\end{aligned}$$ Estimating the first and the second r. h. s. term by (with $r'$ replaced by $s$) and by , respectively, we obtain $$\begin{aligned}
\|(\frac{u}{\sqrt{T}},\nabla u)\|_{p,r'}\lesssim\sqrt{T}^d\|(\frac{g}{\sqrt{T}},f)\|_{p,r,q}
+\|(\frac{g}{\sqrt{T}},f)\|_{p,r}.
\end{aligned}$$ By definition of $q$, we may use on the first r. h. s. term and so obtain , recalling that $T\ge 1$.
Proof of the non-perturbative CZ estimates {#Sec:ProofPropMassCZ}
------------------------------------------
This section contains the proofs of Propositions \[PropmassCZ\] and \[PropCZ\](ii).
As announced in Section \[Sec:Strat\], the proof is divided into four parts.
This part is at the core of our argument. Let $1 \leq \tau \leq T$. We recall the splitting $(u,\nabla u)=(v_>,h_>)+(v_<,h_<)$ from . Then, for exponents $p, r \in (1,\infty)$, $$\begin{aligned}
\label{Ann_orderExp}
&1 \leq s'''' < s''' < s'' < s' <r, \et \quad 1\leq r''<r'<r,
\end{aligned}$$ we claim that the following estimates hold: $$\begin{aligned}
\label{Ann_ae141}
\big\|\big(\frac{v_>}{\sqrt{T}}, h_>\big)\big\|_{p,r''}
&\lesssim \barC_{p,r',r}(\tau)
\big\|\big(\frac{g}{\sqrt{T}} ,f\big)\big\|_{p,r},
\\
\big\|\big( \frac{v_<}{\sqrt{T}}, h_<\big)\big\|_{p,s''''}
&\lesssim \barC_{p,s''',s''}(\tau) \sqrt{\tau}^{-\frac{1}{2}} \barC_{p,s',r}(T) \big\|\big(\frac{g}{\sqrt{T}} ,f\big)\big\|_{p,r}.
\label{Ann_ae17}
\end{aligned}$$
We note that if all the stochastic exponents in and were equal to $r$, then we could deduce from and that $$\begin{aligned}
\big\|\big( \frac{u}{\sqrt{T}},\nabla u\big)\big\|_{p,r} \lesssim \big(\barC_{p,r,r}(\tau) +\barC_{p,r,r}(\tau) \sqrt{\tau}^{-\frac{1}{2}} \barC_{p,r,r}(T)\big) \big\|\big(\frac{g}{\sqrt{T}} ,f\big)\big\|_{p,r},
\end{aligned}$$ which amounts to $$\begin{aligned}
\label{Ann_ee17}
\barC_{p,r,r}(T) \lesssim \barC_{p,r,r}(\tau) +\barC_{p,r,r}(\tau) \sqrt{\tau}^{-\frac{1}{2}} \barC_{p,r,r}(T).
\end{aligned}$$ Thus, if we would know that $\barC_{p,r,r}(\tau) \lesssim \sqrt{\tau}^{\frac{1}{2} \theta}$ for some $\theta<1$, then we would obtain from the above estimate that $\barC_{p,r,r}(T)$ is uniformly bounded in $T \geq 1$. (This motivates assumption of Part 2.) However, the stochastic exponents have to be strictly ordered as in . Therefore, in Part 2, we resort to real interpolation to increase the stochastic exponent of $\nabla u$, and we buckle with an estimate similar to .
We first estimate the high-pass part $(v_>,h_>)$. The right-hand sides of and may be expressed in a more convenient form thanks to , , and , namely $$\begin{aligned}
\frac{1}{T} \ubar_> - \nabla \cdot \abar_\tau \nabla \ubar_>
&=\big(1-\frac{\tau}{T} \big) \big( \nabla \cdot ( a \nabla u_> + f) + \frac{1}{T}g\big),
\label{Ann_ae10}
\\
\label{Ann_ae11}
\frac{1}{T} \ubar_< - \nabla \cdot \abar_\tau \nabla \ubar_<
&=\big(1-\frac{\tau}{T} \big)
\nabla \cdot\big(a \nabla w +( \phi_{\tau,i} a -\sigma_{\tau,i}) \nabla \partial_i \ubar +\frac{1}{\tau} \partial_i \ubar \psi_{\tau,i} \big).
\end{aligned}$$ By equation (recall that $\tau \leq T$), we obtain $$\label{Num:902}
\big\|\big(\frac{u_>}{\sqrt{\tau}},\nabla u_>\big) \big\|_{p,r'}
\overset{\eqref{Ann_ee5}}{\leq}
\barC_{p,r',r}(\tau) \|(\frac{\sqrt{\tau} g}{T} ,f)\|_{p,r}
\leq
\barC_{p,r',r}(\tau) \big\|\big(\frac{g}{\sqrt{T}} ,f\big)\big\|_{p,r},$$ which by implies $$\big\|\big(\frac{u_>}{\sqrt{T}},\nabla u_>\big)\big\|_{p,r''}
\lesssim
\barC_{p,r',r}(\tau) \big\|\big(\frac{g}{\sqrt{T}} ,f\big)\big\|_{p,r}.
\label{Ann_ae13}$$ By estimate on the massive correctors and recalling , we have $$\begin{aligned}
\big\|\big(\frac{\ubar_>}{\sqrt{T}}, \partial_i \ubar_> (e_i + \nabla \phi_{\tau,i})\big)\big\|_{p,r''}
&\overset{\eqref{Ann_ap56}}{\lesssim}
\big\|\big(\frac{\ubar_>}{\sqrt{T}},\nabla \ubar_>\big)\big\|_{p,r'}
\\
&\overset{\eqref{Ann_eee23}}{\lesssim}
\big\|\big(\frac{g}{\sqrt{T}},f,\nabla u_>\big)\big\|_{p,r'}
\\
&\overset{\eqref{Num:902}}{\lesssim} \barC_{p,r',r}(\tau) \big\|\big(\frac{g}{\sqrt{T}} ,f\big)\big\|_{p,r}.
\end{aligned}
\label{Ann_ae14}$$ In view of , estimates and combine to .
We now turn to the low-pass contribution $(v_<,h_<)$. We start with its first constituent $w$ (*cf.* ). Appealing to equation and the correctors estimate , we obtain (recall that $\tau \leq T$) $$\begin{aligned}
\big\|\big(\frac{w}{\sqrt{T}},\nabla w\big)\big\|_{p,s'''}
\leq~~& \|(\frac{w}{\sqrt{\tau}},\nabla w)\|_{p,s'''}
\\
\overset{\eqref{Ann_ee5}}{\lesssim}~&
\barC_{p,s''',s''}(\tau)
\Big\|\Big( \big|(\frac{\psi_{\tau}}{\sqrt{\tau}},\phi_\tau)\big| \frac{\nabla \ubar }{\sqrt{\tau}},|(\phi_{\tau},\sigma_{\tau})|\nabla^2 \ubar \Big)\Big\|_{p,s''}
\\
\overset{\eqref{Ann_ap57}}{\lesssim}~& \barC_{p,s''',s''}(\tau) \mu_d(\sqrt{\tau})
\big\|\big(\frac{\nabla \ubar }{\sqrt{\tau}},\nabla^2 \ubar \big)\big\|_{p,s'}.
\end{aligned}
\label{Ann_ae15}$$ Turning to the second constituent $\ubar_<$ of $(v_<,h_<)$, we obtain once more by : $$\| \partial_i \ubar_<(e_i +\nabla \phi_{\tau,i}) \|_{p,s''''}
\lesssim \| \nabla \ubar_<\|_{p,s'''}.$$ Therefore, recalling equation and the correctors estimate , we thus have $$\begin{aligned}
\begin{aligned}
\big\|\big(\frac{\ubar_<}{\sqrt{T}}, \partial_i \ubar_< (e_i + \nabla \phi_{\tau,i})\big)\big\|_{p,s''''}
&\overset{\eqref{Ann_eee23}}{\lesssim} \Big\| \big(\nabla w ,\frac{|\psi_{\tau}|}{\sqrt{\tau}}\frac{\nabla \ubar }{\sqrt{\tau}},|( \phi_{\tau},\sigma_{\tau})| \nabla^2 \ubar\big) \Big\|_{p,s'''}
\\
&\overset{\eqref{Ann_ap57}}{\lesssim} \| \nabla w \|_{p,s'''} +\mu_d(\sqrt{\tau}) \big\|\big(\frac{\nabla \ubar }{\sqrt{\tau}},\nabla^2 \ubar \big)\big\|_{p,s''}
\\
&\overset{\eqref{Ann_ae15}}{\lesssim} \barC_{p,s''',s''}(\tau) \mu_d(\sqrt{\tau})
\big\|\big(\frac{\nabla \ubar }{\sqrt{\tau}},\nabla^2 \ubar \big)\big\|_{p,s'}.
\end{aligned}
\label{Ann_ae16}
\end{aligned}$$ Last, the third constituent of $(v_<,h_<)$ is estimated by appealing to the correctors estimate : $$\label{Ann_ae16_1}
\big\|\big(\frac{\phi_{\tau,i} \partial_i \ubar }{\sqrt{T}},\phi_{\tau,i} \nabla \partial_i \ubar \big)\big\|_{p,s''''}
\overset{\tau \leq T}{\lesssim} \mu_d(\sqrt{\tau}) \big\|\big(\frac{\nabla \ubar }{\sqrt{\tau}},\nabla^2 \ubar \big)\big\|_{p,s'}.$$ By definition of $(v_<,h_<)$, gathering , , and yields $$\begin{aligned}
\big\|\big(\frac{v_<}{\sqrt{T}},h_<\big)\big\|_{p,s''''}
&\lesssim \barC_{p,s''',s''}(\tau) \mu_d(\sqrt{\tau})
\big\|\big(\frac{\nabla \ubar }{\sqrt{\tau}},\nabla^2 \ubar \big)\big\|_{p,s'}
\\
&\lesssim \barC_{p,s''',s''}(\tau) \frac{\mu_d(\sqrt{\tau})}{\sqrt{\tau}} \| \nabla u\|_{p,s'},
\end{aligned}$$ where we used the annealed estimate for the massive equation with constant coefficients in its differentiated form: $$\frac{1}{\tau} \partial_i \ubar - \nabla \cdot \abar_\tau \nabla \partial_i \ubar = \big(\frac{1}{\tau}-\frac{1}{T}\big)\partial_iu.$$ Finally, using and recalling establishes .
Let $p$, $r$, $s''''$, $s'''$, $s''$, $s'$, $r''$, $r'$ be as in Part 1 (*i.e.* satisfying ). Assume that there exists $\theta \in (0,1)$ such that $$\begin{aligned}
\label{Ann_ap43}
\frac{1}{s'}\;>\frac{1}{s}:=\;\theta\frac{1}{s''''}+(1-\theta)\frac{1}{r''}.
\end{aligned}$$ (See Figure \[Figure\_indices\].) Suppose that we sub-optimally control the massive CZ constants both for high $(r',r)$ and low $(s''',s'')$ stochastic integrability, in the sense of $$\begin{aligned}
\barC_{p,r',r}(\tau)
+\barC_{p,s''',s''}(\tau)
\le\Lambda\sqrt{\tau}^{\frac{1}{2}\theta}&\pourtout \tau\ge 1
\label{Ann_ee7}
\end{aligned}$$ for some fixed constant $\Lambda\ge 1$. We claim the following control of the CZ constant: $$\begin{aligned}
\label{Ann_ee8}
\barC_{p,s',r}(T)\lesssim \Lambda^{\frac{1}{1-\theta}} &\pourtout T \geq 1.
\end{aligned}$$
; ; ;
(1S) at (0,); (1T) at (0,); (1R) at (0,); (2S) at (10,); (2T) at (10,); (2R) at (10,); (S4) at (1.5,); (S3) at (2.25,); (S2) at (3,); (S1) at (4,); (S0) at (9,); (T1) at (4,); (T2) at (5.075,); (T0) at (9,); (R2) at (7,); (R1) at (8,); (R0) at (9,);
(1S)–(2S); (1T)–(2T); (1R)–(2R);
(1S) node [$\bullet$]{}; (1S) node \[below\] [$1$]{}; (1T) node [$\bullet$]{}; (1T) node \[below\] [$1$]{}; (1R) node [$\bullet$]{}; (1R) node \[below\] [$1$]{};
(S4) node [$\bullet$]{}; (S4) node \[below\] [$s''''$]{}; (S3) node [$\bullet$]{}; (S3) node \[below\] [$s'''$]{}; (S2) node [$\bullet$]{}; (S2) node \[below\] [$s''$]{}; (S1) node [$\bullet$]{}; (S1) node \[below\] [$s'$]{}; (S0) node [$\bullet$]{}; (S0) node \[below\] [$r$]{};
(R2) node [$\bullet$]{}; (R2) node \[below right\] [$r''$]{}; (R1) node [$\bullet$]{}; (R1) node \[below\] [$r'$]{}; (R0) node [$\bullet$]{}; (R0) node \[below\] [$r$]{};
(T2) node [$\bullet$]{}; (T2) node \[below right\] [$s$]{}; (T1) node [$\bullet$]{}; (T1) node \[below left\] [$s'$]{}; (T0) node [$\bullet$]{}; (T0) node \[below\] [$r$]{};
(R2)–(S4);
The form of the result arises from a real interpolation argument. We make use of the $K$-method [@BerghLofstrom Chap. 3, p. 38], which relies on the one-parameter family of splittings of $(u,\nabla u)$ studied in the Part 1 into a part $(v_>,h_>)$ with good stochastic integrability (slightly worse than $r'$) and a part $(v_<,h_<)$ with a bad one (slightly worse than $s'''$). Here the length scale $\sqrt{\tau}$ plays the role of the parameter.
We now fix $T\geq 1$. By the $K$-method, in view of , we have $$\label{Ann_interpolv}
\begin{aligned}
&\big\|\big(\frac{u}{\sqrt{T}},\nabla u \big)\big\|_{p,s'}
\\
&\quad \lesssim \sup_{\mu>0}
\inf_{\scriptsize{\begin{array}{c} v_>+v_<=u \\ h_>+h_<=\nabla u \end{array}}}
\Big\{\mu^{-\theta} \big\|\big(\frac{v_>}{\sqrt{T}},h_> \big)\big\|_{p,r''} + \mu^{1-\theta} \big\|\big(\frac{v_<}{\sqrt{T}},h_< \big)\big\|_{p,s''''} \Big\},
\end{aligned}$$ (indeed, by [@BerghLofstrom Th. 5.1.2 & Th. 5.2.1] the above r. h. s. corresponds to a norm on the functional space $\LL^p_{\R^d}(\LL^{s,\infty}_{\langl\cdot\rangl})$, and the Lorentz space $\LL^{s,\infty}_{\langl\cdot\rangl}$ is included in $\LL_{\langl\cdot\rangl}^{s'}$). Next, we will show in Part 2, Step 2 that for any $\mu>0$ and $T \geq 1$ there is a splitting $(u,\nabla u)=(v_>,h_>)+(v_<,h_<)$ such that $$\label{Ann_anycase}
\begin{aligned}
\mu^{-\theta} \big\|\big(\frac{v_>}{\sqrt{T}},h_>\big)\big\|_{p,r''} + \mu^{1-\theta} \big\|\big(\frac{v_<}{\sqrt{T}},h_<\big)\big\|_{p,s''''}
\lesssim \Lambda \barC_{p,s',r}^\theta(T) \big\|\big(\frac{g}{\sqrt{T}},f\big)\big\|_{p,r}.
\end{aligned}$$ Then, inserting into entails $$\barC_{p,s',r}(T) \lesssim \Lambda \barC^{\theta}_{p,s',r}(T),$$ and since $\barC_{p,s',r}(T)$ is finite by Lemma \[Lem\_SsOpti\], we obtain .
For given $\mu>0$, we define $\tau$ by $$\begin{aligned}
\mu=\frac{\sqrt{\tau}^{\frac{1}{2}}}{\barC_{p,s',r}(T)}.
\end{aligned}$$ Thus we have to estimate $$\begin{aligned}
&\mu^{-\theta} \big\|\big(\frac{v_>}{\sqrt{T}},h_>\big)\big\|_{p,r''} + \mu^{1-\theta} \big\|\big(\frac{v_<}{\sqrt{T}},h_<\big)\big\|_{p,s''''}
\\
&\quad= \barC_{p,s',r}^\theta(T) \Big( \sqrt{\tau}^{-\frac{1}{2}\theta}\big\|\big(\frac{v_>}{\sqrt{T}},h_>\big)\big\|_{p,r''} + \frac{\sqrt{\tau}^{\frac{1}{2}(1-\theta)}}{\barC_{p,s',r}(T)} \big\|\big(\frac{v_<}{\sqrt{T}},h_<\big)\big\|_{p,s''''}\Big),
\end{aligned}
\label{Ann_ee18}$$ and distinguish the three cases: either $\tau\leq 1$, $1 \leq \tau \leq T$, or $T \leq \tau$.
If $\tau \leq 1$, we set $(v_>,h_>) = (0,0)$ and $(v_<,h_<)=(u,\nabla u)$ so that by equation $$\begin{aligned}
\big\|\big(\frac{v_<}{\sqrt{T}},h_<\big)\big\|_{p,s''''}
\overset{\eqref{Num:03_bis}}{\leq} \big\|\big(\frac{u}{\sqrt{T}},\nabla u\big) \big\|_{p,s'}
\overset{\eqref{Ann_ee5}}{\lesssim} \barC_{p,s',r}(T) \big\|\big(\frac{g}{\sqrt{T}},f\big)\big\|_{p,r}.
\end{aligned}$$ Inserting this estimate into establishes since $\tau\leq 1 \leq \Lambda$.
In the generic case $1 \leq \tau \leq T$, we define $(v_>,h_>)$ and $(v_<,h_<)$ by . Therefore, and combined with our assumption yield $$\left\{
\begin{aligned}
\big\|\big(\frac{v_>}{\sqrt{T}},h_>\big)\big\|_{p,r''}
&\lesssim
\Lambda \sqrt{\tau}^{\frac{1}{2}\theta}\big\|\big(\frac{g}{\sqrt{T}} ,f\big)\big\|_{p,r},
\\
\big\|\big(\frac{v_<}{\sqrt{T}},h_<\big)\big\|_{p,s''''}
&\lesssim
\Lambda\sqrt{\tau}^{\frac{1}{2}(\theta-1)} \barC_{p,s',r}(T) \big\|\big(\frac{g}{\sqrt{T}} ,f\big)\big\|_{p,r}.
\end{aligned}
\right.$$ Inserting this into entails .
Finally, if $\tau \geq T$ we set $(v_>,h_>)=(u,\nabla u)$ and $(v_<,h_<)=(0,0)$ to the effect of $$\begin{aligned}
\big\|\big(\frac{v_>}{\sqrt{T}},h_>\big)\big\|_{p,r''} \leq \|\big(\frac{u}{\sqrt{T}},\nabla u\big)\|_{p,r'} \lesssim \barC_{p,r',r}(T) \|\big(\frac{g}{\sqrt{T}},f\big)\|_{p,r}.
\end{aligned}$$ Then, invoking (with $\tau$ replaced by $T$) and recalling that $T \leq \tau$, we obtain $$\begin{aligned}
\big\|\big(\frac{v_>}{\sqrt{T}},h_>\big)\big\|_{p,r''}
\lesssim \Lambda \sqrt{T}^{\frac{1}{2} \theta} \|\big(\frac{g}{\sqrt{T}},f\big)\|_{p,r}
\leq \Lambda \sqrt{\tau}^{\frac{1}{2} \theta} \|\big(\frac{g}{\sqrt{T}},f\big)\|_{p,r}.
\end{aligned}$$ Inserting this into and yields also in this case.
We consider two triplets of exponents $(p,r',r)$ and $(q,s',s) \in (1,\infty)\times [2,\infty)^2$ that satisfy $$\begin{aligned}
\label{Ann_ap94_ter}
s'\le r',\quad s\le r,\quad \et \quad
\frac{1}{r'}-\frac{1}{r}>\frac{1}{s'}-\frac{1}{s}>0.
\end{aligned}$$ For any $\theta\in(0,1]$ such that $$\begin{aligned}
\label{Ann_ap94_quad}
\theta>\max\Big\{1-\frac{s}{r},1-\frac{s'}{r'},1-\frac{q}{p},
\frac{\frac{1}{p}-\frac{1}{q}}{1-\frac{1}{q}}\Big\}
\end{aligned}$$ we claim that $$\begin{aligned}
\label{Ann_ap90bis}
\barC_{p,r',r}(\tau)\lesssim \barC_{q,s',s}^{1-\theta}(\tau)\sqrt{\tau}^{d\theta}
\pourtout \tau\ge 1.
\end{aligned}$$ Note that the r. h. s. of is strictly less than $1$ so that such a $\theta$ always exists.
Here comes the argument: We first note that thanks to the constraint on $\theta$, the identities $$\begin{aligned}
\label{fs01}
\frac{1}{r}=(1-\theta)\frac{1}{s}+\theta\frac{1}{\tilde r},\quad
\frac{1}{r'}=(1-\theta)\frac{1}{s'}+\theta\frac{1}{\tilde r'},\quad
\frac{1}{p}=(1-\theta)\frac{1}{q}+\theta\frac{1}{\tilde p}
\end{aligned}$$ define a triple of exponents $(\tilde p,\tilde r',\tilde r)\in(1,\infty)\times[2,\infty)^2$. Indeed, in case of $\tilde r$, we rewrite the implicit as the explicit $\frac{1}{\tilde r}$ $=\frac{1}{\theta}(\frac{1}{r}-(1-\theta)\frac{1}{s})$, and note that this expression is positive thanks to the first constraint on $\theta$ in , rewritten as $1-\theta<\frac{s}{r}$. We also have $\frac{1}{\theta}(\frac{1}{r}-(1-\theta)\frac{1}{s})$ $\le\frac{1}{2}$, or equivalently, $\frac{1}{s}-\frac{1}{r}$ $\ge\theta(\frac{1}{s}-\frac{1}{2})$ because of our assumption $2\le s\le r$, see . The case of $\tilde r'$ is treated analogously. Finally, in the case of $\tilde p$, the lower bound $\frac{1}{\tilde p}$ $=\frac{1}{\theta}(\frac{1}{p}-(1-\theta)\frac{1}{q})>0$ follows as for the two others. The upper bound $\frac{1}{\theta}(\frac{1}{p}-(1-\theta)\frac{1}{q})$ $<1$, which is equivalent to $\frac{1}{p}-\frac{1}{q}<\theta(1-\frac{1}{q})$, follows from the last of the four constraints in .
Note that by the third item of we have $\tilde{r}'< \tilde{r}$, so that by Lemma \[Lem\_SsOpti\] there holds $\barC_{\tilde{p},\tilde{r}',\tilde{r}}(\tau) \lesssim \sqrt{\tau}^d$. On the other hand, by complex interpolation, there holds $\barC_{p,r',r}(\tau)$ $\le \barC_{q,s',s}^{1-\theta}(\tau)$ $ \barC_{\bar p,\tilde{r}',\tilde{r}}^{\theta}(\tau)$. This yields .
;
;
(A0) at (0\*[1]{},[0]{}); (A1) at (10\*[1]{},[0]{}+0); (A2) at (1\*[1]{},[0]{}+0); (A3) at (4\*[1]{},[0]{}+0); (A4) at (6\*[1]{},[0]{}+0);
(A11) at (2.5\*[1]{},[0]{}-0.5); (A12) at (5\*[1]{},[0]{}-0.5);
(A21) at (1\*[1]{},[0]{}-0.25); (A22) at (4\*[1]{},[0]{}-0.25); (A23) at (6\*[1]{},[0]{}-0.25);
(A0)–(A1); in [0,1,...,4]{} [ (A) node [$\bullet$]{}; ]{} (A0) node\[above\] [$0$]{}; (A1) node\[above\] [$1$]{}; (A2) node\[above\] [$\frac{1}{q}$]{}; (A3) node\[above\] [$\frac{1}{p}$]{}; (A4) node\[above\] [$\frac{1}{\tilde{p}}$]{}; (A11) node [$\theta$]{}; (A12) node [$1-\theta$]{};
(A21)–(A22); (A22)–(A23);
;
;
(A0) at (0\*[0.5]{},[-1.5]{}); (A1) at (10\*[0.5]{},[-1.5]{}+0); (A2) at (4\*[0.5]{},[-1.5]{}+0); (A3) at (6\*[0.5]{},[-1.5]{}+0); (A4) at (9\*[0.5]{},[-1.5]{}+0);
(A11) at (5\*[0.5]{},[-1.5]{}-0.5); (A12) at (7.5\*[0.5]{},[-1.5]{}-0.5);
(A21) at (4\*[0.5]{},[-1.5]{}-0.25); (A22) at (6\*[0.5]{},[-1.5]{}-0.25); (A23) at (9\*[0.5]{},[-1.5]{}-0.25);
(A0)–(A1); in [0,1,...,4]{} [ (A) node [$\bullet$]{}; ]{} (A0) node\[above\] [$0$]{}; (A1) node\[above\] [$\frac{1}{2}$]{}; (A2) node\[above\] [$\frac{1}{\tilde{r}'}$]{}; (A3) node\[above\] [$\frac{1}{r'}$]{}; (A4) node\[above\] [$\frac{1}{s'}$]{}; (A11) node [$1-\theta$]{}; (A12) node [$\theta$]{};
(A21)–(A22); (A22)–(A23);
;
;
(A0) at (0\*[0.5]{},[-3.]{}); (A1) at (10\*[0.5]{},[-3.]{}+0); (A2) at (3\*[0.5]{},[-3.]{}+0); (A3) at (5\*[0.5]{},[-3.]{}+0); (A4) at (8\*[0.5]{},[-3.]{}+0);
(A11) at (4\*[0.5]{},[-3.]{}-0.5); (A12) at (6.5\*[0.5]{},[-3.]{}-0.5);
(A21) at (3\*[0.5]{},[-3.]{}-0.25); (A22) at (5\*[0.5]{},[-3.]{}-0.25); (A23) at (8\*[0.5]{},[-3.]{}-0.25);
(A0)–(A1); in [0,1,...,4]{} [ (A) node [$\bullet$]{}; ]{} (A0) node\[above\] [$0$]{}; (A1) node\[above\] [$\frac{1}{2}$]{}; (A2) node\[above\] [$\frac{1}{\tilde{r}}$]{}; (A3) node\[above\] [$\frac{1}{r}$]{}; (A4) node\[above\] [$\frac{1}{s}$]{}; (A11) node [$1-\theta$]{}; (A12) node [$\theta$]{};
(A21)–(A22); (A22)–(A23);
Let us start by rewriting in terms of $\theta\leadsto\frac{\theta}{2d}$ in the more iterable (but equivalent) form of $$\begin{aligned}
\label{fs02}
\begin{array}{c}
(1-\frac{\theta}{2d})\frac{2}{s}<\frac{2}{r},\quad
(1-\frac{\theta}{2d})\frac{2}{s'}<\frac{2}{r'},\quad
(1-\frac{\theta}{2d})\frac{2}{q}<\frac{2}{p},\\[1ex]
\mbox{and}\quad
(1-\frac{\theta}{2d})(2-\frac{2}{q})<2-\frac{2}{p}.
\end{array}
\end{aligned}$$ In a first step, we apply Part 3 with $(q,s',s)=(2,2,2)$, feeding in the pivotal estimate , so that assumes the form $$\begin{aligned}
\label{Ann_ap91}
\barC_{p,r',r}(\tau)\lesssim\sqrt{\tau}^{\frac{1}{2}\theta}\pourtout \tau\ge 1.
\end{aligned}$$ In view of , such a $\theta<1$ exists provided the triplet $(p,r',r)\in(1,\infty)\times[2,\infty)^2$ of exponents is sufficiently to $(2,2,2)$ in the sense of $$\begin{aligned}
\label{Ann_ap92}
2-\big(1-\frac{1}{2d}\big)
=1 + \frac{1}{2d}
>\frac{2}{p}>1-\frac{1}{2d} \et 1 \ge \frac{2}{r'}>\frac{2}{r} >1-\frac{1}{2d},
\end{aligned}$$ We take this as input yielding by the output $$\begin{aligned}
\label{Ann_ap93}
\barC_{p,r',r}(\tau)\lesssim 1\pourtout \tau\ge 1
\end{aligned}$$ in the same range (\[Ann\_ap92\]) of exponents.
In the second step, we feed (\[Ann\_ap93\]) with $(p,r',r)$ playing the role of $(q,s',s)$ into (\[Ann\_ap90bis\]). In view of , this yields and ultimately in the extended range of $$\begin{aligned}
2-\big(1-\frac{1}{2d}\big)^2>\frac{2}{p} >\big(1-\frac{1}{2d}\big)^2 \et 1\geq \frac{2}{r'}>\frac{2}{r} >\big(1-\frac{1}{2d}\big)^2.
\end{aligned}$$ Obviously, we may reach any arbitrary triplet $(p,r',r)$ of exponents in $(1,\infty) \times [2,\infty)^2$ with $r'<r$ by a finite number of such steps. By a duality argument and Jensen’s inequality we finally obtain for any triplet $(p,r',r)$ of exponents in $(1,\infty)^3$ with $1\leq r'<r$.
Given $f$, we consider the sequence of solutions $u_T$ to for $g=0$. By Proposition \[PropmassCZ\], such functions have their gradients uniformly bounded in $\LL^p_{\R^d}\big(\LL^{r'}_{\langl\cdot\rangl}\big)$. Therefore, as $T\uparrow\infty$, there exists $\nabla u_\infty \in \LL^p_{\R^d}\big(\LL^{r'}_{\langl\cdot\rangl}\big)$ such that, for a subsequence, there holds $$\begin{aligned}
\nabla u_T \rightharpoonup \nabla u_\infty \dans \LL^{p}_{\R^d}(\LL^{r'}_{\langl \cdot \rangl}).
\end{aligned}$$ By lower semi-continuity of the norm of $\LL^p_{\R^d}\big(\LL^{r'}_{\langl\cdot\rangl}\big)$, $\nabla u_\infty$ inherits from the estimates and satisfied by $\nabla u_T$. Moreover, by the energy estimate, *cf.* , there holds $$\begin{aligned}
\frac{1}{T} \int u_T^2 + \int |\nabla u_T|^2 \lesssim \int |f|^2,
\end{aligned}$$ which implies $$\begin{aligned}
\frac{1}{T} u_T \rightarrow 0 \dans \LL^2(\R^d).
\end{aligned}$$ As a consequence, the limit $\nabla u_\infty$ of the sequence $\nabla u_T$ is a weak solution to (since $\nabla u_\infty$ is square-integrable, it is the only solution).
Proof of the stochastic estimates of the massive correctors {#Sec_LemMassCorr}
-----------------------------------------------------------
The proof of Lemma \[LemMassCorr\] follows the same scheme as the proof of Proposition \[Propcorr\], with minor modifications due to the presence of the massive term. We introduce the abbreviation $$\begin{aligned}
\label{Num:710}
\Lambda_r:=\Big\langl \Big( \fint_{\Boule_1} \big| \big(\frac{\phi_T}{\sqrt{T}},\nabla \phi_T \big) \big|^2 \Big)^{r} \Big\rangl^{\frac{1}{r}}
\pour r \in [1,\infty).
\end{aligned}$$ In Step 1, we momentarily fix some deterministic function $g$ and vector field $f$ and we derive a representation formula for the Malliavin derivatives of the random variables $F$ and $F^\star$ defined by $$\label{Def:F_T}
F:=\int \big(-\frac{1}{T}g\phi_{T,j}+f\cdot\nabla\phi_{T,j}\big)
\et \quad F^\star:=\int f\cdot(q_{T,j}-\langl q_{T,j}\rangl).$$ (We henceforth omit the index $j$.) Using the spectral gap , we establish in Step 2 that these satisfy the following estimate: $$\begin{aligned}
\langl |(F,F^\star)|^{2r}\rangl^\frac{1}{r}
\lesssim \Lambda_r \int \big|\big(\frac{g}{\sqrt{T}}, f \big) \big|^2
\quad \text{provided}\; r \gg 1.
\label{Massive_Step1}
\end{aligned}$$ In Step 3, we argue that, if the stationary vector fields $u$, $f$, and $g$ are related through , then there exists $\theta=\theta(d,\lambda) >0$ such that, for all $R>0$, the following annealed Caccioppoli estimate holds $$\label{Num:730}
\Big\langl \Big( \fint_{\Boule_R} \big|\big( \frac{u}{\sqrt{T}},\nabla u \big)\big|^2 \Big)^r \Big\rangl
\lesssim
\Big\langl \big|\big( \frac{g}{\sqrt{T}},f \big)\big|^{2r} \Big \rangl
+ \Big\langl \Big| \fint_{\Boule_{\theta R}} \big( \frac{u}{\sqrt{T}},\nabla u \big) \Big|^{2r} \Big \rangl.$$ Then, in Step 4, we buckle on Step 2 and Step 3 (where $\phi_T$ plays the role of $u$) to obtain $$\begin{aligned}
\label{Massive_Step3}
\Lambda_r \lesssim 1,
\end{aligned}$$ which yields that $\nabla \phi_T$ satisfies by local regularity. In Step 5, we argue that and imply $$\begin{aligned}
\label{Ann_as14}
\Big\langl\Big|\int_{\Boule_1}\big(\phi_T,\sigma_T,\frac{\psi_T}{\sqrt{T}}\big)\Big|^{2r}\Big\rangl^\frac{1}{r}
&\lesssim \mu^2_d(\sqrt{T}),
\\
\label{Num:733}
\Big\langl\Big|\int_{\Boule_1}\big(\nabla \sigma_T,\frac{\nabla \psi_T}{\sqrt{T}}\big)\Big|^{2r}\Big\rangl^\frac{1}{r}
&\lesssim
1.
\end{aligned}$$ In Step 6, we show that also $\nabla \sigma_T$ and $\nabla \psi_T$ satisfy . Last, in Step 7, we argue that and imply .
The starting point is provided by the following representation formulas of the Malliavin derivatives: $$\begin{aligned}
\label{Id_1}
\frac{\partial F}{\partial a}&=\nabla v\otimes(e+\nabla\phi_T),
\\
\label{Id_2}
\frac{\partial F^\star}{\partial a}&=(f+\nabla v^\star)\otimes(e+\nabla\phi_T),
\end{aligned}$$ where $v$ and $v^\star$ are defined through $$\begin{aligned}
\frac{1}{T}v-\nabla\cdot a^\star\nabla v&=\frac{1}{T}g+\nabla\cdot f,
\label{Ann_as06}
\\
\frac{1}{T} v^\star-\nabla\cdot a^\star\nabla v^\star&= \nabla \cdot a^\star f.
\nonumber
\end{aligned}$$
Here comes the argument for ( is justified exactly the same way). Given a smooth and compactly supported infinitesimal variation $\delta a$ of $a$, we have for the generated infinitesimal variation of $F$ $$\begin{aligned}
\delta F
&\overset{\eqref{Def:F_T}}{=} \int \big(-\frac{1}{T}g \delta \phi_T +f\cdot\nabla \delta \phi_T\big)
\overset{\eqref{Ann_as06}}{=}-\int \frac{1}{T} v \delta \phi_T + \nabla v \cdot a \nabla \delta \phi_T.
\end{aligned}$$ Now, differentiating the definition of $\phi_T$ entails $$\begin{aligned}
\label{Diff_phiT}
\frac{1}{T} \delta \phi_{T}-\nabla\cdot a \nabla \delta \phi_{T}=\nabla \cdot \delta a (e+\nabla \phi_T),
\end{aligned}$$ which we test against $v$, obtaining in form of $$\begin{aligned}
\delta F=\int \nabla v \cdot \delta a (e+\nabla \phi_T).
\end{aligned}$$
We now derive from and , focusing on $F$, since the argument for $F^\star$ is similar. Since $\phi_T$ is stationary, so that $\nabla \phi_T$ and $\nabla\cdot a(e+\nabla \phi_T)$ have vanishing expectations and thus also $\frac{1}{T}\phi_T$ by , the random variable $F$ is of vanishing expectation. Therefore, by the spectral gap , invoking a duality argument, the Hölder inequality, the stationarity of $\nabla \phi_T$, and Jensen’s inequality, we obtain: $$\label{Estim11}
\begin{aligned}
\langl F^{2r}\rangl^{\frac{1}{r}}
\lesssim~& \sup_{\langl G^{2r^\star}\rangl \leq 1}
\Big\langl \int \Big( \fint_{\Boule_1(x)} | ( G\nabla v) \otimes(e+\nabla\phi_T)| \Big)^2 \dd x \Big\rangl
\\
\leq~& \sup_{\langl G^{2r^\star}\rangl \leq 1}
\int \Big\langl \Big( \fint_{\Boule_1(x)} | G\nabla v|^2 \Big)^{r^\star}
\Big\rangl^{\frac{1}{r^\star}}
\Big\langl \Big(\fint_{\Boule_1(x)} |e+\nabla\phi_T|^2\Big)^r \Big\rangl^{\frac{1}{r}} \dd x
\\
\overset{\eqref{Num:710}}{\leq}& (\Lambda_r+1) \sup_{\langl G^{2r^\star}\rangl \leq 1}
\int \Big\langl \Big( \fint_{\Boule_1(x)} | G\nabla v|^2 \Big)^{r^\star}
\Big\rangl^{\frac{1}{r^\star}} \dd x
\\
\overset{\eqref{Num:556}}{\leq}~& (\Lambda_r+1) \sup_{\langl G^{2r^\star}\rangl \leq 1}
\int \langl | G\nabla v|^{2r^\star} \rangl^{\frac{1}{r^\star}} \dd x.
\end{aligned}$$ Using annealed CZ estimates on or rather on $\frac{1}{T}Gv-\nabla\cdot a^\star\nabla Gv$ $=\frac{1}{T}Gg+\nabla\cdot Gf$ (the perturbative proof of which, based on , is similar to the proof of Proposition \[PropCZ\](i) relying on Lemma \[LemCstMassCZ\]), provided $r^\star-1 \ll 1$, we have $$\begin{aligned}
\int \langl|G\nabla v|^{2r^\star}\rangl^\frac{1}{r^\star}
\lesssim~& \int \Big\langl\big|G\big(\frac{g}{\sqrt{T}},f\big)\big|^{2r^\star}\Big\rangl^\frac{1}{r^\star}
= \langl G^{2r^\star} \rangl^{\frac{1}{r^\star}} \int \big|\big(\frac{g}{\sqrt{T}},f\big)\big|^{2}.
\end{aligned}$$ Inserting this into yields .
Let $R' \leq R$ (to be fixed later). We denote by $u_{R'}$ the mollification of $u$ on scale $R'$ by averaging on balls of radius $R'$ (see ). By the Caccioppoli estimate (see Lemma \[Lem\_Caccio\]), for all constants $c$, we have $$\fint_{\Boule_{R}}\big|\big(\frac{u-c}{\sqrt{T}},\nabla u\big)\big|^2\lesssim
\fint_{\Boule_{2R}}\big|\big(\frac{g-c}{\sqrt{T}},f,\frac{u-c}{R}\big)\big|^2.$$ Choosing $c=\fint_{\Boule_{2R}}u_{R'}$ and invoking (replacing $R \rightsquigarrow R/2$), we obtain $$\begin{aligned}
\fint_{\Boule_R}\big|\big(\frac{u}{\sqrt{T}},\nabla u\big)\big|^2
\lesssim~&
\fint_{\Boule_{4R}}\big|\big(\frac{g}{\sqrt{T}},f,\frac{u_{R'}}{\sqrt{T}},\nabla u_{R'}\big)\big|^2
+\big(\frac{R'}{R}\big)^2\fint_{\Boule_{4R}}|\nabla u|^2,
\end{aligned}$$ of which we take the $r$-th moments: $$\begin{aligned}
&\Big\langl \Big(\fint_{\Boule_R}\big|\big(\frac{u}{\sqrt{T}},\nabla u\big)\big|^2\Big)^r \Big\rangl
\\
&\qquad\lesssim
\Big\langl \Big(\fint_{\Boule_{4R}}\big|\big(\frac{g}{\sqrt{T}},f,\frac{u_{R'}}{\sqrt{T}},\nabla u_{R'}\big)\big|^2\Big)^r \Big\rangl
+\big(\frac{R'}{R}\big)^{2r}\Big\langl \Big(\fint_{\Boule_{4R}}|\nabla u|^2\Big)^r \Big\rangl.
\end{aligned}$$ By the triangle inequality and using the stationarity of $\nabla u$, which implies the estimate $\langl (\fint_{\Boule_{4R}}|\nabla u|^2)^r \rangl$ $\lesssim \langl (\fint_{\Boule_{R}}|\nabla u|^2)^r \rangl$, *cf.* , and choosing ${R'}=\theta R$ with $\theta \ll 1$, we may absorb the second r. h. s. term into the l. h. s. : $$\begin{aligned}
\Big\langl \Big(\fint_{\Boule_R}\big|\big(\frac{u}{\sqrt{T}},\nabla u\big)\big|^2\Big)^r \Big\rangl
\lesssim~&
\Big\langl \Big(\fint_{\Boule_{4R}}\big|\big(\frac{g}{\sqrt{T}},f,\frac{u_{\theta R}}{\sqrt{T}},\nabla u_{\theta R}\big)\big|^2\Big)^r \Big\rangl.
\end{aligned}$$ Finally, by Jensen’s inequality and the stationarity of $u$, $f$, and $g$, we get .
By the same reasoning based on the stationarity of $(\phi_T,\nabla \phi_T)$ as the one leading to , we get $$\Lambda_r
\lesssim
R^{d(1-\frac{1}{r})} \Big\langl \Big(\fint_{\Boule_{2R}} \big|\big(\frac{\phi_T}{\sqrt{T}},\nabla \phi_T\big)\big|^2 \Big)^{r}\Big\rangl^{\frac{1}{r}}
\pour R \gg 1.$$ Inserting , where we replace $(u,g,f) \rightsquigarrow (\phi_T, 0, a e)$, into this estimate yields $$\begin{aligned}
\label{Estim12}
\Lambda_r
\lesssim
R^{d(1-\frac{1}{r})} \Big( \Big\langl \Big| \fint_{\Boule_{\theta R}} \big( \frac{\phi_T}{\sqrt{T}},\nabla \phi_T \big) \Big|^{2r} \Big \rangl^{\frac{1}{r}}
+1\Big).
\end{aligned}$$ On the other hand, setting $g:=\sqrt{T}\mathds{1}_{\Boule_{R}}$ and $f:=\mathds{1}_{\Boule_{R}}e_j$ for $j \in \{1,\cdots,d\}$, we obtain from that $$\begin{aligned}
\Big\langl \Big| \fint_{\Boule_{R}} \big(\frac{\phi_T}{\sqrt{T}},\nabla \phi_T \big) \Big|^{2r} \Big\rangl^{\frac{1}{r}}
\lesssim \Lambda_r R^{-d}.
\end{aligned}$$ Inserting this (for $R \rightsquigarrow \theta R$) into yields $$\begin{aligned}
\Lambda_r
\lesssim R^{-\frac{d}{r}} \Lambda_r +R^{d(1-\frac{1}{r})} \pour R \gg 1.
\end{aligned}$$ Hence, we may absorb the first r. h. s. term and obtain . By Hölder regularity of the coefficient field, namely by Lemma \[LemAppendHold\] (in which we replace the fields $(u,f)$ and the exponents $(q,\alpha)$ by $(\phi_T,a\cdot e)$ and $(2,\alpha/2)$, respectively), the Hölder inequality, and Lemma \[LemSG\], estimate may be upgraded to $$\begin{aligned}
\langl|\nabla\phi_T|^{2r}\rangl^{\frac{1}{r}}
&\overset{\eqref{cw18}}{\lesssim} \Big\langle (1+ \|a \|_{\CC^{\frac{\alpha}{2}}(\Boule_1)})^{\frac{2r}{\alpha}(\frac{d}{2}+1)} \Big(1+\fint_{\Boule_1}\big| \big(\frac{\phi_T}{\sqrt{T}},\nabla \phi_T \big) \big|^2 \Big)^r \Big\rangle^{\frac{1}{r}}
\\
&\leq \big\langle (1+ \|a \|_{\CC^{\frac{\alpha}{2}}(\Boule_1)})^{\frac{4r}{\alpha}(\frac{d}{2}+1)} \big\rangle^{{\frac{\alpha}{4r(\frac{d}{2}+1)}}} \Big\langle \Big(1+\fint_{\Boule_1}\big| \big(\frac{\phi_T}{\sqrt{T}},\nabla \phi_T \big) \big|^2 \Big)^{2r} \Big\rangle^{\frac{1}{2r}}
\\
&\!\!\!\!\!\!\!\!\overset{\eqref{wg03},\eqref{Massive_Step3}}{\lesssim} 1.
\end{aligned}$$
Fixing $i, j, k, l \in \{1,\cdots,d\}$, we introduce the auxiliary functions $$\begin{aligned}
\big(\frac{1}{T}-\Delta\big)\ubar =\mathds{1}_{\Boule_1} \et \quad \big(\frac{1}{T}-\Delta\big)\vbar =-\nabla \cdot (\mathds{1}_{\Boule_1} e_l),
\end{aligned}$$ so that by the symmetry of $\frac{1}{T}-\Delta$ and the definitions and of $\sigma_{T}$ and $\psi_{T}$, we obtain the representations $$\begin{aligned}
&\int_{\Boule_1}\phi_{T}=\int \big(\frac{1}{T}g\phi_T-f\cdot\nabla\phi_T\big)
&&\pour(g,f):=(\ubar ,-\nabla\ubar ),
\\
&\int_{\Boule_1}\sigma_{T,ijk}=\int f\cdot (q_{T,i}-\langl q_{T,i}\rangl)
&&\pour f:=\partial_k\ubar e_j-\partial_j\ubar e_k,
\\
&\int_{\Boule_1}\psi_{T,ij}=\int f\cdot (q_{T,i}-\langl q_{T,i}\rangl-\nabla\phi_{T,i})
&&\pour f:=\ubar e_j,
\\
&\int_{\Boule_1}\partial_l \sigma_{T,ijk}=\int f\cdot (q_{T,i}-\langl q_{T,i}\rangl)
&&\pour f:=\partial_k\vbar e_j-\partial_j\vbar e_k,
\\
&\int_{\Boule_1}\partial_l \psi_{T,ij}=\int f\cdot (q_{T,i}-\langl q_{T,i}\rangl-\nabla\phi_{T,i})
&&\pour f:=\vbar e_j.
\end{aligned}$$ Hence (\[Ann\_as14\]) and follow from and via the standard estimates $$\begin{aligned}
\int \big|\big(\frac{\ubar }{\sqrt{T}},\nabla \ubar \big)\big|^2
\lesssim\mu^2_d(\sqrt{T})
\et \quad
\int \big|\big(\frac{\vbar }{\sqrt{T}},\nabla \vbar \big)\big|^2 \lesssim 1.
\end{aligned}$$
The argument, similar to Part 2, Step 2 of the proof of Proposition \[Propcorr\], is the same for $\sigma_T$ and $\psi_T$; therefore we only show for $\nabla \sigma_T$.
By local CZ estimate (replacing $(p,q) \rightsquigarrow (2r,2)$) for the constant-coefficient operator $\frac{1}{T} - \Delta$ applied to the equation , there holds $$\fint_{\Boule_R} \big|\big( \frac{\sigma_T}{\sqrt{T}}, \nabla \sigma_T \big) \big|^{2r}
\lesssim \fint_{\Boule_{2R}} \lt|q_T\rt|^{2r} + \Big(\fint_{\Boule_{2R}} \big|\big( \frac{\sigma_T}{\sqrt{T}}, \nabla \sigma_T \big) \big|^2\Big)^r.$$ By ergodicity and stationarity of $\sigma_T$ and $q_T$, when $R \uparrow \infty$, each of these spatial averages converges almost surely to the associated expectation (this is a consequence of the Birkhoff theorem). Hence $$\label{Num:951}
\big\langle \big|\big( \frac{\sigma_T}{\sqrt{T}}, \nabla \sigma_T \big) \big|^{2r} \big\rangle
\lesssim \big \langle \lt|q_T\rt|^{2r} \big \rangle
+\big\langle \big|\big( \frac{\sigma_T}{\sqrt{T}}, \nabla \sigma_T \big) \big|^{2} \big\rangle^r.$$ Moreover, since the constant-coefficient operator $\frac{1}{T}-\Delta$ obviously satisfies the weighted energy estimate (for a weight $\omega_T$ defined by for $C \gg_d 1$), there holds $$\int \omega_T \big\langle\big|\big( \frac{\sigma_T}{\sqrt{T}}, \nabla \sigma_T \big) \big|^{2} \big \rangle \lesssim \int \omega_T \big\langle\lt|q_T\rt|^{2}\big \rangle,$$ from which we deduce $$\big\langle \big|\big( \frac{\sigma_T}{\sqrt{T}}, \nabla \sigma_T \big) \big|^{2}\big\rangl \lesssim \big\langle \lt|q_T\rt|^{2}\big\rangle
\overset{\eqref{Def_qT},\eqref{Ann_ap56}}{\lesssim} 1.$$ Inserting this estimate into finally yields the desired estimate for $\nabla \sigma_T$.
We finally pass from to with help of . Since the proof is similar for $\phi_T$, $\sigma_T$ and $\psi_T$, we only treat $\phi_T$. Appealing to a Sobolev embedding in form of $$\begin{aligned}
|\phi_T(0)|^{2r} \lesssim
\fint_{\Boule_1}|\nabla\phi_T|^{2r}
+\Big|\fint_{\Boule_1}\phi_T\Big|^{2r},
\end{aligned}$$ which holds provided $2r>d$, and to the stationarity of $\nabla\phi_T$, we obtain the desired estimate: $$\begin{aligned}
\langl \phi_T^{2r}\rangl^\frac{1}{r}
\lesssim \langl |\nabla\phi_T|^{2r}\rangl^\frac{1}{r}
+\Big\langl \Big|\fint_{\Boule_1}\phi_T\Big|^{2r} \Big\rangl^{\frac{1}{r}} \overset{\eqref{Ann_ap56},\eqref{Ann_as14}}{\lesssim} \mu^2_d(\sqrt{T}).
\end{aligned}$$ [$\qed$]{}
Alternative approach via quenched CZ estimates {#SecLocal}
----------------------------------------------
We describe here another approach in order to obtain annealed CZ estimates; we restrict our explanation to a special case of Proposition \[PropCZ\](ii), namely specifying $r' \rightsquigarrow p$ in : $$\begin{aligned}
\label{Prop3E2}
\int \langl |\nabla u|^p \rangl
\lesssim_{\gamma, p, r}
\int \langl |f|^{r} \rangl^{\frac{p}{r}} \quad\text{provided } 1< p < r < \infty.
\end{aligned}$$ This strategy, which is close to [@DuerinckxOtto_2019 Prop. 6.4] and inspired by periodic homogenization (see the recent monograph [@Shen Th. 4.3.1, p. 83]), is quite different from the one of previous section. However, it involves two important concepts:
- [the minimal radius $\rstar$, above which homogenization kicks in,]{}
- [the quenched CZ estimates (*i.e.* deterministic or pathwise estimates).]{}
The very difference between the strategy exposed here and [@DuerinckxOtto_2019 Prop. 6.4] is that we do not need the full strength of the Lipschitz regularity theory to get quenched estimates; on the contrary, we only rely on the weaker result Proposition \[PropHcon\].
The first idea is to build a random stationary field of minimal radii $\rstar$ above which homogenization kicks in the sense of Proposition \[PropHcon\] (see [@Gloria_Neukamm_Otto_2019 Th. 1]):
For every $\delta >0$, there exists a random stationary field $\rstar \geq 1$ such that:
\(i) For all $R \geq \rstar(x)$, we have $$\frac{1}{R} \Big( \fint_{\Boule_R(x)} | ( \phi,\phi^\star,\sigma,\sigma^\star) - \fint_{\Boule_R(x)} ( \phi,\phi^\star,\sigma,\sigma^\star) |^2 \Big)^{\frac{1}{2}} \leq \delta.$$
\(ii) For all $p<\infty$ there holds $$\langl \rstar^p \rangl \lesssim_{\gamma, p, \delta} 1.$$
\(iii) The function $x \mapsto \rstar(x)$ is (almost surely) $\frac{1}{2}$-Lipschitz.
In establishing (ii), the control of moments of $(\phi,\sigma)$ provided by Proposition \[Propcorr\] plays a crucial role.
Next, appealing to a CZ decomposition [@Shen Th. 4.3.1, p. 83], a global *quenched* $\LL^p$ estimate (see also [@DuerinckxOtto_2019 Prop. 6.4]) may be obtained:
For every $p \in (1,\infty)$, there exists a constant $\delta=\delta(d,\lambda,p)>0$ such that, for the associated $\rstar$ defined in Lemma \[Cor2\], we have:
If $\nabla u$ and $f$ are square-integrable random vector fields related by $$\begin{aligned}
\nabla \cdot ( a \nabla u + f)=0,
\end{aligned}$$ then the following quenched estimates hold: $$\begin{aligned}
\label{Quench}
\int \Big( \fint_{\Boule_{\rstar(x)}(x)} | \nabla u|^2 \Big)^p \dd x
\lesssim_{\gamma, p, \delta}
\int \Big( \fint_{\Boule_{\rstar(x)}(x)} | f|^2 \Big)^p \dd x.
\end{aligned}$$
Finally, Proposition \[PropCZ\](ii) is obtained in two steps:
- [first, the quenched estimate is upgraded to an annealed estimate thanks to the control on the minimal radius $\rstar$ (at the price of a loss in stochastic integrability),]{}
- [then, an interpolation argument between the previous estimate and establishes . ]{}
Acknowledgements
================
The second author wishes to thank his long-term collaborators A. Gloria and M. Duerinckx, the many discussions with whom influenced this text in a way that cannot be done justice by the bibliography. Just to mention one example: The strategy of the proof of Lemma \[LemMassCorr\] originated from discussions with A. Gloria and J. Nolen. F. O. also acknowledges the hospitality of the Institut des Mathématiques de Toulouse, where he gave a lecture series on Sections 1 through 5 (and Section \[SecLocal\]) of this text. Both authors thank J. Sauer for his precious help concerning the theory of UMD spaces and his accurate advice for references.
[10]{}
G. Allaire. , volume 146 of [*Applied Mathematical Sciences*]{}. Springer-Verlag, New York, 2002.
A. Anantharaman, R. Costaouec, C. Le Bris, F. Legoll, and F. Thomines. Introduction to numerical stochastic homogenization and the related computational challenges: some recent developments. In [*Multiscale modeling and analysis for materials simulation*]{}, volume 22 of [*Lect. Notes Ser. Inst. Math. Sci. Natl. Univ. Singap.*]{}, pages 197–272. World Sci. Publ., Hackensack, NJ, 2012.
S. Armstrong and J.-P. Daniel. Calderón-[Z]{}ygmund estimates for stochastic homogenization. , 270(1):312–329, 2016.
S. Armstrong, T. Kuusi, and J.-C. Mourrat. , volume 352 of [*Grundlehren der Mathematischen Wissenschaften \[Fundamental Principles of Mathematical Sciences\]*]{}. Springer, Cham, 2019.
S. Armstrong and C. Smart. Quantitative stochastic homogenization of convex integral functionals. , 49(2):423–481, 2016.
M. Avellaneda and F.-H. Lin. Compactness methods in the theory of homogenization. , 40(6):803–847, 1987.
H. Bahouri, J.-Y. Chemin, and R. Danchin. , volume 343 of [*Grundlehren der Mathematischen Wissenschaften*]{}. Springer, Heidelberg, 2011.
P. Bella, A. Giunti, and F. Otto. Quantitative stochastic homogenization: local control of homogenization error through corrector. In [*Mathematics and materials*]{}, volume 23 of [*IAS/Park City Math. Ser.*]{}, pages 301–327. Amer. Math. Soc., Providence, RI, 2017.
J. Bergh and J. Löfström. . Springer-Verlag, Berlin-New York, 1976.
N. Clozeau, M. Josien, F. Otto, and Q. Xu. Systematic error in computation of the homogenized matrix. in preparation.
J. Conlon and A. Naddaf. On homogenization of elliptic equations with random coefficients. , 5:no. 9, 58, 2000.
M. Duerinckx, A. Gloria, and F. Otto. The structure of fluctuations in stochastic homogenization. , 2019. in press, arXiv preprint arXiv:1602.01717.
M. Duerinckx and F. Otto. Higher-order pathwise theory of fluctuations in stochastic homogenization. , 2019.
J. Fischer. The [C]{}hoice of [R]{}epresentative [V]{}olumes in the [A]{}pproximation of [E]{}ffective [P]{}roperties of [R]{}andom [M]{}aterials. , 234(2):635–726, 2019.
J. Fischer and F. Otto. A higher-order large-scale regularity theory for random elliptic operators. , 41(7):1108–1148, 2016.
M. Giaquinta. , volume 105 of [*Annals of Mathematics Studies*]{}. Princeton University Press, Princeton, NJ, 1983.
D. Gilbarg and N. Trudinger. . Classics in Mathematics. Springer-Verlag, Berlin, 2001.
A. Gloria, S. Neukamm, and F. Otto. Quantification of ergodicity in stochastic homogenization: optimal bounds via spectral gap on [G]{}lauber dynamics. , 199(2):455–515, 2015.
A. Gloria, S. Neukamm, and F. Otto. Quantitative estimates in stochastic homogenization for correlated coefficient fields. , 2019.
A. Gloria, S. Neukamm, and F. Otto. A regularity theory for random elliptic operators. , 2020. arXiv:1409.2678v4.
A. Gloria and F. Otto. An optimal variance estimate in stochastic homogenization of discrete elliptic equations. , 39(3):779–856, 2011.
A. Gloria and F. Otto. The corrector in stochastic homogenization: optimal rates, stochastic integrability, and fluctuations. , 2015.
A. Gloria and F. Otto. Quantitative results on the corrector equation in stochastic homogenization. , 19(11):3489–3548, 2017.
L. Grafakos. , volume 249 of [*Graduate Texts in Mathematics*]{}. Springer, New York, third edition, 2014.
Y. Gu and J.-C. Mourrat. Scaling limit of fluctuations in stochastic homogenization. , 14(1):452–481, 2016.
B. Helffer. Remarks on decay of correlations and [W]{}itten [L]{}aplacians, [B]{}rascamp-[L]{}ieb inequalities and semiclassical limit. , 155(2):571–586, 1998.
T. Hytönen, J. van Neerven, M. Veraar, and L. Weis. , volume 63 of [*Ergebnisse der Mathematik und ihrer Grenzgebiete. 3. Folge. A Series of Modern Surveys in Mathematics*]{}. Springer, Cham, 2016.
V. Jikov, S. Kozlov, and O. Ole[ĭ]{}nik. . Springer-Verlag, Berlin, 1994.
M. Josien and F. Otto. and statistics: a short proof of the annealed calderon-zygmund estimate.
R. B. Kellogg. Higher order singularities for interface problems. In [*The mathematical foundations of the finite element method with applications to partial differential equations ([P]{}roc. [S]{}ympos., [U]{}niv. [M]{}aryland, [B]{}altimore, [M]{}d., 1972)*]{}, pages 589–602, 1972.
F. Legoll and W. Minvielle. Variance reduction using antithetic variables for a nonlinear convex stochastic homogenization problem. , 8(1):1–27, 2015.
D. Marahrens and F. Otto. Annealed estimates on the [G]{}reen function. , 163(3-4):527–573, 2015.
T. McConnell. On [F]{}ourier multiplier transformations of [B]{}anach-valued functions. , 285(2):739–757, 1984.
N. Meyers. An [$L^{p}$]{}-estimate for the gradient of solutions of second order elliptic divergence equations. , 17:189–206, 1963.
J.-C. Mourrat. Efficient methods for the estimation of homogenized coefficients. , 19(2):435–483, 2019.
J.-C. Mourrat and F. Otto. Correlation structure of the corrector in stochastic homogenization. , 44(5):3207–3233, 2016.
S. Müller. Nonlinear partial differential equations [I]{}. Lectures notes, <https://www.iam.uni-bonn.de/users/lienstromberg/teaching/npde1/>.
A. Naddaf and T. Spencer. Estimates on the variance of some homogenization problems. , 1998.
G. C. Papanicolaou and S. R. S. Varadhan. Boundary value problems with rapidly oscillating random coefficients. In [*Random fields, [V]{}ol. [I]{}, [II]{} ([E]{}sztergom, 1979)*]{}, volume 27 of [*Colloq. Math. Soc. János Bolyai*]{}, pages 835–873. North-Holland, Amsterdam-New York, 1981.
Z. Shen. , volume 269 of [ *Operator Theory: Advances and Applications*]{}. Birkhäuser/Springer, Cham, 2018. Advances in Partial Differential Equations (Basel).
M. Stein. . Springer Series in Statistics. Springer-Verlag, New York, 1999. Some theory for Kriging.
L. Tartar. , volume 7 of [*Lecture Notes of the Unione Matematica Italiana*]{}. Springer-Verlag, Berlin; UMI, Bologna, 2009.
V. V. Yurinskiĭ. Averaging of symmetric diffusion in a random medium. , 27(4):167–180, 215, 1986.
Standard regularity theory
==========================
Caccioppoli estimate {#ProofCaccioppoli}
--------------------
For the convenience of the reader, we reproduce the proof of the classical Caccioppoli estimate (see for instance [@Giaquinta Th. 4.4, p. 63] for the massless operator):
\[Lem\_Caccio\] Let $T \in [1,+\infty]$, and $R >0$. Assume that the coefficient field $a$ satisfies . If $u$ solves $$\begin{aligned}
\label{Caccio_Eq}
\frac{1}{T}u-\nabla \cdot a \nabla u=\frac{1}{T}g + \nabla \cdot f \dans \Boule_{R},
\end{aligned}$$ then, for any constant $c$, there holds $$\label{Caccioppoli3}
\fint_{\Boule_{R/2}}\big|\big(\frac{u-c}{\sqrt{T}},\nabla u\big)\big|^2\lesssim
\fint_{\Boule_{R}}\big|\big(\frac{g-c}{\sqrt{T}},f,\frac{u-c}{R}\big)\big|^2.$$
We select a smooth function $\eta$ supported in $\Boule_R$ with values in $[0,1]$ such that $\eta=1$ in $\Boule_{R/2}$ and $|\nabla \eta| \lesssim R^{-1}$. Note that is invariant under the substitution ${(u,g)\rightsquigarrow({u + c}, \break{g+c})}$. Hence, w. l. o. g. we show only for $c=0$. Testing with $\eta^2 u$, we get by Leibniz’ rule in form of $\nabla(\eta^2 u)=\eta^2\nabla u+ 2 \eta u \nabla \eta$ that $$\begin{aligned}
&\frac{1}{T}\fint_{\Boule_R} \eta^2 u^2 + \fint_{\Boule_R} \eta^2 \nabla u \cdot a \nabla u
\\
&\qquad =
-2 \fint_{\Boule_R} \eta u \nabla \eta \cdot a \nabla u
+\frac{1}{T}\fint_{\Boule_R}\eta^2 u g
-\fint_{\Boule_R} \eta^2 f \cdot \nabla u - 2 \fint_{\Boule_R} \eta u f \cdot \nabla \eta.
\end{aligned}$$ Using the uniform ellipticity of $a$ for the l. h. s. and the Cauchy-Schwarz inequality for the r. h. s. (recalling that $\eta \leq 1$ and $|\nabla \eta| \lesssim R^{-1}$), we obtain $$\begin{aligned}
&\fint_{\Boule_R} \big|\big( \frac{\eta u}{\sqrt{T}}, \eta \nabla u \big) \big|^2
\\
&\lesssim
\Big(\fint_{\Boule_R} \big(\frac{u}{R}\big)^2\Big)^{\frac{1}{2}}
\Big(\fint_{\Boule_R} |\eta \nabla u|^2\Big)^{\frac{1}{2}}
+\Big(\fint_{\Boule_R} \big(\frac{g}{\sqrt{T}}\big)^2 \Big)^{\frac{1}{2}} \Big(\fint_{\Boule_R} \big( \frac{\eta u}{\sqrt{T}}\big)^2 \Big)^{\frac{1}{2}}
\\
&~~~~+
\Big(\fint_{\Boule_R} |f|^2\Big)^{\frac{1}{2}}
\Big(\fint_{\Boule_R} |\eta \nabla u|^2\Big)^{\frac{1}{2}}
+\Big(\fint_{\Boule_R} |f|^2\Big)^{\frac{1}{2}}
\Big(\fint_{\Boule_R} \big(\frac{u}{R}\big)^2\Big)^{\frac{1}{2}}
.
\end{aligned}$$ By the Young inequality, we may absorb the r. h. s. integrals $\fint_{\Boule_R} |\eta \nabla u|^2$ and $\fint_{\Boule_R} \big( \frac{\eta u}{\sqrt{T}}\big)^2$ into the l. h. s., obtaining for $c=0$ (since $\eta=1$ in $\Boule_{R/2}$).
A Sobolev estimate {#ProofE}
------------------
The basis for (\[wg04\]) is a Sobolev embedding in the form of $$\begin{aligned}
\label{E1}
\sup_{\Boule_{1}(x)}\zeta^2\lesssim
\Big(\int_{\Boule_2(x)}\sum_{n=0}|\nabla^n\zeta|^2\Big)^\frac{1}{2}
\Big(\int_{\Boule_2(x)}\zeta^2\Big)^\frac{1}{2},
\end{aligned}$$ which by Cauchy-Schwarz, an identity of the type of , and Plancherel leads to $$\begin{aligned}
\label{Estim04}
\int \sup_{\Boule_1(x)}\zeta^2 \dd x\lesssim
\Big(\int(1+|k|)^{2d}|\FF \zeta|^2\dd k\Big)^\frac{1}{2}
\Big(\int|\FF \zeta|^2\dd k\Big)^\frac{1}{2}.
\end{aligned}$$ Under additional conditions on the support of the Fourier transform in form of $\Supp(\FF \zeta)\subset\{1\le R|k|<2\}$, this yields $$\begin{aligned}
\int \sup_{\Boule_1(x)}\zeta^2 \dd x\lesssim
R^{2\alpha}\int(1+|k|)^{d+2\alpha}|\FF \zeta|^2\dd k
\end{aligned}$$ and if $\Supp(\FF \zeta) \subset\{|k|<1\}$, then $$\begin{aligned}
\int \sup_{\Boule_1(x)}\zeta^2 \dd x\lesssim
\int(1+|k|)^{d+2\alpha}|\FF \zeta|^2\dd k.
\end{aligned}$$ Hence (\[wg04\]) follows from decomposing $\zeta$ in Fourier space accordingly (*i.e.* with $R=2^{-n}$, $n=0, 1, \cdots$) and using the triangle inequality on the l. h. s. norm in . The sum over $n$ converges since $\alpha>0$.
Here comes the proof of . Without loss of generality, we may assume that $\zeta$ has a support in $\Boule_2$ (if not, we may just multiply it by a cut-off function and appeal to the Leibniz’ rule). Then, using the Fourier transform[^5] and separating small and large wavelengths according to a parameter $R>0$, we obtain $$\begin{aligned}
\sup_{\Boule_1(x)}|\zeta|
&\leq
\frac{1}{(2\pi)^d}\int |\FF \zeta|
=\frac{1}{(2\pi)^d} \Big(\int_{|k| \leq R} |\FF \zeta| \dd k + \int_{|k|>R} |\FF \zeta| \dd k\Big)
\\
&\lesssim R^{\frac{d}{2}} \Big( \int |\FF \zeta|^2 \dd k\Big)^{\frac{1}{2}} + R^{-\frac{d}{2}} \Big( \int (1+|k|)^{2d} |\FF \zeta|^2 \dd k\Big)^{\frac{1}{2}}
\\
&\lesssim R^{\frac{d}{2}} \Big(\int_{\Boule_2(x)}\zeta^2\Big)^\frac{1}{2}
+ R^{-\frac{d}{2}} \Big(\int_{\Boule_2(x)}(|\nabla^d\zeta|^2+\zeta^2)\Big)^\frac{1}{2}.
\end{aligned}$$ We now obtain by optimizing in $R$.
Small scale regularity results
------------------------------
### Hölder regularity of the Gaussian field {#SecProofLemloc}
We first argue that it is enough to establish $$\begin{aligned}
\label{Besov}
\langl (g(h)-g(0))^2\rangl\lesssim|h|^{2\alpha}.
\end{aligned}$$ Indeed, since $g$ is Gaussian, yields $\langl |g(h)-g(0)|^r\rangl^{\frac{1}{r}}\lesssim|h|^{\alpha}$, likewise, $\langl g^2(0) \rangl$ $ =c(0) \lesssim 1$ yields $\langl |g(0)|^r \rangl^{\frac{1}{r}} \lesssim 1$. Hence, by stationarity, we obtain $$\begin{aligned}
\Big\langl \int_{\Boule_2} \int_{\Boule_2} \Big(\frac{|g(x+h)-g(x)|}{|h|^{\alpha''}}\Big)^{r} \frac{\dd h}{|h|^d} \dd x + \int_{\Boule_2} |g|^r \Big\rangl
\lesssim 1
\end{aligned}$$ for all $r \in [1,\infty)$ and $\alpha''<\alpha$, which with help of a smooth cut-off function $\eta$ for $\Boule_1$ in $\Boule_2$ we rewrite as $$\begin{aligned}
\Big\langl \int_{\R^d} \int_{\Boule_2} \Big(\frac{|(\eta g)(x+h)-(\eta g)(x)|}{|h|^{\alpha''}}\Big)^{r} \frac{\dd h}{|h|^d} \dd x + \int_{\R^d} |\eta g|^r \Big\rangl
\lesssim 1.
\end{aligned}$$ This amounts to an estimate of $\eta g$ in the Besov space $\Besov^{\alpha''}_{r,r}(\R^d)$ *cf.* [@Bahouri Th. 2.36, p. 74], which embeds into a Hölder space $\CC^{\alpha'}(\R^d)$ for any $\alpha' < \alpha ''$ provided $r$ is sufficiently large (see [@Bahouri Prop. 2.71, p. 99] and [@Bahouri p. 99, Examples]). Therefore, implies .
Estimate can be seen as follows: $$\begin{aligned}
\langl (g(x)-g(0))^2\rangl
&~=2c(0)-c(x)-c(-x)\\
&~=\frac{4}{(2\pi)^d}\int \sin^2\big(\frac{k\cdot x}{2}\big)(\FF c) \dd k
\\
&\stackrel{(\ref{wg01})}{\le} \frac{4}{(2\pi)^d}\int \min\Big\{\frac{|k|^2|x|^2}{4},1\Big\}|k|^{-d-2\alpha}\dd k
\\
&~= \frac{4}{(2\pi)^d}|x|^{2\alpha}\int \min\{|\hat k|^2,1\} |\hat k|^{-d-2\alpha}\dd\hat k,
\end{aligned}$$ where the last integral converges because of $\alpha\in(0,1)$.
### Small-scale annealed CZ estimates for ensembles of Hölder continuous coefficients
We recall the norms and and state the following:
\[Lem\_AnnealedCZ\_2\] Let the ensemble $\langl \cdot \rangl$ be defined as in Section \[SecAssumpGauss\] and $T \geq 1$. Assume that the square-integrable functions $u$, $f$, and $g$ are related through $$\label{Num:905}
\frac{1}{T} u -\nabla\cdot a\nabla u=\frac{1}{T}g+\nabla \cdot f.$$ (i) For $q,p\in(1,\infty)$ and $1\le r'<r<\infty$ we have $$\label{CZstar_2}
\begin{aligned}
\big\|\big(\frac{u}{\sqrt{T}},\nabla u\big)\big\|_{p,r'}
\lesssim_{\gamma,q,p,r',r}
& \big\|\big(\frac{u}{\sqrt{T}},\nabla u\big)\big\|_{p,r,q}
+\big\|\big(\frac{g}{\sqrt{T}},f\big)\big\|_{p,r}.
\end{aligned}$$ (ii) For $q\in(1,2]$ and $r>2$ we have $$\label{Num:116}
\begin{aligned}
&\big\|\big(\frac{u}{\sqrt{T}},\nabla u\big)\big\|_{2,2,q}
\lesssim_{\gamma,q,r}
\big\|\big(\frac{g}{\sqrt{T}},f\big)\big\|_{2,r,q}.
\end{aligned}$$
We split the proof of Lemma \[Lem\_AnnealedCZ\_2\] into two independent parts.
Let $q,p, r'$ be as in the statement, and choose $r'' \in (r',r)$. W. l. o. g. we may assume that $q \leq\min\{r',p\}$ (the other case $q \geq\min\{r',p\}$ can be recovered from the former one by Jensen’s inequality).
The proof is divided into three steps. In Step 1, we define a (small) random radius $\rho_\star$ below which we may apply classical regularity theory. In Step 2, we define a splitting of $u$ by means of which we establish , admitting the estimate that for any $F=F(a)$ and $y \in \R^d$ $$\label{Num:019}
\big\langle |F|^{r'} \big\rangle^{\frac{p}{r'}}
\lesssim
\int \big\langle \mathds{1}_{\Boule_{\frac{\rho_\star(x)}{2}}(x)}(y) |F|^{r''} \big\rangle^{\frac{p}{r''}} \dd x,$$ which we derive in Step 3. Let $\delta \leq 1$ (to be fixed in Step 2). We introduce the (local) radius below which $a$ does not vary much: $$\label{Def_rhostar}
\rho_\star(x):= \sup \Big\{ \rho \leq 1, \sup_{y,y' \in \Boule_\rho(x)} |a(y)-a(y')| \leq \delta \Big\}.$$ Note that by we have $0<\rho_\star(x) \leq 1$ $\langl\cdot\rangl$-almost-surely; furthermore, $\rho_\star$ is stationary. We claim that $\rho_\star$ is bounded by below in the sense of $$\label{Num:012}
\langl \rho_\star^{-s} \rangl^{\frac{1}{s}} \lesssim_{\gamma,s,\delta} 1 \pourtout s \in [1,\infty).$$
By stationarity, we may assume that $x=0$ (we omit this variable in the expressions below), and estimate $\rho_\star$ by the easier quantity $$\begin{aligned}
\tilde{\rho}_\star:= \min\Big\{\frac{1}{2}\Big(\frac{\delta}{\|a\|_{\CC^{\frac{\alpha}{2}}(\Boule_1)}} \Big)^{\frac{2}{\alpha}},1\Big\}.
\end{aligned}$$ Indeed we have $\rho_\star \geq \tilde{\rho}_\star$ because, if $y, y' \in \Boule_{\tilde{\rho}_\star} \subset \Boule_{1}$ then $$|a(y)-a(y')|
\leq \|a\|_{\CC^{\frac{\alpha}{2}}(\Boule_1)} |y-y'|^{\frac{\alpha}{2}}
\leq \|a\|_{\CC^{\frac{\alpha}{2}}(\Boule_1)} (2\tilde{\rho}_\star)^{\frac{\alpha}{2}}
\leq \delta.$$ Therefore is a consequence of : $$\begin{aligned}
\langl \rho_\star^{-s} \rangl^{\frac{1}{s}}
\leq \langl \tilde{\rho}_\star^{-s} \rangl^{\frac{1}{s}}
\lesssim_{\gamma,s,\delta} \big\langl \|a\|_{\CC^{\frac{\alpha}{2}}(\Boule_1)}^{\frac{2s}{\alpha}} \big\rangl^{\frac{1}{s}}
\overset{\eqref{wg03}}{\lesssim}_{\gamma,s,\delta} 1.
\end{aligned}$$
We introduce $$\begin{aligned}
&\Boule_\star(x):=\Boule_{\rho_\star(x)}(x),
\quad \tilde{g}_x:=\mathds{1}_{\Boule_1(x)}g,
\quad\tilde{f}_x:= \mathds{1}_{\Boule_1(x)}f,
\\
&\text{and} \qquad
\tilde{a}_x(y):=
\begin{cases}
a(y) & \text{if } y \in \Boule_\star(x),
\\
a(x) & \text{otherwise}.
\end{cases}
\end{aligned}$$ Accordingly, we freeze the variable $x$ and define $\tilde{u}_x$ as the solution to $$\frac{1}{T} \tilde{u}_x -\nabla \cdot \tilde{a}_x \nabla \tilde{u}_x=\frac{1}{T}\tilde{g}_x+ \nabla \cdot \tilde{f}_x,$$ where the operator $\nabla$ acts on the variable $y$. This combines with to $$\label{Num:013}
\frac{1}{T}(\tilde{u}_x-u) - \nabla \cdot a \nabla (\tilde{u}_x-u)=0 \dans \Boule_\star(x).$$
First, we estimate $\tilde{u}_x$. Since by definition of $\rho_\star$, we have $\|\tilde{a}_x-a(x)\|_{\LL^\infty} \leq \delta$, by a perturbative argument *à la Meyers* as in the proof Proposition \[PropCZ\](i), there holds for $\delta \ll_{d,\lambda,p,q,r''} 1$ $$\begin{aligned}
\label{Num:011}
\int \big| \big(\frac{\tilde{u}_x}{\sqrt{T}}, \nabla \tilde{u}_x\big) \big|^q
&\lesssim \int_{\Boule_1(x)} \big|\big(\frac{g}{\sqrt{T}},f\big)\big|^q,
\\
\label{Num:010}
\int \big\langle \big|\big(\frac{\tilde{u}_x}{\sqrt{T}},\nabla \tilde{u}_x\big) \big|^{r''} \big\rangle^{\frac{p}{{r''}}}
&\lesssim \int_{\Boule_1(x)} \big\langle \big|\big(\frac{g}{\sqrt{T}},f\big) \big|^{r''} \big\rangle^{\frac{p}{{r''}}}.
\end{aligned}$$
Second, we estimate $\nabla u -\nabla \tilde{u}_x$. Using Lemma \[LemAppendHold\], more precisely a rescaled version of for $\alpha \rightsquigarrow \frac{\alpha}{2}$, on , the triangle inequality and , there also holds $$\begin{aligned}
&\sup_{\frac{1}{2} \Boule_\star(x)} \Big|\Big(\frac{u-\tilde{u}_x}{\sqrt{T}},\nabla u-\nabla \tilde{u}_x\Big)\Big|
\\
&\qquad\overset{\eqref{Num:031_bis}}{\lesssim}
\big(1+\rho_\star^{\frac{\alpha}{2}}(x) \|a\|_{\CC^{\frac{\alpha}{2}}(\Boule_\star(x))}\big)^{\frac{2}{\alpha}(\frac{d}{q}+1)}
\Big(\fint_{\Boule_\star(x)} \Big|\Big(\frac{u-\tilde{u}_x}{\sqrt{T}},\nabla u-\nabla \tilde{u}_x\Big)\Big|^q \Big)^{\frac{1}{q}}
\\
&\quad\overset{\eqref{Def_rhostar},\eqref{Num:011}}{\lesssim}
\big(1+\|a\|_{\CC^{\frac{\alpha}{2}}(\Boule_\star(x))}\big)^{\frac{2d}{q\alpha}} \rho_\star^{-\frac{d}{q}}(x) \Big(\fint_{\Boule_1(x)} \big|\big(\frac{g}{\sqrt{T}},f,\frac{u}{\sqrt{T}},\nabla u\big)\big|^q \Big)^{\frac{1}{q}}.
\end{aligned}$$ Taking the $\LL^{r''}_{\langl\cdot\rangl}$-norm of this estimate and using the Hölder inequality followed by and yields $$\label{Num:914}
\begin{aligned}
&\Big\langle \sup_{\frac{1}{2} \Boule_\star(x)} \Big|\Big(\frac{u-\tilde{u}_x}{\sqrt{T}},\nabla u-\nabla \tilde{u}_x\Big)\Big|^{r''} \Big\rangle^{\frac{1}{{r''}}}
\\
&\qquad\qquad\lesssim
\Big\langle\Big(\fint_{\Boule_1(x)} \big|\big(\frac{g}{\sqrt{T}},f,\frac{u}{\sqrt{T}},\nabla u\big)\big|^q \Big)^{\frac{r}{q}} \Big \rangle^{\frac{1}{{r}}}.
\end{aligned}$$
Finally, momentarily fixing a $y\in\R^d$, which we suppress in most of the notation, by (for $F:=(\frac{u(y)}{\sqrt{T}},\nabla u(y))$) followed by the triangle inequality, there holds $$\begin{aligned}
\big\langle \big|\big(\frac{u}{\sqrt{T}},\nabla u\big)\big|^{r'} \big\rangle^{\frac{p}{r'}}
&\lesssim
\int\big\langle \mathds{1}_{\frac{1}{2}\Boule_\star(x)} \big|\big(\frac{\tilde{u}_x}{\sqrt{T}},\nabla \tilde{u}_x\big)\big|^{r''} \big\rangle^{\frac{p}{r''}} \dd x
\\
&\qquad+\int\big\langle \mathds{1}_{\frac{1}{2}\Boule_\star(x)} \big|\big(\frac{u-\tilde{u}_x}{\sqrt{T}},\nabla u-\nabla \tilde{u}_x\big)\big|^{r''} \big\rangle^{\frac{p}{r''}} \dd x.
\end{aligned}$$ Integrating this in the $y$ variable (recall that $\rho_\star \leq 1$), and using , , and the definitions of the norms and yields $$\begin{aligned}
\int \big\langle \big|\big(\frac{u}{\sqrt{T}},\nabla u\big)\big|^{r'} \big\rangle^{\frac{p}{r'}}
&\lesssim
\int\int \big\langle \big|\big(\frac{\tilde{u}_x}{\sqrt{T}},\nabla \tilde{u}_x\big)\big|^{r''} \big\rangle^{\frac{p}{r''}} \dd x
\\
&\qquad+\int\int\big\langle \mathds{1}_{\frac{1}{2}\Boule_\star(x)} \big|\big(\frac{u-\tilde{u}_x}{\sqrt{T}},\nabla u-\nabla \tilde{u}_x\big)\big|^{r''} \big\rangle^{\frac{p}{r''}} \dd x
\\
&\!\!\!\!\!\!\!\!\overset{\eqref{Num:010}, \eqref{Num:914}}{\lesssim}
\int \int_{\Boule_1(x)} \big\langle \big|\big(\frac{g}{\sqrt{T}},f\big) \big|^{r''} \big\rangle^{\frac{p}{{r''}}} \dd x
\\
&\qquad+
\int\Big\langle\Big(\fint_{\Boule_1(x)} \big|\big(\frac{g}{\sqrt{T}},f,\frac{u}{\sqrt{T}},\nabla u\big)\big|^q \Big)^{\frac{r}{q}} \Big \rangle^{\frac{p}{{r}}}\dd x.
\end{aligned}$$ By definition of the norms , this assumes the form $$\begin{aligned}
\big\| \big(\frac{u}{\sqrt{T}},\nabla u\big) \big\|_{p,r'}
\overset{\eqref{Num:556}, \eqref{Ann_ap20}}{\lesssim} \big\| \big(\frac{g}{\sqrt{T}},f\big) \big\|_{p,r''}
+ \big\|\big(\frac{g}{\sqrt{T}},f,\frac{u}{\sqrt{T}},\nabla u\big)\big\|_{p,r,q}.
\end{aligned}$$ By the estimates and (recall that $r''\leq r$ and $q \leq \min\{p,r\}$), we have $$\big\| \big(\frac{g}{\sqrt{T}},f\big) \big\|_{p,r''} + \big\| \big(\frac{g}{\sqrt{T}},f\big) \big\|_{p,r,q} \lesssim \big\| \big(\frac{g}{\sqrt{T}},f\big) \big\|_{p,r}.$$ This yields the desired estimate .
By the Hölder inequality in space (based on $\rho_\star \leq 1$) followed by Jensen’s inequality, there holds $$\label{Num:021}
\begin{aligned}
\Big(\int \big\langl \mathds{1}_{\frac{1}{2}\Boule_\star(x)}(y) |F|^{r''} \big\rangl^{\frac{p}{{r''}}}\dd x \Big)^{\frac{1}{p}}
&\gtrsim
\int \big\langl \mathds{1}_{\frac{1}{2}\Boule_\star(x)}(y) |F|^{r''} \big\rangl^{\frac{1}{{r''}}}\dd x
\\
&\geq
\Big\langl \Big( \int \mathds{1}_{\frac{1}{2}\Boule_\star(x)}(y)\dd x |F|\Big)^{r''} \Big\rangl^{\frac{1}{{r''}}}.
\end{aligned}$$ Note that if $x \in \frac{1}{3}\Boule_\star(y)$ then $\Boule_{\frac{2}{3}\rho_\star(y)}(x) \subset \Boule_{\star}(y)$. Indeed, if $z \in \Boule_{\frac{2}{3}\rho_\star(y)}(x)$, then we have by the triangle inequality $|z-y| \leq |z-x| + |x-y|< \rho_\star(y)$. Therefore, by definition of $\rho_\star$, this implies that $\rho_\star(x) \geq \frac{2}{3}\rho_\star(y)$ and in particular $y \in \frac{1}{2}\Boule_\star(x)$. As a consequence, there holds $$\int \mathds{1}_{\frac{1}{2}\Boule_\star(x)}(y)\dd x
\geq \int_{\frac{1}{4}\Boule_\star(y)}\dd x
\gtrsim \rho_\star(y)^d.$$ Inserting this estimate into , we get $$\label{Num:022}
\begin{aligned}
\langl \rho_\star(y)^{{r''}d} |F|^{r''} \rangl^{\frac{1}{{r''}}}
\lesssim
\Big(\int \big\langl \mathds{1}_{\frac{1}{2}\Boule_\star(x)}(y) |F|^{r''} \big\rangl^{\frac{p}{{r''}}}\dd x \Big)^{\frac{1}{p}}.
\end{aligned}$$ Finally, by the Hölder inequality we obtain $$\langl |F|^{r'} \rangl^{\frac{1}{r'}}
=
\langl (\rho_\star(y)^{d} |F|)^{r'}
\rho_\star(y)^{-d r'} \rangl^{\frac{1}{r'}}
\leq \langl \rho_\star(y)^{{r''}d} |F|^{{r''}} \rangl^{\frac{1}{{r''}}} \big\langl \rho_\star(y)^{-d\frac{r'r''}{{r''}-r'}}\big\rangl^{\frac{{r''}-r'}{{r''} r'}},$$ whence, by (for $s \rightsquigarrow d\frac{r'r''}{{r''}-r'}$), $$\langl |F|^{r'} \rangl^{\frac{1}{r'}}
\lesssim \big\langl \rho_\star(y)^{r'' d} |F|^{{r''}} \big\rangl^{\frac{1}{{r''}}}.$$ Inserting into the above estimate, we obtain .
By duality, it is sufficient to establish for $p\in[2,\infty)$ and $r<2$ that $$\begin{aligned}
\label{cw20}
\big\|\big(\frac{u}{\sqrt{T}},\nabla u\big)\big\|_{2,r,p}\lesssim\big\|\big(\frac{g}{\sqrt{T}},f\big)\big\|_{2,2,p}.
\end{aligned}$$ In order to establish this, we appeal to Lemma \[LemAppendHold\], more precisely (\[cw15\]) for $q=2$, with $\alpha$ replaced by some $0<\alpha'<\alpha$ (which we now fix), and with the origin replaced by a general point $x\in\mathbb{R}^d$: $$\begin{aligned}
\lefteqn{\big\|\big(\frac{u}{\sqrt{T}},\nabla u\big)\big\|_{\LL^p(\Boule_1(x))}}\nonumber\\
&\qquad\lesssim
\|a\|_{\CC^{\alpha'}(\Boule_2(x))}^{\frac{1}{\alpha'}(\frac{d}{2}-\frac{d}{p})}
\big\|\big(\frac{u}{\sqrt{T}},\nabla u\big)\big\|_{\LL^2(\Boule_2(x))}
+\big\|\big(\frac{g}{\sqrt{T}},f\big)\big\|_{\LL^p(\Boule_2(x))}.
\end{aligned}$$ We first take the $\LL^r_{\langle\cdot\rangle}$-norm; by in Lemma \[LemSG\] (note that the shift by $x$ is irrelevant by stationarity) in conjunction with Hölder’s inequality (recall $r<2$) for the first r. h. s. term (and Jensen’s inequality on the second) we obtain $$\begin{aligned}
\lefteqn{\big\|\big\|\big(\frac{u}{\sqrt{T}},\nabla u\big)\big\|_{\LL^p(\Boule_1(x))}\big\|_{\LL^r_{\langle\cdot\rangle}}}
\nonumber\\
&\qquad\lesssim
\big\|\big\|\big(\frac{u}{\sqrt{T}},\nabla u\big)\big\|_{\LL^2(\Boule_2(x))}\big\|_{\LL^2_{\langle\cdot\rangle}}
+\big\|\big\|\big(\frac{g}{\sqrt{T}},f\big)\big\|_{\LL^p(\Boule_2(x))}\big\|_{\LL^2_{\langle\cdot\rangle}}.
\end{aligned}$$ We next take the $\LL^2(\mathbb{R}^d)$-norm (in $x$) of this: On the l. h. s. we use the definition and on both r. h. s. terms we use , to the effect of $$\begin{aligned}
\big\|\big(\frac{u}{\sqrt{T}},\nabla u\big)\big\|_{2,r,p}\lesssim
\big\|\big(\frac{u}{\sqrt{T}},\nabla u\big)\big\|_{2,2,2}
+\big\|\big(\frac{g}{\sqrt{T}},f\big)\big\|_{2,2,p}.
\end{aligned}$$ It remains to appeal to the energy estimate , Hölder’s inequality in the inner space variable, in conjunction with a version of to obtain : $$\begin{aligned}
\big\|\big(\frac{u}{\sqrt{T}},\nabla u\big)\big\|_{2,2,2} &\stackrel{\eqref{Num:556}}{=}
\big\|\big(\frac{u}{\sqrt{T}},\nabla u\big)\big\|_{2,2}\\
&\stackrel{\eqref{Ann_ap97}}{\lesssim}
\big\|\big(\frac{g}{\sqrt{T}},f\big)\big\|_{2,2}\stackrel{\eqref{Num:556}}{=}
\big\|\big(\frac{g}{\sqrt{T}},f\big)\big\|_{2,2,2}\lesssim\big\|\big(\frac{g}{\sqrt{T}},f\big)\big\|_{2,2,p}.
\end{aligned}$$
### Local estimates for Hölder continuous coefficients
The interest of the local CZ and Schauder estimates below is the explicit, polynomial dependence on the Hölder norm of the coefficient field $a$. A secondary aspect is the uniformity in the massive parameter $T$. The ingredients are classical; we give an argument for the convenience of the reader.
\[LemAppendHold\] Let $\alpha \in (0,1)$, and let $a \in \CC^{\alpha}(\Boule_1)$ satisfy . For $T\ge 1$, we assume that the functions $u$, $g$, and the vector field $f$ satisfy the relation $$\begin{aligned}
\label{cw09}
\frac{1}{T}u-\nabla\cdot a\nabla u=\frac{1}{T}g+\nabla\cdot f\quad\mbox{in}\;\Boule_2.
\end{aligned}$$ For $1<q\le p<\infty$ we claim that $$\label{cw15}
\begin{aligned}
\lefteqn{\big\|\big(\frac{u}{\sqrt{T}},\nabla u\big)\big\|_{\LL^p(\Boule_1)}}
\\
&\qquad\lesssim_{d,\lambda,\alpha,p,q}
\|a\|_{\CC^\alpha(\Boule_2)}^{\frac{1}{\alpha}(\frac{d}{q}-\frac{d}{p})}
\big\|\big(\frac{u}{\sqrt{T}},\nabla u\big)\big\|_{\LL^q(\Boule_2)}
+\big\|\big(\frac{g}{\sqrt{T}},f\big)\big\|_{\LL^p(\Boule_2)},
\end{aligned}$$ and, provided $g\equiv0$, $$\label{cw18}
\begin{aligned}
\lefteqn{\big\|\big(\frac{u}{\sqrt{T}},\nabla u\big)\big\|_{\CC^{\alpha}(\Boule_1)}}
\\
&\qquad\lesssim_{d,\lambda,\alpha,q}
\|a\|_{\CC^\alpha(\Boule_2)}^{\frac{1}{\alpha}(\frac{d}{q}+1)}
\big\|\big(\frac{u}{\sqrt{T}},\nabla u\big)\big\|_{\LL^q(\Boule_{2})}
+\|a\|_{\CC^\alpha(\Boule_2)}^{\frac{1}{\alpha}-1}\|f\|_{\CC^{\alpha}(\Boule_{2})}.
\end{aligned}$$
In case of $g=0$ and $f=0$, since $T \geq 1$, implies $$\label{Num:031_bis}
\begin{aligned}
\big\|\big(\frac{u}{\sqrt{T}},\nabla u\big)\big\|_{\LL^\infty(\Boule_{1})}
\lesssim_{d,\lambda,\alpha,q}~& \|a \|_{\CC^{\alpha}(\Boule_2)}^{\frac{1}{\alpha}(\frac{d}{q}+1)} \big\|\big(\frac{u}{\sqrt{T}},\nabla u\big)\big\|_{\LL^q(\Boule_2)}.
\end{aligned}$$
The strategy of the proof is the following: First, assuming the additional condition $[a]_{\CC^{\alpha}(\Boule_2)}\le 1$, we prove in Step 1 and 2 that and hold. In Step 3, we make use of a rescaling in order to apply the previous result and finally get and in full generality.
Provided $$\begin{aligned}
\label{cw08}
[a]_{\CC^{\alpha}(\Boule_2)}\le 1
\end{aligned}$$ we claim $$\begin{aligned}
\label{cw01}
\big\|\big(\frac{u}{\sqrt{T}},\nabla u\big)\big\|_{\LL^p(\Boule_1)}\lesssim
\big\|\big(\frac{u}{\sqrt{T}},\nabla u\big)\big\|_{\LL^q(\Boule_2)}+ \big\|\big(\frac{g}{\sqrt{T}},f\big)\big\|_{\LL^p(\Boule_2)},
\end{aligned}$$ and, in case of $g\equiv0$, $$\begin{aligned}
\label{cw02}
\big\|\big(\frac{u}{\sqrt{T}},\nabla u\big)\big\|_{\CC^{\alpha}(\Boule_1)}\lesssim
\big\|\big(\frac{u}{\sqrt{T}},\nabla u\big)\big\|_{\LL^q(\Boule_2)}+ \|f\|_{\CC^{\alpha}(\Boule_2)}.
\end{aligned}$$
Here comes the argument: We split $u=v+w$ with help of the Dirichlet problem $$\begin{aligned}
\label{cw10}
\frac{1}{T}v-\nabla\cdot a\nabla v=\frac{1}{T}g+\nabla\cdot f\;\;\mbox{in}\;\Boule_2,\quad
v=0\;\;\mbox{on}\;\partial \Boule_2,
\end{aligned}$$ so that in view of (\[cw09\]), $w$ satisfies $$\begin{aligned}
\label{cw07}
\frac{1}{T}w-\nabla\cdot a\nabla w=0\quad\mbox{in}\;\Boule_2.
\end{aligned}$$ By standard CZ and Schauder theory for (\[cw10\]) (see [@Giaquinta Th. 7.2 p. 140], and [@GT Chap. 8 p. 177-218] or [@Giaquinta Chap. 5 p. 75-96], respectively), which is not affected by the presence of the massive term with $T\ge 1$ (as for instance can be seen by expressing the solution operator in terms of the semi-group and appealing to the parabolic version of the theory), we have in case of $g\equiv 0$, $$\begin{aligned}
\label{cw11}
\|\nabla v\|_{\LL^p(\Boule_2)}&\lesssim\|f\|_{\LL^p(\Boule_2)}\quad\mbox{and}\quad
\|\nabla v\|_{\CC^{\alpha}(\Boule_2)}\lesssim\|f\|_{\CC^{\alpha}(\Boule_2)}.
\end{aligned}$$ Note that thanks to our assumption (\[cw08\]), the implicit constant depends only on $d,\lambda,p,\alpha$. In the presence of $g$, we solve the auxiliary Dirichlet problem $-\triangle v'=\frac{1}{T}g$ on $\Boule_2$, so that $f':=-\nabla v'$ satisfies $\nabla\cdot f'=\frac{1}{T}g$. By constant-coefficient CZ theory, we have $\|\nabla f'\|_{\LL^p(\Boule_2)}\lesssim\|\frac{1}{T}g\|_{\LL^p(\Boule_2)}$. In conjunction with the Poincaré estimate for $v$, and the Poincaré-Wirtinger estimate for $f'$ (which only matters up to a constant), we may upgrade the first item in (\[cw11\]) to $$\begin{aligned}
\lefteqn{\|(\frac{v}{\sqrt{T}},\nabla v)\|_{\LL^p(\Boule_2)}
\stackrel{T\ge 1}{\le}\|(v,\nabla v)\|_{\LL^p(\Boule_2)}\lesssim\|\nabla v\|_{\LL^p(\Boule_2)}}
\\
&\qquad\lesssim\|(f',f)\|_{\LL^p(\Boule_2)}\lesssim\|(\frac{g}{T},f)\|_{\LL^p(\Boule_2)}
\stackrel{T\ge 1}{\le}\big\|\big(\frac{g}{\sqrt{T}},f\big)\big\|_{\LL^p(\Boule_2)}.
\end{aligned}$$ Since $\|\nabla v\|_{\CC^0(\Boule_2)}$ controls the Lipschitz constant of $v$, the second item in (\[cw11\]) obviously upgrades to $$\begin{aligned}
\|(\frac{v}{\sqrt{T}},\nabla v)\|_{\CC^{\alpha}(\Boule_2)}&\lesssim\|f\|_{\CC^{\alpha}(\Boule_2)}.
\end{aligned}$$ It remains to appeal to (\[cw06\]) in Step 2 to obtain (\[cw01\]) and (\[cw02\]) by the triangle inequality.
Suppose that $w$ satisfies (\[cw07\]). Then under the assumption (\[cw08\]) we claim that for any $p,q\in(1,\infty)$ $$\begin{aligned}
\label{cw06}
\big\|\big(\frac{w}{\sqrt{T}},\nabla w\big)\big\|_{\LL^p(\Boule_1)}\lesssim
\big\|\big(\frac{w}{\sqrt{T}},\nabla w\big)\big\|_{\CC^\alpha(\Boule_1)}\lesssim
\big\|\big(\frac{w}{\sqrt{T}},\nabla w\big)\big\|_{\LL^q(\Boule_2)}.
\end{aligned}$$ The first estimate is trivial. For the second estimate, the presence of the massive term in requires a classical bootstrap argument that we outline for the convenience of the user. We split $w=w'+(w-w')$ once more according to $$\begin{aligned}
\label{cw05}
-\nabla\cdot a\nabla w'=-\frac{1}{T}w\;\;\mbox{in}\;\Boule_2,\quad
w'=0\;\;\mbox{on}\;\partial \Boule_2.
\end{aligned}$$ We have by CZ theory that $\|\nabla w'\|_{\LL^{q'}(\Boule_2)}\lesssim\|\frac{w}{T}\|_{\LL^q(\Boule_2)}$, where the prime on an exponent indicates the improvement coming from Sobolev’s embedding, that is, $$\begin{aligned}
q'=\frac{dq}{d-q}
\end{aligned}$$ (which we only use as long as $q<d$). For the second contribution $w-w'$, which by (\[cw07\]) and (\[cw05\]) satisfies the homogeneous equation $-\nabla\cdot a\nabla(w-w')=0$ in $\Boule_2$, we have by CZ-based inner regularity theory $\|\nabla(w-w')\|_{\LL^{q'}(\Boule_1)}\lesssim \|\nabla(w-w')\|_{\LL^q(\Boule_2)}$. Appealing to the triangle inequality and Hölder’s inequality in form of $\|\nabla w'\|_{\LL^q(\Boule_2)}$ $\lesssim\|\nabla w'\|_{\LL^{q'}(\Boule_2)}$, both estimates add up to $\|\nabla w\|_{\LL^{q'}(\Boule_1)}\lesssim\|\frac{w}{T}\|_{\LL^q(\Boule_2)}+\|\nabla w\|_{\LL^q(\Boule_2)}$. In combination with the Poincaré estimate $\|w\|_{\LL^{q'}(\Boule_1)}\lesssim \|\nabla w\|_{\LL^{q'}(\Boule_1)}+\|w\|_{\LL^q(\Boule_1)}$, and using $T\ge 1$, this yields the iterable form $$\begin{aligned}
\big\|\big(\frac{w}{\sqrt{T}},\nabla w\big)\big\|_{\LL^{q'}(\Boule_1)}\lesssim
\big\|\big(\frac{w}{\sqrt{T}},\nabla w\big)\big\|_{\LL^q(\Boule_2)}.
\end{aligned}$$ We now consider an exponent $p>d$ and set $$\begin{aligned}
\alpha'=1-\frac{d}{p}
\end{aligned}$$ (which we use only for $\alpha'\le\alpha$). In this situation, in the splitting introduced in (\[cw05\]), we may pass from CZ-theory to Schauder theory to obtain $\|\nabla w'\|_{\CC^{\alpha'}(\Boule_2)}$ $\lesssim\|\frac{w}{T}\|_{\LL^p(\Boule_2)}$ and $\|\nabla(w-w')\|_{\CC^{\alpha'}(\Boule_1)}\lesssim\|\nabla(w-w')\|_{\LL^q(\Boule_2)}$. This combines to $\|\nabla w\|_{\CC^{\alpha'}(\Boule_1)}$ $\lesssim\|\frac{w}{T}\|_{\LL^p(\Boule_r)}+\|\nabla w\|_{\LL^q(\Boule_r)}$, and as above upgrades to $$\begin{aligned}
\big\|\big(\frac{w}{\sqrt{T}},\nabla w\big)\big\|_{\CC^{\alpha'}(\Boule_1)}\lesssim
\big\|\big(\frac{w}{\sqrt{T}},\nabla w\big)\big\|_{\LL^q(\Boule_2)}.
\end{aligned}$$ In one final step of Schauder theory, applied to both contributions of the splitting (\[cw05\]), we obtain $$\begin{aligned}
\big\|\big(\frac{w}{\sqrt{T}},\nabla w\big)\big\|_{\CC^{\alpha}(\Boule_1)}\lesssim
\big\|\big(\frac{w}{\sqrt{T}},\nabla w\big)\big\|_{\CC^{\alpha'}(\Boule_2)}.
\end{aligned}$$ An inspection of the above arguments shows that the pair of balls $(\Boule_1,\Boule_2)$ may be replaced by $(\Boule_r,\Boule_R)$ with $1\le r<R\le 2$, at the expense of a constant that depends on $R-r$. Hence a finite iteration indeed yields (\[cw06\]).
For any radius $0<R\le 1$ and point $z\in\mathbb{R}^d$, the change of variables $x= R\hat x+z$, alongside with $\hat T=R^{-2}T\ge 1$, shows the statement of Step 1 generalizes to: Provided $$\begin{aligned}
\label{cw12}
R^\alpha[a]_{\CC^{\alpha}(\Boule_{2R}(z))}\le 1
\end{aligned}$$ we have $$\label{cw13}
\begin{aligned}
&\big\|\big(\frac{u}{\sqrt{T}},\nabla u\big)\big\|_{\LL^p(\Boule_R(z))}
\\
&\qquad\qquad\lesssim
R^{\frac{d}{p}-\frac{d}{q}}\big\|\big(\frac{u}{\sqrt{T}},\nabla u\big)\big\|_{\LL^q(\Boule_{2R}(z))}
+\|(\frac{1}{\sqrt{T}},f)\|_{\LL^p(\Boule_{2R}(z))},
\end{aligned}$$ and, in case of $g\equiv0$, $$\label{cw17}
\begin{aligned}
\big\|\big(\frac{u}{\sqrt{T}},\nabla u\big)\big\|_{\CC^{\alpha}(\Boule_R(z))}
\lesssim
R^{-\alpha-\frac{d}{q}}\big\|\big(\frac{u}{\sqrt{T}},\nabla u\big)\big\|_{\LL^q(\Boule_{2R}(z))}
+\|f\|_{\CC^{\alpha}(\Boule_{2R}(z))}.
\end{aligned}$$
Selecting $R\le\frac{1}{2}$ to be a small fraction of $\|a\|_{\CC^\alpha(\Boule_2)}^{-\frac{1}{\alpha}}\lesssim \lambda^{\frac{1}{\alpha}}$ (recall ), we may cover $\Boule_1$ by the union of $N$ balls $\Boule_R(z_n)$, $n=1,\cdots,N$, with $$\begin{aligned}
\label{cw14}
R^{-1}\sim\|a\|_{\CC^\alpha(\Boule_2)}^{\frac{1}{\alpha}}
\end{aligned}$$ in such a way that (\[cw12\]) is satisfied for every $z=z_n\in \Boule_1$. Taking the $\ell^p$-norm of (\[cw13\]) (with $z=z_n$) in $n=1,\cdots,N$, we obtain $$\begin{aligned}
\label{cw16}
\big\|\big(\frac{u}{\sqrt{T}},\nabla u\big)\big\|_{\LL^p(\Boule_1)}
\lesssim
R^{\frac{d}{p}-\frac{d}{q}}\big\|\big(\frac{u}{\sqrt{T}},\nabla u\big)\big\|_{\LL^q(\Boule_2)}
+\big\|\big(\frac{g}{\sqrt{T}},f\big)\big\|_{\LL^p(\Boule_2)},
\end{aligned}$$ where on the first r. h. s. term we used that (recall $p\ge q$) $$\begin{aligned}
\big(\sum_{n=1}^N\|f\|_{\LL^q(\Boule_{2R}(z_n))}^p\big)^\frac{1}{p}
\le\big(\sum_{n=1}^N\|f\|_{\LL^q(\Boule_{2R}(z_n))}^q\big)^\frac{1}{q}
\lesssim \|f\|_{\LL^q(\Boule_2)}.
\end{aligned}$$ Now (\[cw15\]) follows from inserting (\[cw14\]) into (\[cw16\]).
We now turn to the Schauder estimate . Using a telescoping sum we learn the norm relation $$\begin{aligned}
\|u\|_{\CC^\alpha(\Boule_1)}\lesssim R^{\alpha-1}\max_{n=1,\cdots,N}\|u\|_{\CC^\alpha(\Boule_R(z_n))}.
\end{aligned}$$ Hence taking the maximum of (\[cw17\]) (with $z=z_n$) in $n=1,\cdots,N$ yields $$\begin{aligned}
\big\|\big(\frac{u}{\sqrt{T}},\nabla u\big)\big\|_{\CC^{\alpha}(\Boule_1)}
\lesssim
R^{-1-\frac{d}{q}}\big\|\big(\frac{u}{\sqrt{T}},\nabla u\big)\big\|_{\LL^q(\Boule_{2})}
+R^{\alpha-1}\|f\|_{\CC^{\alpha}(\Boule_{2})}.
\end{aligned}$$ Finally, (\[cw18\]) follows from inserting (\[cw14\]) into the above estimate.
[^1]: Here, the Fourier transform is normalized in the following way: $\FF u (k):= \int \ee^{-\ii k x} u(x) \dd x$.\[fref\]
[^2]: The argument for this identity is formal, since $\nabla \phi$ is only defined for $\langl\cdot\rangl$-a.e. $a$, see Section \[SecAssumpGen\]. However, this is fully justified in [@GNO_2019_Correlatedfields p. 21], and we also refer to where we work with the massive corrector which is defined for every $a$.
[^3]: Indeed, for $d >3$, this follows from the energy identity $\int |\nabla \ubar|^2 = \fint_{\Boule_1(x)} \ubar - \fint_{\Boule_1(0)} \ubar$ via the Sobolev inequality $(\int \ubar^{\frac{2d}{d-2}} )^{\frac{d-2}{2d}} \lesssim ( \int |\nabla \ubar|^2 )^{\frac{1}{2}}$. In dimension $d=2$, we may appeal to the explicit representation $\nabla \ubar(y)=\fint_{\Boule_1} ( \nabla G(y-x-z)-\nabla G(y-z) ) \dd z$ via the fundamental solution $\nabla G(y):=\frac{1}{2\pi} \frac{y}{|y|^2}$.
[^4]: Even for $d=2$, such a configuration may be obtained by using the singularities induced by coefficients piecewise constant on sectors. In such a case (see [@Kellogg_1972]), there exist $a$-harmonic functions $u$ such that $\nabla u \in \LL^2(\R^d)$ but $\nabla u \notin \LL^p(\R^d)$ for $p \gg_\lambda 1$. These counterexamples are available for the continuum framework and their large-scale behavior transmits to the discrete setting by interpreting it as a discretization of the continuum one. The latter argument is in the spirit of [@Fischer_Otto_2015 Prop. 22].
[^5]: see footnote \[fref\] p. for the normalization of the Fourier transform
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We study quantum graphs corresponding to isotropic lattices with quasiperiodic coupling constants given by the same expressions as the coefficients of the discrete surface Maryland model. The absolutely continuous and the pure point spectra are described. It is shown that the transition between them is governed by the Hill operator corresponding to the edge potential.'
address: 'Laboratoire de Mathématiques, Université Paris Sud, Bâtiment 425, 91405 Orsay Cedex, France'
author:
- Konstantin Pankrashkin
title: |
Quasiperiodic surface Maryland models\
on quantum graphs
---
Introduction
============
The present paper is devoted to the spectral analysis of a special class of quasiperiodic interactions on quantum graphs. We are going to show how some the theory of discrete quasiperiodic operators can be transferred to the quantum graph case using the tools of the operator theory.
The paper [@GFP] provided the first explicit example of a difference quasiperiodic operator having a pure point spectrum dense everywhere; this operator is often referred to as the Maryland model. Later the class of such Hamiltonians was considerably extended in several directions, e.g. to the multidimensional case and to more general coefficients, see e.g. [@B1; @FP]. The papers [@BBP; @BBP1; @Kh] studied interactions similar to the Maryland model but supported by a subspace (surface Maryland model). In this case the quasiperiodic perturbation leaves unchanged the absolutely continuous spectrum of the unperturbed operator but produces a dense pure point spectrum of the rest of the real line.
On the other hand, discrete operators are closely related to the quantum graph models, i.e. differential operators acting on geometric configuration consisting of segments, see eg. [@Ku1; @Ku2]. The aim of the present paper is to provide an analog of the surface Maryland model for quantum graphs and to study its spectral properties. The work is a natural continuation of our previous paper [@KP2] where we considered full-space Maryland quantum graph model.
The model operator {#secmod}
==================
Let us describe first some basic constructions for quantum graphs. For a detailed discussion see e.g. [@GS; @Ku1; @Ku2]. There are many approaches to the study of the spectra of quantum graphs, we use the one from [@BGP; @KP1] based on the theory of self-adjoint extensions.
We consider a quantum graph whose set of vertices is identified with ${\mathbb{Z}}^d$, $d\ge 2$ (i.e. we explicitly need a multidimensional lattice). By ${{\mathbf h}}_j$, $j=1,\dots,d$, we denote the standard basis vectors of ${\mathbb{Z}}^d$. Two vertices ${{\mathbf m}}$, ${{\mathbf m}}'$ are connected by an oriented edge ${{\mathbf m}}\to {{\mathbf m}}'$ iff ${{\mathbf m}}'={{\mathbf m}}+{{\mathbf h}}_j$ for some $j\in\{1,\dots,d\}$; this edge is denoted as $({{\mathbf m}},j)$ and one says that ${{\mathbf m}}$ is the initial vertex and ${{\mathbf m}}'\equiv {{\mathbf m}}+{{\mathbf h}}_j$ is the terminal vertex.
Replace each edge $({{\mathbf m}},j)$ by a copy of the segment $[0,1]$ in such a way that $0$ is identified with ${{\mathbf m}}$ and $1$ is identified with ${{\mathbf m}}+{{\mathbf h}}_j$. In this way we arrive at a certain topological set carrying a natural metric structure. The quantum state space of the system is $${\mathcal{H}}:=\bigoplus_{({{\mathbf m}},j)\in{\mathbb{Z}}^d\times\{1,\dots,d\}} {\mathcal{H}}_{{{\mathbf m}},j},\quad
{\mathcal{H}}_{{{\mathbf m}},j}={\mathcal{L}}^2[0,1],$$ and the vectors $f\in{\mathcal{H}}$ will be denoted as $f=(f_{{{\mathbf m}},j})$, $f_{{{\mathbf m}},j}\in{\mathcal{H}}_{{{\mathbf m}},j}$, ${{\mathbf m}}\in{\mathbb{Z}}^d$, $j=1,\dots,d$.
Let us introduce a Schrödinger operator acting in ${\mathcal{H}}$. Fix a real-valued potential $q\in {\mathcal{L}}^2[0,1]$ and some real constants $\alpha({{\mathbf m}})$, ${{\mathbf m}}\in{\mathbb{Z}}^d$. Set $A:={\mathop{\mathrm{diag}}}\big(\alpha({{\mathbf m}})\big)$; this is a self-adjoint operator in $\ell^2({\mathbb{Z}}^d)$. Denote by $H_A$ the operator acting as
\[eq-sch\] $$\label{eq-act}
(f_{{{\mathbf m}},j})\mapsto \Big(
-f''_{{{\mathbf m}},j}+q f_{{{\mathbf m}},j}\Big),$$ on functions $f=(f_{{{\mathbf m}},j})\in\bigoplus_{{{\mathbf m}},j} H^2[0,1]$ satisfying the following boundary conditions: $$\label{eq-cont1}
f_{{{\mathbf m}},j}(0)=f_{{{\mathbf m}}-{{\mathbf h}}_k,k}(1)=:f({{\mathbf m}}),\quad j,k=1,\dots,d,\quad {{\mathbf m}}\in{\mathbb{Z}}^d,$$ (which means the continuity at all vertices) and $$f'({{\mathbf m}})=\alpha({{\mathbf m}})f({{\mathbf m}}),\quad {{\mathbf m}}\in{\mathbb{Z}}^d,$$ where $$f'({{\mathbf m}}):= \sum_{j=1}^d f'_{{{\mathbf m}},j}(0)-\sum_{j=1}^d f'_{{{\mathbf m}}-h_j,j}(1).$$
The constants $\alpha({{\mathbf m}})$ are usually referred to as *Kirchhoff coupling constants* and interpreted as the strengths of zero-range impurity potentials placed at the corresponding vertices. The zero coupling constants correspond hence to the ideal couplings and are usually referred to as the standard boundary conditions.
We are going to study the above operator $H_A$ for a special choice of the coefficients $\alpha({{\mathbf m}})$. Namely, take $d_1\in\{1,\dots, d-1\}$ and set $d_2:=d-d_1$. In what follows one represents any ${{\mathbf m}}\in{\mathbb{Z}}^d$ as ${{\mathbf m}}=({{\mathbf m}}_1,{{\mathbf m}}_2)$ with ${{\mathbf m}}_1\in{\mathbb{Z}}^{d_1}$ and ${{\mathbf m}}_2\in{\mathbb{Z}}^{d_2}$
Pick $g\ne 0$, ${{\boldsymbol \omega}}\in{\mathbb{R}}^{d_2}$, $\varphi\in{\mathbb{R}}$ with $$\label{eq-vf1}
\varphi\ne {{{\boldsymbol \omega}}{{\mathbf m}}_2} \mod \dfrac12, \quad {{\mathbf m}}_2\in{\mathbb{Z}}^{d_2}$$ and set $$\alpha({{\mathbf m}}):=g\tan\pi \big({{\boldsymbol \omega}}{{\mathbf m}}_2+\varphi\big),\quad m\in{\mathbb{Z}}^d.$$ This operator will be noted simply by $H$.
To formulate the results we need some additional constructions. Denote by $s$ and $c$ the solutions to $-y''+q y=zy$ satisfying $s(0;z)=c'(0;z)=0$ and $s'(0;z)=c(0;z)=1$, $z\in{\mathbb{C}}$, and set $\eta(z):=s(1;z)+c'(1;z)$. Consider an auxiliary one-dimensional Hill operator $$\label{eq-hill}
L=-\dfrac{d^2}{dx^2}+Q,\quad Q(x+n)=q(x), \quad (x,n)\in[0,1)\times {\mathbb{Z}}.$$ It is known that ${\mathop{\mathrm{spec}}}L=\eta^{-1}\big([-2,2]\big)$.
For any ${{\boldsymbol \omega}}$ and $\varphi$ one has ${\mathop{\mathrm{spec}}}L\subset {\mathop{\mathrm{spec}}}H$. If the components of ${{\boldsymbol \omega}}$ are rationally independent, then the spectrum of $H$ in $\eta^{-1}\big((-2,2)\big)$ is purely absolutely continuous. If ${{\boldsymbol \omega}}$ satisfies additionally the Diophantine condition $$\label{eq-vf2}
\text{there are }C,\beta>0 \text{ with }|{{\boldsymbol \omega}} {{\mathbf m}}_2-r|\ge C|{{\mathbf m}}_2|^{-\beta}
\text{ for all }{{\mathbf m}}_2\in{\mathbb{Z}}^{d_2}\setminus\{0\},\, r\in{\mathbb{Z}},$$ then the spectrum of $H$ outside ${\mathop{\mathrm{spec}}}L$ is dense pure point.
The above theorem is a combination of propositions \[spec1\], \[spec2\], \[spec3\], \[spec4\] whose proof will be given below.
As easily seen, the location of the absolutely continuous spectrum of $H$ is completely determined by the spectrum of the periodic operator $L$, and (under some additional assumptions) the rest of the spectrum is pure point. It is interesting to mention that a similar interlaced spectrum was found recently in a completely different model involving singular potentials [@EF].
Some construction from the theory of self-adjoint extensions {#sec-ext}
============================================================
In this section we recall the operator-theoretical machinery which will be used to study the spectrum of $H$. For detailed discussion we refer to [@BGP Section 1].
Let $S$ be a closed linear operator in a separable Hilbert space ${\mathcal{H}}$ with the domain ${\mathop{\mathrm{dom}}}S$. Assume that there exist an auxiliary Hilbert space ${\mathcal{G}}$ and two linear maps $\Gamma,\Gamma':{\mathop{\mathrm{dom}}}S\to {\mathcal{G}}$ such that
- for any $f,g\in{\mathop{\mathrm{dom}}}S$ there holds $\langle f,Sg\rangle-\langle Sf,g\rangle=\langle \Gamma f,\Gamma' g\rangle-\langle\Gamma'f,\Gamma g\rangle$,
- the map $(\Gamma,\Gamma'):{\mathop{\mathrm{dom}}}S\to{\mathcal{G}}\oplus{\mathcal{G}}$ is surjective,
- the set $\ker\,(\Gamma,\Gamma')$ is dense in ${\mathcal{H}}$.
A triple $({\mathcal{G}},\Gamma, \Gamma')$ with the above properties is called a *boundary triple* for $S$. If $S^*$ is symmetric, boundary triples deliver an effective description of all self-adjoint restrictions of $S$. If $A$ is a self-adjoint operator in ${\mathcal{G}}$, then the restriction of $S$ to the vectors $f$ satisfying the abstract boundary conditions $\Gamma' f=A\Gamma f$ is a self-adjoint operator in ${\mathcal{H}}$, which we denote by $H_A$. Another example is the “distinguished” restriction $H^0$ corresponding to the boundary conditions $\Gamma f=0$. One can show that a self-adjoint restriction $H'$ of $S$ can be represented as $H_A$ with a suitable $A$ iff ${\mathop{\mathrm{dom}}}H'\cap {\mathop{\mathrm{dom}}}H^0= {\mathop{\mathrm{dom}}}S^*$; such restrictions are called *disjoint to $H^0$*. The resolvents of $H^0$ and $H_A$ as well as their spectral properties are connected by Krein’s resolvent formula, which will be described now.
Let $z\notin{\mathop{\mathrm{spec}}}H^0$. For $g \in {\mathcal{G}}$ denote by $\gamma(z) g$ the unique solution to the abstract boundary value problem $(S-z)f=0$ with $\Gamma f=g$. Clearly, $\gamma(z)$ is a linear map from ${\mathcal{G}}$ to ${\mathcal{H}}$ and an isomorphism between ${\mathcal{G}}$ and $\ker(S-z)$; it is sometimes referred to as the Krein $\gamma$-field associated with the boundary triple. Denote also by $M(z)$ the bounded linear operator on ${\mathcal{G}}$ given by $M(z)g=\Gamma' \gamma(z)g$; this operator will be referred to as the Weyl function (or the abstract Dirichlet-to-Neumann map) corresponding to the boundary triple $({\mathcal{G}},\Gamma,\Gamma')$. The operator-valued functions $\gamma$ and $M$ are analytic outside ${\mathop{\mathrm{spec}}}H^0$, and $M(z)$ is self-adjoint for real $z$. If these maps are known, one can relate the operators $H_A$ and $H^0$ as follows,:
\[krein\] For $z\notin {\mathop{\mathrm{spec}}}H^0\cup{\mathop{\mathrm{spec}}}H_A$ the operator $M(z)-A$ acting on ${\mathcal{G}}$ has a bounded inverse defined everywhere, and $$\label{eq-krein}
(H_A-z)^{-1}=(H^0-z)^{-1}-\gamma(z)\big(M(z)-A\big)^{-1}\gamma(\bar z)^*.$$ In particular, the set ${\mathop{\mathrm{spec}}}H_A\setminus{\mathop{\mathrm{spec}}}H^0$ consists exactly of $z\in{\mathbb{R}}\setminus{\mathop{\mathrm{spec}}}H^0$ such that $0\in{\mathop{\mathrm{spec}}}\big(M(z)-A\big)$. The same correspondence holds for the eigenvalues, i.e. $z\in {\mathbb{R}}\setminus{\mathop{\mathrm{spec}}}H^0$ is an eigenvalue of $H_A$ iff $0$ is an eigenvalue of $M(z)-A$, and $\gamma(z)$ is an isomorphism of the corresponding eigensubspaces.
The maps $\gamma$ and $M$ satisfy a number of important properties. In particular, $\gamma$ and $M$ depend analytically on their argument (outside of ${\mathop{\mathrm{spec}}}H^0$), $M(z)$ satisfies $M(\bar z)=M(z)^*$ and $$\begin{gathered}
\label{eq-meg}
\text{for any non-real $z$ there is $c_z>0$ with}
\dfrac{\Im M(z)}{\Im z}\ge c_z, \text{ and}\\
\label{eq-meg2}
M'(\lambda)=\gamma(\lambda)^*\gamma(\lambda)>0 \text{ for } \lambda\in{\mathbb{R}}\setminus{\mathop{\mathrm{spec}}}H^0.\end{gathered}$$ Furthermore, $$\label{eq-gz}
\gamma(z)^* f=0 \text{ for any }f\in\ker(S-z)^\perp\equiv{\mathop{\mathrm{ran}}}\gamma(z)^\perp,$$ see [@BGP Section 1] for more details.
Below it will be useful to have a certain relationship between the resolvent of $H_A$ and that of the operator $H_0$ (i.e. $H_A$ with $A=0$). Clearly, $({\mathcal{G}},{\widetilde}\Gamma,{\widetilde}\Gamma')$ with ${\widetilde}\Gamma:=\Gamma'$ and ${\widetilde}\Gamma':=-\Gamma$ is a new boundary triple for $S$. With respect to this boundary triple $H_0$ is the distinguished extension, and one can easily calculate (at least for non-real $z$) the corresponding Krein $\gamma$-field ${\widetilde}\gamma(z):=\gamma(z) M(z)^{-1}$ and the Weyl function ${\widetilde}M(z):=-M(z)^{-1}$ which then extend by the analyticity to ${\mathbb{R}}\setminus{\mathop{\mathrm{spec}}}H_0$. The operator $H_A$ corresponds now to the boundary conditions: ${\widetilde}\Gamma f =0$ for $\Gamma f\in\ker A$ and ${\widetilde}\Gamma' f =-A^{-1} {\widetilde}\Gamma f$ otherwise. Hence, if the operator $A$ is not invertible, one cannot use the above proposition \[krein\] for the resolvents. One can avoid this difficulty as follows. Denote ${\mathcal{G}}':=\ker A^{\perp}$; clearly, ${\mathcal{G}}'$ is a closed linear subspace of ${\mathcal{G}}$. Denote by $P$ the orthogonal projection from ${\mathcal{G}}$ to ${\mathcal{G}}'$ and set $\Pi:=P{\widetilde}\Gamma P$ and $\Pi':=P{\widetilde}\Gamma' P$ and $S':=S|_{{\widetilde}\Gamma^{-1}({\mathcal{G}}')}$, then $({\mathcal{G}}',\Pi,\Pi')$ is a boundary triple for $S'$ with the Krein field $\nu(z):=P{\widetilde}\gamma(z)P$ and the Weyl function $N(z)=P{\widetilde}M(z)P$, and $H_A$ and $H_0$ become disjoint self-adjoint restrictions of $S'$. Hence, one has the following resolvent formula $$\label{eq-nz}
(H_0-z)^{-1}-(H_A-z)^{-1}=\nu(z)\big(N(z)-B\big)\nu^*(\Bar z), \quad B=-PA^{-1}P,$$ or $$\begin{gathered}
(H_0-z)^{-1}-(H_A-z)^{-1}\\
=P\gamma(z) M(z)^{-1}\big(PA^{-1}P-PM(z)^{-1}P\big)M(z)^{-1}P\gamma^*(\Bar z).\end{gathered}$$
Resolvents for the quantum graph
================================
We are going to show now how the constructions of section \[sec-ext\] apply to the quantum graph Hamiltonian $H$ (see Section \[secmod\]).
Denote by $S$ the operator acting as on the functions $f$ satisfying only the boundary conditions . On the domain of $S$ one can define linear maps $$f\mapsto \Gamma f:= \big(f({{\mathbf m}})\big)_{{{\mathbf m}}\in{\mathbb{Z}}^d}\in \ell^2({\mathbb{Z}}^d),\quad
f\mapsto \Gamma' f:= \big(f'({{\mathbf m}})\big)_{{{\mathbf m}}\in{\mathbb{Z}}^d}\in \ell^2({\mathbb{Z}}^d).$$ One can show that $({\mathbb{Z}}^d,\Gamma,\Gamma')$ form a boundary triple for $S$. The distinguished restriction to $\ker\Gamma$, $H^0$, acts as on functions $(f_{{{\mathbf m}},j})$ with $f_{{{\mathbf m}},j}\in H^2[0,1]$ satisfying the Dirichlet boundary conditions, $f_{{{\mathbf m}},j}(0)=f_{{{\mathbf m}},j}(1)=0$ for all $m,j$, and the spectrum of $H^0$ is just the Dirichlet spectrum of $-\dfrac{d^2}{dt^2}+q$ on the segments $[0,1]$; we will refer to ${\mathop{\mathrm{spec}}}H^0$ as to the *Dirichlet spectrum* of the graph.
Let us construct the maps $\gamma(z)$ and $M(z)$ for the above boundary triple. In in terms of these functions $s$ and $c$ one has obviously $$\begin{gathered}
\big(\gamma(z)\xi\big)_{{{\mathbf m}},j}(t)=
\dfrac{1}{s(1;z)}\Big(\xi({{\mathbf m}}+{{\mathbf h}}_j)s(t;z)\\
+\xi({{\mathbf m}})\big(
s(1;z) c(t;z)-c(1;z)s_j(t;z)
\big)\Big),\\
t\in[0,1],\quad ({{\mathbf m}},j)\in{\mathbb{Z}}^d\times\{1,\dots,d\}.\end{gathered}$$ The corresponding Weyl function $M(z):\ell^2({\mathbb{Z}}^d)\to \ell^2({\mathbb{Z}}^d)$ is given by $$M(z)\xi({{\mathbf m}})=
\dfrac{1}{s(1;z)} \sum_{j=1}^d \cdot\Big( \xi({{\mathbf m}}-{{\mathbf h}}_j)+\xi({{\mathbf m}}+{{\mathbf h}}_j)
-\eta(z)\,\xi({{\mathbf m}})\Big),\quad \xi\in\ell^2({\mathbb{Z}}^d),$$ where $\eta(z):=c(1;z)+s'(1;z)$ is the Hill discriminant associated with the potential $q$. It is useful to introduce the discrete Hamiltonian $\Delta_d$ in $\ell^2({\mathbb{Z}}^d)$ by $$\Delta_d \xi({{\mathbf m}})=\sum_{{{\mathbf m}}':|{{\mathbf m}}- {{\mathbf m}}'|=1} \xi({{\mathbf m}}'),$$ then one has obviously $$\label{eq-lapl}
M(z):= a(z) \big(\Delta_d-d \eta(z)\big), \quad a(z):=\dfrac{1}{s(1;z)}.$$ By proposition \[krein\], outside of the discrete set ${\mathop{\mathrm{spec}}}H^0$, the spectrum of $H$ consists of the real $z$ satisfying $0\in{\mathop{\mathrm{spec}}}\big( M(z)-A\big)$. It is important to emphasize that, for real $z$, the operator $ M(z)-A$ is exactly the surface Maryland model studied in [@BBP]. The results of [@BBP1] imply that
- the operator $M(z)-A$ has no bounded inverse if $\big|\eta(z)\big|\le 2$,
- if the components of ${{\boldsymbol \omega}}$ are rationally independent, then the spectrum in the interval $a(z)\big(2d-\eta(z),2d+\eta(z) \big)$ is purely absolutely continuous,
- for Diophantine ${{\boldsymbol \omega}}$ the rest of the real line is covered by the dense pure point spectrum.
Using first of these properties and proposition \[krein\] one immediately obtains
\[spec1\] ${\mathop{\mathrm{spec}}}L\subset {\mathop{\mathrm{spec}}}H$.
(It is sufficient to recall that the set $\big|\eta(z)\big|\le 2$ coincides with spectrum of $L$.) Nevertheless, one is not able to conclude about the spectral nature of $H$ from that of $M(z)-A$ using just the general theory of self-adjoint extensions [@BrMN]. We are going to use some additional considerations from [@BBP1; @FP] in order to understand completely the spectral properties of $H$.
We will also use the formula relating the resolvent of $H_A$ and $H_0$. Denoting by $P$ the orthogonal projection from $\ell^2({\mathbb{Z}}^d)$ to $\ell^2({\mathbb{Z}}^{d_2})$ on obtains $N(z)=-PM(z)P$. The parameter operator $B:=-(PAP)^{-1}$ is the multiplication by $-g^{-1}\cot\pi({{\boldsymbol \omega}} {{\mathbf m}}_2+\varphi)\equiv g^{-1}\tan\pi({{\boldsymbol \omega}} {{\mathbf m}}_2+\varphi+1/2)$. It is also important to emphasize that, as shown in [@KP1], one has ${\mathop{\mathrm{spec}}}H_0={\mathop{\mathrm{spec}}}L\cup{\mathop{\mathrm{spec}}}H^0$ where $L$ is the one-dimensional Hill operator , and each point of ${\mathop{\mathrm{spec}}}H^0$ is an infinitely degenerate eigenvalue.
The absolutely continuous spectrum
==================================
Below we will use actively the Fourier transform. Denote ${\mathbb{S}}^1:=\{z\in{\mathbb{C}},\,|z|=1\}$ and ${\mathbb{T}}^n:=\underbrace{{\mathbb{S}}^1\times\dots\times{\mathbb{S}}^1}_{n \text{ times}}\subset{\mathbb{C}}^n$. For ${{\boldsymbol \theta}}=(\theta_1,\dots,\theta_n)\subset{\mathbb{C}}^n$ and ${{\mathbf p}}=(p_1,\dots,p_n)\in{\mathbb{Z}}^n$ we write ${{\boldsymbol \theta}}^{{\mathbf p}}:=\theta_1^{p_1}\dots\theta_n^{p_n}$, and in this context $k\in{\mathbb{Z}}$ will be identified with the vector $(k,\dots,k)\in{\mathbb{Z}}^n$, i.e. ${{\boldsymbol \theta}}^{-1}:=\theta_1^{-1}\dots\theta_l^{-1}$ etc. We denote by $F_n$ the Fourier transform carrying $\ell^2({\mathbb{Z}}^n)$ to ${\mathcal{L}}^2({\mathbb{T}}^n)$, $$F_n\psi({{\boldsymbol \theta}})=\sum_{{{\mathbf n}}\in{\mathbb{Z}}^n} \psi({{\mathbf n}}) {{\boldsymbol \theta}}^{{\mathbf n}},\quad
F_n^{-1}f({{\mathbf n}})=\dfrac{1}{(2\pi i)^n} \int_{{\mathbb{T}}^n} f({{\boldsymbol \theta}}) {{\boldsymbol \theta}}^{-{{\mathbf n}}-1}d{{\boldsymbol \theta}}.$$ Each ${{\boldsymbol \theta}}\in{\mathbb{T}}^d$ will be represented ${{\boldsymbol \theta}}=({{\boldsymbol \theta}}_1,{{\boldsymbol \theta}}_2)$ with ${{\boldsymbol \theta}}_1\in{\mathbb{T}}^{d_1}$ and ${{\boldsymbol \theta}}_2\in{\mathbb{T}}^{d_2}$.
We will repeat first some algebraic manipulations in the spirit of [@BBP]. Without loss of generality assume $g>0$ (otherwise one can change the signs of ${{\boldsymbol \omega}}$ and $\varphi$).
Consider the operator $L(z):=M(z)-A=M(z)+PvP$, where $v$ is an operator in $\ell^2({\mathbb{Z}}^{d_2})$ acting as $v f({{\mathbf m}}_2)=-g\tan({{\boldsymbol \omega}}{{\mathbf m}}_2+\varphi)f({{\mathbf m}}_2)$, ${{\mathbf m}}_2\in{\mathbb{Z}}^{d_2}$. For $\Im z\ne 0$ the operator $M(z)$ is invertible (as its imaginary part is non-degenerate) and one has $$L(z)^{-1}=M(z)^{-1}-M(z)^{-1} T(z) M(z)^{-1}, \quad
T(z)=v-T(z) M(z)^{-1} v.$$ Obviously one can write $T(z)=P t(z)P$ where the operator $t(z)$ acting in $\ell^2({\mathbb{Z}}^{d_2})$ satisfies $t(z)=v+t(z) N(z) v$. Formally one has $t(z)=v(1-N(z) v)^{-1}$, and it is needed to show that the operator in question is really invertible.
Let $U$ be the unitary operator in $\ell^2({\mathbb{Z}}^{d_2})$ defined by the relation $$(u f)({{\mathbf m}}_2)=e^{-2\pi i {{\boldsymbol \omega}} {{\mathbf m}}_2} f({{\mathbf m}}_2),$$ then, denoting $\chi:=e^{-2\pi i \varphi}$, one can write $$v=-\dfrac{g}{i} \dfrac{1-\chi U}{1+\chi U}.$$ As $\Im N(z)\ge 0$ for $\Im z\ge 0$, the operator $i+g N(z)$ is invertible for such $z$. Hence, for $\Im z\ge 0$ after a simple algebra one obtains $$1-N(z) v = \big(g N(z)+i\big)(1-b(z)\chi U)\big( i (1+\chi U)\big)^{-1}$$ where $b(z)=\big(g N(z)-i\big)\big(i+g N(z)\big)^{-1}$. In order to represent the inverse operator in terms of the Neumann series it is sufficient to show that $|b(z)|<1$ for some $z$. To see this, it is useful to pass to the Fourier representation.
For $\lambda\in{\mathbb{C}}$ denote $G_d(\lambda):=(\Delta_d-\lambda)^{-1}$. Recall that in the Fourier representation $\Delta_d$ becomes the multiplication by the function $\Delta_d({{\boldsymbol \theta}})=\sum_j (\theta_j+\theta_j^{-1})$, hence the matrix of $G_d(\lambda)$ is given by $$G_d({{\mathbf m}}-{{\mathbf m}}';\lambda)=\dfrac{1}{(2\pi i)^d}\int_{{\mathbb{T}}^d} \dfrac{\theta^{-({{\mathbf m}}-{{\mathbf m}}')-1} d{{\boldsymbol \theta}}}{\Delta_d({{\boldsymbol \theta}})-\lambda}.$$ On the other hand, the matrix of the operator $N(z)$ is $N({{\mathbf m}}_2-{{\mathbf m}}'_2;z)=-a(z)^{-1}G_d\big((0,{{\mathbf m}}_2)-(0,{{\mathbf m}}'_2);d\eta(z)\big)$, hence $$\begin{gathered}
\label{eq-nzint}
N({{\mathbf m}}_2-{{\mathbf m}}'_2;z)=-a(z)^{-1}\dfrac{1}{(2\pi i)^d}
\int_{{\mathbb{T}}^{d_2}} {{\boldsymbol \theta}}_2^{-({{\mathbf m}}_2-{{\mathbf m}}'_2)-1} d{{\mathbf \theta}}_2
\int_{{\mathbb{T}}^{d_1}} \dfrac{{{\boldsymbol \theta}}_1^{-1} d{{\boldsymbol \theta}}_1}{\Delta_d({{\boldsymbol \theta}})-d\eta(z)}\\
=-a(z)^{-1}
\dfrac{1}{(2\pi i)^{d_2}}
\int_{{\mathbb{T}}^{d_2}} G_{d_1}\big({{\mathbf 0}};d\eta(z)-\Delta_{d_2}({{\boldsymbol \theta}}_2)\big){{\boldsymbol \theta}}_2^{-({{\mathbf m}}_2-{{\mathbf m}}'_2)-1} d{{\boldsymbol \theta}}_2.\end{gathered}$$ In particular, it is clear that in the Fourier representation $N(z)$ is the multiplication by the function $$\label{eq-nfun}
N({{\boldsymbol \theta}}_2;z)=-a(z)^{-1}G_{d_1}\big({{\mathbf 0}};d\eta(z)-\Delta_{d_2}({{\boldsymbol \theta}}_2)\big).$$ As $\Im N(z)>0$ for $\Im z>0$, the imaginary part $\Im N({{\boldsymbol \theta}}_2;z)$ is positive for such $z$. The operator $b(z)$ in the Fourier representation becomes the multiplication by the function $$b({{\boldsymbol \theta}}_2,z)=\dfrac{g N({{\boldsymbol \theta}}_2,z)-i}{g N({{\boldsymbol \theta}}_2,z)+i},$$ hence $\|b(z)\|\equiv \sup_{{{\boldsymbol \theta}}_2} |b({{\boldsymbol \theta}}_2,z)|<1$ for $\Im z>0$. Therefore, one can represent $$\begin{gathered}
t(z)=v(1-N(z) v)^{-1}\\
=-g(1-\chi U)\big(1-b(z)\chi U\big)^{-1}\big(gN(z)+i\big)^{-1}\\
=-g(1-\chi U) \sum_{m=0}^\infty \chi^m \big(b(z)U\big)^m\\
=-g\big(gN(z)+i\big)^{-1}\Big(1-
2i \sum_{m=1}^\infty \big(gN(z)+i\big)^{-1} U \big(b(z)U\big)^{m-1}\Big).\end{gathered}$$ and one has $$\big(M(z)-A\big)^{-1}=M(z)^{-1} -M(z)^{-1} P t(z) P M(z)^{-1}.$$ After these preparations we can prove
\[spec2\] Denote $I:=\eta^{-1}\big((-2,2)\big)$. If the vector ${{\boldsymbol \omega}}$ has rationally independent components, then the operator $H$ has only absolutely continuous spectrum in $I$.
According to the general spectral theory we need to show that there exists a dense subset ${\mathcal{L}}$ of ${\mathcal{H}}$ such that the limit $\lim_{\varepsilon\to 0+}\Im\langle f, (H-\lambda-i\varepsilon)^{-1} f\rangle$ exists and is finite for all $g\in{\mathcal{L}}$ and $\lambda\in I.$
Represent $$\label{eq-h01}
{\mathcal{H}}={\mathcal{H}}_0+{\mathcal{H}}_1,\quad
{\mathcal{H}}_0:=\Big(\bigcup_{\Im z \ne 0} \gamma(z)\big(\ell^2({\mathbb{Z}}^d)\big)\Big)^\perp,
\quad
{\mathcal{H}}_1:={\mathcal{H}}_0^\perp;$$ in other words, ${\mathcal{H}}_1$ is the closure of the linear hull of the set $\{\gamma(z)\varphi:\, \Im z\ne 0,\,\varphi\in\ell^2({\mathbb{Z}}^d)\}$.
By the Krein resolvent formula, for any $f\in{\mathcal{H}}_0$ and any $z$ with $\Im z\ne 0$ one has $\gamma^*(z)f=0$. Hence, by , there holds $(H-z)^{-1}f=(H^0-z)^{-1}f$, hence $\lim_{\varepsilon\to 0+}\Im\langle f, (H-\lambda-i\varepsilon)^{-1} f\rangle=0$n because $(H^0-\lambda)^{-1}$ is a bounded self-adjoint operator.
Consider the vectors $f=\gamma(\zeta)h$ for $h=\big(M(\zeta)-A\big)^{-1}\xi$, $\Im\zeta\ne0$. These vectors form a dense subset in ${\mathcal{H}}_1$ as $\xi$ runs over a dense subset of $\ell^2({\mathbb{Z}}^d)$. By elementary calculations (see e.g. section 3 in [@BGP]) one can write $$(H-\lambda-i\varepsilon)^{-1}f=
\dfrac{1}{\zeta-\lambda-i\varepsilon}\,\Big(
f-\gamma(\lambda+i\varepsilon) \big(M(\lambda+i\varepsilon)-A\big)^{-1}\xi\Big).$$ Hence it is sufficient to show that $\lim_{\varepsilon\to 0+}\Im\langle \delta_{{{\mathbf m}}}, (M(\lambda+i\varepsilon)-A)^{-1} \delta_{{{\mathbf m}}}\rangle$ exists and is finite for any ${{\mathbf m}}\in{\mathbb{Z}}^d$. In view of the series representation for $(M(z)-A)^{-1}$ it is sufficient to show that the series converges for real $z\in I$ and not only for $\Im z>0$. On the other hand, $(M(z)-A)^{-1}=a(z)^{-1}(\Delta_d-d\eta(z)-a(z)^{-1}A)^{-1}$, and it is shown in [@BBP1 Theorem 3.1] that $\lim_{\varepsilon\to 0+}\Im\langle \delta_{{{\mathbf m}}}, (\Delta_d-b-c A)^{-1} \delta_{{{\mathbf m}}}\rangle$ exists and is finite for any $b\in(-2d,2d)$ and any $c\in{\mathbb{R}}$. This completes the proof.
The pure point spectrum
=======================
In this section we will use the second version of the resolvent formula, Eq. . Hence for $z\notin{\mathop{\mathrm{spec}}}H_0$ we have the equivalence $z\in {\mathop{\mathrm{spec}}}H$ iff $0\in {\mathop{\mathrm{spec}}}(N(z)-B)$. Here $N$ is a translationally invariant operator in $\ell^2({\mathbb{Z}}^{d_2})$ whose matrix elements are given by , and the operator $B$, as already mentioned below, in the multiplication by the sequence $g^{-1}\tan\pi({{\boldsymbol \omega}} {{\mathbf m}}_2+\varphi+1/2)$. It is useful to set $g'=-g$, ${{\boldsymbol \omega}}':=-{{\boldsymbol \omega}}$, $\varphi':=-\varphi-1/2$, then $B$ becomes a multiplication by $-g'\tan\pi({{\boldsymbol \omega}}' {{\mathbf m}}_2+\varphi')$ with $g'>0$.
To alleviate the notation, below we will write $d$ instead of $d_2$ and drop the indices for $g'$, ${{\boldsymbol \omega}}'$ and $\varphi'$ as this does not lead to confusions.
Introduce the operators $$D(z):=\big(N(z)-ig\big)^{-1}, \quad
C(z):=-\big(N(z)+ig\big)\big(N(z)-ig\big)^{-1};$$ they are defined at least for $z$ with $\Re z\notin{\mathop{\mathrm{spec}}}H_0$ and $|\Im z|$ sufficiently small. One can write for such $z$ the identity $$\label{eq-MA}
N(z)-B=D(z)^{-1}\big(1-\chi C(z)U\big)(1+\chi U)^{-1} .$$ Recall that under the Fourier transform $N(z)$ becomes the multiplication by the function $N({{\boldsymbol \theta}},z)$ given by , the operators $D(z)$ and $C(z)$ become the multiplications by $D({{\boldsymbol \theta}},z):=\big(N({{\boldsymbol \theta}},z)-ig\big)^{-1}$ by $C({{\boldsymbol \theta}},z):=-\big(N({{\boldsymbol \theta}},z)+ig\big)\big(N({{\mathbf \theta}},z)-ig\big)^{-1}$, respectively, and $U$ becomes a shift operator, $U k({{\boldsymbol \theta}})=k(e^{-2\pi i \omega_1}\theta_1,\dots,e^{-2\pi i \omega_d}\theta_d)$.
Consider an arbitrary segment $[a,b]\subset{\mathbb{R}}\setminus{\mathop{\mathrm{spec}}}H_0$. Recall that the spectrum of $H_0$ coincides with the spectrum of $L$ up to the discrete set ${\mathop{\mathrm{spec}}}H^0$. Eq. , the analyticity of $\gamma$, and the self-adjointness of $N(z)$ for real $z$ imply the existence of $\delta'>0$ such that $\|\Im N(z)\|\le g/2$ for $z\in Z:=\{z\in{\mathbb{C}}:\, |\Im z|\le\delta',\ \Re z\in[a,b]\}$. At the same time, this means that $|\Im N({{\boldsymbol \theta}},z)|\le g/2$ for $z\in Z$. As follows from the integral representation, $N({{\boldsymbol \theta}},z)$ can be continued to an analytic function in $Z\times \Theta$, $\Theta:=\{{{\boldsymbol \theta}}\subset {\mathbb{C}}^d: r<|\theta_j|<R\}$, $0<r<1<R<\infty$. Choosing $r$ and $R$ sufficiently close to $1$ one immediately sees that the function $$C({{\boldsymbol \theta}},z):=\dfrac{g^2-\big(\Im N({{\boldsymbol \theta}},z)\big)^2-\big(\Re N({{\boldsymbol \theta}},z)\big)^2-2ig\Re N({{\boldsymbol \theta}},z)}{|N({{\boldsymbol \theta}},z)-ig\big|^2}$$ does not take values in $(-\infty,0)$ for $({{\boldsymbol \theta}},z)\in \Theta\times Z$. Therefore, the function $f({{\boldsymbol \theta}},z):=\log C({{\boldsymbol \theta}},z)$ is well-defined and analytic in $\Theta\times Z$, where $\log$ denotes the principal branch of the logarithm. The Diophantine property implies (see [@FP Lemma 3.2]) that the operator $1-U$ is a bijection on the set of functions $v$ analytic in $\Theta$ with $$\int_{{\mathbb{T}}^d} v({{\boldsymbol \theta}}){{\boldsymbol \theta}}^{-1}d{{\boldsymbol \theta}}=0.$$ Hence the function $t({{\boldsymbol \theta}},z):=(1-U)^{-1}\big(f({{\boldsymbol \theta}},z)-f_0(z)\big)$ is well-defined and analytic in $Z\times\Theta$, where $$\label{eq-f0}
f_0(z):=
\dfrac{1}{(2\pi i)^d}\int_{{\mathbb{T}}^d} f({{\boldsymbol \theta}},z){{\boldsymbol \theta}}^{-1}d{{\boldsymbol \theta}}.$$
\[prop-1\] The function $f_0$ is analytic in $Z$, $$\begin{gathered}
\label{eq-f01}
\Re f_0(z)<0 \quad \text{for} \quad\Im z>0,\\
\label{eq-f02}
\Re f({{\boldsymbol \theta}},z)=\Re t({{\boldsymbol \theta}},z)=\Re f_0(z)=0 \quad \text{for} \quad \Im z=0.\end{gathered}$$ For real $\lambda$ one has $f_0(\lambda)=2i \sigma(\lambda)$, where $$\sigma(\lambda)=\dfrac{1}{(2\pi i)^d}\int_{{\mathbb{T}}^d} \arctan \dfrac{N({{\boldsymbol \theta}},\lambda)}{g}\,{{\boldsymbol \theta}}^{-1}d{{\boldsymbol \theta}}.$$ The function $\sigma$ is real-valued, strictly increasing, and continuously differentiable on $[a,b]$.
The analyticity of $f_0$ follows from its integral representation. Eq. follows from if one takes into account the inequalities $\Im N({{\boldsymbol \theta}},z)>0$ for $\Im z>0$ and $\Re\log z<0$ for $|z|<1$. Equalities follows from from and the real-valuedness of $N({{\boldsymbol \theta}},z)$ for real $z$.
By elementary calculations, for $x\in{\mathbb{R}}$ and $y>0$ one has $$\label{eq-hh2}
g_1(x):=\dfrac{1}{2i}\log \dfrac{iy+x}{iy-x}\equiv \arctan \dfrac{x}{y}=:g_2(x).$$ In fact, this follows from $$\label{eq-hh}
g'_1(x)=g'_2(x)= \dfrac{y}{x^2+y^2}$$ and $g_1(0)=g_2(0)=0$. Eq. obviously implies $f_0(\lambda)=2i\sigma(\lambda)$ for $\lambda\in{\mathbb{R}}$. Furthermore, as follows from , $$\sigma'(\lambda)= \dfrac{1}{(2\pi i)^d}\int_{{\mathbb{T}}^d} \dfrac{g N'_\lambda({{\boldsymbol \theta}},\lambda)}{N({{\boldsymbol \theta}},\lambda)^2+g^2} \,{{\boldsymbol \theta}}^{-1}d{{\boldsymbol \theta}},$$ and, by , $\sigma'(\lambda)>0$.
An immediate corollary of the analyticity of $f_0$ and of is
\[lem-eps\] There exists $\varepsilon_0>0$ such that $\big|e^{f_0(\lambda)}\xi-1\big|\le 2\big|e^{f_0(\lambda+i\varepsilon)}\xi-1\big|$ for all $\xi\in{\mathbb{S}}^1$, $\lambda\in[a,b]$, and $\varepsilon\in[0,\varepsilon_0]$.
Denote by $t(z)$ and $f(z)$ the multiplication operators by $t({{\boldsymbol \theta}},z)$ and $f({{\boldsymbol \theta}},z)$ in ${\mathcal{L}}^2({\mathbb{T}}^d)$, respectively. By definition of $t({{\boldsymbol \theta}},z)$ for any $\\xi\in{\mathcal{L}}^2({\mathbb{T}}^d)$ $$\begin{gathered}
\label{eq-comm}
e^{t(z)} e^{f_0(z)} U e^{-t(z)}\xi({{\boldsymbol \theta}})\\
=
e^{t({{\boldsymbol \theta}},z)} e^{f_0({{\boldsymbol \theta}},z)} \exp\big({}-t(z,e^{-2\pi i \omega_1}\theta_1,\dots,e^{-2\pi i \omega_d}\theta_d)\big) U \varphi({{\boldsymbol \theta}})\\
=\exp\big(t({{\boldsymbol \theta}},z)-Ut({{\boldsymbol \theta}},z)+f_0({{\boldsymbol \theta}},z)\big) U\varphi({{\boldsymbol \theta}},z)=e^{f(z)}U\varphi({{\boldsymbol \theta}})=C(z)U\varphi({{\boldsymbol \theta}}). \end{gathered}$$ Therefore, one can rewrite Eq. as $$\label{eq-MA2}
N(z)-B=D(z)^{-1} e^{t(z)}\big(1-e^{f_0(z)}\chi U\big)e^{-t(z)}\big(1+\chi U\big)^{-1}.$$
\[prop-ev\] The set of the eigenvalues of $H$ in $[a,b]$ is dense and coincides with the set of solutions $\lambda$ to $$\label{eq-mlx}
\sigma(\lambda)=\pi({{\boldsymbol \omega}}{{\mathbf m}} +\varphi\big)\mod \pi,\quad {{\mathbf m}}\in{\mathbb{Z}}^d.$$ Each of these eigenvalues is simple, and for any fixed ${{\mathbf m}}\in{\mathbb{Z}}^d$ Eq. has at most one solution $\lambda({{\mathbf m}})$, and $\lambda({{\mathbf m}})\ne\lambda({{\mathbf m}}')$ for ${{\mathbf m}}\ne {{\mathbf m}}'$.
As follows from proposition \[krein\] and the resolvent formula , the eigenvalues $\lambda$ of $H$ outside ${\mathop{\mathrm{spec}}}H_0$ are determined by the condition $\ker\big(N(\lambda)-B\big)\ne 0$, an their multiplicity coincides with the dimension of the corresponding kernels. Eq. shows that the condition $(N(\lambda)-B)u=0$ is equivalent to $\big(1-e^{f_0(\lambda)}\chi U\big)e^{-t(\lambda)}\big(1+\chi U\big)^{-1}u=0$ or, denoting $v:=e^{-t(\lambda)}\big(1+\chi U\big)^{-1}u$, $(1-e^{f_0(\lambda)}\chi U\big)v=0$, which can be rewritten as $$\label{eq-uv}
\chi Uv=e^{-f_0(\lambda)}v,\quad v\ne 0.$$ As $\chi U$ has the simple eigenvalues $e^{-2\pi i ({{\boldsymbol \omega}}{{\mathbf m}}+\varphi)}$, ${{\mathbf m}}\in{\mathbb{Z}}^d$, and the corresponding eigenvectors form a basis, Eq. implies if one takes into account the identity $f_0(\lambda)=2i\sigma(\lambda)$ proved in lemma \[prop-1\]. The rest follows from the monotonicity of $\sigma$, the inclusion ${\mathop{\mathrm{ran}}}\sigma\subset (-\pi/2,\pi/2)$, and the arithmetic properties and .
As $[a,b]$ was an arbitrary interval from ${\mathbb{R}}\setminus{\mathop{\mathrm{spec}}}H_0$, one has an immediate corollary
\[spec3\] The pure point spectrum of $H$ is dense in ${\mathbb{R}}\setminus {\mathop{\mathrm{spec}}}H_0$.
Now it remains to show that the spectrum of $H$ in the interval considered is pure point.
Take some $\alpha>0$. For any $\delta>0$ we denote $${\mathbb{S}}^1_\delta=\bigcup_{m\in{\mathbb{Z}}^d} \Big\{
\xi\in {\mathbb{S}}^1: |{\mathop{\mathrm{Arg}}}\xi-{\mathop{\mathrm{Arg}}}e^{2\pi i{{\boldsymbol \omega}}{{\mathbf m}}}|\le \delta\big(1+|{{\mathbf m}}|\big)^{-d-\alpha}
\Big\},\quad
\widetilde{\mathbb{S}}^1_\delta:={\mathbb{S}}^1\setminus {\mathbb{S}}^1_\delta.$$ Clearly, there holds $$\label{eq-ss2}
|1-\xi e^{-2\pi i{{\boldsymbol \omega}} {{\mathbf m}}}|\ge 2\pi^{-1}\delta \big(1+|{{\mathbf m}}|\big)^{-d-\alpha},\quad \xi\in \widetilde{\mathbb{S}}^1_\delta,
\quad m\in{\mathbb{Z}}^d.$$
Let $\Delta\subset[a,b]$ be an interval whose ends are not eigenvalues of $H$. Consider the mapping $h:\lambda\mapsto \chi e^{f_0(\lambda)}$. By lemma \[prop-1\], $h$ is a diffeomorphism between $\Delta$ and $h(\Delta)$. By proposition \[prop-ev\] one has $h\big(\lambda({{\mathbf m}})\big)=e^{2\pi i{{\boldsymbol \omega}}{{\mathbf m}}}$. Take an arbitrary $\delta>0$ and denote $$\Delta_\delta:=\Delta\cap h^{-1}({\mathbb{S}}^1_\delta),\quad
\widetilde\Delta_\delta:=\Delta\cap h^{-1}(\widetilde{\mathbb{S}}^1_\delta)\equiv\Delta\setminus\Delta_\delta.$$ Clearly, $\Delta_\delta$ is a countable union of intervals, and the limit set $\bigcap_{\delta>0}\Delta_\delta$ coincides with the set of all the eigenvalues $\bigcup_m\{\lambda(m)\}$.
\[lem-fin\] There exists $\varepsilon_0>0$ such that for any $\delta>0$ and any ${{\mathbf n}}\in{\mathbb{Z}}^d$ there exists $C>0$ such that $$\label{eq-mla}
\big\|\big(N(\lambda+i\varepsilon)-B\big)^{-1}\delta_{{{\mathbf n}}}\big\|\le C$$ for all $\lambda\in\widetilde\Delta_\delta$, and $\varepsilon\in(0,\varepsilon_0)$.
Rewrite Eq. in the form $$\big(N(z)-B\big)^{-1}=\big(1+\chi U\big)e^{t(z)}\big(1-e^{f_0(z)}\chi U\big)^{-1}e^{-t(z)} D(z).$$ Note that the Fourier transform of $\delta_{{{\mathbf n}}}$ is the function ${{\boldsymbol \theta}}\mapsto {{\boldsymbol \theta}}^{{{\mathbf n}}}$. Denote $\Psi(z;{{\boldsymbol \theta}}):=e^{-t({{\boldsymbol \theta}},z)}B({{\boldsymbol \theta}},z){{\boldsymbol \theta}}^{{{\mathbf n}}}$. Due to the analyticity one can estimate uniformly in $Z$: $$|\psi_z({{\mathbf m}})|\le C' e^{-\rho|{{\mathbf m}}|},\quad C',\rho>0, \quad \psi_z:=F_d^{-1}\Psi,\quad
\|(1+\chi U)e^{t(z)}\|\le C'.$$ Therefore, follows from the inequality $$\label{eq-fini}
\big\|
(1-e^{f_0(\lambda+i\varepsilon)}\chi U)^{-1}\Psi
\big\|\le C.$$ Assume that $\varepsilon_0$ satisfies the conditions of lemma \[lem-eps\], then uniformly for $\lambda\in\Delta$ and $\varepsilon\in(0,\varepsilon_0)$ one has $$\begin{gathered}
\big|\big(F_d^{-1}(1-e^{f_0(\lambda+i\varepsilon)}\chi U)^{-1}\Psi\big)({{\mathbf m}})\big|
=\big|
(1-e^{f_0(\lambda+i\varepsilon)}\chi e^{2\pi i {{\boldsymbol \omega}} {{\mathbf m}}})^{-1}\psi_{\lambda+i\varepsilon}({{\mathbf m}})
\big|\\
\le 2
\big|
(1-e^{f_0(\lambda)}\chi e^{2\pi i {{\boldsymbol \omega}} {{\mathbf m}}})^{-1}\big| \cdot|\psi_{\lambda+i\varepsilon}({{\mathbf m}})|.\end{gathered}$$ As in our case $h(\lambda)\equiv\chi e^{f_0(\lambda)}\in \widetilde{\mathbb{S}}^1_\delta$, due to we have $$\big|
(1-e^{f_0(\lambda)}\chi e^{-2\pi i {{\boldsymbol \omega}} {{\mathbf m}}})^{-1}\big|\le \dfrac{\pi}{2\delta}\,\big(1+|{{\mathbf m}}|\big)^{d+\alpha}.$$ Finally, $$\begin{gathered}
\big\|
(1-e^{f_0(\lambda+i\varepsilon)}\chi U)^{-1}\Psi
\big\|^2=\sum_{{{\mathbf m}}\in{\mathbb{Z}}^d} \big|\big(F_d^{-1}(1-e^{f_0(\lambda+i\varepsilon)}\chi U)^{-1}\Psi\big)({{\mathbf m}})\big|^2\\
\le \Big(\dfrac{\pi C'}{\delta}\Big)^2\,\sum_{{{\mathbf m}}\in{\mathbb{Z}}^d}\big(1+|{{\mathbf m}}|\big)^{2(d+\alpha)}e^{-2\rho|{{\mathbf m}}|}<\infty,\end{gathered}$$ and is proved.
Now we are able to estimate the spectral projections corresponding to $H$.
\[lem-fff\] For any $f\in{\mathcal{H}}$ and any $\delta>0$ one has $$\label{eq-rrr}
\lim_{\varepsilon\to 0+} \varepsilon \int_{\widetilde\Delta_\delta} \|(H-\lambda-i\varepsilon)^{-1}f\|^2d\lambda=0.$$
Here we are going to use proposition \[krein\]. First note that due to $\widetilde \Delta_\delta\subset{\mathbb{R}}\setminus{\mathop{\mathrm{spec}}}H_0$ one has $$\label{eq-h0}
\lim_{\varepsilon\to 0} \varepsilon \int_{\widetilde\Delta_\delta} \|(H_0-\lambda-i\varepsilon)^{-1}f\|^2d\lambda=0
\text{ for any } f\in{\mathcal{H}}.$$ Similar to let us consider the decomposition $${\mathcal{H}}={\mathcal{H}}_0+{\mathcal{H}}_1,\quad
{\mathcal{H}}_0:=\Big(\bigcup_{\Im z \ne 0} \nu(z)\big(\ell^2({\mathbb{Z}}^d)\big)\Big)^\perp,
\quad
{\mathcal{H}}_1:={\mathcal{H}}_0^\perp;$$ As previously, by , for any $f\in{\mathcal{H}}_0$ and any $z$ with $\Im z\ne 0$ one has $\nu^*(z)f=0$. Hence, by , there holds $(H-z)^{-1}f=(H_0-z)^{-1}à$, and implies for $f\in{\mathcal{H}}_0$.
Now it is sufficient to show for vectors $f=\nu(\zeta)h$ for $h=\big(N(\zeta)-B\big)^{-1}\delta_{{{\mathbf m}}}$, ${{\mathbf m}}\in{\mathbb{Z}}^d$, $\Im\zeta\ne0$. The operators $\big(N(\zeta)-B\big)^{-1}$ have dense range (coinciding with ${\mathop{\mathrm{dom}}}B$), hence the linear hull of such vectors $f$ is dense in ${\mathcal{H}}_1$. By elementary calculations (see e.g. section 3 in [@BGP]) one rewrites Eq. as $$\label{eq-res}
(H-\lambda-i\varepsilon)^{-1}f=
\dfrac{1}{\zeta-\lambda-i\varepsilon}\,\Big(
f-\nu(\lambda+i\varepsilon) \big(N(\lambda+i\varepsilon)-B\big)^{-1}\delta_{{{\mathbf m}}}\Big).$$ Due to lemma \[lem-fin\] we have $\big\|\big(N(\lambda+i\varepsilon)-B\big)^{-1}\delta_{{{\mathbf m}}}\big\|\le C$ with some $C>0$, for all $\lambda\in\widetilde\Delta_\delta$ and sufficiently small $\varepsilon$, and implies $$\|(H-\lambda-i\varepsilon)^{-1}f\|\le \dfrac{\|f\|+C\|\nu(\lambda+i\varepsilon)\|}{|\zeta-\lambda-i\varepsilon|},$$ and due to the analyticity of $\gamma$, one can estimate $\|(H-\lambda-i\varepsilon)^{-1}f\|\le C'$ with some $C'>0$ for all $\lambda\in\widetilde\Delta_\delta$ and sufficiently small $\varepsilon$. This obviously implies .
\[spec4\] The spectrum of $H$ outside ${\mathop{\mathrm{spec}}}L$ is pure point.
We are going to show that for any $f\in{\mathcal{H}}$ and any interval $\Delta\subset{\mathbb{R}}\setminus{\mathop{\mathrm{spec}}}H_0$ the spectral measure $\mu_f$ associated with $H$ and $f$ satisfies $\mu_f(\Delta)=\mu_f\big(\Delta\cap\bigcup_m\{\lambda(m)\}\big)$; this proves that all the spectral measures are pure point.
By the Stone formula, for any set $X$ which is a countable union of intervals whose ends are not eigenvalues of $H$ one has $$\mu_f(X)=\lim_{\varepsilon\to0+}\dfrac{\varepsilon}{\pi}\int_X\|(H-\lambda-i\varepsilon)f\|^2d\lambda.$$ Using lemma \[lem-fff\], for any $\delta>0$ we estimate $$\begin{gathered}
\mu_f(\Delta)=\lim_{\varepsilon\to0+}\dfrac{\varepsilon}{\pi}\int_\Delta\|(H-\lambda-i\varepsilon)f\|^2d\lambda\\
=\lim_{\varepsilon\to0+}\dfrac{\varepsilon}{\pi}\int_{\Delta_\delta}\|(H-\lambda-i\varepsilon)f\|^2d\lambda
+\lim_{\varepsilon\to0+}\dfrac{\varepsilon}{\pi}\int_{\widetilde\Delta_\delta}\|(H-\lambda-i\varepsilon)f\|^2d\lambda\\
=\lim_{\varepsilon\to0+}\dfrac{\varepsilon}{\pi}\int_{\Delta_\delta}\|(H-\lambda-i\varepsilon)f\|^2d\lambda=\mu_f(\Delta_\delta).\end{gathered}$$ As $\delta$ is arbitrary and $\bigcap_{\delta>0}\Delta_\delta=\bigcup_m\{\lambda(m)\}$, the theorem is proved.
Acknowledgments {#acknowledgments .unnumbered}
===============
The work was supported by the Marie Curie Intra-European Fellowship PIEF-GA-2008-219641 during the stay at the University Paris Nord in July-Septembre 2008.
[999]{}
J. Bellissard, R. Lima, E. Scoppola: *Localization in $\nu$-dimensional incommensurate structures.* Commun. Math. Phys. [**88**]{} (1983) 465–477.
F. Bentosela, Ph. Briet, L. Pastur: *On the spectral and wave propagation properties of the surface Maryland model.* J. Math. Phys. [**44**]{} (2003) 1–35.
F. Bentosela, Ph. Briet, L. Pastur: *Spectral analysis of the generalized surface Maryland model.* St. Petersbg. Math. J. [**16**]{} (2005) 923–942.
J. F. Brasche, M. M. Malamud, H. Neidhardt: *Weyl function and spectral properties of self-adjoint extensions.* Integr. Equations Operator Theory [**43**]{} (2002) 264–289.
J. Brüning, V. Geyler, K. Pankrashkin: *Spectra of self-adjoint extensions and applications to solvable Schrödinger operators.* Rev. Math. Phys. [**20**]{} (2008) 1–70.
P. Exner, M. Fraas: *On the dense point and absolutely continuous spectrum for Hamiltonians with concentric $\delta$ shells.* Lett. Math. Phys. [**82**]{} (2007) 25–37.
A. L. Figotin, L. A. Pastur: *An exactly solvable model of a multidimensional incommensurate structure.* Commun. Math. Phys. [**95**]{} (1984) 401–425.
S. Gnutzmann, U. Smilansky: *Quantum graphs Applications to quantum chaos and universal spectral statistics.* Adv. Phys. [**55**]{} (2006) 527-–625.
D. R. Grempel, S. Fishman, R. E. Prange: *Localization in an incommensurate potential an exactly solvable model.* Phys. Rev. Lett. [**49**]{} (1982) 833–836.
B. A. Khoruzhenko, L. A. Pastur, *The localization of surface states: an exactly solvable model.* Phys. Rep. [**288**]{} (1997) 109–126.
P. Kuchment: *Quantum graphs* I. *Some basic structures.* Waves Random Media [**14**]{} (2004) S107–S128.
P. Kuchment: *Quantum graphs* II. *Some spectral properties of quantum and combinatorial graphs.* J. Phys. A: Math. Gen. [**38**]{} (2005) 4887–4900.
K. Pankrashkin: *Spectra of Schrödinger operators on equilateral quantum graphs.* Lett. Math. Phys. [**77**]{} (2006) 139–154.
K. Pankrashkin: *Localization in a quasiperiodic model on quantum graphs.* In the book P. Exner et al. (Eds): *Analysis on graphs and its applications* (Proc. Symp. Pure Math., vol. 77, AMS, Provdence, 2008) 459–467.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Modern astrophysics is undergoing a revolution. As detector technology has advanced, and astronomers have been able to study the sky with finer temporal detail, a rich diversity of sources which vary on timescales from years down to a few nanoseconds has been found. Among these are Fast Radio Bursts, with pulses of millisecond duration and anomalously high dispersion compared to Galactic pulsars, first seen a decade ago. Since then, a new research community is actively working on a variety of experiments and developing models to explain this new phenomenon, and devising ways to use them as astrophysical tools. In this article, I describe how astronomers have reached this point, review the highlights from the first decade of research in this field, give some current breaking news, and look ahead to what might be expected in the next few years.'
author:
- 'Duncan R. Lorimer$^{1,2}$'
bibliography:
- 'refs.bib'
title: A Decade of Fast Radio Bursts
---
1. Department of Physics and Astronomy, West Virginia University, P.O. Box 6315, Morgantown, WV 26506, USA
2. Center for Gravitational Waves and Cosmology, Chestnut Ridge Research Building, Morgantown, WV 26505, USA
Astronomers have long predicted the existence of short-duration pulses that might be present in high-time resolution surveys of the radio sky[@1]. In 2007, we published the first such example — a bright, highly dispersed pulse lasting 15 ms and of astrophysical but unknown origin[@2]. A decade on, with over 60 further examples[@3] of what are now known as Fast Radio Bursts (FRBs), astronomers are now probing a population within the landscape of transient phenomena. We now think about the Universe both in terms of sources we see from epoch to epoch, and transients like FRBs which provide an ever-changing landscape rich in possibilities for future study.
The story of FRBs can be traced back to the serendipitous discovery of pulsars by Jocelyn Bell and Anthony Hewish in 1967[@4]. Pulsars are characterized by the short duration (few 10s of ms) pulses which are extremely regularly spaced. Today, with over 2600 pulsars known[@5], the currently accepted picture of a pulsar is that it is a rapidly rotating highly magnetized neutron star which formed from the collapsed core of a normal star after it underwent a supernova explosion. The pulses detected on Earth are caused by the rotation of the magnetic field which sweeps across our line of sight in much the same way as a lighthouse. The first pulsars were found through their individual pulses, but subsequent searches relied on the improved sensitivity that comes from Fourier analyses of long data streams[@6].
Among the first pulsars to be discovered was PSR B0531+21 in the Crab nebula[@7] which was originally seen through the emission of exceptionally bright individual pulses. The Crab and other energetic pulsars are now known to emit what are now known as giant pulses with a rate that follows a power-law distribution of amplitudes[@8]. Crab-like giant pulses could only be seen in the nearest galaxies using the largest radio telescopes[@9].
The promise of detecting pulses from energetic sources inspired a number of searches over the years. Following Hawking’s prediction in 1974[@10] that black holes formed early in the Universe’s history would actually be evaporating today, Rees suggested them as a potential population of sources emitting bright radio pulses[@11]. Colgate also predicted radio pulses produced during supernova explosions[@12]. Motivated by these predictions, Phinney and Taylor carried out a search for bursts in data taken with the Arecibo radio telescope[@13]. Although unsuccessful, this search placed the first constraints on the rates of bursting radio sources in the Universe. Other searches have been carried out over the intervening years but, despite unconfirmed claims of pulses from the giant elliptical galaxy M87[@14], no further events were found[@14a].
By the end of the 1990s, as telescope instrumentation and the availability of high-speed computing resources improved, the interest in searching for radio pulses was renewed[@15; @16]. McLaughlin et al.[@17] applied the single-pulse search techniques to a large-scale survey with the Parkes radio telescope and found a new population of sporadically emitting sources known as “Rotating Radio Transients” (RRATs). While RRATs were quickly identified with Galactic neutron stars, during further searches of data collected in a survey of the Magellanic clouds, we serendipitously discovered a bright and highly anomalous new type of pulse[@2].
![Waterfall plot showing radio frequency versus time (lower panel) for the original FRB 010724 in the main beam of the telescope. The upper panel shows the de-dispersed pulse after appropriately delaying the filterbank channels to account for the inverse frequency squared behaviour seen below. Also evident in these figures is the offset level of the baseline noise prior and following the pulse. This is due to nature of the integrating circuit employed in the single-bit digitizers by this extremely bright pulse and was not shown in the original discovery paper [@2]. Figure credit: Evan Keane.](F1){width="\textwidth"}
Fig. 1 shows the pulse we found (now known as the Lorimer burst) in terms of radio frequency versus arrival time for a band spanning 288 MHz centred around 1374 MHz. Due to the frequency-dependent refractive index of the ionized interstellar medium, the highest frequency components of the signal travel faster than their lower frequency counterparts and arrive earlier. The signature of this effect is the inverse quadratic delay of the pulse with observing frequency[@4]. The total delay across an observing band is proportional to the dispersion measure $${\rm DM} = \int_0^d n_e \, {\rm d}l.$$ Here $d$ is the distance to the source and $n_e$ is the number of electrons per unit volume along the line of sight. It is also noticeable that the pulse gets broader towards the lower frequencies. This is also a hallmark that is expected for a pulse propagating through a turbulent medium and is well-known from studies of pulsars. Using models for the electron density[@18], astronomers routinely estimate the distance to pulsars and RRATs from their observed DM values. Applying the same technique to the Lorimer burst, it is immediately apparent that the delay is too large (by about an order of magnitude) to be attributed to free electrons in the Milky Way. In addition, since the location of the pulse is about three degrees south of the Small Magellanic Cloud, we concluded that the source was located well beyond the Milky Way and Magellanic Clouds. Despite almost 100 hours of follow-up observations, no further pulses were found. Making parallels with gamma-ray bursts we proposed that this event was a prototype of a new class of astrophysical transients, and that the rate of similar bright bursts over the sky could be at least as large as a few hundred per day.
In spite of a great deal of effort that went into trying to identify further examples in other archival data sets, the small fields of view of most radio telescopes and the short duration of the bursts meant that subsequent detections were not immediately forthcoming. Bursts were detected that were clearly of terrestrial origin (referred to as Perytons[@19]), that have subsequently been traced to emission from microwave ovens[@20]. In 2012, however, a second convincing burst of astrophysical origin from a different sky location was found in Parkes data taken of a survey of the Milky Way[@21]. Like the Lorimer burst, its DM was larger than that expected from electrons in the Milky Way and was likely much further away. For many researchers, a game-changing discovery in the field came in 2013 with the announcement of four more bursts by Thornton et al.[@22] which clearly revealed a population of highly dispersed anomalous pulses. At this point the convention of naming the phenomenon Fast Radio Bursts (FRBs) began. The designation for each FRB followed that of the gamma-ray bursts as year, month and day of its arrival at the telescope. For example, the FRB discovered on February 20, 2011 is called FRB 110220. Presently over 60 FRBs are known, and an up-to-date list of source parameters can be found online[@3]. In addition to further discoveries at Parkes, FRBs have also been found at Arecibo[@23], Green Bank[@24], Molonglo[@25; @26], ASKAP – the Australian Square Kilometre Array Pathfinder[@27; @28] and, most recently the Canadian HI Intensity Mapping Experiment (CHIME)[@29]. As shown in Fig. 2, the anomalously high dispersion of this sample of FRBs is a defining trait, clearly representative of a different population compared to Galactic pulsars and RRATs.
{width="\textwidth"}
In an analogous way to how distances to pulsars are estimated from their DMs, which make use of a model for the free electrons in the Galaxy[@18], redshifts of FRBs, z, can be estimated from a model of the distribution of electrons which turns out to be dominated by those spanning intergalactic distances. As a very crude rule of thumb[@30], $z \sim$ (DM/1000 cm$^{-3}$ pc), so for typical FRBs with DMs in the range 200–2600 cm$^{-3}$ pc, we infer redshifts in the range $0.2 < z <2.6$. We stress that this calculation uses an estimate of the redshift rather than a direct measurement which has up to now only been possible for one source, FRB 121102, discussed further below. However, with these caveats in mind, for a “canonical FRB” with ${\rm DM} = 1000$ cm$^{-3}$ pc with a peak flux density of a jansky and a width of 5 ms, one can infer a co-moving distance of 3 Gpc and equivalent energy release of $10^{32}$ joules[@22]. For reference, the Sun emits $10^{26}$ W, so this corresponds to the amount of energy released by the Sun in two weeks! Although this seems like a lot, a typical Gamma-Ray Burst releases something like $10^{45}$ joules — more energy than the Sun will emit in its ten billion year projected lifetime!
By the beginning of 2016, due to the lack of repeat pulses among FRBs known at the time, in spite of hundreds of hours of follow-up time[@2; @22], a suspicion among many researchers was that FRBs do not repeat and are likely one-time events. FRB 121102, originally found at Arecibo[@23], has now been observed to repeat on multiple occasions[@31; @32; @33; @34]. Individual pulses are quite variable in nature. A major implication of this discovery, however, is that for FRB 121102, whatever is producing these bursts, it cannot be catastrophic in nature. Models involving one-off explosions, or stellar mergers, for this source at least are ruled out.
Further constraints on the nature of FRB 121102 have now been found thanks to the precise localization of the source in 2017 using the Karl Jansky Very Large Array[@35], and the measurement of a significant amount of polarization in the source by the Arecibo telescope[@36]. The positional determination[@35] allowed the unambiguous association of FRB 121102 with a low-metallicity star-forming dwarf galaxy at redshift $z = 0.192$[@37]. This is the first, and currently only, identification of a counterpart and the direct measurement of distance to an FRB. The luminosity distance corresponding to this redshift is 972 Mpc. The implied isotropic energy release of individual bursts is variable and ranges between $10^{30}$ and $10^{33}$ joules[@33].
Further radio observations using the European Very Long Baseline Interferometry Network and Arecibo[@38] provide compelling evidence for an association of FRB 121102 with a persistent radio source within this dwarf galaxy. This could be explained as being a signature of the emission from either the nucleus of the galaxy, or of a young pulsar energizing a supernova remnant therein[@38; @39]. Subsequent polarization measurements presented by Michilli et al.[@36] show that the radio emission is almost entirely linearly polarized and the amount of Faraday rotation of the electromagnetic signal is variable, but extremely high: in the range 130,000–150,000 radians per square metre. Such large rotation measures are only observed in the extremely highly magnetized environments found close to supermassive black holes, such as Sagittarius A\* at the center of our Galaxy[@40].
FRB 121102 is presently unique in that it shows multiple pulses which are sporadic in nature. Unlike the case for pulsars and RRATs, there is no discernable periodicity. The shortest separation between bursts currently observed is 34 ms[@32]. Do all FRBs repeat? While we cannot definitively rule out the existence of repeat bursts on timescales of years, there is now growing evidence that burst rates of the kind only seen in FRB 121102 are the exception, rather than the rule. At the time of writing, most FRBs have been observed for hundreds of hours, without any evidence for repeat pulses. A recent analysis[@41] based on a sample of 21 FRBs found at Parkes, if all FRBs are like 121102, finds the probability of non-detections in the sample is less than 0.1%. A novel feature of the initial survey by ASKAP[@27; @28] is that the same fields on the sky are repeatedly observed. Each of the 20 FRBs found in this survey has typically hundreds of hours of follow-up time with comparable sensitivity. An analysis of these non-detections[@28] also concludes that it is unlikely that they have amplitude distributions similar to FRB 121102. Although further constraints from other surveys and more sensitive telescopes will be more useful, it is becoming increasingly likely that, much like gamma-ray bursts, that there are two distinct classes of FRBs. In what follows, we will proceed under the assumption that this is the case, i.e. that there are both repeating FRBs and one-time FRBs.
So far, I have talked very little about what FRBs could be. This is largely because the nature of the sources is still unclear. A stringent constraint on the emission mechanism can be made by noting that, for a pulse with peak flux density $S_{\rm peak}$, and width $W$ at a distance $d$ from the Earth, observing a some frequency $f$, the equivalent temperature of a blackbody source of the same isotropic luminosity, i.e., the brightness temperature $$T_{\rm b} = \left(\frac{S_{\rm peak}}{2\pi k}\right)
\left(\frac{fW}{d}\right)^2,$$ where $k$ is Boltzmann’s constant. Inserting the measured values now available for FRB 121102, we find $T_{\rm b} \sim 10^{35}$ K which clearly implies a coherent non-thermal radiation mechanism that all explanations summarized below must satisfy.
Very much like in the early days of gamma-ray burst astrophysics[@42] there are currently more theories for the emission mechanisms than actual FRBs known. Those put forward thus far fall broadly into three main categories: extraterrestrial, Galactic and extragalactic. The first of these, that FRBs represent some sort of message from extra-terrestrial civilizations, can be largely discounted by the preponderance of them over the whole sky. Similar arguments were made following the discovery of pulsars[@4]. An intriguing possibility is that some fraction of FRBs could be produced “accidentally” by extra-terrestrial intelligence such as beamed emission from light sails[@43]. Any source of FRBs that originates in the second category, i.e. as natural sources from within our Galaxy, have to find a way of explaining the extremely high DMs observed for FRBs when compared with the pulsar population that is now well studied. Clearly FRB 121102 is inconsistent with a Galactic origin[@37]. Attempts to explain the rest of the FRBs as coming from anomalously high dispersion produced by Galactic flare stars[@44], remains less attractive due to fact that the high implied plasma densities would not result in an inverse frequency squared dispersion[@45].
Within the extragalactic category, theories consistent with the observed energetics involve compact objects (i.e. black holes, neutron stars, strange stars and white dwarfs) which are necessary, from light-travel-time considerations, to explain the short duration of the pulses. Among the models currently suggested are: collapsing neutron stars forming black holes[@46; @47], coalescing neutron star binaries[@48; @49], coalescing neutron star—black hole binaries[@50], coalescing white dwarf binaries[@51], primordial black holes falling into neutron stars[@52], evaporating black holes[@53], giant pulses from young neutron stars[@54; @55], nascent and highly energetic magnetars[@56; @57], white holes[@58]. While these scenarios involve compact stars, non-stellar options involving cosmic strings, large-scale defects in the structure of spacetime thought to have been formed in early Universe, have also been proposed[@59; @60]. Such theories, while in principle still viable, do need to be further developed so that they can make concrete and testable observational predictions.
Regardless of the origin of FRBs, predictions for ongoing and future experiments can be made by relatively simple modeling of them as a cosmological population. To demonstrate how a self-consistent model can be set up, one can mimic the $S_{\rm peak}$—DM distribution seen in the Parkes pulsars using a very simple Monte Carlo simulation. In this model, synthetic FRBs are uniformly distributed in co-moving volume out to $z = 2.5$. A redshift is computed for each one and a DM is assigned which assumes a contribution of 25 cm$^{-3}$ pc from the Milky Way, $1000 z$ cm$^{-3}$ pc from the intergalactic medium, and 150 cm$^{-3}$ pc from the host. A radio luminosity is chosen from a normal distribution with a mean of 180 Jy Gpc$^2$ and standard deviation of 100 Jy Gpc$^2$. The flux density of each FRB is then computed. To simulate the uncertain position of each FRB within the Parkes telescope beam, these fluxes are additionally multiplied by a dithering factor from a Gaussian with full-width half maximum of 11 arcminutes. The result of this simulation is shown as the background of small points in Fig. 3. This simple procedure demonstrates that the sample requires a distribution of luminosities, and that FRBs are unlikely to be standard candles. More detailed analyses can be found elsewhere[@61; @62]. With larger sample sizes and more detailed modeling in the future, it should be possible to better constrain the luminosity distribution and spatial distribution of the population.
![Peak flux density versus dispersion measure (DM) for observed (large stars) and simulated (dots) FRBs. The simulation assumes FRBs are uniformly distributed in co-moving volume with a Gaussian luminosity function (see text). Although the simulated events span the DM and flux densities of the observed sample, refinements to the model will be required to produce the correct proportion of events as a function of DM and flux density. More sophisticated simulations are now being carried out[@61; @62]. ](F3){width="\textwidth"}
I have summarized the discovery and salient properties of a population of mysterious radio bursts of currently unknown origin that appear to originate from random locations on the sky. Perhaps the most remarkable thing about FRBs is their inferred rate of occurrence. Based on the fields of view and time spent in surveys to date, current estimates of their all-sky rates range between $10^3$ and $10^4$ FRBs per sky per day[@22; @63]. Expressed in another way, somewhere over the sky, an FRB goes off every 30 seconds! Although all the models we reviewed previously can be tuned to accommodate this rate, many of them could be ruled out if we knew the distance scale for the FRBs and, hence, their rate per unit volume in the Universe. To emphasize the extent of the problem, FRBs are currently not too far below the most violent known explosions in the Universe – supernovae, which occur somewhere in the visible universe at a rate of about one per second. If we are missing a significant fraction of faint FRBs, or if many of them are not beamed towards us, the rate that we infer could be as high or even higher than the supernova rate. If, as it seems likely, there are multiple sub-populations of FRBs (e.g. repeating and non-repeating sources), what are the relative rates of these? Only through the discovery of a larger sample of these objects will we be able to answer these questions.
Currently, new FRBs continue to be found with existing instrumentation. ASKAP is making great inroads into the low DM (100–1000 cm$^{-3}$ pc) population, with over 20 discoveries so far[@28]. ASKAP has also found the first FRB that directly probes the line of sight of a galaxy cluster. In a 300-hr survey with ASKAP, Agarwal et al.[@64] have found one bright FRB that probes the ionized medium of the Virgo cluster. Further examples of such pulses could be found by dedicated searches of nearby galaxy clusters[@65]. Very recently, CHIME has begun operations in the 400–800 MHz frequency band and has already found several FRBs, including some that are visible at frequencies as low as 400 MHz[@29].
The next decade is set to be very promising for FRB hunting[@66]. New instruments coming online at Arecibo, Green Bank, and elsewhere promise to find more of these enigmatic sources. Predictions for CHIME are extremely promising, with rates of up to dozens of sources per day being possible[@67]! A new giant telescope in China – the Five-Hundred-Meter Aperture Spherical Telescope (FAST) – also recently completed, will have unprecedented sensitivity for faint and distant FRBs[@68]. At the time of writing, FAST is about to begin surveys with a 19-beam system, while plans are underway to upgrade Arecibo’s throughput to 40 beams. Predictions for FAST and Arecibo FRBs using the simulations described above suggest that they could see sources out to redshift of 5 with DMs approaching $10,000$ cm$^{-3}$ pc.
A new radio telescope in South Africa, MeerKAT will soon begin scanning the skies for FRBs[@meertrap] and an optical partner, known as MeerLICHT[@meerlicht] will simultaneously observe them at optical wavelengths. Such an experiment would probe the optical transient sky in an entirely new way, and place further constraints on the FRB emission mechanism[@69]. MeerKAT is a precursor to the next generation radio telescope known as the Square Kilometer Array (SKA) which, upon its completion in the mid 2020s[@70], would be able to locate and study FRBs to unprecedented precision. There will be no doubt many surprises in store as we enter the second decade of study of FRBs. I predict that by 2030 FRBs will become standard tools to study the large-scale structure and magnetoionic content of the cosmos. Only time will tell, of course, but if the first ten years is anything to go by, FRBs still have plenty of surprises in store for astrophysics in the next decade.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'There have been a number of studies on sparse signal recovery from one-bit quantized measurements. Nevertheless, little attention has been paid to the choice of the quantization thresholds and its impact on the signal recovery performance. This paper examines the problem of one-bit quantizer design for sparse signal recovery. Our analysis shows that the magnitude ambiguity that ever plagues conventional one-bit compressed sensing methods can be resolved, and an arbitrarily small reconstruction error can be achieved by setting the quantization thresholds close enough to the original unquantized measurements. Note that the unquantized data are inaccessible by us. To overcome this difficulty, we propose an adaptive quantization method that iteratively refines the quantization thresholds based on previous estimate of the sparse signal. Numerical results are provided to collaborate our theoretical results and to illustrate the effectiveness of the proposed algorithm.'
author:
- 'Jun Fang, Yanning Shen, and Hongbin Li, [^1] [^2] [^3]'
bibliography:
- 'newbib.bib'
title: 'One-Bit Quantization Design and Adaptive Methods for Compressed Sensing '
---
One-bit compressed sensing, adaptive quantization, quantization design.
Introduction
============
Compressive sensing is a recently emerged paradigm of signal sampling and reconstruction, the main purpose of which is to recover sparse signals from much fewer linear measurements [@CandesTao05; @Donoho06] $$\begin{aligned}
\boldsymbol{y}=\boldsymbol{Ax}\end{aligned}$$ where $\boldsymbol{A}\in\mathbb{R}^{m\times n}$ is the sampling matrix with $m\ll n$, and $\boldsymbol{x}$ denotes the $n$-dimensional sparse signal with only $K$ nonzero coefficients. Such a problem has been extensively studied and a variety of algorithms that provide consistent recovery performance guarantee were proposed, e.g. [@CandesTao05; @Donoho06; @TroppGilbert07; @Wainwright09]. Conventional compressed sensing assumes infinite precision of the acquired measurements. In practice, however, signals need to be quantized before further processing, that is, the real-valued measurements need to be mapped to discrete values over some finite range. Besides, in some sensing systems (e.g. distributed sensor networks), data acquisition is expensive due to limited bandwidth and energy constraints [@AkyildizSu02]. Aggressive quantization strategies which compress real-valued measurements into one or only a few bits of data are preferred in such scenarios. Another benefit brought by low-rate quantization is that it can significantly reduce the hardware complexity and cost of the analog-to-digital converter (ADC). As pointed out in [@LaskaWen11], low-rate quantizer can operate at a much higher sampling rate than high-resolution quantizer. This merit may lead to a new data acquisition paradigm which reconstructs the sparse signal from over-sampled but low-resolution measurements.
Inspired by practical necessity and potential benefits, compressed sensing based on low-rate quantized measurements has attracted considerable attention recently, e.g. [@ZymnisBoyd10; @WimalajeewaVarshney12; @LaskaBaraniuk12]. In particular, the problem of recovering a sparse or compressible signal from one-bit measurements $$\begin{aligned}
\boldsymbol{b}=\text{sign}(\boldsymbol{y})=\text{sign}(\boldsymbol{A}\boldsymbol{x})
\label{eq1}\end{aligned}$$ was firstly introduced by Boufounos and Baraniuk in their work [@BoufounosBaraniuk08], where “sign” denotes an operator that performs the sign function element-wise on the vector, the sign function returns $1$ for positive numbers and $-1$ otherwise. Following [@BoufounosBaraniuk08], the reconstruction performance from one-bit measurements was more thoroughly studied and a variety of one-bit compressed sensing algorithms such as binary iterative hard thresholding (BIHT) [@JacquesLaska11], matching sign pursuit (MSP) [@Boufounos09], and many others [@PlanVershynin11; @YanYang12; @ShenFang13] were proposed. Despite all these efforts, very little is known about the optimal choice of the quantization thresholds and its impact on the recovery performance. In most previous work (e.g [@BoufounosBaraniuk08; @JacquesLaska11; @Boufounos09; @PlanVershynin11; @LaskaWen11; @YanYang12; @ShenFang13]), the quantization threshold is set equal to zero. Nevertheless, such a choice does not necessarily lead to the best reconstruction accuracy. In fact, in this case, only the sign of the measurement is retained while the information about the magnitude of the signal is lost. Hence an exact reconstruction of the sparse signal is impossible without additional information regarding the sparse signal. In some other scenarios such as intensity-based source localization in sensor networks, since all sampled measurements are non-negative, comparing the original measurements with zero yields all ones wherever the sources are located, which makes identifying the true locations impossible. From the above discussions, we can see that quantization is an integral part of the sparse signal recovery and is critical to the recovery performance.
There have been some interesting work [@SunGoyal09; @KamilovGoyal11; @KamilovBourguard12] on quantizer design for compressed sensing. Specifically, the work [@SunGoyal09] utilizes the high-resolution distributed functional scalar quantization theory for quantizer design, where the quantization error is modeled as random noise following a certain distribution. Nevertheless, such modeling holds valid only for high-resolution quantization, and may bring limited benefits when a low-rate quantization strategy is adopted. In [@KamilovGoyal11; @KamilovBourguard12], authors proposed a generalized approximate message passing (GAMP) algorithm for quantized compressed sensing, and studied the quantizer design under the GAMP reconstruction. The method [@KamilovGoyal11; @KamilovBourguard12] was developed in a Bayesian framework by modeling the sparse signal as a random variable and minimizing the average reconstruction error with respective to all possible realizations of the sparse signal. This method, however, involves high-dimensional integral or multidimensional search to find the optimal quantizer, which makes a close-form expression of the optimal quantizer design difficult to obtain. In [@KamilovBourguard12], an adaptive algorithm was proposed which adjusts the thresholds such that the hyperplanes pass through the center of the mass of the probability distribution of the estimated signal. Albeit intuitive, no rigorous theoretical guarantee was available to justify the proposed adaptive method. The problem of reconstructing image from one-bit quantized data is considered in [@BourquardAguet10], where the quantization thresholds are set such that the binary measurements are acquired with equal probability. Such a scheme has been shown to be empirically effective. Nevertheless, it is still unclear why it leads to better reconstruction performance.
In this paper, we investigate the problem of one-bit quantization design for recovery of sparse signals. The study is conducted in a deterministic framework by treating the sparse signal as a deterministic parameter. We provide an quantitative analysis which examines the choice of the quantization thresholds and quantifies its impact on the reconstruction performance. Analyses also reveal that the reconstruction error can be made arbitrarily small by setting the quantization thresholds close enough to the original unquantized measurements $\boldsymbol{y}$. Since the original unquantized data samples $\boldsymbol{y}$ are inaccessible by us, to address this issue, we propose an adaptive quantization approach which iteratively adjusts the quantization thresholds such that the thresholds eventually come close to the original data samples and thus a small reconstruction error is achieved.
The rest of the paper is organized as follows. In Section \[sec:problem-formulation\], we introduce the one-bit compressed sensing problem. The main results of this paper are presented in Section \[sec:analysis\], where we show that, by properly selecting the quantization thresholds, sparse signals can be recovered from one-bit measurements with an arbitrarily small error. A rigorous proof of our main result is provided in Section \[sec:proof\]. An adaptive quantization method is developed in \[sec:algorithm\] to iteratively refine the thresholds based on previous estimate. In Section \[sec:numerical-results\], numerical results are presented to corroborate our theoretical analysis and illustrate the effectiveness of the proposed algorithm, followed by concluding remarks in Section \[sec:conclusion\].
Problem Formulation {#sec:problem-formulation}
===================
We consider a coarse quantization-based signal acquisition model in which each real-valued sample is encoded into one-bit of information $$\begin{aligned}
\boldsymbol{b}=\text{sign}(\boldsymbol{y}-\boldsymbol{\tau})
=\text{sign}(\boldsymbol{A}\boldsymbol{x}-\boldsymbol{\tau})
\label{eq1}\end{aligned}$$ where $\boldsymbol{b}=[b_1\phantom{0}b_2\phantom{0}\ldots\phantom{0}b_m]^T$ are the binary observations, and $\boldsymbol{\tau}=[\tau_1\phantom{0}\tau_2\phantom{0}\ldots\phantom{0}\tau_m]^T$ denotes the quantization threshold vector. As mentioned earlier, in most previous one-bit compressed sensing studies, the quantization thresholds are set equal to zero, i.e. $\boldsymbol{\tau}=\boldsymbol{0}$. A major issue with the zero quantization threshold is that it introduces a magnitude ambiguity which cannot be resolved without additional information regarding the sparse signal. The reason is that when $\boldsymbol{\tau}=\boldsymbol{0}$, multiplying $\boldsymbol{x}$ by any arbitrary nonzero scaling factor will result in the same quantized data $\{b_n\}$. To circumvent the magnitude ambiguity issue, in [@JacquesLaska11; @PlanVershynin11], a unit-norm constraint is imposed on the sparse signal and it was shown that unit-norm signals can be recovered with a bounded error from one bit quantized data. Nevertheless, the choice of the zero threshold is not necessarily the best. In this paper, we are interested in examining the choice of the quantization thresholds and its impact on the reconstruction performance.
Specifically, we consider the following canonical form for sparse signal reconstruction $$\begin{aligned}
\min_{\boldsymbol{z}}\|\boldsymbol{z}\|_0\quad
\text{s.t.}\phantom{0}\text{sign}(\boldsymbol{A}\boldsymbol{z}-\boldsymbol{\tau})=\boldsymbol{b}\label{opt-1}\end{aligned}$$ Such an optimization problem, albeit non-convex, is more amenable for theoretical analysis than its convex counterpart (\[opt-2\]). Suppose $\boldsymbol{\hat{x}}$ is a solution of the above optimization. In the following, we examine the choice of $\boldsymbol{\tau}$ and its impact on the reconstruction error $\|\boldsymbol{x}-\boldsymbol{\hat{x}}\|_2$.
One-Bit Quantization Design: Analysis {#sec:analysis}
=====================================
To facilitate our analysis, we decompose the quantization threshold vector $\boldsymbol{\tau}$ into a sum of two terms: $$\begin{aligned}
\boldsymbol{\tau}=\boldsymbol{y}-\boldsymbol{\delta} \label{eq2}\end{aligned}$$ where $\boldsymbol{\delta}=[\delta_1\phantom{0}\delta_2\phantom{0}\ldots\phantom{0}\delta_m]^T$ can be treated as a constrained deviation from the original unquantized measurements $\boldsymbol{y}$. Substituting (\[eq2\]) into (\[eq1\]), we get $$\begin{aligned}
\boldsymbol{b}=\text{sign}(\boldsymbol{A}\boldsymbol{x}-\boldsymbol{\tau})=\text{sign}(\boldsymbol{\delta})
\label{eq3}\end{aligned}$$ Suppose $\hat{\boldsymbol{x}}$ is the solution of (\[opt-1\]). Let $\boldsymbol{h}\triangleq\hat{\boldsymbol{x}}-\boldsymbol{x}$ be the residual (reconstruction error) vector. Clearly, $\boldsymbol{h}$ is a $2K$-sparse vector which has at most $2K$ nonzero entries since $\hat{\boldsymbol{x}}$ has at most $K$ nonzero coefficients. Also, the solution $\hat{\boldsymbol{x}}$ yields estimated measurements that are consistent with the observed binary data, i.e. $$\begin{aligned}
\boldsymbol{b}=\text{sign}(\boldsymbol{A}\boldsymbol{\hat{x}}-\boldsymbol{\tau})=
\text{sign}(\boldsymbol{Ah}+\boldsymbol{\delta}) \label{eq4}\end{aligned}$$ Combining (\[eq3\]) and (\[eq4\]), the residual vector $\boldsymbol{h}$ has to satisfy the following constraint $$\begin{aligned}
\text{sign}(\boldsymbol{Ah}+\boldsymbol{\delta})=\text{sign}(\boldsymbol{\delta})\label{condition-1}\end{aligned}$$ In the following, we will show that the residual vector $\boldsymbol{h}$ can be bounded by the deviation vector $\boldsymbol{\delta}$ if the number of measurements are sufficiently large and $\boldsymbol{A}$ satisfies a certain condition. Thus the reconstruction error $\boldsymbol{h}$ can be made arbitrarily small by setting $\boldsymbol{\delta}$ close to zero. Our main results are summarized as follows.
\[theorem1\] Let $\boldsymbol{x}\in\mathbb{R}^n$ be an $K$-sparse vector. $\boldsymbol{A}\in\mathbb{R}^{m\times n}$ is the sampling matrix. Suppose there exist an integer $\kappa\geq 2K$ and a positive parameter $\mu$ such that any $\kappa\times n$ submatrix $\boldsymbol{\bar{A}}$ constructed by selecting certain rows of $\boldsymbol{A}$ satisfies $$\begin{aligned}
\|\boldsymbol{\bar{A}}\boldsymbol{u}\|_2^2\geq
\mu\|\boldsymbol{u}\|_2^2 \label{theorem1:condition1}\end{aligned}$$ for all $2K$-sparse vector $\boldsymbol{u}$. Also, assume that each entry of $\boldsymbol{\delta}$ is independently generated from a certain distribution with equal probabilities being positive or negative. Let $\boldsymbol{\hat{x}}$ denote the solution of the optimization problem (\[opt-1\]). For any arbitrarily small value $\eta>0$, we can ensure that the following statement is true with probability exceeding $1-\eta$: if the number of measurements, $m$, is sufficiently large and satisfies $$\begin{aligned}
&m-2K\log (m-\kappa+1)-(\kappa-1)\log m \nonumber\\
&\geq 2K (\log (ne^2)-2\log(2K)+1)+\log(1/\eta)+c
\label{theorem1:condition2}\end{aligned}$$ then the sparse signal can be recovered from (\[opt-1\]) with the reconstruction error bounded by $$\begin{aligned}
\|\boldsymbol{\hat{x}}-\boldsymbol{x}\|_2\leq
\frac{\epsilon}{\sqrt{\mu}}\triangleq\lambda\epsilon
\label{theorem1:errorbound}\end{aligned}$$ where $\epsilon\triangleq\|\boldsymbol{\delta}\|_2$, $\lambda\triangleq 1/\sqrt{\mu}$, $e$ in (\[theorem1:condition2\]) represents the base of the natural logarithm, and $c$ in (\[theorem1:condition2\]) is defined as $$\begin{aligned}
c\triangleq(\kappa-1)(\log(e/(\kappa-1))+1)\end{aligned}$$ which is a constant only dependent on $\kappa$.
We have the following remarks regarding Theorem \[theorem1\].
*Remark 1:* Clearly, the condition (\[theorem1:condition1\]) is guaranteed if there exists a constant $\gamma_{2K}<1$ such that any $\kappa\times n$ submatrix of $\boldsymbol{A}$ satisfies $$\begin{aligned}
(1+\gamma_{2K})\|\boldsymbol{u}\|_2^2\geq\|\boldsymbol{\bar{A}u}\|_2^2\geq(1-\gamma_{2K})\|\boldsymbol{u}\|_2^2\end{aligned}$$ for all $2K$-sparse vector $\boldsymbol{u}$. As indicated in previous studies [@CandesTao05], when the sampling matrix is randomly generated from a Gaussian or Bernoulli distribution, this restricted isometry property is met with an overwhelming probability, particularly when $\kappa$ is large. Hence it is not difficult to find an integer $\kappa$ and a positive $\mu$ such that the condition (\[theorem1:condition1\]) holds.
*Remark 2:* Note that the term on the left hand side of (\[theorem1:condition2\]) is a monotonically increasing function of $m$ when $m$ is greater than a certain value. Thus for fixed values of $K$, $\kappa$, $n$, and $\eta$, we can always find a sufficiently large $m$ to ensure the condition (\[theorem1:condition2\]) is guaranteed. In addition, a close inspection of (\[theorem1:condition2\]) reveals that, for a constant $\eta$, the number of measurements required to guarantee (\[theorem1:condition2\]) is of order $\mathcal{O}(K\log(n))$, which is the same as that for conventional compressed sensing.
*Remark 3:* From (\[theorem1:errorbound\]), we see that the sparse signal can be recovered with an arbitrarily small error by letting $\boldsymbol{\delta}\rightarrow\boldsymbol{0}$, or in other words, by setting the threshold vector $\boldsymbol{\tau}$ sufficiently close to the unquantized data $\boldsymbol{y}$. This result has two important implications. Firstly, sparse signals can be reconstructed with negligible errors even from one-bit measurements. We note that [@JacquesLaska11; @PlanVershynin11] also showed that sparse signals can be recovered with a bounded error from one bit quantized data. But their analysis ignores the magnitude ambiguity and confines sparse signals on the unit Euclidean sphere. Such a sphere constraint is no longer necessary for our analysis. Secondly, our result suggests that the best thresholds should be set as close to the original data samples $\boldsymbol{y}$ as possible. So far there has no theoretical guarantee indicating that other choice of thresholds can also lead to a stable recovery with an arbitrarily small reconstruction error.
Proof of Theorem \[theorem1\] {#sec:proof}
=============================
Let $\boldsymbol{a}_i^T$ denote the $i$th row of the sampling matrix $\boldsymbol{A}$. Clearly, to ensure the sign consistency (\[condition-1\]) for each component, we should either have $$\begin{aligned}
\text{sign}(\boldsymbol{a}_i^T\boldsymbol{h})=\text{sign}(\delta_i)
\label{condition-2}\end{aligned}$$ or $$\begin{aligned}
|\boldsymbol{a}_i^T\boldsymbol{h}|<|\delta_i| \label{condition-3}\end{aligned}$$ Note that in the latter case the sign of $(\boldsymbol{a}_i^T\boldsymbol{h}+\delta_i)$ does not flip regardless of the sign of $\boldsymbol{a}_i^T\boldsymbol{h}$. Therefore if there exists a $2K$-sparse residual vector $\boldsymbol{h}$ such that each component of $\boldsymbol{Ah}$ satisfies either (\[condition-2\]) or (\[condition-3\]), then $\boldsymbol{\hat{x}}=\boldsymbol{x}+\boldsymbol{h}$ is the solution of (\[opt-1\]). Our objective in the following is to show that such a residual vector $\boldsymbol{h}$ is bounded by the deviation vector $\boldsymbol{\delta}$ with an arbitrarily high probability.
Without loss of generality, we decompose $\boldsymbol{A}$ and $\boldsymbol{\delta}$ into two parts: $\boldsymbol{A}=[\boldsymbol{A}_{1}^T\phantom{0}\boldsymbol{A}_{2}^T]^T$, $\boldsymbol{\delta}=[\boldsymbol{\delta}_{1}^T\phantom{0}\boldsymbol{\delta}_{2}^T
]^T$, according to the two possible relationships between $\boldsymbol{a}_i^T\boldsymbol{h}$ and $\delta_i$: $$\begin{aligned}
\left.
\begin{array}{cc}
\text{sign}(\boldsymbol{A}_{1}\boldsymbol{h})=\text{sign}(\boldsymbol{\delta}_{1})
\\
|\boldsymbol{A}_{2}\boldsymbol{h}|< |\boldsymbol{\delta}_{2}|
\end{array} \right. \label{event}\end{aligned}$$ where $\boldsymbol{A}_{1}\in\mathbb{R}^{m_1\times n}$, $\boldsymbol{A}_{2}\in\mathbb{R}^{m_2\times n}$, $m_1+m_2=m$, and in the second equation, both the absolute value operation $|\cdot|$ and the inequality symbol “$<$” applies entrywise to vectors.
We now analyze the probability of the residual vector $\boldsymbol{h}$ being greater than $\lambda\epsilon$, given that the condition (\[event\]) is satisfied. This conditional probability can be denoted as $P(\|\boldsymbol{h}\|_2>\lambda\epsilon|\text{E})$, where we use “E” to denote the event (\[event\]). Clearly, $P(\|\boldsymbol{h}\|_2>\lambda\epsilon|\text{E})$ can also be explained as the probability that the event $\|\boldsymbol{h}\|_2>\lambda\epsilon$ will occur, when the event $\text{E}$ has occurred. To facilitate our analysis, we divide the event (\[event\]) into two disjoint sub-events which are defined as $$\begin{aligned}
&\text{E1: The event (\ref{event}) holds true for $m_2\geq
\kappa$} \nonumber
\\
&\text{E2: The event (\ref{event}) holds true for $m_2<\kappa$}
\nonumber\end{aligned}$$ Clearly, the union of these two sub-events is equal to the event (\[event\]). Utilizing Bayes’ Theorem, the probability $P(\|\boldsymbol{h}\|_2>\lambda\epsilon|\text{E})$ can be expressed as $$\begin{aligned}
P(\|\boldsymbol{h}\|_2>\lambda\epsilon|\text{E})
\stackrel{(a)}{=}&P(\|\boldsymbol{h}\|_2>\lambda\epsilon,\text{E})
\nonumber\\
=&P(\|\boldsymbol{h}\|_2>\lambda\epsilon,\text{E1}\cup\text{E2})
\nonumber\\
\stackrel{(b)}{=}&P(\|\boldsymbol{h}\|_2>\lambda\epsilon,\text{E1})+P(\|\boldsymbol{h}\|_2>\lambda\epsilon,\text{E2})
%\nonumber\\
%=&P(\|\boldsymbol{h}\|_2>\lambda\epsilon|\text{Ea})P(\text{Ea})\nonumber\\
%&+ P(\|\boldsymbol{h}\|_2>\lambda\epsilon|\text{Eb})P(\text{Eb})
\label{eq8}\end{aligned}$$ where $(a)$ holds because the events (\[event\]) is a prerequisite condition that is always met to ensure the sign consistency, i.e. $P(\text{E})=1$, and $(b)$ follows from the fact that the probability of the union of two disjoint events is equal to the sum of their respective probabilities.
Let us first examine the probability $P(\|\boldsymbol{h}\|_2>\lambda\epsilon,\text{E1})$. We show that the events $\|\boldsymbol{h}\|_2>\lambda\epsilon$ and E1 are two mutually exclusive events which cannot occur at the same time. To see this, note that when the event E1 occurs, we should have $$\begin{aligned}
|\boldsymbol{A}_{2}\boldsymbol{h}|<
|\boldsymbol{\delta}_{2}|\Rightarrow\|\boldsymbol{A}_{2}\boldsymbol{h}\|_2^2<\epsilon^2
\label{eq5}\end{aligned}$$ in which $\boldsymbol{A}_{2}\in\mathbb{R}^{m_2\times n}$ and $m_2\geq \kappa$. On the other hand, recalling that any $\kappa\times n$ sub-matrix formed by selecting certain rows of $\boldsymbol{A}$ satisfies the condition (\[theorem1:condition1\]), we have $$\begin{aligned}
\|\boldsymbol{A}_{2}\boldsymbol{h}\|_2^2
\geq\mu\|\boldsymbol{h}\|_2^2 \label{eq6}\end{aligned}$$ Combining (\[eq5\]) and (\[eq6\]), we arrive at $$\begin{aligned}
\|\boldsymbol{h}\|_2<\frac{\epsilon}{\sqrt{\mu}}\triangleq\lambda\epsilon\end{aligned}$$ which is contradictory to the event $\|\boldsymbol{h}\|_2>\lambda\epsilon$. Hence the events E1 and $\|\boldsymbol{h}\|_2>\lambda\epsilon$ cannot occur simultaneously, which implies $$\begin{aligned}
P(\|\boldsymbol{h}\|_2>\lambda\epsilon,\text{E1})=0 \label{eq7}\end{aligned}$$
Substituting (\[eq7\]) into (\[eq8\]), the probability $P(\|\boldsymbol{h}\|_2>\lambda\epsilon|\text{E})$ is simplified as $$\begin{aligned}
P(\|\boldsymbol{h}\|_2>\lambda\epsilon|\text{E}) =
P(\|\boldsymbol{h}\|_2>\lambda\epsilon,\text{E2}) \label{eq20}\end{aligned}$$ The probability $P(\|\boldsymbol{h}\|_2>\lambda\epsilon,\text{E2})$, however, is still difficult to analyze. To circumvent this difficulty, we, instead, derive an upper bound on $P(\|\boldsymbol{h}\|_2>\lambda\epsilon,\text{E2})$: $$\begin{aligned}
&P(\|\boldsymbol{h}\|_2>\lambda\epsilon,\text{E2}) \nonumber\\
=&P(\|\boldsymbol{h}\|_2>\lambda\epsilon|\text{E2}) P(\text{E2})
\leq P(\text{E2}) \nonumber\\
\stackrel{(a)}{=}&
P(\text{sign}(\boldsymbol{A}_{1}\boldsymbol{h})=\text{sign}(\boldsymbol{\delta}_{1}),
|\boldsymbol{A}_{2}\boldsymbol{h}|< |\boldsymbol{\delta}_{2}|, m_2<\kappa) \nonumber\\
\leq&
P(\text{sign}(\boldsymbol{A}_{1}\boldsymbol{h})=\text{sign}(\boldsymbol{\delta}_{1}),
m_2<\kappa) \nonumber\\
=&
P(\text{sign}(\boldsymbol{A}_{1}\boldsymbol{h})=\text{sign}(\boldsymbol{\delta}_{1}),
m_1> m-\kappa) \nonumber\\
\stackrel{(b)}{=}&P\bigg(\bigcup_{i=0}^{\kappa-1}\Omega_i\bigg)
\label{eq10}\end{aligned}$$ where $(a)$ comes from the definition of the event E2, and in $(b)$, the event $\Omega_i$ is defined as $$\begin{aligned}
\Omega_i:\phantom{0}
&\text{sign}(\boldsymbol{A}_{1}\boldsymbol{h})=\text{sign}(\boldsymbol{\delta}_{1})
\nonumber\\
&\text{where $\boldsymbol{A}_{1}\in\mathbb{R}^{m_1\times n}$, and
$m_1=m-i$}\end{aligned}$$ which means that there exist at least $m-i$ components in $\boldsymbol{Ah}$ whose signs are consistent with the corresponding entries in $\boldsymbol{\delta}$. Note that since the rest $i$ components are not explicitly specified in $\Omega_i$, the event $\Omega_i$ include all possibilities for the rest $i$ components. Based on the definition of $\Omega_i$, we can infer the following relationship: $\Omega_{i_1}\supset\Omega_{i_2}$ for $i_1>i_2$. This is because $\Omega_{i_2}$ can be regarded as a special case of the event $\Omega_{i_1}$ with some of the unspecified $i_1$ components also meeting the sign consistency requirement. With this relation, the upper bound derived in (\[eq10\]) can be simplified as $$\begin{aligned}
P(\|\boldsymbol{h}\|_2>\lambda\epsilon,\text{E2})\leq
P\bigg(\bigcup_{i=0}^{\kappa-1}\Omega_i\bigg)
=P(\Omega_{\kappa-1}) \label{eq19}\end{aligned}$$ We note that for the event $\Omega_{\kappa-1}$, $\boldsymbol{A}_{1}$ is not specified and can be any submatrix of $\boldsymbol{A}$. Considering selection of $m-\kappa+1$ rows (out of $m$ rows of $\boldsymbol{A}$) to construct $\boldsymbol{A}_{1}$, the event $\Omega_{\kappa-1}$ can be expressed as a union of a set of sub-events $$\begin{aligned}
\Omega_{\kappa-1}=\bigcup_{j=1}^J \Omega_{\kappa-1}^{j}
\label{eq11}\end{aligned}$$ where $J\triangleq C(m,m-\kappa+1)$, $C(m,k)$ denotes the number of $k$ combinations from a given set of $m$-elements, and each sub-event $\Omega_{\kappa-1}^{j}$ is defined as $$\begin{aligned}
\Omega_{\kappa-1}^{j}:
\text{sign}(\boldsymbol{A}_{1}\boldsymbol{h})=\text{sign}(\boldsymbol{\delta}_{1})\phantom{0}
\text{where $\boldsymbol{A}_{1}=\boldsymbol{A}[I_j,:]$}\end{aligned}$$ in which $I_j$ is an unique index set which consists of $m-\kappa+1$ non-identical indices selected from $\{1,2,\ldots,m\}$, $\boldsymbol{A}[I_j,:]$ denotes a submatrix of $\boldsymbol{A}$ constructed by certain rows from $\boldsymbol{A}$, and the indices of the selected rows are specified by $I_j$. From (\[eq11\]), we have $$\begin{aligned}
P(\Omega_{\kappa-1})=&P\bigg(\bigcup_{j=1}^J
\Omega_{\kappa-1}^{j}\bigg) \stackrel{(a)}{\leq} \sum_{j=1}^J
P(\Omega_{\kappa-1}^{j}) \label{eq12}\end{aligned}$$ where the inequality $(a)$ follows from the fact that the probability of a union of events is no greater than the sum of probabilities of respective events. The inequality becomes an equality if the events are disjoint. Nevertheless, the sub-events $\{\Omega_{\kappa-1}^{j}\}$ are not necessarily disjoint and may have overlappings due to the $\kappa-1$ unspecified components.
We now analyze the probability $P(\Omega_{\kappa-1}^{j})$. To begin with our analysis, we introduce the concept of orthant originally proposed in [@JacquesLaska11] for analysis of one-bit compressed sensing. An orthant in $\mathbb{R}^m$ is a set of vectors that share the same sign pattern, i.e. $$\begin{aligned}
\mathcal{O}_{\tilde{\boldsymbol{u}}}=\{\boldsymbol{u}\in\mathbb{R}^m|\text{sign}(\boldsymbol{u})=\tilde{\boldsymbol{u}}\}\end{aligned}$$ A useful result concerning intersections of orthants by subspaces is summarized as follows.
\[lemma1\] Let $\mathcal{S}$ be an arbitrary $k$-dimensional subspace in an $m$-dimensional space. Then the number of orthants intersected by $\mathcal{S}$ can be upper bounded by $$\begin{aligned}
I(m,k)\leq 2^{k} C(m,k)\end{aligned}$$ where $C(m,k)$ denotes the number of $k$-combinations from a set of $n$-elements.
See [@JacquesLaska11a Lemma 8].
The probability $P(\Omega_{\kappa-1}^{j})$ of our interest can be interpreted as, the probability of the vector $\boldsymbol{A}_{1}\boldsymbol{h}$ lying in the same orthant as $\boldsymbol{\delta}_{1}$ for a given $\boldsymbol{A}_1$. We first examine the number of sign patterns the vector $\boldsymbol{A}_{1}\boldsymbol{h}$ could possibly have. Let $S$ denote the set of all possible sign patterns for $\boldsymbol{A}_{1}\boldsymbol{h}$, i.e. $$\begin{aligned}
S=\{\boldsymbol{u}=\text{sign}(\boldsymbol{A}_{1}\boldsymbol{h})|&\text{$\boldsymbol{h}\in
\mathbb{R}^{n}$ is a $2K$-sparse vector}, \nonumber\\
& \boldsymbol{A}_{1}=\boldsymbol{A}[I_j,:]\}\end{aligned}$$ Also, let $T$ denote the support of the sparse residual vector $\boldsymbol{h}$, we can write $$\begin{aligned}
\boldsymbol{g}\triangleq\boldsymbol{A}_{1}\boldsymbol{h}=\boldsymbol{A}_{1}[:,T]\boldsymbol{h}_T\end{aligned}$$ where $\boldsymbol{g}\in\mathbb{R}^{m-\kappa+1}$, and $\boldsymbol{A}_1[:,T]$ denotes a submatrix of $\boldsymbol{A}_1$ obtained by concatenating columns whose indices are specified by $T$. We see that $\boldsymbol{g}$ is a linear combination of $2K$ columns of $\boldsymbol{A}_{1}$, and thus $\boldsymbol{g}$ lies in an $2K$-dimensional subspace spanned by the columns of $\boldsymbol{A}_{1}[:,T]$. Recalling Lemma \[lemma1\], we know that the number of orthants intersected by this subspace is upper bounded by $2^{2K}C(m-\kappa+1,2K)$. Therefore the vector $\boldsymbol{g}$ which lies in this subspace has at most $2^{2K}C(m-\kappa+1,2K)$ possible sign patterns. Note that this result is for a specific choice of the index set $T$. The selection of the support $T$ from $n$ entries has at most $C(n,2K)$ combinations. Therefore, in summary, the number of sign patterns in the set $S$ is upper bounded by $$\begin{aligned}
N_{\text{SP}}\leq 2^{2K}C(m-\kappa+1,2K)C(n,2K) \label{eq9}\end{aligned}$$ The probability $P(\Omega_{\kappa-1}^{j})$ can be calculated as $$\begin{aligned}
P(\Omega_{\kappa-1}^{j})=&
P(\text{sign}(\boldsymbol{\delta}_{1})\in S) \nonumber\\
\stackrel{(a)}{=}&\frac{N_{\text{SP}}}{2^{(m-\kappa+1)}}\leq\frac{2^{2K}C(m-\kappa+1,2K)C(n,2K)}{2^{(m-\kappa+1)}}
\label{eq13}\end{aligned}$$ where $(a)$ comes from the fact that $\boldsymbol{\delta}$ is a vector whose entries are independently generated from a certain distribution with equal probabilities being positive and negative, and $\boldsymbol{\delta}_{1}$ has $2^{(m-\kappa+1)}$ possible sign patterns.
Combining (\[eq20\]), (\[eq19\]), (\[eq12\]) and (\[eq13\]), we arrive at $$\begin{aligned}
&P(\|\boldsymbol{h}\|_2>\lambda\epsilon|\text{E})=P(\|\boldsymbol{h}\|_2>\lambda\epsilon|\text{E2})
\leq P(\Omega_{\kappa-1}) \nonumber\\
\leq&
\frac{2^{2K}C(m-\kappa+1,2K)C(n,2K)C(m,\kappa-1)}{2^{(m-\kappa+1)}}\end{aligned}$$ where the last inequality comes from $J=C(m,m-\kappa+1)=C(m,\kappa-1)$. Utilizing the following inequality [@JacquesLaska11] $$\begin{aligned}
C(a,b)\leq \left(\frac{ae}{b}\right)^b\end{aligned}$$ in which $e\approx2.718$ denotes the base of the natural logarithm, the probability $P(\|\boldsymbol{h}\|_2>\lambda\epsilon|\text{E})$ can be further bounded by $$\begin{aligned}
P(\|\boldsymbol{h}\|_2>\lambda\epsilon|\text{E})\leq\frac{a}{b}
\label{eq15}\end{aligned}$$ where $$\begin{aligned}
a\triangleq\left(\frac{n(m-\kappa+1)e^2}{(2K)^2}\right)^{2K}
\left(\frac{me}{\kappa-1}\right)^{(\kappa-1)} \nonumber\end{aligned}$$ $$\begin{aligned}
b\triangleq 2^{(m-\kappa+1-2K)} \nonumber\end{aligned}$$ Examine the condition which ensures that $a/b$ is less than a specified value $\eta$, where $0<\eta<1$. Taking the base-2 logarithm on both sides of $(a/b)\leq \eta$ and rearranging the equation, we obtain $$\begin{aligned}
&m-2K\log (m-\kappa+1)-(\kappa-1)\log m \nonumber\\
&\geq 2K (\log (ne^2)-2\log(2K)+1)+\log(1/\eta)+c \label{eq16}\end{aligned}$$ where $$\begin{aligned}
c\triangleq(\kappa-1)(\log(e/(\kappa-1))+1) \nonumber\end{aligned}$$ is a constant only dependent on $\kappa$. In summary, for a specified $\eta$, if the condition (\[eq16\]) is satisfied, then we can ensure that the probability of the residual vector $\boldsymbol{h}$ being greater than $\lambda\epsilon$ is smaller than $\eta$, i.e. $$\begin{aligned}
P(\|\boldsymbol{h}\|_2>\lambda\epsilon|\text{E})\leq \eta\end{aligned}$$ or $$\begin{aligned}
P(\|\boldsymbol{h}\|_2\leq\lambda\epsilon|\text{E})\geq 1-\eta\end{aligned}$$ The proof is completed here.
Quantization Design: Adaptive Methods {#sec:algorithm}
=====================================
In this section, we aim to develop a practical algorithm for one-bit compressed sensing. Previous analyses show that a reliable and accurate recovery of sparse signals is possible even based on one-bit measurements. This theoretical result is encouraging but we are still confronted with two practical difficulties while trying to recover the sparse signal via solving (\[opt-1\]). Firstly, the optimization (\[opt-1\]) is a non-convex and NP hard problem that has computational complexity growing exponentially with the signal dimension $n$. Hence alternative optimization strategies which are more computationally efficient in finding the sparse solution are desirable. Secondly, our theoretical analysis suggests that the quantization thresholds should be set as close as possible to the original unquantized measurements $\boldsymbol{y}$. However, in practice, the decoder does not have access to the unquantzed data $\boldsymbol{y}$. To overcome this difficulty, we will consider a data-dependent adaptive quantization approach whereby the quantization threshold vector is dynamically adjusted from one iteration to next, in a way such that the threshold come close to the desired values.
Computational Issue
-------------------
To circumvent the computational issue of (\[opt-1\]), we can replace the $\ell_0$-norm with alternative sparsity-promoting functionals. The most popular alternative is the $\ell_1$-norm. Replacing the $\ell_0$-norm with this sparsity-encouraging functional leads to the following optimization $$\begin{aligned}
\min_{\boldsymbol{z}}\|\boldsymbol{z}\|_1\quad
\text{s.t.}\phantom{0}\text{sign}(\boldsymbol{A}\boldsymbol{z}-\boldsymbol{\tau})=\boldsymbol{b}\label{opt-2}\end{aligned}$$ which is convex and can be recast as a linear programming problem that can be solved efficiently. Although a rigorous theoretical justification for $\ell_1$-minimization based optimization is still unavailable[^4], our simulation results indeed suggest that (\[opt-2\]) is an effective alternative to $\ell_0$-minimization and is able to yield a reliable and accurate reconstruction of sparse signals.
In addition to the $\ell_1$-norm, another alternative sparse-promoting functional is the log-sum penalty function. The optimization based on the log-sum penalty function can be formulated as $$\begin{aligned}
\min_{\boldsymbol{z}}\sum_{i=1}^n \log (|z_i|+\epsilon)\quad
\text{s.t.}\phantom{0}\text{sign}(\boldsymbol{A}\boldsymbol{z}-\boldsymbol{\tau})=\boldsymbol{b}\label{opt-3}\end{aligned}$$ where $\epsilon>0$ is a parameter ensuring that the penalty function is well-defined. Log-sum penalty function was originally introduced in [@CoifmanWickerhauser92] for basis selection and has gained increasing attention recently [@CandesWakin08; @WipfNagaranjan10]. Experiments and theoretical analyses show that log-sum penalty function behaves more like $\ell_0$-norm than $\ell_1$-norm, and has the potential to present superiority over the $\ell_1$-minimization based methods. The optimization (\[opt-3\]) can be efficiently solved by resorting to a bound optimization technique [@LangeHunter00; @CandesWakin08; @WipfNagaranjan10]. The basic idea is to construct a surrogate function $Q(\boldsymbol{z}|\boldsymbol{\hat{z}}^{(t)})$ such that $$\begin{aligned}
Q(\boldsymbol{z}|\boldsymbol{\hat{z}}^{(t)})-L(\boldsymbol{z})\geq
0\end{aligned}$$ where $L(\boldsymbol{z})$ is the objective function and the minimum is attained when $\boldsymbol{z}=\boldsymbol{\hat{z}}^{(t)}$, i.e. $Q(\boldsymbol{\hat{z}}^{(t)}|\boldsymbol{\hat{z}}^{(t)})=L(\boldsymbol{\hat{z}}^{(t)})$. Optimizing $L(\boldsymbol{z})$ can be replaced by minimizing the surrogate function $Q(\boldsymbol{z}|\boldsymbol{\hat{z}}^{(t)})$ iteratively. An appropriate choice of such a surrogate function for the objective function (\[opt-3\]) is given by [@CandesWakin08] $$\begin{aligned}
Q(\boldsymbol{z}|\boldsymbol{\hat{z}}^{(t)})=\sum_{i=1}^n\left\{\frac{|z_i|}{|z_i^{(t)}|+\epsilon}+
\log
(|z_i^{(t)}|+\epsilon)-\frac{|z_i^{(t)}|}{|z_i^{(t)}|+\epsilon}\right\}\end{aligned}$$ Therefore optimizing (\[opt-3\]) can be formulated as reweighted $\ell_1$-minimization which iteratively minimizes the following weighted $\ell_1$ function: $$\begin{aligned}
\min_{\boldsymbol{z}}\quad&
Q(\boldsymbol{z}|\boldsymbol{\hat{z}}^{(t)})=\sum_{i=1}^n
w_i^{(t)}|z_i|+\text{constant}\nonumber\\
\text{s.t.}\quad&
\phantom{0}\text{sign}(\boldsymbol{A}\boldsymbol{z}-\boldsymbol{\tau})=\boldsymbol{b}
\label{opt-4}\end{aligned}$$ where the weighting parameters are given by $w_i^{(t)}=1/(|z_i^{(t-1)}|+\epsilon),\forall i$. The optimization (\[opt-4\]) is a weighted version of (\[opt-2\]) and can also be recast as a linear programming problem. By iteratively minimizing (\[opt-2\]), we can guarantee that the objective function value $L(\boldsymbol{x})$ is non-increasing at each iteration. In this manner, the reweighted iterative algorithm eventually converges to a local minimum of (\[opt-3\]).
![Schematic of one-bit adaptive quantization for compressed sensing.[]{data-label="fig8"}](AQCS){width="9cm"}
Adaptive Quantization
---------------------
As indicated earlier, in addition to the computational issue, the other difficulty we face in developing a practical algorithm is that the suggested optimal quantization thresholds are dependent on the original unquantized data samples $\boldsymbol{y}$ which are inaccessible to the decoder. To overcome this difficulty, we consider an adaptive quantization strategy in which the thresholds are iteratively refined based on the previous estimate/reconstruction.
The basic idea of one-bit adaptive quantization is described as follows. At iteration $t$, we compute the estimated measurements $\boldsymbol{\hat{y}}^{(t)}$ at the decoder based on the sparse signal $\boldsymbol{\hat{x}}^{(t-1)}$ recovered in the previous iteration: $\boldsymbol{\hat{y}}^{(t)}=\boldsymbol{A}\boldsymbol{\hat{x}}^{(t-1)}$. This estimate is then used to update the quantization thresholds: $\boldsymbol{\tau}^{(t)}=\boldsymbol{\hat{y}}^{(t)}+\boldsymbol{\delta}^{(t)}$, where $\boldsymbol{\delta}^{(t)}$ is a vector randomly generated from a certain distribution. In practice, the deviation $\boldsymbol{\delta}^{(t)}$ should be gradually decreased to ensure that the thresholds will eventually come close to $\boldsymbol{y}$. The updated thresholds $\boldsymbol{\tau}^{(t)}$ are then fed back to the encoder. At the encoder, we compare the unquantized measurements $\boldsymbol{y}$ with the updated thresholds $\boldsymbol{\tau}^{(t)}$, and obtain a new set of one-bit measurements $\boldsymbol{b}^{(t)}$, which are sent to the decoder. Based on $\boldsymbol{\tau}^{(t)}$ and the new data $\boldsymbol{b}^{(t)}$, at the decoder, we compute a new estimate of the sparse signal $\boldsymbol{\hat{x}}^{(t)}$ via solving the optimization (\[opt-2\]) or (\[opt-3\]). A schematic of the proposed adaptive quantization scheme is shown in Fig. \[fig8\]. Note that throughout this iterative process, the unquantized measurements $\boldsymbol{y}$ are unchanged. For clarity, the one-bit adaptive quantization scheme is summarized as follows.
One-bit adaptive quantization scheme
1. Given an initial estimate $\boldsymbol{\hat{y}}^{(0)}$, and randomly generate an initial deviation vector $\boldsymbol{\delta}^{(0)}$ according to a certain distribution.
2. At iteration $t\geq 0$, let $\boldsymbol{\hat{y}}^{(t)}=\boldsymbol{\hat{y}}^{(0)}$ if $t=0$; otherwise compute $\boldsymbol{\hat{y}}^{(t)}$ as $\boldsymbol{\hat{y}}^{(t)}=\boldsymbol{A}\boldsymbol{\hat{x}}^{(t-1)}$. Based on $\boldsymbol{\hat{y}}^{(t)}$, update the thresholds as: $\boldsymbol{\tau}^{(t)}=\boldsymbol{\hat{y}}^{(t)}+\boldsymbol{\delta}^{(t)}$, where $\boldsymbol{\delta}^{(t)}$ for $t>0$ is randomly generated according to a certain distribution with a decreasing variance. Compare $\boldsymbol{y}$ with the updated thresholds $\boldsymbol{\tau}^{(t)}$ and obtain a new set of one-bit measurements $\boldsymbol{b}^{(t)}$.
3. Based on $\boldsymbol{\tau}^{(t)}$ and $\boldsymbol{b}^{(t)}$, compute a new estimate of the sparse signal $\boldsymbol{\hat{x}}^{(t)}$ via solving the optimization (\[opt-2\]) or (\[opt-3\]).
4. Go to Step 2 if $\|\boldsymbol{\hat{x}}^{(t)}-\boldsymbol{\hat{x}}^{(t-1)}\|>\omega$, where $\omega$ is a prescribed tolerance value; otherwise stop.
As indicated in [@LaskaWen11; @BourquardAguet10], an important benefit brought by the one-bit design is the significant reduction of the hardware complexity. One-bit quantizer which takes the form of a simple comparator is particularly appealing in hardware implementations, and can operate at a much higher sampling rate than the high-resolution quantizer. Also, one-bit measurements are much more amiable for large-scale parallel processing than high-resolution data. With these merits, the proposed adaptive architecture enables us to develop data acquisition devices with lower-cost and faster speed, meanwhile achieving reconstruction performance similar to that of using multiple-bit quantizer.
The proposed adaptive quantization scheme shares a similar architecture with the method [@KamilovBourguard12] in that both methods iteratively refine the thresholds using estimates obtained in previous iteration. Nevertheless, the rationale behind these two methods are different. Our proposed adaptive quantization scheme is based on Theorem \[theorem1\] which suggests that an arbitrarily small reconstruction error can be attained if the thresholds are set close enough to the unquantized measurements $\boldsymbol{y}$, whereas for [@KamilovBourguard12], a similar theoretical guarantee is unavailable.
Numerical Results {#sec:numerical-results}
=================
We now carry out experiments to corroborate our previous analysis and to illustrate the performance of the proposed adaptive quantization scheme. In our simulations, the $K$-sparse signal is randomly generated with the support set of the sparse signal randomly chosen according to a uniform distribution. The signals on the support set are independent and identically distributed (i.i.d.) Gaussian random variables with zero mean and unit variance. The measurement matrix $\boldsymbol{A}\in\mathbb{R}^{m\times n}$ is randomly generated with each entry independently drawn from Gaussian distribution with zero mean and unit variance. To circumvent the difficulty of solving (\[opt-1\]), we replace the $\ell_0$ norm with alternative sparsity-encouraging functionals, namely, $\ell_1$ norm and the log-sum penalty function. The new formulated optimization problems (c.f. (\[opt-2\]) and (\[opt-3\])) can be efficiently solved.
Performance under Different Threshold Choices
---------------------------------------------
We first examine the impact of the quantization design on the reconstruction performance. The knowledge of the original unquantized measurements $\boldsymbol{y}$ is assumed available in order to validate our theoretical results. The thresholds are chosen to be the sum of the unquantized measurements $\boldsymbol{y}$ and a deviation term $\boldsymbol{\delta}$, i.e. $\boldsymbol{\tau}=\boldsymbol{y}+\boldsymbol{\delta}$, where $\boldsymbol{\delta}$ is a vector with its entries being independent discrete random variables with $P(\delta_i=-a)=0.5$ and $P(\delta_i=a)=0.5$, in which the parameter $a>0$ controls the deviation of $\boldsymbol{\tau}$ from $\boldsymbol{y}$. Fig. \[fig1\] depicts the reconstruction normalized mean squared error (NMSE), $E[\frac{\|\boldsymbol{x}-\boldsymbol{\hat{x}}\|^2}{\|\boldsymbol{x}\|^2}]$, vs. the number of measurements $m$ for different choices of $a$, where we set $n=50$, and $K=3$. Results are averaged over $10^4$ independent runs. From Fig. \[fig1\], we see that the reconstruction performance can be significantly improved by reducing the deviation parameter $a$. In particular, a NMSE as small as $10^{-6}$ can be achieved when $a$ is set $0.001$. This corroborates our theoretical analysis that sparse signals can be recovered with an arbitrarily small error by letting $\boldsymbol{\delta}\rightarrow\boldsymbol{0}$. Also, as expected, the reconstruction error decreases with an increasing number of measurements $m$. Nevertheless, the performance improvement due to an increasing $m$ is mild when $m$ is large. This fact suggests that the choice of quantization thresholds is a more critical factor than the number of measurements in achieving an accurate reconstruction. From Fig. \[fig1\], we also see that $\ell_1$-minimization and log-sum minimization provide similar reconstruction performance. In Fig. \[fig2\], we plot the root mean squared error (RMSE), $E[\|\boldsymbol{x}-\boldsymbol{\hat{x}}\|_2]$, as a function of the deviation magnitude $\epsilon=\|\boldsymbol{\delta}\|_2=\sqrt{m}a$, where we set $m=100$, $n=120$, $K=2$ and $a$ varies from $10^{-3}$ to $1$. It can be observed that the RMSE decreases proportionally with the value $\epsilon$, which coincides with our theoretical analysis (\[theorem1:errorbound\]).
To further corroborate our analysis, we consider a different way to generate the deviation vector $\boldsymbol{\delta}$, with its entries randomly generated according to a Gaussian distribution with zero mean and variance $\sigma^2$. Fig. \[fig3\] depicts the NMSE vs. the number of measurements $m$ for different values of $\sigma$. Again, we observe that a more accurate estimate is achieved when the thresholds get closer to the unquantized measurements $\boldsymbol{y}$.
Experiments are also carried out on real world images in order to validate our theoretical results. As it is well-known, images have sparse structures in certain over-complete basis, such as wavelet or discrete cosine transform (DCT) basis. In our experiments, we sample each column of the $256\times 256$ image using a randomly generated measurement matrix $\boldsymbol{A}\in
\mathbb{R}^{m\times 256}$. We then quantize each real-valued measurement into one bit of information by using the threshold $\tau_i=y_i +\delta_i$, in which $\delta_i$ is drawn from a Gaussian distribution $\mathcal{N}(0, 0.01)$. Fig. \[fig4\] shows the original image and the reconstructed images based on $m\times 256$ one-bit measurements, where $m$ is equal to 150, 200, and 250, respectively. We see that the images restored from one-bit measurements still provide a decent effect, given that the thresholds are well-designed.
Performance of Adaptive Quantization Scheme
-------------------------------------------
We now carry out experiments to illustrate the performance of the proposed adaptive quantization algorithm. For simplicity, the algorithm uses the optimization (\[opt-2\]) at Step 3 of each iteration. In our experiments, we set $n=50$, $K=2$, $\boldsymbol{\tau}^{(0)}=\boldsymbol{\delta}^{(0)}$ and the threshold vector is updated as $\boldsymbol{\tau}^{(t)}=\boldsymbol{\hat{y}}^{(t)}+\xi^{(t)}\boldsymbol{\delta}^{(t)}$ for $t>0$, where $\boldsymbol{\delta}^{(t)}, \forall t$ is a random vector with its entries following a normal distribution with zero mean and unit variance. The parameter $\xi^{(t)}$ controls the magnitude of the deviation error. We set $\xi^{(0)}=1$, and to gradually decrease the deviation error, $\xi^{(t)}$ is updated according to $\xi^{(t+1)}=\xi^{(t)}/10$. The NMSE vs. the number of iterations is plotted in Fig. \[fig5\], where we set $m=40$. Results are averaged over $10^3$ independent runs, with the sampling matrix and the sparse signal randomly generated for each run. From Fig. \[fig5\], we see that the adaptive algorithm provides a consistent performance improvement through iteratively refining the quantization thresholds, and usually provides a reasonable reconstruction performance within only a few iterations. Fig. \[fig6\] depicts the NMSEs as a function of the number of measurements for the one-bit adaptive quantization scheme and a non-adaptive one-bit scheme which uses $\boldsymbol{\tau}^{(0)}$ as its thresholds. For the adaptive scheme, the iterative process stops if $\|\boldsymbol{\hat{x}}^{(t)}-\boldsymbol{\hat{x}}^{(t-1)}\|_2<0.01$. Numerical results show that the adaptive scheme usually converges within ten iterations. We observe from Fig. \[fig6\] that the adaptive scheme presents a clear performance advantage over the non-adaptive method.
We examine the effectiveness of the adaptive quantization scheme for image recovery. In the experiments, we sample each column of the $128\times 128$ image using a randomly generated measurement matrix $\boldsymbol{A}\in \mathbb{R}^{m\times 128}$, and then quantize each real-valued measurement into one bit of information. The initial threshold vector is set to be $\boldsymbol{\tau}^{(0)}=\boldsymbol{y}+\xi^{(0)}\boldsymbol{\delta}^{(0)}$, where $\xi^{(0)}$ is chosen as $\|\boldsymbol{y}\|_l/m$ such that the deviation is comparable to the magnitude of entries in $\boldsymbol{y}$. For $t>0$, the threshold is then updated as $\boldsymbol{\tau}^{(t)}=\boldsymbol{\hat{y}}^{(t)}+\xi^{(t)}\boldsymbol{\delta}^{(t)}$, with $\xi^{(t)}=\xi^{(t-1)}/10$. Fig. \[fig8\] plots the images reconstructed by the proposed one-bit adaptive scheme and the non-adaptive one-bit scheme which uses $\boldsymbol{\tau}^{(0)}$ as its thresholds, where we set $m=256$. Fig. \[fig8\] demonstrates that the adaptive quantization scheme improves the reconstruction of the image significantly.
![Root mean-squared error versus $\epsilon$.[]{data-label="fig2"}](rmsevsepsilon){width="9cm"}
----------------------------------- ------------------------------------------------------------------------- ------------------------------------- --
{width="4.8cm"} {width="4.8cm"} {width="4.8cm"} {width="4.8cm"}
----------------------------------- ------------------------------------------------------------------------- ------------------------------------- --
![Normalized mean-squared error versus number of iterations.[]{data-label="fig5"}](msevsitr){width="9cm"}
![Normalized mean-squared errors versus the number of measurements.[]{data-label="fig6"}](AQmsevsm-2){width="9cm"}
------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- --
![Reconstruction of images from one-bit measurements: (a) Image reconstructed by non-adaptive one-bit scheme; (b) Image reconstructed by adaptive one-bit scheme.[]{data-label="fig8"}](image_NQ "fig:"){width="4.9cm"} ![Reconstruction of images from one-bit measurements: (a) Image reconstructed by non-adaptive one-bit scheme; (b) Image reconstructed by adaptive one-bit scheme.[]{data-label="fig8"}](image_AQ "fig:"){width="4.9cm"}
------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- --
Conclusion {#sec:conclusion}
==========
We studied the problem of one-bit quantization design for compressed sensing. Specifically, the following two questions were addressed: how to choose quantization thresholds, and how close can the reconstructed signal be to the original signal when the quantization thresholds are well-designed? Our analysis revealed that sparse signals can be recovered with an arbitrarily small error by setting the thresholds close enough to the original unquantized measurements. The unquantized measurements, unfortunately, are inaccessible to us. To address this issue, we proposed an adaptive quantization method which iteratively refines the quantization thresholds based on previous estimate. Simulation results were provided to corroborate our theoretical analysis and to illustrate the effectiveness of the proposed adaptive quantization scheme.
[^1]: Jun Fang, and Yanning Shen are with the National Key Laboratory on Communications, University of Electronic Science and Technology of China, Chengdu 611731, China, Emails: [email protected], [email protected]
[^2]: Hongbin Li is with the Department of Electrical and Computer Engineering, Stevens Institute of Technology, Hoboken, NJ 07030, USA, E-mail: [email protected]
[^3]: This work was supported in part by the National Science Foundation of China under Grant 61172114, and the National Science Foundation under Grant ECCS-0901066.
[^4]: A theoretical guarantee for $\ell_1$-minimization is under study and will be provided in our future work.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We present the source catalog and the properties of the $B-, R-$, and $I-$band images obtained to support the [*AKARI*]{} North Ecliptic Pole Wide (NEP-Wide) survey. The NEP-Wide is an [*AKARI*]{} infrared imaging survey of the north ecliptic pole covering a 5.8 deg$^2$ area over 2.5 – 6 $\micron$ wavelengths. The optical imaging data were obtained at the Maidanak Observatory in Uzbekistan using the Seoul National University 4k $\times$ 4k Camera on the 1.5m telescope. These images cover 4.9 deg$^2$ where no deep optical imaging data are available. Our $B-, R-$, and $I-$band data reach the depths of $\sim$23.4, $\sim$23.1, and $\sim$22.3 mag (AB) at 5$\sigma$, respectively. The source catalog contains 96,460 objects in the $R-$band, and the astrometric accuracy is about 0.15$\arcsec$ at 1$\sigma$ in each RA and Dec direction. These photometric data will be useful for many studies including identification of optical counterparts of the infrared sources detected by [*AKARI*]{}, analysis of their spectral energy distributions from optical through infrared, and the selection of interesting objects to understand the obscured galaxy evolution.'
author:
- 'Yiseul Jeon, Myungshin Im, Mansur Ibrahimov, Hyung Mok Lee, Induk Lee, and Myung Gyoon Lee'
title: |
Optical Images and Source Catalog\
of [*AKARI*]{} North Ecliptic Pole Wide Survey Field
---
INTRODUCTION
==============
Recently, the observational study of the cosmic star formation history has been at the center of our efforts to understand the evolution and the formation of galaxies. This is because the observational data play a key role in constraining physical processes governing the evolutionary history of galaxies on which outcomes of the theoretical predictions are strongly dependent. Many studies have been carried out to derive star formation rates in distant galaxies [e.g., @goto10; @lebo09; @lefl09; @redd08; @shim07; @im05], however, there is a major obstacle in such efforts - the dust extinction of the light coming from star formation.
Since the discovery of ultra-luminous infrared galaxies (ULIRGs) and luminous infrared galaxies (LIRGs) by the InfraRed Astronomical Satellite (IRAS), it has been recognized that a significant portion of the light from the star formation is missing in the optical/ultraviolet but emitted in infrared (IR). Therefore, statistical studies of infrared luminous galaxies to reveal the hidden cosmic star formation activities were major themes of the more recent infrared space telescope missions such as the Infrared Space Observatory (ISO), the [*Spitzer*]{} Space Telescope, and the [*AKARI*]{} Space Telescope. Such studies found that the star formation at higher redshifts ($z \sim 1$) are more obscured than at $z \sim 0$ [e.g., @lefl05; @bell07; @goto10] underlining the importance of infrared-based studies of galaxies. However, due to the lack of efficient coverage between 10 to 20 $\micron$ in the [*Spitzer*]{} imaging, comparison of the same rest-frame mid-infrared (MIR) quantities over a wide redshift range is not possible to further examine the usefulness of the MIR star formation indicators at various redshifts and the correlation between the star formation and the relative strengths of MIR spectroscopic features made by molecules such as PolyAromatic Hydrocarbon (PAH) and silicates. The additional MIR coverage at 10 to 20 $\micron$ is also advantageous in studying the MIR-excess emission in early-type galaxies [@ko09; @shim10], and the identification of Active Galactic Nuclei [AGN; e.g., @lee07; @huan09].
The [*AKARI*]{} satellite is an infrared space telescope launched in February 2006. By offering a contiguous, wide-field imaging capability in the wavelength range between 2.5 and 26$\micron$, [*AKARI*]{} complements the capability of the [*Spitzer*]{} space telescope, and provides an unprecedented view of the universe, especially at 11 and 15$\micron$. Because of the Sun-synchronous orbit of [*AKARI*]{}, deep observations can be done only at the North and the South ecliptic poles. Therefore, the North Ecliptic Pole (NEP) survey is one of the important extragalactic surveys of [*AKARI*]{}. The NEP survey has two major programs, NEP-Wide and NEP-Deep surveys. The NEP-Wide field covers 5.8 deg$^2$ centered at NEP, whereas the NEP-Deep field covers 0.38 deg$^2$ with deeper exposures than the NEP-Wide. Due to the wide field, the contiguous MIR wavelength coverage, and the low extinction value of $E(B-V)=0.047$ at the center of NEP [@schl98], the NEP-Wide field is unique for the studies of the extragalactic objects like infrared luminous galaxies and AGNs. For further information about NEP surveys, the readers can refer to @mats06, @wada07, and @lee07 [@lee09].
Optical imaging data can enhance the value of the unique [*AKARI*]{} NEP survey. Since [*AKARI*]{} has a low spatial resolution (the full width at half maximum (FWHM) of the point spread function (PSF) is about 6$\arcsec$ at 15$\micron$), the optical imaging with a higher resolution can help improving the astrometric accuracy and help deblending confused objects. Also, the optical bands are important to construct multi-wavelength spectral energy distributions (SEDs) of infrared sources such as galaxies and AGNs to understand their properties. There already exist deep optical data covering 2 deg$^2$ of the central region of the NEP-Wide field, obtained by the Canada-France-Hawaii Telescope (CFHT) with $u^*, g', r', i'$, and $z'$ filters [@hwan07]. However, there is no deep optical data for the remaining NEP-Wide area. The Digitized Sky Survey images are available, but neither their depths nor spatial resolutions are sufficient (the depth is about 20 mag at $R$ and the seeing is about 3$\arcsec$ with pixel scale of 1.0$\arcsec$ at the NEP area) for identifying optical counterparts of many infrared sources. Hence, we observed this remaining NEP-Wide area in optical $B, R$, and $I$ filters. The $BRI$ filters were chosen because we needed a multiple color set at least three colors to construct the optical part of the SED of galaxy; one blue part, one red part, and a middle part between red and infrared bands. Also, those Bessell filters are wider than the Sloan Digital Sky Survey (SDSS) filters, enabling us to achieve a higher signal to noise ratio (S/N) per unit exposure time than the SDSS filters. Throughout the paper, we use a cosmology with $\Omega_M=0.3,\Omega_\Lambda=0.7,$ and $H_0$ = 70 km s$^{-1}$Mpc$^{-1}$ [e.g., @im97]. We also use the AB magnitude system.
OBSERVATION
===========
The [*AKARI*]{} NEP-Wide field is centered at $\alpha = 18^h00^m00^s, \delta = +66\degr36\arcmin00\arcsec $ covering a circular area of 5.8 deg$^2$. Our optical $BRI$ imaging survey covers 4.9 deg$^2$ outside the central 2 deg$^2$ area of the [*AKARI*]{} NEP-Wide. Figure \[fig1\] shows the field coverage of different components of the NEP-Wide survey. The thick, solid circle indicates the [*AKARI*]{} NEP-Wide (5.8 deg$^2$) coverage. The dashed squares are the CFHT optical survey area. The small squares show the Maidanak optical imaging area. It consists of 70 subfields, each of them having an $18\arcmin\times18\arcmin$ field of view. These Maidanak fields are grouped into 4 different sections, East (8 fields), West (7 fields), North (23 fields), and South (32 fields), and we name each subfield with E01–E08, W01–W08, N01–N23, and S01–S32. The names of the subfields and their central positions are listed in Table \[tbl0\].
The images were obtained in Bessell $B$, $R$, and $I$ filters from 2007 June 12 to August 5 at the Maidanak Observatory in Uzbekistan, using the Seoul National University 4k $\times$ 4k Camera (SNUCAM) on the 1.5 m Richey-Chretien, AZT-22 telescope [@im10]. The CCD camera has a 4096 $\times$ 4096 format and a pixel scale of 0.266$\arcsec$ at f/7.75. We used the four amplifier mode for the CCD image readout.
Twilight flats were taken at the start and the end of each night when possible. Since the $I-$band images are found to have complicated fringe patterns, all of the $I-$band images were dithered by $\sim$20$\arcsec$ in random directions. These dithered images were used to construct the master fringe image. For photometric calibration, we observed photometric standard fields from @land92 at two different altitudes at least. In most subfields, standard star images were obtained on the same night, but in a few cases no corresponding standard star observations were carried out due to bad weather. In such cases, the photometric calibration was done using a different way, as described in section \[sec:calib\].
Table \[tbl1\] summarizes our observations. For each filter, typical total exposure times are 21 min in $B$, 30 min in $R$, and 20 min in $I$ filter. The images for a given subfield and a given filter were obtained mostly on a single night so that images are in homogeneous quality at least for a single filter of a given subfield. There are a few exceptions to this rule (3 subfields for $B-$band and 13 subfields for $I-$band). This happened when we needed to take images of a given subfield and a given filter over two nights due to observational constraints such as the lack of the observing time (e.g., the night was ending before finishing all the images of the subfield), or a rapidly deteriorating weather or seeing. In some cases, the total exposure times and the number of images of each subfield were adjusted to the weather conditions. For example, when there were thin clouds in the sky, the exposure time or the number of images were increased to compensate for the loss of the S/N. The frame times for individual exposures were also adjusted. Shorter exposure times (2 minutes) were used when there were no guide stars, or the weather condition was varying too rapidly. The $B-$ and $R-$band images were mostly taken during the same nights under dark/grey sky to provide images with similar seeing values, but the $I-$band images were taken mostly during the bright time since the depth of $I-$band images is the least influenced by the presence of the moon among the $BRI$ filters. The seeing condition for each subfield was determined as the median value of FWHM of non-saturated point sources with the magnitudes between 13 and 15 mag. In general, the seeing FWHM values of the stacked $R-$band and $I-$band images are stable at $\sim1.2\arcsec$ and $\sim1.1\arcsec$, respectively, but there were more variations in the seeing conditions in $B-$band and the $B-$band images have the worst seeing FWHM values in general among the $BRI$ filters. The maximum and the minimum seeing values among the $BRI$ filter images of a given subfield differ by a factor of 1.5 in median, but in extreme cases, the difference could be as large as a factor of $\sim$2 due to deteriorated seeing conditions in the $B-$band. The 5$\sigma$ detection limits of each subfield were determined from the sky background sigma, within an aperture diameter of 3 times FWHM value. The detection limits vary from subfield to subfield, but the median 5$\sigma$ detection limit in $R-$band is 23.1 mag. Table \[tbl2\] lists the seeings and 5$\sigma$ depths for each subfield. Figure \[fig2\] shows the seeing and the 5$\sigma$ detection limit of the three filters for each subfield.
DATA REDUCTION
==============
Preprocessing
-------------
We used IRAF[^1] `noao.imred.ccdred` package to process raw images. The data were processed with usual reduction procedures such as bias subtractions using a combined bias image and flat-field correction using twilight flat of each night. Dark correction was unnecessary since there is virtually no dark current in the images taken with our CCD camera.
Because the bias subtraction and the flat fielding could not remove the amplifier bias level variations perfectly due to small temporal variations in the bias level, some of the preprocessed images show slightly different bias levels in different quadrants, each representing a section of image read by a single amplifier. To remove this pattern, we determined the background values of the 4 quadrants, and subtracted the differences between the background levels of the 4 quadrant images from each quadrant. Then the background levels are normalized to the background level of a quadrant with the lowest background value.
We removed the fringe pattern in $I-$band following the steps below. First, we created a master fringe image combining all $I-$band images (716 images) in median. Then, before subtracting the master fringe frame, we re-scaled the master fringe frame by an appropriate factor. This scaling factor is taken as the ratio of the median background value of each science image to the master fringe frame background value. Finally, we subtracted the re-scaled master fringe frame from each $I-$band image. We checked the difference in image quality before and after the fringe removal. Before the fringe pattern removal, the large scale fringe pattern varies $\pm3\%$ of the background level which is comparable to the sky sigma value per pixel. After the fringe subtraction, the large scale variation is reduced to $< 0.5\%$ of the background or more than 6 times less than the sky sigma value, showing that the fringe pattern is removed successfully well below the sky sigma values.
After the fringe correction and the registration, all images were combined in median. Cosmic rays are removed during the stacking of multiple images with a sigma clipping adopting +3-sigma for rejecting outlier pixels.
Astrometry
----------
To derive astrometry solutions, we used a software named SCAMP which finds astrometry solution by cross-matching automatically an input reference catalog containing a set of accurate RA and Dec values of sources against a list of object positions obtained from an input FITS image [@bert06]. In our case, we used USNO-B1 catalog [@mone03] as the reference catalog for the astrometry calibration. Astrometry solutions were derived for all the stacked, subfield images in each filter separately. The degree of the polynomial for distortion correction is adopted to be 4, since it gave the best solution among other options. SCAMP can refine the astrometry solution by comparing positions of objects in overlapping images, but the procedure does not apply to our case.
To check the accuracy of the astrometry solution, we compared our catalog with the USNO-B1 catalog stars that had the positional errors of $<$ 0.2$\arcsec$ and the apparent magnitudes of 14 $< B1$ mag $<$ 19. Figure \[fig3\]-(a) shows the distribution of the differences in position between the USNO-B1 stars and the corresponding objects in our catalog in RA and Dec directions. Figures \[fig3\]-(b) and (c) show the histograms of the astrometric differences between our catalog and USNO-B1 catalog of RA and Dec, respectively. To derive the astrometry error, we fit the histogram of the astrometric differences with Gaussian functions. The dotted lines indicate the best-fit Gaussian function to the distributions. After the fitting, we find that the 1$\sigma$ astrometric errors are 0.16$\arcsec$ for RA and 0.15$\arcsec$ for Dec.
Photometric Calibration {#sec:calib}
-----------------------
We carried out the photometry calibration using three methods depending on the availability of the standard star data. The first method uses the Landolt standard stars [@land92]. This procedure applies to 25 fields in $B-$band, 25 fields in $R-$band, and 18 fields in $I-$band, where the standard star data were taken at two different altitudes at the least. The adopted standard calibration formula is $m_{ins}-M_{std}=k_0+k_1X$, where $m_{ins}$ is an instrumental magnitude ( $m_{ins} = 25 - 2.5$ $\log$ $(DN/sec)$ ), $M_{std}$ is a standard magnitude, $k_0$ is a zero point, $k_1$ is an airmass coefficient, and $X$ is an airmass. The instrumental magnitudes were measured within an aperture of a 12$\arcsec$ diameter on the standard stars. We checked with growth curves that this aperture is large enough to contain virtually all photons from the star irrespective of the seeing of our data. We chose the standard stars that have similar colors for deriving the coefficients of the above formula. To confirm the effect of not including the color terms, we checked the difference between the listed magnitudes of standard stars with various colors (those not included in the derivation of the photometric zero-points) and the magnitudes of these standard stars derived with our photometry solutions. We find that the rms difference is less than 0.05 mag. That is, the magnitudes derived with our photometry solutions are independent of the color terms well within the error 0.05 mag. This is consistent with the expected variation in the photometric zero points due to neglecting the color term, based on an observational campaign of photometry standards using the SNUCAM at Maidanak [@lim09]. Therefore we decided to neglect the color term of the formula. We found $k_0$ and $k_1$ values for each night and each filter, and derived the standard magnitudes of the objects using these values. The $k_0$ has the values in the range of 1.67 $\pm$ 0.13 at $B-$band, 1.70 $\pm$ 0.11 at $R-$band, and 2.07 $\pm$ 0.08 at $I-$band. Also, the $k_1$ have the values of 0.33 $\pm$ 0.05 at $B-$band, 0.10 $\pm$ 0.03 at $R-$band, and 0.04 $\pm$ 0.01 at $I-$band. These values are consistent with values derived from a campaign observation of standard stars [@lim09]. The zero-point errors were estimated from the rms magnitude differences between the known magnitudes of the standard stars and the magnitudes of the same standard stars derived from our zero-point solution. The zero-point error is about 0.032 mag using this method.
The second photometric calibration method is adopted for the data taken during the nights which have only one altitude data for standard stars (34, 40, and 27 fields at $B-$, $R-$, and $I-$band, respectively.). In such cases, we used other night’s $k_1$ value since the night-to-night variation in $k_1$ is not significant. This produces an additional zero-point error of 0.028 mag, at most. Consequently, the total error associated with the zero-point determination of the second photometric calibration method is less than 0.040 mag.
The third method of the photometric calibration is applied to the subfields without associated standard star observations. For such fields (11, 5, and 25 fields in $B-$, $R-$, and $I-$band, respectively.), we used photometry of sources in the overlapping area of a neighbor subfield. The corresponding sources that have magnitudes between 12 and 16 mag were used to measure the photometric zero-point. For this method, the zero-point error was derived from the standard deviation of the magnitude offset between the sources in the field where we want to determine the photometric zero-point and the matched sources in the neighbor fields. In this case, the zero-point error is about 0.045 mag with respect to the photometry of the neighbor field. Therefore, the total error in the zero-point for the third method is about 0.032–0.045 mag, depending on the method used for the photometry calibration of the neighbor field.
We checked the consistency of our photometric calibration using an overlap area between two neighboring fields (those that were not used for the photometric calibration of the other field). In the case of the first method, the mean photometry differences are $-0.006 \pm$ 0.028 mag in $B$ (30 fields for overlaps), 0.002 $\pm$ 0.033 mag in $R$ (23 fields for overlaps), and 0.001 $\pm$ 0.017 mag in $I$ magnitudes (17 fields for overlaps). In the case of the second method, the mean photometry differences are about 0.002 $\pm$ 0.030 mag in $B$ (37 fields for overlaps), $-0.006 \pm$ 0.040 mag in $R$ (44 fields for overlaps), and $-0.022 \pm$ 0.035 mag in $I$ magnitudes (20 fields for overlaps). Note that the consistency of our photometric calibration between different nights are well below 0.05 mag for all filters which is expected from the zero-point errors.
The $Z_p$ for each subfield are listed in Table \[tbl2\]. For the nights which have the $k_0$ and the $k_1$ values, the $Z_p$ means $25-(k_0+k_1X)$, i.e. the zero point for converting DN per exposure time to magnitude, including the airmass correction. Therefore we can determine the apparent magnitude $M$, through $M=Z_p - 2.5\log$ $(DN/sec)$. The $Z_p$ can be used for the purpose of checking the depth of each subfield. Figure \[fig4\] shows several depths of each subfield for the three filters. For each subfield, the depth are checked using the 5$\sigma$ detection limit, the 50$\%$ completeness (see section \[sec:comple\]), the 99$\%$ reliability (see section \[sec:reli\]), or the $Z_p$, as well as the total exposure time.
CATALOGS
========
Object Detection and Photometry {#sec:flags}
-------------------------------
SExtractor [@bert96] was used for the detection and the photometry of objects in the images. We stacked $B-, R-$, and $I-$band images together, and used the $BRI$ stacked image as our object detection image. After many trials of SExtractor parameter sets and examining the detected and the non-detected sources by eye, we chose a set of 1.2$\sigma$ for DETECT\_THRESH, 5 connected pixels for DETECT\_MINAREA, DEBLEND\_NTHRESH of 64, DEBLEND\_MINCONT of 0.005, 200 for BACK\_SIZE, and 3 for BACK\_FILTERSIZE as the optimal set of detection parameters. Lowering the detection limit produced too many spurious detections and increasing the detection limit tend to miss obvious sources. The adopted detection limit corresponds to the S/N of about 4.5 which is usually an optimal S/N limit for detection of objects [e.g., @shim06].
To cross-identify and perform the photometry of objects in each band, we used the ASSOC parameter in SExtractor with a matching radius of 0.658$\arcsec$, 3$\sigma$ value of the astrometric error. The above process may miss faint objects which are brighter in a particular filter than the others. To supplement the above source detection, we also ran SExtractor on images in each filter separately. This process reveals faint sources detected in a single filter image, but not detected in the $BRI$ stacked image. These sources detected in a single filter image were also matched with detections in other single filter images with the matching radius of 0.658$\arcsec$. Adding these faint sources, we detected 63,333 for $B$, 96,460 for $R$ and 70,492 sources for $I$ filter over the reliability of 99$\%$ (see section \[sec:reli\]). Among these objects, 16$\%$ in $B$, 33$\%$ in $R$, and 15$\%$ in $I$ are detected only in the single filter images. We derived aperture magnitudes using an aperture of diameter of 3 times FWHM and auto magnitudes with Kron-like elliptical apertures [@kron80] which are taken as the total magnitudes . The aperture correction value is 0.10 $\pm$ 0.02 mag for all the filters and all the subfields.
The magnitude errors are computed by combining the standard SExtractor estimates and the zero-point errors we estimated. The standard SExtractor error estimates are based on the Poisson statistics, and could underestimate the noise values when the sky noises in adjoining pixels are correlated due to the stacking of the dithered frames [@gawi06]. Using a similar procedure described in @gawi06, we calculated how the error within a square aperture of $N \times N$ pixels, $\sigma_{N}$, changes as a function of $N$. In the case where each pixel noise is uncorrelated with each other, $\sigma_{N} \propto N$. When the noise is correlated, the exponent of $N$ is greater than 1. In our data, we find $\sigma_{N} \propto N^{1.14}$. For a magnitude measured within an aperture with 3.0$\arcsec$ diameter (3 times a typical FWHM of our image), there are about 110 pixels (or $N \simeq 10$). Therefore, the photometric errors of faint objects ($R \gtrsim 22$ mag) could be underestimated by as much as about a factor of 1.4. For the brighter objects ($R \lesssim 20$ mag), the photometric errors are not affected by this effect, since their errors are dominated by the object noise.
For the same object that exist on two or more fields in the overlapping areas, we chose an object that had the smallest FLAGS value from SExtractor (the FLAGS has a nonzero value if the object is blended, saturated, and/or truncated at an edge of a image) and the smallest photometric error. Among the same objects, we only included the one with a better S/N or image quality in the final catalog. The examples of such cases are objects on the edge of an image or objects elongated due to astigmatism near the edge of the CCD in one image, but do not suffer such a problem in the other. At the bright end, the photometry is free from saturation and non-linearity above $R$ $\simeq$ 12 mag. We cut off the faint sources over the magnitude of false detection rate 1$\%$ (see section \[sec:reli\]).
We found spurious detections due to the reflection halos caused by the bright objects and due to faint, uncorrected fringe patterns in $R-$band images. In case of the reflection halos around bright stars, we divided the bright stars into three classes, and set a radius to define the region where spurious detections lie within it (120$\arcsec$ at 8 $\leq B$ mag $<$ 10 , 80$\arcsec$ at 10 $ \leq B$ mag $<$ 12 , 40$\arcsec$ at 12 $ \leq B$ mag $<$ 14). We flag detections inside these radii as ‘near$\_$bright$\_$obj’, since they are likely to be spurious detections or their photometry are not reliable. We also found many spurious detections in the outer regions of stacked $R-$band images, thus we flag objects that are not matched with $B-$ or $I-$band catalogs and exist within 30$\arcsec$ from the edge as ‘near$\_$edge’. We found 5,186 (5$\%$) of ‘near$\_$bright$\_$obj’ at $R-$band and 14,833 (15$\%$)of ‘near$\_$edge’ sources, and checked those spurious sources by eyes. These flags are included in the $BRI$ merged catalog. Also we flag non-stellar sources as ‘galaxy’ (see, section \[sec:galcut\]). The flags are marked at the last column of the catalog (Table \[tbl3\]).
Star-Galaxy Separation {#sec:galcut}
----------------------
To distinguish stellar sources from other extragalactic sources, we used a $B-H$ vs. $H-N2$ color-color diagram and the SExtractor stellarity index. Here, the stellarity indices measure the likelihood of a source being a point source or an extended source, with the value of 1 for a perfect point source, and 0 for a diffuse, extended source. The stellarities are measured from the single-filter stacked images. However, when we determine whether a source is point-like or extended, we adopted the stellarity value measured in a stacked image in the filter where the photometric error of the source is the smallest. Note that the stellarity values are nearly identical among different filter images when an object is sufficiently bright ($R < 19$ mag). We also note that the stellarity values are determined from the $R-$band images for most of the objects, since the $R-$band images generally have the highest S/N among the three filters.
The $H-$band magnitudes come from the catalog made from the $J-$ and $H-$band images of the NEP-Wide field which were took with FLAMINGOS on the 2.1m telescope at Kitt Peak National Observatory (Jeon et al., in preparation), and the $N2$ magnitudes come from the [*AKARI*]{} observation, of which the effective wavelength is 2.43$\micron$ [@lee07]. We matched the objects detected in $H-$ and $N2-$band with the objects detected in $B-, R-$, and $I-$band using a matching radius of 1.5$\arcsec$ (for $H-$band) and 3$\arcsec$ (for $N2-$band), respectively.
Figure \[fig5\] shows the $B-H$ vs. $H-N2$ diagram. @bark_01 used $B-J$ vs. $J-K$ colors to distinguish quasars from stars. Instead of $J-K$ color, we used a similar selection method using $H-N2$. On the left panel of Figure \[fig5\], the dots and the contours are made from the object detected in $B-$band with the magnitude error less than 0.1 and matched with $H$ and $N2$ detections. We can notice that there is a stellar locus at $H-N2$ $\simeq -0.5$ mag with 0 $<$ $B-H$ $<$ 5. Because most of stars have similar slopes at wavelength greater than 1$\micron$ on their SEDs, stars have $H-N2 \sim -0.5$ mag. We checked the stellar locus using stars (the crosses of the right panel of Figure \[fig5\]) in NEP-Wide field which had the stellarity greater than 0.8 and the magnitude of 10 $< H <$ 13. Most of the stars are located at the stellar locus with a few exceptions.
We also checked the location of galaxies using redshift tracks of galaxies. The dashed line is for a star-forming galaxy, M51 [@silv98] and the dotted line is for a passively evolving galaxy (see figure caption for the model used for this). We considered the intergalactic medium (IGM) attenuation [@mada96] for all redshift tracks. Most of the galaxy track lie in the area of $H-N2$ greater than $-0.2$ mag, but galaxies with z $\lesssim$ 0.1 lie at $H-N2 < -0.2$ mag. The PSF of galaxy at low redshift (z $<$ 0.1) is in general larger than that of star, so such sources can be identified as extended sources in the image. However, if the low redshift galaxies are faint and small, they may get misclassified as stars with the color cut.
Since the magnitude limit of the $H$ and $N2-$band are different from those in the optical, this method can be applied only to a relatively bright magnitude limit. Also, in order to identify bright, extended sources (namely, low redshift galaxies) in the stellar region on the color-color plot, we apply a stellarity cut where stellarity index has a value between 1 (point sources) and 0 (extended sources).
In summary, we define stellar sources as the following:
Objects with $H-N2 < -0.2$ mag and the stellarity index $>$ 0.2 or\
Objects with $H-N2 \geq -0.2$ mag and the stellarity index $>$ 0.8 or\
if there is no $H-N2$ color available, stellarity index $>$ 0.8
The first condition weeds out extended objects at low redshift, while the second condition allows us to select clear, point sources in the region of galaxies on the color-color magnitude. The third criterion applies to a small number of objects near the detection limit of optical images, since $H$ and $N2$ images show detections of 98$\%$ and 77$\%$ of objects brighter than the magnitude cut of the catalog at 50$\%$ completeness limit.
Catalog Format
--------------
The $B-, R-$, and $I-$band merged catalog is presented in Table \[tbl3\]. The catalog contains objects that are deemed to be reliable above 99$\%$ confidence where the confidence values are derived independently for each subfield in each filter (see section \[sec:reli\]). Our catalog contains each object’s identification number (ID), RA, Dec, the total magnitude and its error in each band, the aperture magnitude within a circular aperture with a diameter of 3 times FWHM and its error in each band, the stellarity value from SExtractor, the galactic extinction value from the extinction map of @schl98 for each band, and the flags described in section \[sec:flags\]. Note that the listed magnitudes are not corrected for the galactic extinction. The conversion formulas of Vega to AB magnitudes are $B($Vega$)=B($AB$)+0.09$, $R($Vega$)=R($AB$)-0.22$ and $I($Vega$)=I($AB$)-0.45$. The offset values between Vega and AB magnitude are computed from the Vega spectrum [@bohl04] and $B$, $R$, and $I$ filter response functions [@bess90] coupled with the CCD quantum efficiency curve. The magnitude error is the square root of the quadratic summation of the photometric measurement error and the zero-point error from the photometric calibration. We use a dummy value of 99.00 for non-detection.
Figure \[fig6\] shows the stellarity distribution (the left three panels) and the photometric error distribution (the right three panels) for all sources detected in $B-, R-$, and $I-$ band images. The white lines on the right panels show the mode values for each magnitude bins, with the 25$\%$ and 75$\%$ quartiles indicated with error bars. The spread of the data points in the magnitude versus error plots are caused mainly by varying depths of the data from different subfields. On average, the photometric error is less than 0.1 mag at $B$ $\lesssim$ 22.5 mag, $R$ $\lesssim$ 22 mag, and $I$ $\lesssim$ 21 mag.
Completeness {#sec:comple}
------------
Since the source detection near the detection limit is dominated by the detections performed on the single filter image ($\simeq90\%$ of the faintest sources of $R-$band are detected only in the single filter image) rather than the $BRI$ stacked images, we computed the completeness using the $B-$, $R-$, and $I-$band single filter images of each subfield, to examine the completeness of the source detection. To estimate the completeness, we created artificial objects for each magnitude bin from 12 to 23.5 mag and placed them randomly in the background regions of the single filter images. The artificial objects are made of 100 point sources and 100 extended sources where the elliptical galaxies account for 40$\%$ of the extended sources. These artificial objects are created using the `noao.artdata` package of the IRAF. After performing detection and photometry of these artificial sources with the same parameter setting as those applied to the actual images, we calculated the completeness, which is defined as the ratio of the number of detected artificial sources to the number of input objects for each magnitude bin. Figure \[fig7\] shows the completeness of $B-, R-$, and $I-$ band data. The gray solid lines in Figure \[fig7\] represent the completeness of all the subfields calculated with this method. We overploted the deepest (S05 in $B$, S05 in $R$ and N23 in $I$) and the shallowest (N22 in $B$, N18 in $R$, and S15 in $I$) fields in the 5$\sigma$ depths with the black solid lines. The galaxy number counts (see section \[sec:galnum\]) after the corrections based on these completeness result match the number counts from deeper surveys, meaning that our completeness estimates are reasonable. Figure \[fig4\] shows the 50$\%$ completeness as the asterisks for each subfield, as well as the 5$\sigma$ depth, the 99$\%$ reliability (see section \[sec:reli\]), the $Z_p$, and the total exposure time. The magnitudes of completeness 50$\%$ for each subfield are presented in Table \[tbl2\]. We can see that the magnitudes at 50$\%$ completeness are brighter than the ones of 5$\sigma$ detection limits by a few tenths of magnitude.
Reliability {#sec:reli}
-----------
We examined the reliability of our detections using false detection rates for each filter and for each subfield. Here, we define false detection rate as the ratio of the number of objects detected in a negative image to the number of objects detected in the original image at a given magnitude bin. The negative image means the original image multiplied by $-1$. By this definition, an image with noises only would have the false detection rate of 1, while an image containing objects but with no noise would have the false detection rate of 0. When estimating the false detection rates, we excluded areas where objects are marked with ‘near\_edge’, since the inclusion of such area could disproportionately overestimate the false detection rate in the other areas which occupy most of the image.
Figure \[fig8\] shows the false detection rates of $B-, R-$, and $I-$band data. The magnitude of the false detection rates 1$\%$ (or the reliability 99$\%$) of all the subfields are indicated in Table \[tbl2\]. Also, Figure \[fig4\] shows the 99$\%$ reliability as diamond symbols for each subfield, as well as the 5$\sigma$ depth, 50$\%$ completeness, the $Z_p$, and the total exposure time. Figure \[fig4\] show that the magnitude limit of reliability 99$\%$ is greater than the 50 $\%$ completeness limit for $B-$ and $R-$band, except $\simeq$20 fields. The 99$\%$ reliability limit in $I-$band is consistently brighter than the 50$\%$ completeness limit by 0.3 mag. The systematic trend arises from false detections of the fringe patterns left in $I-$band image even after the fringe subtraction. Field-to-field variations of false detection rates and the completeness limits are much larger for $R$-band than the other filters. It is due to the large difference of the image depths of each subfield in $R-$band, where some fields (from N02 to N23) have relatively small depth with the 5$\sigma$ limit $<$ 22.5. It is because of the moon phase since these images were taken during full moon phase (29 July – 1 August).
As a separate exercises, we examined the magnitude distribution of objects marked as ‘near\_edge’ or ‘near$\_$bright$\_$obj’. Note that most of these objects are likely to be spurious detections. 80$\%$ of these sources are found to be fainter than the magnitude limit of the 99$\%$ reliability. The 99$\%$ reliability cut of the catalog removes most of the spurious detections, but we note that objects marked with ‘near\_edge’ or ‘near$\_$bright$\_$obj’ should be treated with caution.
PROPERTIES OF THE DATA
======================
Galaxy Number Counts {#sec:galnum}
--------------------
To obtain the depth and homogeneity of this survey, we show the galaxy number count as a function of magnitude. The total area in which we used the galaxy number count of bright galaxies is about 3.8 deg$^2$ where $H-$ and $N2-$band data are also available. We take the Poisson error as the error estimate of number counts.
We calculated the number of galaxies per square degree at each magnitude bin. Since the depth of each subfield is different from another, the combined number counts of all the NEP survey fields were computed following the steps below. First, the number count was computed for each subfield up to the completeness limit of 50$\%$ as determined in section \[sec:comple\]. Then, the completeness correction was applied to the number count of each subfield. Finally, we combined the number counts from all the subfields to derive the final number count result. With this method, the bright end of the number counts, usually susceptible to field-to-field variations, include the data from all the 70 fields, while the deepest number counts include the data from the deepest subfields only. We also calculated the number count of stars using the stellar objects of magnitude brighter than the completeness limit of $\sim50\%$.
Figure \[fig9\] displays the galaxy and the stellar number counts in $B-, R-$, and $I-$band. The lower solid line with crosses represents raw number count and the upper solid line with crosses is the number count after applying correction for the completeness. The solid line with asterisks shows the number count of stars. Also we compare our number counts with those in the literature [@arno99; @arno01; @bert97; @capa04; @couc93; @hogg97; @infa86; @jone91; @kron78; @kumm01; @lill91; @mamo98; @mccr03; @metc95; @metc01; @pica91; @post98; @smai95; @stev86; @tyso88; @yasu01]. Since some of the filter systems in the literature are slightly different from ours, we applied the magnitude transformation by @metc95 [@metc01] to the literature values. We can see that our number counts match with others. The galaxies and stars number count data are provided in Table \[tbl4\].
Color-Magnitude Diagram {#sec:cmd}
------------------------
Figure \[fig10\] displays the color-magnitude diagrams of $R$ vs. $B-R$ (left) and its histograms (right). The upper panel shows the diagram for galaxies with $R$ magnitude error less than 0.1, and the lower panel shows for the stars as the same conditions with galaxies in the upper panel. We checked positions of galaxies on the color-magnitude diagram traced by the redshifts. The dotted line represents a redshift track of a star-forming galaxy, M51 [@silv98], and the dashed line is for a passively evolving galaxy. The SED of a passively evolving galaxy is for a 5 Gyr old stellar population, created with the @bruz03 model, assuming a spontaneous burst of single stellar population with metallicity of Z=0.02 and the Salpeter IMF. We included the attenuation of galaxy light due to the IGM when computing the colors [@mada96]. Two different lines represent galaxies with characteristic absolute magnitudes: for an early type galaxy, the characteristic absolute magnitude, $M^*(r^*)$, is taken as $-20.75+5 \log h$, and for a late type galaxy, $-20.30+5 \log h$ is adopted for $M^*(r^*)$ [@naka03]. To convert the $r^*-$band magnitude to the $R-$band magnitude, we used the magnitude transformation relation of @jest05, and $g-r \simeq 0.8 $ as the mean color of galaxies. The k corrections are calculated from the templates of two galaxies in a standard manner. The $B-R$ colors of elliptical and spiral galaxies are 1–1.5 mag at z=0.1. For 0.2 $\leq$ z $\leq$ 0.6, the elliptical galaxies have the $B-R$ values of 1.5–3 mag and the spiral galaxies have the $B-R$ values of 1–1.5 mag. The color-magnitude plot shows that our $B$ and $R-$band images detect galaxies at $z=0$ through $z=0.6$, and galaxies brighter than $M^*(r^*)$ to higher redshifts ($z \sim 1$). On the other hand, we find bimodality in the color-magnitude diagrams of stellar objects (bottom plot, Figure \[fig10\]). The bimodality originates from G dwarfs that dominate the color at $B-R$ $\simeq$ 0.5 and M Giant stars at $B-R$ $\simeq$ 2. Figure \[fig11\] displays the color-magnitude diagrams in $I$ vs. $R-I$ (left) and its histograms (right) with the sources of the $I$ magnitude error less than 0.1. The symbols and the tracks are the same as in Figure \[fig10\]. The $R-I$ color of elliptical and spiral galaxies is 0.3–0.5 at z $\lesssim$ 0.2 and 0.5–0.9 at z $\simeq$ 0.5. Like Figure \[fig10\], the distribution of objects in the $I$ versus $R-I$ plot confirm that $I-$band images have the depth similar to the $R-$band images in terms of detecting galaxies at different redshifts. We can notice that the model colors of $B-R$ and $R-I$ are consistent with the distributions of galaxies. We expect that the combination of the optical data and the [*AKARI*]{} infrared data to provide a unique view on the evolution of galaxies, such as the morphological transformation process through studying MID excess emission in early-type galaxies [@ko09].
Discussion
==========
Here, we investigate how our data compare with other similar optical datasets, and explore the potential of selecting high redshift galaxies and quasars using the data.
First, many extragalactic surveys have been performed with various depths and area coverage. The 5.8 deg$^2$ coverage of the NEP-Wide survey is wider than that of the Cosmic Evolution Survey [COSMOS, @scov07], or other deeper surveys such as DEEP2 [@davi03] and the Great Observatories Origins Deep Survey [GOODS, @dick03]. On the other hand, the area coverage is smaller than that of the [*Spitzer*]{} Wide-Area Infrared Extragalactic Survey (SWIRE) survey [@Lons03] by a factor of 10 with a similar depth in infrared as well as optical ancillary datasets. In terms of the depth and area coverage, our survey is most similar to the [*Spitzer*]{} First Look Survey [FLS, @fadd04]. Our infrared depth is a bit shallower than those of the FLS, while we provide a contiguous wavelength coverage including 11 to 15 micron. Consequently, our optical imaging depths are designed to match the infrared survey depths, thus about a magnitude shallower than the $R-$band depth of the FLS ancillary data.
Next, we discuss the usefulness of our datasets for selecting high redshift objects such as Lyman Break Galaxies (LBGs) and quasars. Figure \[fig12\] shows color-color diagrams in $B-R$ vs. $R-I$. We plotted sources with $R$ and $I$ magnitude errors less than 0.07 mag as gray points which correspond to objects with $R \lesssim$ 21.0–23.0 mag, and $I \lesssim$ 20.4–21.8 mag. The dashed line with diamonds represents a redshift track for a star-forming galaxy [M51, @silv98] and the dotted line with squares is a redshift track of a median composite spectrum of SDSS quasars from @vand01. The attenuation of the source light below Lyman $\alpha$ due to IGM is calculated using the model of @mada96. Using the redshift tracks as a guide, we define a selection box in the color-color diagram where objects at z $>$ 3.5 can be found [$B-$band dropouts; e.g., @stei99; @giav04; @kang09; @glik10]. The selection box is defined with the following equations:
$$\begin{aligned}
B-R & \geq & 2 \\
B-R & \geq & 4 (R-I) +0.8\end{aligned}$$
Although LBGs could be found in this selection box in principle, the expected number of the $B-$dropout galaxies in our data is very low. Extrapolating the surface number density of z $\sim$ 4 LBGs from @stei99 assuming an exponential decrease of the brightest objects as in the Schechter function, we expect about one or less than one $B-$dropout galaxies at $I <$ 21.5 mag. On the other hand, the number density could be higher by a factor of 10 or so, if the bright end of the luminosity function flattens due to AGN contribution as seen in the luminosity functions of $U-$band dropout objects at $z \sim 3$ [e.g., @hunt04; @shim07]. In our data, we identify $\sim$800 $B-$dropout candidates in the selection box, but we believe that most of these are interlopers such as stars and low redshift galaxies due to the large mismatch with the expected number.
The expected number of quasars over the 4.9 deg$^2$ is $\lesssim$ 10 at $I <$ 20 mag [@rich06], or 10 and 20 at $R <$ 21.5 mag and 22.5 mag, respectively [@glik10]. After inspecting images of the $B$-dropout candidates visually and examining their multi-wavelength SEDs from the optical through the MIR 11$\micron$-band of [*AKARI*]{}, we identify at least several objects that have the SED shapes consistent with those of high redshift quasars. Judging from the shapes of the SEDs, most of the interlopers are low redshift galaxies or late-type stars. The selection of quasar candidates at $z > 3.5$ is ongoing, and we expect that the identification of z $\sim$ 4 quasars at these magnitude limits will be valuable for constraining the faint-end slope of quasars at high redshift [e.g., @glik10].
SUMMARY
=======
We have shown the characteristics of $B-, R-$, and $I-$band data at Maidanak Observatory from the follow-up imaging observation of the [*AKARI*]{} NEP-Wide field. Using the SNUCAM on the 1.5m telescope, we covered 4.9 deg$^2$ with the 5$\sigma$ depth of $\sim$23.4, $\sim$23.1, and $\sim$22.3 magnitudes (AB) at $B-, R-$, and $I-$band, respectively. We detected 63,333 sources in $B-$band, 96,460 objects in $R-$band, and 70,492 in $I-$band. These data are now being used to identifying optical counterparts for accurate astrometry of infrared sources and deblending of confused sources. Also, our data provide a multi-wavelength coverage of the NEP-Wide field, enabling detailed SED analysis. Through these SEDs, we can get key physical properties such as stellar mass and photometric redshifts, and select interesting objects such as high redshift quasars. Spectroscopic follow-up surveys are being conducted on infrared luminous objects using the information gathered from this imaging survey.
We would like to thank the observers in Maidanak Observatory who performed the service observation over 2 months to obtain the data. This work was supported by the Korea Science and Engineering Foundation (KOSEF) grant No. 2009-0063616, funded by the Korea government (MEST), also in part by the Korea Research Foundation Grant funded by the Korean Government (MOEHRD), grant number KRF-2007-611-C00003.
Arnouts, S., et al. 2001, , 379, 740 Arnouts, A., D¡Odorico, S., Cristiani, S., Zaggia, S., Fontana, A., & Giallongo, E., 1999, , 341, 641 Barkhouse, W. A. & Hall, P. B. 2001, , 121, 2843 Bell, E. F., Zheng, X. Z., Papovich, C., Borch, A., Wolf, C., & Meisenheimer, K. 2007, , 663, 834 Bertin, E. 2006, Astronomical Data Analysis Software and Systems XV, 351, 112 Bertin, E. & Arnouts, S. 1996, , 117, 393B Bertin, E. & Dennefeld, M., 1997, , 317, 43 Bessell, M. S. 1990, , 102, 1181 Bohlin, R. C. & Gilliland, R. L. 2004, , 127, 3508 Bruzual, G., & Charlot, S. 2003, , 344, 1000 Capak, P., et al. 2004, , 127, 180 Couch, W.J., Jurcevic, J., & Boyle, B.J., 1993, , 260, 241 Davis, M., et al. 2003, , 4834, 161 Dickinson, M., & Giavalisco, M. 2003, The Mass of Galaxies at Low and High Redshift, ed. R. Bender & A. Renzini (New York: Springer), 324 Fadda, D., Jannuzi, B. T., Ford, A., & Storrie-Lombardi, L. J. 2004, , 128, 1 Gawiser, E., et al. 2006, , 162, 1 Giavalisco, M., et al. 2004, , 600, L93 Glikman, E., Bogosavljevi[ć]{}, M., Djorgovski, S. G., Stern, D., Dey, A., Jannuzi, B. T., & Mahabal, A. 2010, , 710, 1498 Goto, T., et al. 2010, , 514, A6 Hogg, D.W., Pahre, M.A., McCarthy, J.K., Cohen, J.G., Bland-ford, R., Smail, I., & Soifer, B.T., 1997, , 288, 404 Huang, J.-S., et al. 2009, , 700, 183 Hunt, M. P., Steidel, C. C., Adelberger, K. L., & Shapley, A. E. 2004, , 605, 625 Hwang, N. et al. 2007, , 172, 583 Im, M., Griffiths, R. E., & Ratnatunga, K. U. 1997, , 475, 457 Im, M. 2005, J. Korean Astron. Soc, 38, 135 Im, M., Ko, J., Cho, Y., Choi, C., Jeon, Y., Lee, I., & Ibrahimov, M. 2010, J. Korean Astron. Soc, 43, 75 Infante, L., Pritchet, C., & Quintana, H., 1986, , 91, 217 Jester, S. et al. 2005, , 130, 873 Jones, L.R., Fong, R., Shanks, T., Ellis, R.S., & Peterson, B.A., 1991, , 249, 481 Kang, E., & Im, M. 2009, , 691, L33 Ko, J., et al. 2009, , 695, L198 Kron, R. G. 1980, , 43, 305 Kron, R. G. 1978, PhD thesis, University of California at Berkeley. Kummel, M. W. & Wagner, S. J., 2001, , 370, 384K Landolt, A. U. 1992, , 104, 340 Lee, H. M. et al. 2007, , 59S, 529 Lee, H. M. et al. 2009, , 61, 375 Le Floc’h, E., et al. 2005, , 632, 169 Le Floc’h, E., et al. 2009, , 703, 222 Le Borgne, D., Elbaz, D., Ocvirk, P., & Pichon, C. 2009, , 504, 727 Lim, B. et al. 2009, JKAS, 42, 161 Lilly, S.J., Cowie, L.L., & Gardner, J.P., 1991, , 369, 79 Lonsdale, C. J., et al. 2003, , 115, 897 Madau, P., Ferguson, H. C., Dickinson, M. E., Giavalisco, M., Steidel, C. C., & Fruchter, A. 1996, , 283, 1388 Mamon, Gary A. 1998, in Wide Field Surveys in Cosmology, ed., S. Colombi, Y. Mellier, & B. Raban (Gif-sur-Yvette : Editions Frontie‘res), 323 Matsuhara, H. et al. 2006, , 58, 673 McCracken, H. J. et al. , 2003, , 410, 17 Metcalfe, N., Shanks, T., Campos, A., McCracken, H. J., & Fong, R. 2001, , 323, 795 Metcalfe, N., Shanks, T., Fong, R., & Roche, N. 1995, , 273, 257 Monet, D. G. et al. 2003, , 125, 984 Nakamura, O. et al. 2003, , 125, 1682N Picard, A., 1991, , 102, 445 Postman, M., Lauer, T.R., Szapudi, I., & Oegerle, W., 1998, , 506, 33 Reddy, N. A., Steidel, C. C., Pettini, M., Adelberger, K. L., Shapley, A. E., Erb, D. K., & Dickinson, M. 2008, , 175, 48 Richards, G. et al. 2006, , 131, 2766 Schlegel, D. J., Finkbeiner, D. P., & Davis, M. 1998, , 500, 525 Scoville, N., et al. 2007, , 172, 1 Shim, H., Im, M., Pak, S., Choi, P., Fadda, D., Helou, G., & Storrie-Lombardi, L. 2006, , 164, 435 Shim, H., Im, M., Choi, P., Yan, L., & Storrie-Lombardi, L. 2007, , 669, 749 Shim, H. et al. 2010, , submitted Silva, L. et al. 1998, , 509, 103 Smail, I., Hogg, D.W., Yan, L., & Cohen, J.G., 1995, , 449, L105 Steidel, C. C., Adelberger, K. L., Giavalisco, M., Dickinson, M., & Pettini, M. 1999, , 519, 1 Stevenson P. R. F., Shanks T., & Fong R., 1986, in Chiosi C., Renzini A., eds, Spectral Evolution of Galaxies. Reidel, Dordrecht p. 439 Tyson, J.A., 1988, , 96, 1 Vanden Berk, D. E. 2001, , 122, 549 Wada, T. et al. 2007, , 59S, 515 Yasuda, N. et al. 2001, , 122, 1104
[cccccc]{} E01 & 18:12:44.2 & 66:59:38.6 & N21 & 17:55:28.5 & 67:08:38.6\
E02 & 18:10:23.5 & 66:59:38.6 & N22 & 17:52:47.6 & 67:08:38.6\
E03 & 18:12:44.2 & 66:43:38.6 & N23 & 17:50:06.7 & 67:08:38.6\
E04 & 18:10:23.5 & 66:43:38.6 & S01 & 18:11:54.0 & 65:58:38.6\
E05 & 18:12:44.2 & 66:27:38.6 & S02 & 18:09:13.1 & 65:58:38.6\
E06 & 18:10:23.5 & 66:27:38.6 & S03 & 18:06:32.2 & 65:58:38.6\
E07 & 18:12:44.2 & 66:11:38.6 & S04 & 18:03:51.3 & 65:58:38.6\
E08 & 18:10:23.5 & 66:11:38.6 & S05 & 18:01:10.4 & 65:58:38.6\
W01 & 17:49:36.5 & 67:01:38.6 & S06 & 17:58:29.5 & 65:58:38.6\
W03 & 17:49:36.5 & 66:45:38.6 & S07 & 17:55:48.6 & 65:58:38.6\
W04 & 17:47:15.8 & 66:45:38.6 & S08 & 17:53:07.7 & 65:58:38.6\
W05 & 17:49:36.5 & 66:29:38.6 & S09 & 17:50:26.8 & 65:58:38.6\
W06 & 17:47:15.8 & 66:29:38.6 & S10 & 17:47:45.9 & 65:58:38.6\
W07 & 17:49:36.5 & 66:13:38.6 & S11 & 18:10:53.6 & 65:44:38.6\
W08 & 17:47:15.8 & 66:13:38.6 & S12 & 18:08:12.7 & 65:44:38.6\
N01 & 18:06:22.1 & 67:40:38.6 & S13 & 18:05:31.8 & 65:44:38.6\
N02 & 18:03:41.2 & 67:40:38.6 & S14 & 18:02:51.0 & 65:44:38.6\
N03 & 18:01:00.3 & 67:40:38.6 & S15 & 18:00:10.1 & 65:44:38.6\
N04 & 17:58:19.4 & 67:40:38.6 & S16 & 17:57:29.2 & 65:44:38.6\
N05 & 17:55:38.5 & 67:40:38.6 & S17 & 17:54:48.3 & 65:44:38.6\
N06 & 17:52:57.7 & 67:40:38.6 & S18 & 17:52:07.4 & 65:44:38.6\
N07 & 18:10:03.4 & 67:24:38.6 & S19 & 17:49:26.5 & 65:44:38.6\
N08 & 18:07:22.5 & 67:24:38.6 & S20 & 18:09:23.1 & 65:30:38.6\
N09 & 18:04:41.6 & 67:24:38.6 & S21 & 18:06:42.2 & 65:30:38.6\
N10 & 18:02:00.7 & 67:24:38.6 & S22 & 18:04:01.3 & 65:30:38.6\
N11 & 17:59:19.8 & 67:24:38.6 & S23 & 18:01:20.4 & 65:30:38.6\
N12 & 17:56:38.9 & 67:24:38.6 & S24 & 17:58:39.6 & 65:30:38.6\
N13 & 17:53:58.0 & 67:24:38.6 & S25 & 17:55:58.7 & 65:30:38.6\
N14 & 17:51:17.1 & 67:24:38.6 & S26 & 17:53:17.8 & 65:30:38.6\
N15 & 18:11:33.9 & 67:13:38.6 & S27 & 17:50:36.9 & 65:30:38.6\
N16 & 18:08:53.0 & 67:08:38.6 & S28 & 18:05:21.8 & 65:16:38.6\
N17 & 18:06:12.1 & 67:08:38.6 & S29 & 18:02:40.9 & 65:16:38.6\
N18 & 18:03:31.2 & 67:08:38.6 & S30 & 18:00:00.0 & 65:16:38.6\
N19 & 18:00:50.3 & 67:08:38.6 & S31 & 17:57:19.1 & 65:16:38.6\
N20 & 17:58:09.4 & 67:08:38.6 & S32 & 17:54:38.2 & 65:16:38.6\
[cccccc]{} $B$ & 20–45 (21) & 2–15 & 0.8–2.1 (1.4) &22.56–24.34 (23.40) & 22.05–23.86 (22.98)\
$R$&18–40 (30) & 2–10 & 0.8–1.7 (1.2) &22.07–24.14 (23.09) & 21.75–23.57 (22.60)\
$I$ & 16–64 (20) & 2 & 0.8–1.5 (1.1) &21.55–22.90 (22.33) & 21.16–22.37 (21.85)\
[cccccccccccccccccc]{} E01&22.88&0.84&23.95&23.63&23.62&&23.42&1.59&22.83&22.25&22.87&&23.32&1.01&22.16&21.79&21.53\
E02&22.90&0.97&23.73&23.37&23.73&&23.39&1.63&22.95&22.29&23.17&&23.32&1.11&21.94&21.61&21.53\
E03&22.90&0.94&23.75&23.38&23.71&&23.21&1.13&23.01&22.48&22.67&&23.32&1.11&21.98&21.69&21.12\
E04&22.87&1.19&23.85&23.51&23.45&&23.36&1.42&22.73&22.10&22.92&&23.32&1.09&22.12&21.81&21.02\
E05&22.85&1.49&23.18&22.80&23.54&&23.32&1.48&22.74&22.31&22.70&&23.32&0.97&22.11&21.66&20.91\
E06&22.84&1.91&23.07&22.56&23.00&&23.34&1.49&22.79&22.30&22.66&&23.32&1.02&22.05&21.79&20.82\
E07&22.86&1.44&23.40&22.88&23.22&&23.28&1.68&22.79&22.04&22.82&&23.32&1.15&21.84&21.16&20.78\
E08&22.85&1.52&23.40&22.98&23.57&&23.42&1.69&23.29&22.82&23.03&&23.32&1.18&21.71&21.29&20.79\
W01&22.88&1.97&22.97&22.41&22.89&&23.40&1.72&22.58&22.02&22.31&&23.37&0.92&22.44&21.97&21.72\
W03&22.87&1.73&23.19&22.89&22.80&&23.39&1.60&22.65&22.28&22.45&&23.37&1.00&22.42&21.88&21.41\
W04&22.84&1.73&23.14&22.77&22.81&&23.39&1.32&23.09&22.57&22.78&&23.37&1.02&22.33&21.84&21.55\
W05&22.81&1.20&23.46&23.08&23.00&&23.39&1.34&23.01&22.70&22.53&&23.37&0.97&22.43&21.96&21.73\
W06&22.78&1.42&23.16&22.81&23.01&&23.41&1.42&22.63&22.23&22.74&&23.36&1.00&22.36&21.75&21.48\
W07&22.84&1.55&23.21&22.81&23.09&&23.41&1.33&22.98&22.55&22.13&&23.36&0.96&22.40&22.11&21.26\
W08&22.82&1.47&23.06&22.75&22.28&&23.40&1.43&22.95&22.60&22.16&&23.36&0.95&22.38&22.11&21.54\
N01&22.80&1.67&22.65&22.23&22.02&&23.40&1.45&23.17&22.75&22.95&&23.36&0.98&22.36&22.04&21.63\
N02&22.76&1.71&23.00&22.48&22.79&&23.39&1.03&22.78&22.34&21.87&&23.36&1.08&22.14&21.79&21.52\
N03&22.68&1.34&23.35&22.95&23.11&&23.39&1.09&22.40&21.92&22.01&&23.40&1.10&22.34&21.84&21.95\
N04&22.65&0.99&23.29&23.06&22.41&&23.38&1.03&22.68&22.02&21.35&&23.40&1.17&22.29&21.83&21.79\
N05&22.61&1.35&23.15&22.82&23.17&&23.38&0.86&22.91&22.32&21.80&&23.40&1.05&22.48&22.03&21.91\
N06&22.57&1.32&23.23&22.91&22.58&&23.37&1.02&22.62&22.31&21.37&&23.40&1.03&22.56&22.23&21.98\
N07&22.71&1.74&23.08&22.57&22.95&&23.36&1.04&22.51&22.19&21.46&&23.40&1.07&22.26&21.85&21.96\
N08&22.69&1.44&23.22&22.79&22.72&&23.32&1.36&22.28&21.81&21.19&&23.40&0.98&22.34&21.82&22.08\
N09&22.80&1.25&23.40&22.88&23.36&&23.31&1.15&22.43&21.79&21.17&&23.40&1.11&22.27&21.89&21.45\
N10&22.78&1.69&22.98&22.51&23.24&&23.31&1.26&22.38&21.89&20.77&&23.38&1.18&22.60&22.09&22.11\
N11&22.74&1.32&23.28&22.95&22.96&&23.30&1.30&22.32&21.92&20.86&&23.38&1.06&22.52&22.11&22.17\
N12&22.70&1.32&23.36&23.03&22.99&&23.29&1.17&22.26&21.78&20.34&&23.38&1.13&22.31&21.87&22.22\
N13&22.84&1.51&23.21&22.75&23.20&&23.30&1.24&22.37&21.93&21.77&&23.38&1.06&22.60&22.25&22.38\
N14&22.80&2.13&22.83&22.30&22.71&&23.34&1.19&22.65&22.26&22.00&&23.38&1.04&22.50&22.00&21.84\
N15&22.77&2.13&22.66&22.05&22.79&&23.34&1.26&22.43&21.75&21.77&&23.31&1.03&22.49&22.23&21.82\
N16&22.85&1.52&23.32&22.84&23.29&&23.33&1.19&22.53&22.15&21.34&&23.31&0.96&22.33&22.02&21.26\
N17&22.80&1.52&23.25&22.85&22.70&&23.32&1.21&22.35&21.84&21.75&&23.37&1.27&22.36&21.74&21.23\
N18&22.81&1.49&23.38&23.05&23.15&&23.31&1.14&22.07&21.76&21.46&&23.34&1.18&22.34&21.74&21.71\
N19&22.72&1.34&22.73&22.36&21.52&&23.34&1.29&22.88&22.31&22.45&&23.33&1.13&22.18&21.87&21.46\
N20&22.72&1.43&23.06&22.76&22.70&&23.33&1.27&22.73&22.25&22.40&&23.37&1.13&22.33&22.10&20.99\
N21&22.77&1.84&23.02&22.54&23.31&&23.33&1.27&22.78&22.31&22.06&&23.37&1.07&22.43&21.97&21.46\
N22&22.77&1.71&22.56&22.17&22.43&&23.32&1.23&22.82&22.42&21.90&&23.36&1.06&22.44&22.13&20.61\
N23&22.77&1.86&22.91&22.45&22.96&&23.31&1.27&22.63&22.19&21.79&&23.39&0.90&22.90&22.23&21.96\
S01&22.96&1.58&23.72&23.18&23.76&&23.54&1.07&23.89&23.48&23.16&&23.42&0.88&22.84&22.37&20.77\
S02&22.94&1.45&23.80&23.28&23.82&&23.55&1.04&23.82&23.26&23.65&&23.14&1.52&22.20&21.79&21.93\
S03&23.10&1.23&24.10&23.43&24.18&&23.55&1.01&23.80&23.04&23.61&&23.43&0.80&22.85&21.94&19.94\
S04&22.96&1.32&24.02&23.62&23.62&&23.55&0.93&23.90&23.28&23.78&&23.12&1.41&22.03&21.52&21.36\
S05&23.09&1.19&24.34&23.86&23.10&&23.56&0.88&24.14&23.53&23.19&&23.42&1.23&22.40&21.78&21.39\
S06&23.10&1.05&24.30&23.77&24.16&&23.56&0.84&24.06&23.38&23.82&&23.12&1.18&22.33&21.76&21.53\
S07&23.10&1.44&24.00&23.56&24.12&&23.56&0.91&23.86&23.14&23.85&&23.06&1.40&22.23&21.81&21.60\
S08&22.99&1.55&23.82&23.34&24.12&&23.56&0.92&23.92&23.57&23.86&&23.03&1.50&21.61&21.28&21.25\
S09&23.00&1.35&23.93&23.44&24.18&&23.55&1.01&23.91&23.37&23.72&&23.14&1.06&22.37&22.09&21.97\
S10&23.00&1.24&24.09&23.75&23.96&&23.55&1.08&23.81&23.23&23.60&&23.03&1.08&21.81&21.57&21.55\
S11&23.00&1.39&24.08&23.71&23.78&&23.48&0.99&23.63&23.18&23.30&&22.73&1.07&21.58&21.28&20.86\
S12&23.00&1.36&24.00&23.36&24.03&&23.48&0.92&23.80&23.27&23.70&&22.68&1.10&22.06&21.78&20.73\
S13&23.00&1.37&24.04&23.48&23.50&&23.48&0.98&23.77&23.04&23.61&&22.78&1.09&21.89&21.35&19.30\
S14&23.00&1.41&23.97&23.62&23.96&&23.45&1.08&23.67&23.25&23.17&&23.29&1.04&22.51&22.08&21.58\
S15&22.98&1.31&23.81&23.37&23.97&&23.45&0.90&23.73&23.17&23.50&&22.18&1.18&21.55&21.21&20.28\
S16&22.97&1.36&23.91&23.44&23.61&&23.45&0.84&23.82&23.37&23.50&&23.28&1.09&22.60&22.24&21.77\
S17&22.98&1.29&23.79&23.41&23.68&&23.45&0.85&23.60&23.14&23.45&&23.27&1.10&22.51&21.86&21.00\
S18&22.98&1.53&23.63&23.19&23.98&&23.45&0.92&23.75&23.23&23.11&&23.21&1.11&22.40&22.00&21.51\
S19&22.97&1.59&23.57&23.11&23.44&&23.44&1.02&23.45&23.06&23.35&&23.21&1.01&22.40&21.98&21.64\
S20&22.97&1.45&23.71&23.30&23.55&&23.38&0.95&23.60&23.01&23.24&&23.16&1.32&21.87&21.36&21.40\
S21&22.96&1.68&23.45&22.96&23.35&&23.37&0.93&23.35&22.93&23.15&&23.24&1.39&21.83&21.32&21.51\
S22&23.05&1.49&23.72&23.26&23.72&&23.37&0.91&23.41&22.94&22.91&&23.33&1.34&21.93&21.51&21.82\
S23&23.04&1.14&23.89&23.34&23.87&&23.38&1.09&23.50&22.92&23.51&&23.04&1.32&21.74&21.29&21.30\
S24&22.91&1.15&23.66&23.26&23.98&&23.38&1.13&23.55&23.22&23.18&&23.35&1.05&22.29&21.75&21.55\
S25&22.89&1.06&23.72&23.31&23.66&&23.37&1.26&23.46&23.03&23.08&&23.36&1.07&22.12&21.77&21.63\
S26&22.62&1.58&23.29&22.87&23.01&&23.38&1.17&23.24&22.82&22.74&&23.36&1.07&22.24&21.94&21.23\
S27&22.83&1.00&23.63&23.23&23.36&&23.47&1.25&23.62&22.99&23.40&&23.36&1.13&22.23&21.78&21.04\
S28&22.82&0.96&23.64&23.28&23.50&&23.33&1.46&23.21&22.59&23.19&&23.36&1.08&22.25&21.95&21.39\
S29&22.98&1.97&22.92&22.41&23.57&&23.37&1.29&23.19&22.83&23.08&&23.36&1.11&22.22&21.92&21.26\
S30&22.52&1.88&23.02&22.42&23.07&&23.48&1.48&23.25&22.76&23.47&&23.35&1.03&22.24&21.93&21.40\
S31&22.94&1.47&23.18&22.82&22.81&&23.34&1.37&23.21&22.64&23.25&&23.35&1.00&22.33&21.80&21.61\
S32&22.92&1.49&23.24&22.87&22.42&&23.42&1.32&23.11&22.65&22.85&&23.35&1.03&22.27&21.95&21.58\
[cccccccccccccccccccc]{}
E01\_00001&18:11:06.3&+66:57:11.5&22.88& 0.12&23.00& 0.10&99.00&99.00&99.00&99.00&99.00&99.00&99.00&99.00&0.89&0.18&0.11&0.08&\
E01\_00002&18:11:08.9&+66:51:35.9&21.67& 0.10&22.35& 0.08&99.00&99.00&99.00&99.00&99.00&99.00&99.00&99.00&0.00&0.18&0.11&0.08&\
E01\_00003&18:11:09.3&+67:07:40.4&23.07& 0.16&23.32& 0.14&99.00&99.00&99.00&99.00&99.00&99.00&99.00&99.00&0.03&0.19&0.12&0.08&near\_bright\_obj\
E01\_00004&18:11:09.3&+66:51:19.9&23.59& 0.33&24.10& 0.27&99.00&99.00&99.00&99.00&99.00&99.00&99.00&99.00&0.18&0.18&0.11&0.08&galaxy\
E01\_00005&18:11:09.6&+67:07:54.8&23.39& 0.25&23.96& 0.24&99.00&99.00&99.00&99.00&99.00&99.00&99.00&99.00&0.64&0.19&0.12&0.08&near\_bright\_obj\
E01\_00006&18:11:09.6&+66:51:22.9&23.03& 0.21&24.02& 0.25&99.00&99.00&99.00&99.00&99.00&99.00&99.00&99.00&0.69&0.18&0.11&0.08&galaxy\
E01\_00007&18:11:10.2&+67:07:32.0&20.85& 0.08&22.52& 0.09&99.00&99.00&99.00&99.00&99.00&99.00&99.00&99.00&0.00&0.19&0.12&0.08&near\_bright\_obj\
E01\_00008&18:11:10.5&+67:07:51.7&21.90& 0.11&23.19& 0.13&99.00&99.00&99.00&99.00&99.00&99.00&99.00&99.00&0.01&0.19&0.12&0.08&near\_bright\_obj\
E01\_00009&18:11:11.2&+67:05:29.6&22.82& 0.18&23.59& 0.18&99.00&99.00&99.00&99.00&99.00&99.00&99.00&99.00&0.32&0.19&0.12&0.08&\
E01\_00010&18:11:11.2&+66:50:56.2&99.00&99.00&99.00&99.00&22.15& 0.12&22.21& 0.12&99.00&99.00&99.00&99.00&0.36&0.18&0.11&0.08&near\_edge\
E01\_00011&18:11:11.3&+66:50:53.4&99.00&99.00&99.00&99.00&21.91& 0.11&22.04& 0.12&99.00&99.00&99.00&99.00&0.47&0.18&0.11&0.08&near\_edge\
E01\_00012&18:11:11.4&+67:05:27.2&22.35& 0.14&23.33& 0.15&99.00&99.00&99.00&99.00&99.00&99.00&99.00&99.00&0.21&0.19&0.12&0.08&galaxy\
E01\_00013&18:11:11.4&+67:07:10.7&20.55& 0.06&20.60& 0.06&99.00&99.00&99.00&99.00&99.00&99.00&99.00&99.00&0.98&0.19&0.12&0.08&near\_bright\_obj\
E01\_00014&18:11:11.4&+66:50:51.2&99.00&99.00&99.00&99.00&22.77& 0.16&22.27& 0.14&99.00&99.00&99.00&99.00&0.46&0.18&0.11&0.08&near\_edge\
E01\_00015&18:11:11.5&+66:51:19.2&99.00&99.00&99.00&99.00&22.69& 0.17&22.46& 0.16&99.00&99.00&99.00&99.00&0.43&0.18&0.11&0.08&near\_edge\
E01\_00016&18:11:11.6&+67:05:27.1&22.95& 0.17&23.66& 0.19&99.00&99.00&99.00&99.00&99.00&99.00&99.00&99.00&0.71&0.19&0.12&0.08&\
E01\_00017&18:11:11.7&+67:07:12.8&20.50& 0.06&20.62& 0.06&99.00&99.00&99.00&99.00&99.00&99.00&99.00&99.00&0.96&0.19&0.12&0.08&near\_bright\_obj\
E01\_00018&18:11:11.7&+66:51:04.3&99.00&99.00&99.00&99.00&22.81& 0.19&22.53& 0.18&99.00&99.00&99.00&99.00&0.40&0.18&0.11&0.08&near\_edge\
E01\_00019&18:11:11.9&+67:07:22.9&23.59& 0.79&23.24& 0.14&99.00&99.00&99.00&99.00&99.00&99.00&99.00&99.00&0.00&0.19&0.12&0.08&near\_bright\_obj\
E01\_00020&18:11:12.1&+66:53:23.3&99.00&99.00&99.00&99.00&18.10& 0.05&19.19& 0.05&99.00&99.00&99.00&99.00&0.03&0.18&0.11&0.08&\
[ccccccccccccccc]{} 15.75&3.68E+00&9.85E-01&3.68E+00&1.46E+02&&1.13E+01&1.73E+00&1.13E+01&2.24E+02&&9.21E+00&1.56E+00&9.21E+00&2.72E+02\
16.25&3.42E+00&9.49E-01&3.42E+00&1.74E+02&&1.71E+01&2.12E+00&1.71E+01&2.96E+02&&1.76E+01&2.15E+00&1.76E+01&3.35E+02\
16.75&6.58E+00&1.32E+00&6.58E+00&2.19E+02&&2.58E+01&2.61E+00&2.58E+01&3.27E+02&&3.30E+01&4.28E+00&3.31E+01&4.03E+02\
17.25&1.26E+01&1.82E+00&1.26E+01&2.49E+02&&4.24E+01&3.34E+00&4.24E+01&3.91E+02&&5.93E+01&5.91E+00&5.95E+01&5.04E+02\
17.75&2.13E+01&2.37E+00&2.13E+01&2.92E+02&&8.58E+01&4.75E+00&8.58E+01&4.58E+02&&1.03E+02&9.62E+00&1.03E+02&5.84E+02\
18.25&2.97E+01&2.80E+00&2.97E+01&3.42E+02&&1.15E+02&8.29E+00&1.15E+02&5.25E+02&&1.95E+02&1.42E+01&1.97E+02&7.06E+02\
18.75&5.55E+01&3.82E+00&5.55E+01&3.73E+02&&2.19E+02&1.27E+01&2.20E+02&5.99E+02&&3.34E+02&3.08E+01&3.38E+02&7.96E+02\
19.25&8.05E+01&4.60E+00&8.05E+01&3.91E+02&&3.73E+02&2.81E+01&3.78E+02&7.24E+02&&5.07E+02&7.27E+01&5.30E+02&8.81E+02\
19.75&1.27E+02&5.77E+00&1.27E+02&4.78E+02&&6.00E+02&5.98E+01&6.20E+02&8.28E+02&&7.29E+02&1.02E+02&7.87E+02&9.12E+02\
20.25&1.99E+02&2.62E+01&2.02E+02&5.38E+02&&9.25E+02&9.37E+01&9.78E+02&9.36E+02&&1.08E+03&1.24E+02&1.24E+03&9.83E+02\
20.75&3.76E+02&5.74E+01&3.89E+02&6.09E+02&&1.26E+03&1.20E+02&1.38E+03&1.02E+03&&1.54E+03&1.49E+02&2.01E+03&1.05E+03\
21.25&6.59E+02&8.96E+01&7.10E+02&7.69E+02&&1.48E+03&1.38E+02&1.73E+03&1.00E+03&&2.02E+03&1.70E+02&3.12E+03&9.80E+02\
21.75&9.65E+02&1.12E+02&1.11E+03&8.38E+02&&1.71E+03&1.47E+02&2.21E+03&8.55E+02&&&&&\
22.25&1.41E+03&1.35E+02&1.88E+03&9.24E+02&&2.36E+03&1.71E+02&3.37E+03&7.24E+02&&&&&\
22.75&2.01E+03&1.58E+02&2.97E+03&9.31E+02&&3.37E+03&2.02E+02&5.56E+03&6.31E+02&&&&&\
23.25&3.13E+03&1.95E+02&5.09E+03&7.14E+02&&6.12E+03&2.73E+02&1.12E+04&3.98E+02&&&&&\
[^1]: IRAF is distributed by the National Optical Astronomy Observatory, which is operated by the Association of Universities for Research in Astronomy, Inc., under cooperative agreement with the National Science Foundation.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'In tomographic imaging, anatomical structures are reconstructed by applying a pseudo-inverse forward model to acquired signals. Geometric information within this process is usually depending on the system setting only, i.e., the scanner position or readout direction. Patient motion therefore corrupts the geometry alignment in the reconstruction process resulting in motion artifacts. We propose an appearance learning approach recognizing the structures of rigid motion independently from the scanned object. To this end, we train a siamese triplet network to predict the *reprojection error* (RPE) for the complete acquisition as well as an approximate distribution of the RPE along the single views from the reconstructed volume in a multi-task learning approach. The RPE measures the motion-induced geometric deviations independent of the object based on virtual marker positions, which are available during training. We train our network using 27 patients and deploy a 21-4-2 split for training, validation and testing. In average, we achieve a residual mean RPE of 0.013mm with an inter-patient standard deviation of 0.022mm. This is twice the accuracy compared to previously published results. In a motion estimation benchmark the proposed approach achieves superior results in comparison with two state-of-the-art measures in nine out of twelve experiments. The clinical applicability of the proposed method is demonstrated on a motion-affected clinical dataset.'
author:
- 'Alexander Preuhs, Michael Manhart, Philipp Roser, Elisabeth Hoppe, Yixing Huang, Marios Psychogios, Markus Kowarschik, and Andreas Maier, [^1] [^2] [^3]'
bibliography:
- '99lib.bib'
title: 'Appearance Learning for Image-based Motion Estimation in Tomography'
---
[Preuhs : Appearance Learning for Image-based Motion Estimation]{}
rigid motion compensation, reconstruction, interventional CBCT, autofocus, appearance learning
Introduction {#sec:introduction}
============
modeling [@Cootes1998] for interpreting images is a well examined problem in the field of computer vision. An appearance model is trained to extract invariant representations of an object of interest [@comaniciu2003kernel; @grabner2006real], e.g., for the tracking of faces [@babenko2010robust] or event detection [@xu2015learning]. Recently, Preuhs etal.[@preuhs2019image] have applied the strategy of appearance learning for motion detection in tomographic imaging.
The key concept of tomographic imaging is the reconstruction of internal patient anatomy from a series of measured signals. This can be the relaxation properties of hydrogen atoms in *magnetic resonance imaging* (MRI) or photon attenuation in X-ray *computed tomography* (CT). When reconstructing a tomographic image from measured signals, the geometry associated with each signal only depends on the system setting, i.e., source-detector orientation of a CT system or readout position for MRI scanners. The object itself is assumed to be static during the acquisition. As a consequence, patient motion corrupts the geometry alignment and results in motion artifacts within the reconstructed tomographic image.
Many efforts have been devoted to the problem of non-static objects, which are mainly splitted into non-rigid and rigid approaches. Rigid approaches reduce the number of unknowns to a 6 dimensional vector per measured signal, i.e., the respective rigid patient pose. However, complex movements, as apparent in heart imaging, are not reducible to such a simple model. In these cases non-rigid motion estimation must be deployed.
Non-Rigid Motion Compensation {#subsec:nonrigidmotioncompensation}
-----------------------------
Lauritsch etal.[@lauritsch2006towards] presented a gating approach, where the signal is binned to different motion states. Only similar motion states are used for reconstruction. This is extended by Taubmann etal.[@taubmann2016convex] who developed a primal-dual optimization scheme based on a spatial and temporal *total variation* (TV). Gating approaches were also presented by Larson etal.[@larson2004self] and Hoppe etal.[@hoppefree] for cardiac cine MRI, where the motion bin is deduced from the k-space center of each readout. Similar to gating, Fischer etal.[@fischer2017mr] devised an MRI-based model for X-ray fluoroscopy overlays. By binning of 4-D volumes to cardiac and respiratory phases, the motion field is estimated using 3-D/3-D registration.
Recent approaches deploy image-to-image translation from motion-affected reconstructions to such without motion artifacts. Here, prior knowledge on the expected manifold of motion free reconstructions is learned [@zhu2018image]. Kustner etal.[@kustner2019retrospective] and Latif etal.[@latif2018automating] propose a *conditional generative adversarial network* (cGAN) to synthesize motion free MRI reconstructions from a motion degenerated one. The same approach was presented for X-ray imaging by Xiao etal.[@xiao2019x].
Rigid Motion Compensation {#subsec:rigidmotioncompensation}
-------------------------
For many anatomical objects, the structure of the expected motion is already known a priori. The head, for example, is restricted by the skull to move as a rigid object. Further, many anatomies move in an approximate rigid structure during interventions, e.g., the knees or the hands. As the focus of this article is rigid motion compensation, we give a detailed overview of published methods which can be clustered into three categories.
### Projection Consistency {#subsubsec:projectionconsistency}
A computationally fast approach is projection consistency, where only the projection raw data are used, without the need for intermediate reconstructions. The main idea is that information is redundantly sampled by the forward operator with each acquired signal, e.g., the mass of the object. Powerful conditions are the *Helgason-Ludwig consistency conditions* (HLCC) [@helgasonradon] describing the relation between polynomials of degree $n$ and the respective $n$^th^ moment of the projections for parallel-beam CT or radial sampled MR. This was devised by Yu etal.[@yu2007data] to compensate motion in fan-beam geometry. A more broadly applicable approach based on the zero order HLCC and Grangeat’s theorem is epipolar consistency which was applied for geometric jitter and motion compensation in *cone-beam computed tomography* (CBCT) [@preuhs2019symmetry; @preuhs2018double; @Maass2014; @Frysch2015]. Similar approaches have been deployed in MRI motion compensation, where propeller trajectories measure the k-space center redundantly and compensate the motion based on this data redundancy [@pipe1999motion].
### Reconstruction Consistency {#subsubsec:reconstructionconsistency}
Contrary to projection consistency, reconstruction consistency solely uses tomographic images to estimate a rigid motion trajectory and is therefore often related as autofocus. The key idea is similar to the image-to-image translation approaches presented above: a motion-free reconstruction reveals some inherent properties which can be measured using an *image-quality metric* (IQM). In contrast to cGAN-based approaches, however, a motion trajectory is estimated by iterative optimization of the IQM. This ensures data integrity, which is of high importance in a clinical setting. The first application of this autofocus principle was presented by Atkinson etal.[@atkinson1997autofocus] for MRI reconstructions. They optimize a motion trajectory to find a reconstructed image with low entropy of the intensity histogram favoring images with high contrast structures and without motion ghosting or blur. Kingston etal.[@Kingston2011] presented a similar approach based on TV minimization. Subsequently various extensions were proposed [@wicklein2012image; @rohkohl2013improving; @sisniega2017motion; @herbst2019misalignment], including a combination of metrics as well as additional smoothness constraints.
### Data Consistency {#subsubsec:dataconsisntency}
The last category is based on enforcing data fidelity, which is the consistency of the reconstruction domain with the signal domain. In CBCT this approach is used for calibrating the system geometry by minimizing a *reprojection error* (RPE) of 3-D spheres on a calibration phantom and their respective 2-D projections [@strobel2003improving]. Using markers attached to the patient, this strategy was also investigated for motion compensation in MRI [@ooi2009prospective] and CT [@muller2015fully].
A second approach to enforce data fidelity is the virtual application of the forward model to an intermediate reconstruction and comparing these virtual data with the actually acquired data. Haskell etal.[@haskell2018targeted] used a SENSE forward model to maximize the data consistency with the acquired k-space data. For transmission imaging, *digitally rendered radiographs* (DRR) are commonly used to enforce consistency with the acquired projections $\cite{berger2016marker,Ouadah2016}$. In this context, Dennerlein etal.[@Dennerlein2012] exploit directly and indirectly filtered projections to compensate for geometric misalignment.
Potentials and Limitations in the State of the Art {#subsec:limitationsandprotentials}
--------------------------------------------------
Non-rigid approaches (seeSec\[subsec:nonrigidmotioncompensation\]) seem to be unfitting if the problem is known to be rigid. Image-to-image-based approaches do not exploit the full problem knowledge. Furthermore, their clinical applicability is limited because the consistency of the reconstructed image to the acquired data is not guaranteed and anatomic malformations can vanish [@huang2019data].
Consistency conditions (seeSec.\[subsubsec:projectionconsistency\]) have been used for the compensation of various other image artifacts as beam hardening, scatter correction or truncation correction [@abdurahman2018beam; @hoffmannempirical; @wurfl2017epipolar; @punzet2018extrapolation]. This is due to consistency being deduced from a physical model, which only holds on an approximate basis for real applications [@preuhs2019maximum]. Additionally, they are insensitive to certain motion directions and their application is limited to motion patterns outside the acquisition plane [@preuhs2019symmetry; @Frysch2015].
Image-based methods (seeSec.\[subsubsec:reconstructionconsistency\]) currently use hand-crafted features not particularly designed for the specific task. As a consequence, they are object dependent with each object revealing a different histogram entropy or TV.
A robust approach is based on reducing the RPE using markers (seeSec.\[subsubsec:dataconsisntency\]). However, this approach depends on additional marker placements, which has not found its way to clinical routine yet. Marker-free registration approaches are only working robustly if a prior reconstruction is available. Otherwise, the optimization becomes ill-posed, as the intermediate reconstruction, on which the forward model is applied, inherently reveals motion artifacts.
Deep learning has high potential to overcome some of those limitations by replacing bottlenecks of traditional methods with data-driven algorithms. For example, Bier etal.[@bier2018detecting] tackled the problem of manual marker placement by learning anatomical landmarks directly from the projection images. The presented cGAN-based approaches potentially have the risk of vanishing anatomical malformations, however, they may solve the chicken-egg problem for marker-free registration approaches. Additionally, many applications emerged for learning-based registration [@liao2019multiview; @toth20183d; @chou20132d]. They could potentially be extended for motion compensation scenarios.
Contribution {#subsec:contribution}
------------
Despite its great potential in improving rigid motion compensation algorithms, deep learning methods have caught limited attention from the research community. In Preuhs etal.[@preuhs2019image], we have presented the concept of learning image artifacts from a single axial reconstructed slice using a simplified motion model and a vanilla network architecture. The key concept is that a certain motion state is regressed to an object-independent measure defined by the RPE. We extend this line of thinking by developing a new data-driven approach for appearance learning capable of compensating motion artifacts. Our network architecture for motion appearance learning is based on a siamese triplet network trained in a multi-task scenario. Therefore, we incorporate not only a single axial slice but make use of information from 9 slices, extracted from axial, sagittal and coronal orientations. Using a multi-task loss, we estimate both (1) an overall motion score of the reconstructed volume similar to [@preuhs2019image] and (2) a prediction which projections are affected by the motion. To stabilize the network prediction, we deploy a novel pre-processing scheme to compensate for training data variability. These extensions allow us to learn realistic motion appearance, composed of three translation and three rotation parameters per acquired view. We evaluate the accuracy of the motion appearance learning in dependence of the patient anatomy and also the motion type. In a rigid motion estimation benchmark, we demonstrate the performance of the appearance learning approach in comparison to state-of-the-art methods. Finally, we demonstrate its applicability to real clinical data using a motion-affected clinical scan.
We devise the proposed framework for CBCT, however, by exchanging the backward model and training data, this approach is seamlessly applicable to radial sampled MRI or *positron emission tomography* (PET). In addition, by adjusting the regression target, also for Cartesian sampled MRI.
Rigid Motion Model for CBCT
===========================
Cone-Beam Reconstruction {#subsec:conebeamreconstruction}
------------------------
In tomographic reconstruction we compute anatomical structures denoted by $\atten$ from measurements $\meas$ produced with a forward model $\sysmat$ by $\sysmat\atten=\meas$. For X-ray transmission imaging $\atten$ are attenuation coefficients and $\meas$ are the attenuation line integrals measured at each detector pixel. The system geometry — e.g., pixel spacing, detector size and source-detector orientation — is part of the forward model $\sysmat$. Using the pseudo-inverse $$\atten = \sysmat^\top (\sysmat\sysmat^\top)^{-1} \meas$$ we get an analytic solution to this inverse problem, which consists of the back-projection $\sysmat^\top$ of filtered projection data $(\sysmat\sysmat^\top)^{-1} \meas$ [@maier2019learning], commonly known as *filtered back-projection* (FBP). For CBCT with circular trajectories, an approximate solution is provided by the *Feldkamp-Davis-Kress* (FDK) algorithm [@feldkamp1984practical]. The algorithm is regularly used for autofocus approaches [@Kingston2011; @sisniega2017motion] (seeSec.\[subsubsec:reconstructionconsistency\]) due to its low computational costs. Ritetal.[@rit2009comparison] have further shown that even due to its approximate nature, an FDK-based motion-compensated CBCT reconstruction is capable of correcting most motion artifacts. Thus, we use the FDK reconstruction algorithm, having the benefit of only filtering the projection images once and thereafter only altering the back-projection operator for motion trajectory estimation.
It is possible to formulate the FDK algorithm using a tuple of projection matrices $\pm = (\pm_0, \pm_1, ...,\pm_N)$ describing the geometry of operator $\sysmat$. The measurements $\meas$ are reshaped to a tuple of projection images $\proj = (\proj_0, \proj_0, ...,\proj_N)$. In analogy to [@feldkamp1984practical], we implement the FDK for a short scan trajectory using Parker redundancy weights $W_i(u,v)$ [@parker1982optimal], where $i \in [1,2, ..., N]$ describes the projection index and $(u,v)$ denotes a 2-D pixel. The first step is a weighting and filtering of the projection images $$\proj^\prime_i (u,v) = W_i(u,v) \int_{\mathbb{R}} \mathcal{F} \, \tilde{\proj}_i(\eta,v) \, \textnormal{e}^{i 2 \pi u v} \frac{|\eta|}{2} \, \textnormal{d} \eta \enspace,$$ with $\mathcal{F} \, \tilde{\proj}_i$ being the 1-D Fourier transform of the $i$^th^ cosine weighted projection image along the tangential direction of the scan orbit. Thereafter, a distance-weighted voxel-based back-projection is applied mapping a homogeneous world point $\worldpoint \in \mathbb{P}^3$ to a detector pixel described in the projective two-space $\mathbb{P}^2$ $$f_\textnormal{FDK}(\worldpoint,\pm,\proj) = \sum_{i \in N} U({\mathbf}{P}_i,\worldpoint) \proj^\prime_i(\phi_u(\pm_i \worldpoint), \phi_v(\pm_i \worldpoint))
\label{eq:fdk}$$ with $\pm_i$ describing the system calibration associated with $\proj_i$. (seeFig.\[fig:rpevisualization\]). The mapping function $\phi_\diamond : \mathbb{P}^2 \rightarrow \mathbb{R}$ is a dehomogenization $$\phi_\diamond((x,y,w)^\top) =
\begin{cases}
\frac{x}{w} & \textnormal{if $\diamond = u$} \\
\frac{y}{w} & \textnormal{if $\diamond = v$}
\end{cases} \enspace,
\label{eq:dehomogenization}$$ and $U({\mathbf}{P}_i,\worldpoint)$ is the distance weighting according to [@feldkamp1984practical].
Rigid Motion Model {#subsec:rigidmotionmodel}
------------------
![Visualization of the geometry for a point ${\mathbf}{a}$ and two geometries ${\mathbf}{P}_i$ and $\tilde{{\mathbf}{P}}_i$. The $L_2$ distance between the two projected points on the 2-D detector defines the RPE of the scene.[]{data-label="fig:rpevisualization"}](fig_1.pdf){width="\linewidth"}
We assume the rigid motion to be discrete w.r.t. the acquired projections. To this end, we define the motion trajectory ${\mathbf}{M}$ as a tuple of motion states ${\mathbf}{M}_i \in \mathbb{SE}(3)$ describing the orientation of the patient during the acquisition of the $i$^th^ projection $\proj_i$. Each motion state is associated to a projection matrix ${\mathbf}{P}_i$. The motion modulated trajectory is obtained by $${\mathbf}{P} \circ {\mathbf}{M} = ({\mathbf}{P}_0{\mathbf}{M}_0, \, {\mathbf}{P}_1{\mathbf}{M}_1, \, ..., \, {\mathbf}{P}_N{\mathbf}{M}_N) \enspace,$$ where $\circ$ is the element-wise matrix multiplication of two tuples. Typically, the motion trajectory is unknown and the task of motion compensation is to find a tuple of matrices ${\mathbf}{C}_i \in \mathbb{SE}(3)$ annihilating the resulting geometry corruption produced by ${\mathbf}{M}$. The compensation is successful if an annihilating trajectory ${\mathbf}{C} = ({\mathbf}{C}_0, ...{\mathbf}{C}_N)$ is found that suffices ${\mathbf}{C} \circ {\mathbf}{M} = {\mathbf}{1}$, with ${\mathbf}{1}$ being a tuple of identities.
Each motion matrix defined in $\mathbb{SE}(3)$ is parameterized by 3 rotations ($r_x,r_y,r_z$) and 3 translations ($t_x,t_y,t_z)$, describing Euler angles and translations along the corresponding coordinate axis, respectively. Therefore, the annihilating trajectory has $6N$ free parameters for an acquisition with $N$ projections. To reduce the high dimensionality, we model the trajectory using Akima splines [@akima1970new]. This reduces the free parameters to $6M$, where $M$ is the number of nodes typically chosen as $M \ll N$. Based on the expected frequency of the motion the number of spline nodes can be adapted.
Appearance Learning
===================
Conventionally, autofocus approaches are based on hand-crafted features, selected due to their correlation with an artifact-free reconstruction. For example, entropy gives a measure on contingency. As the human anatomy consists of mostly homogeneous tissues, entropy of the gray-value histogram can be expected to be minimal if all structures are reconstructed correctly. Motion blurs the anatomy or produces ghosting artifacts distributing the gray values more randomly. A similar rational is arguable for TV, which is also regularly used for constraining algebraic reconstruction [@taubmann2016convex]. Contrary to algebraic reconstruction, the motion estimation scenario is non-convex and optimization of a cost function based on hand-crafted image features is hardly solvable for geometric deviations exceeding a certain bound [@herbst2019misalignment].
We aim to overcome this problem by designing a tailored image-based metric, which reflects the appearance of the motion structure independent of the object.
Object-Independent Motion Measure
---------------------------------
Several metrics have been proposed to quantify image quality of motion affected reconstructions based on a given ground truth: the *structural similarity* (SSIM) [@silvia2019towards], the $L_2$ distance [@braun2018motion] or binary classification to motion-free and -affected [@meding2017automatic]. However, they were not used for the compensation of motion, but merely for the assessment of image quality, which is of high relevance in the field of MRI to automize prospective motion compensation techniques.
We choose the object-independent RPE for motion quantification. Its geometric interpretation is schematically illustrated in Fig.\[fig:rpevisualization\]. The RPE measures reconstruction relevant deviations in the projection geometry and is defined by a 3-D marker position $\worldpoint \in \mathbb{P}^3$ and two projection geometries $\pm_i, \tilde{\pm}_i$. We consider $\pm_i$ as the system calibration and $\tilde{\pm}_i = \pm_i {\mathbf}{M}_i$ as the actual geometry due to the patient motion. Accordingly, the RPE for a patient movement at projection $i$ is defined by $$d_\textnormal{RPE}({\mathbf}{P}_i,\tilde{\pm}_i,\worldpoint) = \left\Vert \, \begin{pmatrix}
\phi_u(\pm_i \worldpoint) \\[0.3em]
\phi_v(\pm_i \worldpoint) \\[0.3em]
\end{pmatrix}
-
\begin{pmatrix}
\phi_u({\tilde{\pm}_i} \worldpoint) \\[0.3em]
\phi_v({\tilde{\pm}_i} \worldpoint) \\[0.3em]
\end{pmatrix}\, \right\Vert_2^2
\label{eq:projectionrpe}$$ where $\phi_\diamond$ denotes the dehomogenization described in Eq.. Using a single marker, the RPE is insensitive to a variety of motion directions. Therefore, we use $K=90$ virtual marker positions $\worldpoint_k$, distributed homogeneously at three sphere surfaces with the radii $30$mm, $60$mm, and $90$mm. The high number of markers ensures that the RPE is view-independent, i.e., a displacement of a projection at the beginning of the trajectory has the same effect on the RPE as a displacement of a projection at the end of the trajectory. Accordingly, the overall RPE for a single view is $$d_\textnormal{RPE}({\mathbf}{P}_i,\tilde{{\mathbf}{P}}_i) = \frac{1}{K} \sum_{k=1}^K d_\textnormal{RPE}({\mathbf}{P}_i,\tilde{{\mathbf}{P}}_i,\worldpoint_k) \enspace.
\label{eq:rpe}$$ As shown in Strobel etal. [@strobel2003improving], Eq. can be rewritten to a measurement matrix ${\mathbf}{X}$ containing the marker positions, a vector ${\mathbf}{p}$ containing the elements of $\tilde{{\mathbf}{P}}_i$ and a vector ${\mathbf}{d}$ containing the respective marker positions. Given at least six markers, the components of ${\mathbf}{p}$ are estimated as the solution to $\Vert{\mathbf}{X} {\mathbf}{p} - {\mathbf}{d} \Vert^2_2$. Direct application of this method for motion compensation is non-trivial, as the accurate estimation of $\worldpoint_k$ is challenging. The 3-D marker positions must be estimated from projection images with corrupted geometry alignment.
Thus, we follow a different approach: we train a neural network to predict the RPE directly from the reconstructed images. To generate training data, we simulate rigid motion on real clinical acquisitions and compute the corresponding ground truth RPE via the virtual marker positions and their corresponding projections using Eq.. Thus, we aim to approximate Eq. from reconstructed slice images using a neural network.
Network Architecture {#subsec:networkarchitectur}
--------------------
{width="\linewidth"}
Our network architecture depicted in Fig.\[fig:networkarchitecture\] consists of two stages, a feature extraction stage followed by a regression stage. The feature extraction is driven by a siamese triplet network architecture consisting of three weight-sharing feed forward networks denoted by $\mathcal{\featurenet}$. The output of the three networks is concatenated and fed to the regression network $\regressionnet_t$. The feed forward network is almost identical to the ResNet-18 architecture [@he2016deep] upto the last global average pooling. We devise the network to our task by removing the last *fully connected* (FC) layer. Since the input of our network is always a tomographic reconstruction, we also remove all *batch normalization* (BN) layers. Expecting three-channel input images ranging from $\mathbb{R}^{70\times216}$ to $\mathbb{R}^{256\times216}$ for the different anatomical orientations, the final $7\times7$ average pooling is replaced by a $3\times3$ average pooling. The resulting feature maps are concatenated and represent the input to the regression network.
The regression network $\regressionnet_t$ is composed of a $1\times1$ convolution mapping the $1536\times3\times3$ feature maps to $2048 \times3\times3$ feature maps followed by an $1\times1$ global average pooling. The resulting feature maps are fed to four FC layers, each representing a different task $t \in \{1,2,3,4\}$. The first FC layer maps to a single scalar output $\regressionnet_1$, the other three FC layers map to $N$ dimensional outputs $\regressionnet_2$, $\regressionnet_3$ and $\regressionnet_4$, where $N$ represents the number of projections.
Data Generation {#subsec:datageneration}
---------------
Motion-affected reconstructions with corresponding ground truth motion patterns are rarely available. First, contrary to spiral CT, CBCT patient data are not available from public sources and therefore difficult to obtain in general. Second, the only robust motion compensation is based on external markers, which is not used in clinical practice. The only feasible possibility is the generation of artificial motion based on motion-free acquisitions. To this end, our data-base consists of 27 clinical head CBCTs, each being ensured to have no motion artifacts by a medical expert. The data are acquired with a clinical CBCT system (Artis Q, Siemens Healthcare GmbH, Forchheim, Germany). After filtering, the high resolution projection images are down-sampled to low resolution projection images $\proj_i \in \mathbb{R}^{320\times413}$ using an average filter. This improves the computational performance of the method and matches the voxel size of the volume reconstructed from these images. Down-sampling does barely affect the accuracy of the autofocus method [@atkinson1997autofocus].
The first step of the data generation is an alignment process of the 27 clinical CBCT scans to a mean shape. We perform this semi-automatically based on a symmetry plane alignment [@preuhs2019symmetry]. The result of the alignment process is a single rigid transformation which is incorporated into the system trajectory ${\mathbf}{P}$.
The second step is the generation of rigid motion which can be realized by two approaches: (1) the reconstructed volume is reprojected on a motion modulated trajectory using DRRs and then reconstructed again using ${\mathbf}{P}$ or (2) the calibrated system trajectory ${\mathbf}{P}$ is virtually altered by a motion trajectory and reconstructed. As DRRs are simulated projections, they cannot model the complexity of a real system and alter resolution and noise characteristics, where the latter is known to be critical for *convolutional neural network*s (CNN) [@huang2019data; @maier2019learning]. Therefore, we decide to choose strategy (2) where projection characteristics of a real clinical setting are preserved. Further, note that $\mathbb{P}^3$ is diffeomorphic to $\mathbb{SE}(3)$, as a consequence a rigid motion can be analogously expressed by a transformation of the system geometry $({\mathbf}{P}_i{\mathbf}{T}_i)\,\worldpoint$ or a transformation of the object ${\mathbf}{P}_i({\mathbf}{T}_i \worldpoint)$.
The motion generation is applied as follows: First, the system calibration ${\mathbf}{P}$ is altered by a motion trajectory ${\mathbf}{M}$, giving the effective trajectory ${\mathbf}{E} = {\mathbf}{P}\circ{\mathbf}{M}$. The motion trajectory contains a random misalignment in one of the 6 motion splines. The length of misaligned motions is chosen to be distributed over a third of the trajectory and restricted to views unaffected by the Parker redundancy weighting (seeSec.\[subsec:conebeamreconstruction\]). The redundancy weighting alters the appearance of motion artifacts — e.g., any translations of the last few views would barely affect the reconstruction quality as those projections are mostly outfaded — making a consistent mapping of artifact pattern to RPE infeasible. Secondly, two values are computed, the corresponding RPE per view using Eq. and the motion affected reconstruction using Eq.. Note that the RPE is only computed based on virtual marker positions which is possible because we know the system calibration ${\mathbf}{P}$ and motion trajectory ${\mathbf}{M}$ during training. The volume is reconstructed on nine slices distributed in triplets of axial ($\mathbb{R}^{216\times256}$), coronal ($\mathbb{R}^{70\times216}$) and sagittal ($\mathbb{R}^{70\times256}$) slices with an isotropic voxel size of $0.84$mm. The respective slices are distributed in a volume $(\mathbb{R}^{256\times216\times70})$. The field-of-view is selected such that no truncation artifacts in longitudinal direction are present if reconstructed from a typical clinical scan.
Motion Learning {#subsec:motionlearning}
---------------
\#1[\_[FDK]{.nodecor}\^[\#1]{}([E]{})]{}
The overall goal of the motion learning process is to train a network that is capable of approximating Eq. only based on a tomographic reconstruction. Therefore, nine slices of the tomographic reconstruction are used as input to the network. To keep the computational effort on a minimum and still capture all types of motion artifacts, the input of the network are triplets of three slices oriented in *axial* (ax), *coronal* (co) and *sagittal* (sa) direction. We denote the respective coordinates of the nine slices by ${\mathbf}{V} = ({\mathbf}{V}_\textnormal{ax},{\mathbf}{V}_\textnormal{co},{\mathbf}{V}_\textnormal{sa})$, where ${\mathbf}{V}_\diamond$ is a set of coordinates defined in $\mathbb{P}^3$. Thus, $f_\textnormal{FDK}({\mathbf}{V}_\textnormal{ax},{\mathbf}{E},\proj)$ will denote the reconstruction of three slices in axial direction with effective trajectory ${\mathbf}{E}$. Let us define a triplet of reconstructed slice images for a set of projections $\proj$ reconstructed with the effective trajectory ${\mathbf}{E}$ as $\reco{\diamond} = f_\textnormal{FDK}({\mathbf}{V}_\diamond,{\mathbf}{E},\proj)$ and further, let the tuple of all reconstructed slices be $\reco{} = (\reco{\textnormal{ax}}, \reco{\textnormal{co}}, \reco{\textnormal{sa}})$. Then, the input to the regression network is computed as $${\mathbf}{y}_\textnormal{lat}(\reco{}) = \cup_{\diamond \in {\{\textnormal{ax},\textnormal{co},\textnormal{sa}\}}} \featurenet(\reco{\diamond}) \enspace,$$ where $\cup$ denotes concatenation. Thus, each feed forward network processes three slices of the same anatomical orientation and the result is concatenated representing the latent space ${\mathbf}{y}_\textnormal{lat}(\reco{})$. The loss function $l$ is based on a multi-task loss $$l(\reco{}, {\mathbf}{E}) =
\sum_{t=1}^4
\Vert
\regressionnet_t({\mathbf}{y}_\textnormal{lat}(\reco{}) )
\\ - \mathcal{L}_t(\efftraj)
\Vert_2^2 \enspace,$$ with $$\mathcal{L}_t(\efftraj) =
\begin{cases}
\frac{1}{N} \sum_{i=1}^N d_\textnormal{RPE}(\efftraj_i) & \textnormal{if $t = 1$} \\
( d_\textnormal{RPE}(\efftraj_1),...,d_\textnormal{RPE}(\efftraj_N)) & \textnormal{if $t = 2$} \\
( d_\textnormal{RPE}(\efftraj^{\textnormal{ip}}_1),...,d_\textnormal{RPE}(\efftraj^{\textnormal{ip}}_N)) & \textnormal{if $t = 3$} \\
( d_\textnormal{RPE}(\efftraj^{\textnormal{op}}_1),...,d_\textnormal{RPE}(\efftraj^{\textnormal{op}}_N)) & \textnormal{if $t = 4$} \\
\end{cases} \enspace.$$ Here, we assume that $\efftraj$ is implemented such that it can be decomposed into ${\mathbf}{P}$ and ${\mathbf}{M}$ allowing to compute the RPE using Eq.. $\efftraj^{\textnormal{op}}$ and $\efftraj^{\textnormal{ip}}$ refer to in-plane an out-plane motion. Assuming the system is rotating around the ${\mathbf}{z}$ axis, in-plane motion is within the acquisition plane and represented by parameters ($r_z,t_x,t_y$) and out-plane motion is stepping out the acquisition plane and represented by parameters ($r_y,r_z,t_z$). We use this distinction, because in-plane motion is better detectable in axial slices, whereas out-plane motion is better detectable in coronal and sagittal slices.
For optimization we use the ADAM optimizer with a learning rate of $10^{-4}$ and a batch size of 32. To avoid over-fitting, we use the validation set for early stopping. The residual network $\mathcal{S}$ is initialized using pre-trained weights learned for the ImageNet classification task. The regression network $\mathcal{R}_k$ is randomly initialized.
Experiments and Results
=======================
In this section we evaluate the network performance w.r.t. three aspects: (1) the behavior of the network in its core task, i.e., the regression of the RPE, (2) the performance of the network in a motion compensation benchmark in comparison to state-of-the-art methods, and (3) the applicability of the proposed method to motion-affected clinical data.
Network Accuracy {#subsec:networkaccuracy}
----------------
Using the data generation proposed in Sec.\[subsec:datageneration\], we generate 9001 different motion affected reconstructions. The amplitude of the applied motion is in the range of $0$ $^\circ$\\mm to $15$ $^\circ$\\mm, i.e., *mean RPE*s (mRPE) are in a range of $0$mm to $0.74$mm. Using a split, this provides us with a total of 189021 samples for training, 36004 samples for validation and 18002 samples for testing. The number of spline nodes is set to $M=20$. Following the training described in Sec.\[subsec:motionlearning\], we achieve the optimal validation loss after $12\times10^3$ iterations (seeTab.\[tab:networkperformance\]).
### Ablation Study
\[tab:networkperformance\]
$\mathcal{L}_1$ $\mathcal{L}_2$ $\mathcal{L}_3$ $\mathcal{L}_4$
-------------------------- ----------------- ----------------- ----------------- -----------------
Proposed **0.0098** 0.2493 **0.1615** **0.1981**
Proposed with BN 0.0138 0.2835 0.2145 0.2412
Proposed no alignment 0.0436 0.5644 0.5815 0.8192
Proposed no pre-training 0.0171 0.4350 0.2941 0.3279
Proposed dual task 0.0146 **0.2481** x x
Proposed with DenseNet 0.0205 0.3331 0.1929 0.2818
: Best performing validation loss for different network settings.
To inspect the network performance as well as the influence of the pre-processing, Tab.\[tab:networkperformance\] displays the respective best validation loss values for alterations in the network structure or input data. The most important performance boost is obtained by the pre-processing step of aligning the respective reconstructions and slight improvements are obtained by removing the BN. Further, a pre-training of $\mathcal{S}$ on ImageNet increases the accuracy. Without the distinction of in-plane and out-plane motion (dual-task learning), the accuracy of $\mathcal{L}_2$ decreases slightly, however, the mRPE ($\mathcal{L}_1$) accuracy increases by $\approx50$ %. A replacement of the residual network architecture $\mathcal{S}$ with a pre-trained DenseNet [@huang2017densely] worsens the accuracy.
### Patient and Motion Variability {#subsubsec:paitentandmotionvariability}
![Boxplots showing the deviation of the mRPE ($\mathcal{L}_1$) from the ground truth for each patient. The boxplots are grouped to training (Train Pat.), validation (Val. Pat.) and test (Test Pat.) patients. All outliers are displayed as circles.[]{data-label="fig:boxplotmeanrpepatient"}](fig_3.pdf){width="\linewidth"}
An important aspect of motion appearance learning is the independence to the patient anatomy, similar motions applied to different patients should be predicted alike. Therefore, following the data generation presented in Sec.\[subsec:datageneration\] we generate 300 additional motion shapes per patient ranging from mRPEs of $0$mm to $0.7$mm. Note, that the simulated motion is random and therefore not part of the training set. Consequently, the applied motion was never seen by the network. The results depicted in Fig.\[fig:boxplotmeanrpepatient\] show the patient-wise accuracy in predicting the mRPE ($\mathcal{L}_1$). Most of the outliers are within a range of $0.2$mm, and no outlier is exceeding an error of $0.35$mm. While the accuracy is high with an mRPE of 0.013mm, there is a slight tendency of overestimating the mRPE. The inter-patient variability of the estimation is small with a standard deviation of 0.022mm. From the patients never seen during training, we can observe a good generalization of the learned features. The tendency to overestimate the mRPE is even slightly less observable.
![Boxplots visualizing the deviations to the ground truth for the in-plane ($\mathcal{L}_3$), out-plane ($\mathcal{L}_4$) and combined ($\mathcal{L}_2$) vRPE and the mRPE ($\mathcal{L}_1$). The evaluation is based on the four validation and two test patients.[]{data-label="fig:boxplotrpeplots"}](fig_4.pdf){width="\linewidth"}
Besides the mRPE the network further predicts three *view-wise RPE*s (vRPE) split to in-plane motion ($\mathcal{L}_2$), out-plane motion ($\mathcal{L}_3$) and both ($\mathcal{L}_4$). The accuracy for this task is depicted in Fig.\[fig:boxplotrpeplots\]. Comparing the accuracy of the vRPE estimations to the mRPE we observe higher deviations and a higher number of outliers in the vRPE estimations. The accuracy of the in-plane vRPE is higher than for the out-plane vRPE. In-plane motion is mostly distributed in axial slices, which can be reconstructed without significant cone-beam artifacts and are best suitable for motion prediction. The patient-wise deviations are more pronounced compared to the mRPE, however, still on a reasonable low level.
![Boxplots showing the mRPE deviations from the ground truth for motions clustered to the 6 motion directions. The evaluation is based on the four validation and two test patients. All outliers are displayed as circles.[]{data-label="fig:boxplotmeanrpemovement"}](fig_5.pdf){width="\linewidth"}
Figure \[fig:boxplotmeanrpemovement\] shows the mRPE clustered w.r.t. the motion directions for all patients. All motion directions can be predicted with similar accuracy, however, a slight tendency is observable that out-plane motion is predictable with less deviations.
In conclusion, the experiments have shown the patient independence of the proposed appearance learning approach, as well as the independence of the motion direction. This provides us a method that has no inherent limitations to certain motion patterns as apparent, e.g., in epipolar consistency [@Frysch2015], which is sensitive to out-plane motion, but barely applicable to in-plane motion.
### RPE Trajectory Prediction {#subsubsec:correlationstudy}
![Binary soft-classification results plotted over the respective views. FN relates to a motion-affected region that is classified by the network as motion-free. FP relates to a misclassification to the motion-affected class. TP and TN are correctly predicted views.[]{data-label="fig:precision"}](fig_6.pdf){width="\linewidth"}
Using the same data as in Sec.\[subsubsec:paitentandmotionvariability\], we investigate in this experiment the performance of the estimated vRPE for motion classification. The predicted vRPEs are interpreted as soft-classifiers, where we define a view to be in the motion-free class (negative) if the average predicted value for view $i$ satisfies $$\frac{1}{2} \mathcal{R}_2 + \frac{1}{4} \mathcal{R}_3 +\frac{1}{4} \mathcal{R}_4 \leq 0.1 \enspace.
\label{eq:softclassifier}$$ In Fig.\[fig:precision\] the accuracy is displayed encoded to *false negative* (FN), *false positive* (FP), and the combination of FP, *true positive* (TP) and *true negative* (TN). If the predicted value is used as an indicator function (seeSec.\[subsec:motionestimation\]) in a motion estimation scenario a low FN rate is important. Regions classified as motion free will receive little attention within the optimization. On the opposite, a FP classification does not worsen the result and is therefore non-critical. These properties are satisfied as observable from Fig.\[fig:precision\]. The FN rate is $\approx 0$% and the FP rate is $\approx 25$%. Note that the peaks in the FP curve are due to the spline nodes, where transitions from motion-affected to motion-free views arise with increased frequency.
Motion Estimation Benchmark {#subsec:motionestimation}
---------------------------
### Autofocus
The motion estimation benchmark is based on the four validation patients and two test patients. We apply a known motion trajectory ${\mathbf}{M}$ to the projection matrices ${\mathbf}{P}$ and evaluate the performance of six metrics (see\[subsubsec:imagequalitymetrics\]) to find the annihilating trajectory ${\mathbf}{C}$ (see\[subsec:rigidmotionmodel\]). We describe the trajectory as a function of six motion spline nodes ${\mathbf}{m} = ({\mathbf}{m}_{t_x},{\mathbf}{m}_{t_y},{\mathbf}{m}_{t_z},{\mathbf}{m}_{r_x},{\mathbf}{m}_{r_y},{\mathbf}{m}_{r_z})$. Each element of ${\mathbf}{m} \in \mathbb{R}^{6\times M}$ describes the respective spline node within the trajectory. Thus, ${\mathbf}{m}_{r_y,420}$ describes the rotation around the $y$-axis at acquisition view $420$. Then, the motion curve vector ${\mathbf}{t}({\mathbf}{m}) = ({\mathbf}{t}_x,{\mathbf}{t}_y,{\mathbf}{t}_z,{\mathbf}{r}_x,{\mathbf}{r}_y,{\mathbf}{r}_z)$ is computed by evaluating the spline for each acquired view. For example ${\mathbf}{t}_x = (\eta_{{\mathbf}{m}_{t_x}}(0),\eta_{{\mathbf}{m}_{t_x}}(1), ...,\eta_{{\mathbf}{m}_{t_x}}(N))^\top$, with $\eta_{{\mathbf}{m}_{t_x}}(i)$ denoting the spline evaluation at position $i$ based on the spline nodes ${\mathbf}{m}_{t_x}$ as proposed in [@akima1970new]. From the six motion curves described by ${\mathbf}{t}$ we can directly compute the annihilating trajectories denoted by ${\mathbf}{C}({\mathbf}{t}({\mathbf}{m}))$. Note, that the motion trajectory itself is generated in an equal way.
The components of ${\mathbf}{m}$ are found by optimizing the IQM $f_\textnormal{IQM}$ $$\hat{{\mathbf}{m}} = \operatorname*{argmin}_{\mathbf}{m} f_\textnormal{IQM}\left(f_\textnormal{FDK}({\mathbf}{V}, {\mathbf}{P}\circ {\mathbf}{C}({\mathbf}{m}), \proj) \right) \enspace.
\label{eq:optimization}$$
### Optimization {#sec:optimization}
Equation is optimized using the gradient free downhill simplex algorithm [@olsson1975nelder]. We optimize only one node at a time iterating over all nodes in sequential order. We use a coarse to fine strategy by defining 5 stages. In the first three stages we define a starting stepsize of 1 for the simplex and set the number of iterations to 2. This allows a rough estimate of the trajectory. In the last two stages, we set the number of iterations to 100 with initial stepsize of 0.5. The optimization is finished if either the maximum number of iterations is exceeded or the improvement in ${\mathbf}{m}$ is below 0.001 .
### Image Quality Metrics {#subsubsec:imagequalitymetrics}
We define three IQMs denoted by , and . and refer to the histogram entropy and TV of the slice images, respectively. We implement following the methodology of Herbst etal.[@herbst2019misalignment] and as proposed in Kingston etal.[@Kingston2011]. We selected these two metrics due to their popularity in the literature. was found to be superior for geometry alignment in a study by Wicklein etal.[@wicklein2012image]. is our proposed method. In addition we define Ent+, Tv+ and Cnn+, each denoting an initial optimization with either Ent, Tv or Cnn followed by a fine-tuning of the annihilating trajectory with an additional optimization stage using (for Tv+ and Cnn+) or Tv (for Ent+). For and the implementation of $f_\textnormal{IQM}$ is straightforward: the histogram entropy or total variation of the nine reconstructed slices are calculated. Following Wicklein etal.[@wicklein2012image] we use a bone window for the histogram calculation. Their studies showed that restricting the histogram calculation to values within a bone window improves the method’s performance, because only relevant image features are captured. The implementation of $f_\textnormal{IQM}$ for and uses an additional indicator function ${\mathbf}{1}_\mathbb{M}$, where $\mathbb{M}$ describes the set of views satisfying Eq.. Thus, the IQM for is defined by $$f_{\textnormal{IQM}} = \regressionnet_1({\mathbf}{y}_\textnormal{lat}({\mathbf}{V},{\mathbf}{P}\circ {\mathbf}{C},\proj) ) \enspace \textnormal{s.\,t.} \enspace {\mathbf}{1}_\mathbb{M} |{\mathbf}{t}({\mathbf}{m})| = {\mathbf}{0} \enspace,$$ with $|\cdot|$ denoting element wise absolute value. Note that ${\mathbf}{1}_\mathbb{M}$ is not updated during the iterations.
### Motion Scenarios
![The motion trajectories for both motion scenarios. The motion is applied respectively to the motion axis under investigation. The curves are generated using 20 spline nodes and 17 spline nodes for scenario A and B, respectively.[]{data-label="fig:motionshapes"}](fig_7.pdf){width="\linewidth"}
We design two motion scenarios (scenario A and scenario B) differing in their motion shapes and the number of spline nodes used for both, the motion trajectory and annihilating trajectory. Scenario A uses 20 spline nodes for the motion curves and the same number of spline nodes for the annihilating curves. Scenario B uses 17 spline nodes for the motion curves and 40 spline nodes for the annihilating curves. The two motion amplitudes are depicted in Fig.\[fig:motionshapes\]. For both scenarios the respective motion is applied to one of the 6 motion axes, respectively. In each scenario, we optimize only for the axis which is affected by the motion.
### Motion Estimation Results
\[tab:ssima\] \[tab:ssima2\]
{width="\linewidth"} \[fig:boxplot\_compensation\_a\]
\[tab:ssimb\] \[tab:ssimb2\]
{width="\linewidth"} \[fig:boxplot\_compensation\_b\]
To quantify the performance, we measure the mismatch of the respective motion curves and estimated annihilating curves. For a complete compensation, we need ${\mathbf}{M}\circ{\mathbf}{C} = {\mathbf}{1}$, which is the case if the motion curve and the annihilating curve add up to zero. The second metric measures the reconstruction quality by computing the SSIM of the respective motion-compensated reconstruction and the ground truth.
Figure \[fig:boxplot\_compensation\_a\] and Fig.\[fig:boxplot\_compensation\_b\] show the misalignment of the motion curve for motion scenario A and B, respectively, averaged over the 4 validation and 2 test patients for Ent, Tv, Cnn and Cnn+. Numeric results showing the misalignment for all six metrics are displayed in Tab.\[tab:ssima\] and Tab.\[tab:ssimb\] and numeric results showing the SSIM values for all six metrics are displayed in Tab.\[tab:ssima2\] and Tab.\[tab:ssimb2\], respectively. Selected reconstructions for both motion scenarios are presented in Fig.\[fig:recos\].
The proposed method performs well in both scenarios. In motion scenario A, the state-of-the-art methods perform similar to the network-based solution. In the majority of cases, the network-based results are superior. However, in almost 50% of the experiments, either or achieves the best results. A fine-tuning of the traditional metrics (Ent+, Tv+) barely improves the results and can lead to a degeneration of the performance. The margin by which the network-based method outperforms both state-of-the-art metrics is much higher than vice-versa. Figure \[fig:recos\] shows that the network-based approach is further capable of dealing with metal artifacts. In this case, the network without post-optimization using the entropy () achieves the best results. Note, that our training set also includes patients with metal artifacts.
For scenario B, our method is constantly outperforming the state-of-the-art metrics, both in terms of SSIM and measured misalignment. By an additional post-optimization with the entropy-based compensation, the best results are achieved with . As can be seen from Fig.\[fig:recos\], the network is capable of approximating the true motion curve but ignores small high-frequent motions. These motions are then eliminated by the entropy. However, deploying entropy alone produces mediocre results because the optimization gets stuck in local minima. Tv performs worst w.r.t. the misalignment of the annihilating and motion curves, however, the SSIM is comparable to the entropy-based procedure.
{width="\linewidth"}
Motion-Affected Clinical Data
-----------------------------
### Data and Preprocessing
To demonstrate the effectiveness of the proposed method in clinical practice, we apply it on a motion-affected clinical dataset. Similar to the acquired data used for the network training and evaluation, the patient was scanned with a clinical CBCT system (Artis Q, Siemens Healthcare GmbH, Forchheim). The projections were downsampled and aligned following the same procedure (i.e., step 1 of data generation) as presented in Sec.\[subsec:datageneration\].
### Motion Compensation Scheme
We model the annihilating trajectory with an Akima spline consisting of 20 spline nodes equally distributed over the trajectory. We adapt the optimization scheme from Sec.\[sec:optimization\]. We sequentially optimize for all six motion parameters in the following sequence ($t_z$,$t_x$,$t_y$,$r_x$,$r_y$,$r_z$). To optimize for motion we use and .
### Motion Compensation Results
Figure \[fig:clinicalresults\] displays reconstructed slice images from the motion-affected clinical dataset () as well as motion-compensated reconstructions (, ). As ground-truth reconstructions are not available, only a qualitative inspection is possible. In slice 1 we observe motion artifacts especially at the boarders of the temporal bones as well as near the nasal cavities and ethmoid bone. The anatomy contours can be well recovered using or . As can be observed from the difference images (d, d), streaks at the bone contours are eliminated. In slice 2 the motion artifacts are severe in the orbital bone structures. The streak artifacts are reduced in both motion-compensated reconstructions, restoring the homogeneous regions. A small residual motion is still observable with the motion-compensated reconstructions, however, the image quality could be improved substantially. From the difference images between the -based and -based compensated reconstructions (d) we see that the entropy-based fine-tuning barely affects the reconstruction quality.
{width="\linewidth"}
Discussion {#sec:discussionandconclusion}
==========
We propose an appearance learning approach that can be deployed for image-based motion compensation. For that purpose, we devise a framework that learns the mRPE as well as vRPEs from reconstructed slice images. Exact computation of the RPE allows for geometric calibration for high-quality CBCT [@strobel2003improving]. Hence, given a 100% network accuracy, minimizing the network-predicted mRPE would yield highly accurate motion parameters. The axis with lowest accuracy in predicting the mRPE is also the axis with lowest performance in the motion compensation benchmark.
We further show that we can learn general features applicable to all three types of translations and rotations. The learned features are even less dependent on the motion axis than traditional methods. For example, Tv shows superior results in compensating translation along the y-axis as observable from motion experiment A.
Autofocus methods are characterized by optimizing an IQM in the reconstruction domain. Inherently, information from which part of the trajectory a misalignment is expected can only be deduced from the gradients within the optimization. We aim to overcome this by learning an initial estimate about the distribution of the expected motion. The view-wise prediction can be used as a soft-classifier to steer the optimization. The FN range is close to $0$%, ensuring that the optimization cannot be worsen by applying the soft-classifier for optimization steering.
We use randomly generated motions that are limited in their frequency due to the deployed splines. However, the network is capable to generalize to unseen motion frequencies. In motion scenario B we compensate the motion with a spline controlled by 40 spline nodes, whereas in training, only motion patterns generated with 20 spline nodes were shown to the network. Hence, the network is capable to generalize to higher frequency motion patterns. However, we can observe, that very small but high-frequent motion artifacts are barely accounted for by the network (see Fig.\[fig:recos\], scenario B). These types of motion patterns where not part of the network training. Besides those small motion artifacts, the overall motion trajectory can be estimated well by the network. This property is synergistic with traditional IQMs. After fine-tuning with the entropy-based IQM the fine motion artifacts are eliminated.
Traditional IQMs measure the artifact strength by a limited set of image features. If the reconstruction is not corrupted by motion, the image shows homogeneous soft-tissue areas and clear bone-boundaries. This results in a low TV value and low histogram entropy. Motion corrupts the homogeneous regions and blur bone-edges increasing the histogram entropy and TV value. However, the image features recognized by TV and entropy are not directly linked to the patient motion strength and therefore are susceptible to local minima. Thus, both metrics are successful if the motion is small but fail if the motion is large. This is shown by the experiments, where both metrics, and , perform well in scenario A. For larger motions as apparent in motion scenario B, both metrics fail. In those scenarios, the learned metric outperforms the traditional methods in all six experiments.
Besides being only trained on synthetically generated motion, the network generalizes well to real clinical motion. We demonstrate this using a clinical motion-affected scan. Due to the rigid structure of the motion, a transformation of the object can be equivalently described by a static object and a transformation of the system geometry. This allows realistic generation of motion artifacts from artifact-free CBCT acquisitions.
Conclusion and Outlook
======================
Our proposed method can be used in a variational manner for image-based autofocus techniques. The result is always based on the acquired raw-data and ensures data integrity. This is a strong advantage to all other learning-based approaches found in our literature review. Current learning-based approaches perform an image-to-image translation, without any guarantee for the consistency with the acquired raw data. In contrast, using the proposed method the images are always reconstructed from the raw data minimizing the risk for generating clinical images leading to improper diagnosis.
The experiments show that motion artifacts can be learned by a neural network and that our learning-based approach can outperform state-of-the-art IQMs in a motion estimation benchmark. We devised the approach based on the FDK algorithm and artificial motion. Using a motion-affected clinical dataset, we further demonstrate that the method translates to real clinical motion. The FDK is suitable for autofocus approaches [@sisniega2017motion; @Kingston2011] due to its computational efficiency. A possible extension, however, would be a reconstruction algorithm, capable of reconstructing arbitrary trajectories [@Defrise1994]. The FDK assumes two fundamental properties: (1) homogeneous object in the direction perpendicular to the acquisition plane and (2) equally sampled trajectories along an arc. If any of those assumptions are not met, the reconstruction reveals cone-beam artifacts or intensity inhomogeneities. Therefore, it can only compute approximate solutions for motion compensation.
Although our experiments are tailored for head CBCT, the concept is neither limited to rigid head motion nor to transmission imaging. By replacing the filtered back-projection with the inverse model for MRI — e.g., non-uniform Fourier transform [@fessler2007nufft] — the approach can be directly trained for propeller trajectories in MRI. By additionally replacing the RPE-based regression metric with an appropriate metric (e.g., energy of a spline deformation field), also Cartesian sampled MRI can be tackled. Similar strategies are thinkable for PET.\
\
**Disclaimer:** The concepts and information presented in this article are based on research and are not commercially available.
[^1]: Copyright (c) 2019 IEEE. Personal use of this material is permitted. However, permission to use this material for any other purposes must be obtained from the IEEE by sending a request to [email protected]. A. Preuhs, P. Roser, E. Hoppe, Y. Huang and A. Maier are with the Pattern Recognition Lab, Friedrich-Alexander-Universität Erlangen-Nürnberg, 91058 Erlangen, Germany e-mail: [email protected].
[^2]: M. Psychogios is with the Neuroradiology Department, Universitätsspital Basel, 4031 Basel, Switzerland.
[^3]: M. Manhart and M. Kowarschik are with Siemens Healthcare GmbH, 91301 Forchheim, Germany.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Revised Hipparcos parallaxes for classical Cepheids are analysed together with 10 HST-based parallaxes (Benedict et al.). In a reddening-free $V,I$ relation we find that the coefficient of $\log P$ is the same within the uncertainties in our Galaxy as in the LMC, contrary to some previous suggestions. Cepheids in the inner region of NGC4258 with near solar metallicities (Macri et al.) confirm this result. We obtain a zero-point for the reddening-free relation and apply it to the Cepheids in galaxies used by Sandage et al. to calibrate the absolute magnitudes of SNIa and to derive the Hubble constant. We revise their result for $H_{0}$ from 62 to $ \rm 70 \pm 5\, km\,s^{-1}\,Mpc^{-1}$. The Freedman et al. 2001 value is revised from 72 to $\rm 76 \pm 8\, km\,s^{-1}\,Mpc^{-1}$. These results are insensitive to Cepheid metallicity corrections. The Cepheids in the inner region of NGC4258 yield a modulus of $29.22 \pm 0.03$ (int.) compared with an maser-based modulus of $29.29\pm 0.15$. Distance moduli for the LMC, uncorrected for any metallicity effects, are: $18.52 \pm 0.03$ from a reddening-free relation in $V,I$; $18.47 \pm 0.03$ from a period-luminosity relation at $K$; $18.45 \pm 0.04$ from a period-luminosity-colour relation in $J,K$. Adopting a metallicity correction in $V,I$ from Macri et al. leads to a true LMC modulus of $18.39 \pm 0.05$.'
author:
- |
Floor van Leeuwen$^{1}$, Michael W. Feast$^{2}$, Patricia A. Whitelock$^{2,3,4}$, Clifton D. Laney$^{3}$\
$^{1}$ Institute of Astronomy, Madingley Rd, Cambridge, England\
$^{2}$ Astronomy Dept., University of Cape Town, Rondebosch, 7701, South Africa\
$^{3}$ South African Astronomical Observatory, P.O. Box 9, Observatory, 7935, South Africa\
$^{4}$ National Astrophysics and Space Science Programme, Department of Mathematics and Applied Mathematics, University of Cape\
Town, Rondebosch, 7701, South Africa\
title: Cepheid Parallaxes and the Hubble Constant
---
astrometry, Cepheids, distance scale, cosmological parameters, Magellanic Clouds, supernovae:general
Introduction
============
The success of the Hipparcos astrometric satellite in obtaining a large number of absolute stellar parallaxes with much greater accuracy than had previously been possible (ESA 1997), allowed a first estimate to be made of the zero-point of a Cepheid Period-Luminosity (PL) relation directly from Cepheid parallaxes (Feast & Catchpole 1997). Investigations of the Hipparcos data continued after the publication of the catalogue, and led to the identification of a number of repairable problems associated with the reconstruction of the satellite’s attitude (van Leeuwen 2005). This ultimately led to a completely new reduction of the astrometric data (van Leeuwen & Fantino 2005), the results of which are soon to be published (van Leeuwen 2007) and have been incorporated in the current study. The main impact of the reductions is for the brightest stars, where improvements of up to a factor of four in parallax accuracy can be reached. For many of the Cepheids improvements by up to a factor of two have been achieved.
In the present paper we discuss and analyse the revised Hipparcos parallaxes of (classical) Cepheids. Recently, the parallaxes of 10 Cepheids, measured with the Hubble Space Telescope (HST) and corrected to an absolute reference frame using ground-based observations, have been published and discussed (Benedict et al. 2002, Benedict et al. 2007 henceforth FB2007) and we have incorporated these data in our analysis.
Since the discussion of the original Hipparcos data there has been much work on Cepheid photometry, period-luminosity (PL) relations, metallicity effects etc. and we have been able to take advantage of this in the present work. There has also been a great deal of theoretical work on Cepheids. We do not discuss this since our aim has been to obtain empirically-based conclusions. We establish a reddening-free PL relation in $V$ and $I$ (PL($W$)) as well as PL and Period-Luminosity-Colour (PLC) relations in $J$ and $K$. These relations should be of use in a variety of contexts both Galactic and extragalactic and for testing theoretical models. In the present paper we confine ourselves to the implication of our results for Cepheid-based distances of galaxies including the LMC and the estimation of the Hubble constant.
Data
====
The basic data used here are tabulated in the appendix together with notes on individual stars. The table contains only those stars which had all the data required for the analyses and which are considered to be classical Cepheids.
Table A1 contains:\
[**1. Hipparcos number (no).**]{}\
[**2. Hipparcos parallax ($\pi$).**]{} Here and throughout in milliarcsec (mas)\
[**3. Hipparcos parallax standard error ($\sigma_\pi$).**]{} In mas.\
[**4. Star name.**]{}\
[**5. Logarithm of fundamental period.**]{} Where the star is an overtone pulsator the fundamental period, $P_{0}$, was derived from the observed period, $P_{1}$, using the relation (Alcock et al. 1995, Feast & Catchpole 1997): $$P_{1}/P_{0} = 0.716 -0.027\log P_{1}.$$ Such stars are denoted by “O" in the notes. Here and throughout the periods are in days.\
[**6,7,8. Intensity mean magnitudes $<B>,<V>,<I>$.**]{} The $I$ magnitudes are on the Cousins system.\
The $BVI$ photometry used was from Berdnikov (private communication 2005) and is an update of Berdnikov et al. (2000). Where values of $I$ were not available in this source they were taken from Groenewegen (1999) who transformed other workers’ data to the Cousins system. These stars are marked “G" in the notes.\
[**9,10,11. Intensity mean magnitudes $<J>,<H>,<K>$ on the SAAO system (Carter 1990).**]{} Where possible these are from multiple observations. They are mainly from the papers by Laney & Stobie (1992, 1993, 1994) and previously unpublished SAAO data. The individual observations on which these latter intensity means depend will be published separately. In a limited number of cases intensity means are from Groenewegen’s (1999) tabulation. All these stars are indicated by “L" in the notes. Where no source is specified the mean magnitude has been estimated from the single 2MASS measure transformed to the SAAO system (Carpenter 2001) and corrected for phase by the procedure outlined by Soszyński et al. (2005). Since in many cases the phases used are old (GCVS) the accuracy of these corrections can be poor and this is taken into account in the later work.\
[**12. $ET$ is the value of $E(B-V)$ given by Tammann et al. (2003).**]{} If this is not available then this column contains $EF \times 0.951$ (see Tammann et al. 2003) and is indicated by “F" in the notes.\
[**13. $\Delta$ is the full $V$ amplitude of the star**]{}.\
[**14. $EF$ is the value of $E(B-V)$,**]{} as estimated by Fernie at al. (1995) and given in the DDO data base as $FE1$. If this is not available $FE2$ is given.\
[**15. Notes using the following symbols:**]{}\
$O?$, possible overtone, but considered a fundamental pulsator.\
$B, B?$, binary or possible binary.\
$VB$, visual binary.\
$SB$, spectroscopic binary.\
$SB2$, spectroscopic binary with both spectra observed.\
The data on the binaries are mainly from the Szabados data base (see Szabados 2003). In cases of specific stars, additional references are in the notes. Many of the binaries listed by Szabados are, or possibly are, single line spectroscopic binaries. In the case of the original Hipparcos data the effects of binaries on the zero-point was discussed in Feast (1998) both as regards the photometry and the astrometry. Where the photometry may have been affected by a companion this is mentioned in the notes and if thought appreciable the star was omitted from the analysis.\
$DM$, double mode pulsator; the fundamental period is listed.\
[**16. The type of astrometric solution using the following codes:**]{}\
5: standard 5-parameter solution for single stars.\
25: standard 5-parameter solution for a variable double star.\
65: standard 5-parameter solution for the photo-centre of a variable double star.\
105: standard 5-parameter solution for the secondary in a resolved binary with a variable component (a difficult solution).\
7: single star with time-dependent proper motion (accelerated solution, which may indicate long-term orbital motion).\
3: variability induced mover, a probably spurious indication of duplicity depending on the phase of the lightcurve.\
1: stochastic solution, too much unexplained noise left in the data, generally unreliable, and an indication of short-term orbital motion.\
Initial tests showed that with the Hipparcos parallaxes and listed photometry, the following stars lay 5$\sigma$ or more from any reasonable PL relation: Y Sgr, V350 Sgr, GQ Ori, SY Nor, HL Pup. The Hipparcos data for all of these stars were omitted from the analysis, although only Y Sgr would have sufficient weight to make a significant contribution to the solutions discussed. In some cases erroneous classification may be the cause of the discrepancy. The reason for the significant discrepancy in the case of Y Sgr is not fully understood since the star has an HST parallax (FB2007) in good accord with other Cepheid data. In the combined solutions discussed below we use only the HST parallax for Y Sgr.
Leaving aside Y Sgr there are nine Cepheids in common with the HST parallax work. These parallaxes are listed with the Hipparcos results in Table 1 and the weighted means are also shown. In the solutions below we use these weighted means unless otherwise indicated.
$\pi_{HST}$ $\pi_{mean}$
-------- -------------- ------------------ ----------------- -----------------
47854 $l$ Car $2.06 \pm 0.27$ $2.01 \pm 0.20$ $2.03 \pm 0.16$
34088 $\zeta$ Gem $2.71 \pm 0.17$ $2.78 \pm 0.18$ $2.74 \pm 0.12$
26069 $\beta$ Dor $3.64 \pm 0.28$ $3.14 \pm 0.16$ $3.26 \pm 0.14$
88567 W Sgr $2.59 \pm 0.75$ $2.28 \pm 0.20$ $2.30 \pm 0.19$
87072 X Sgr $3.39 \pm 0.21$ $3.00 \pm 0.18$ $3.17 \pm 0.14$
110991 $\delta$ Cep $3.81 \pm 0.20$ $3.66 \pm 0.15$ $3.71 \pm 0.12$
93124 FF Aql $2.05 \pm 0.34$ $2.81 \pm 0.18$ $2.64 \pm 0.16$
102949 T Vul $2.31 \pm 0.29$ $1.90 \pm 0.23$ $2.06 \pm 0.22$
30827 RT Aur $-0.23 \pm 1.01$ $2.40 \pm 0.19$ $2.31 \pm 0.19$
89968 Y Sgr $3.73 \pm 0.32$ $2.13 \pm 0.29$
: Hipparcos and HST Cepheids in common.
Methods using $VI$ photometry
=============================
The existence of a Cepheid PL relation goes back of course to Leavitt (1908, 1912). It is known that the PL($V$) relation in the LMC has significant width and that this is greatly reduced, probably to within the observational errors, by the use of a period-luminosity-colour (PLC) relation (e.g. Martin et al. 1979). However, the value of the coefficient of the colour term and whether it varies with period has been a matter of uncertainty and debate.
In recent times there has been considerable discussion on the possibility that some PL relations might be non-linear. Thus, it has been suggested that the PL($V$) relation in the LMC is non-linear (Sandage et al. 2004, Ngeow et al. 2005). However, it is not entirely clear whether this effect is real or, for instance, due to systematic errors in the adopted reddenings varying with period. In any case, there seems to be good evidence (Ngeow & Kanbur 2005) that in the LMC, a “reddening-free" relation (Madore 1976) such as, $$M_{W} = \alpha \log P + \beta (V - I) + \gamma,$$ is linear. Here the coefficient, $\beta$, is an adopted ratio of total to selective extinction ($A_{V}/(A_{V} - A_{I})$). We have restricted our discussion at optical wavelengths to a relation of this form because of this and because of its importance in extragalactic work.
In a very detailed analysis of the data available to them, Sandage et al. (2004) have suggested that the slopes of PL relations in the optical, and hence the $\log P$ coefficient in equation 2, differ between the LMC and our Galaxy, presumably due to metallicity effects. In the case of the Galaxy the slopes were derived by combining Cepheid distances derived from Baade-Wesselink (pulsation parallax) type analyses with those from Cepheids in clusters with distances from main-sequence fitting. Their results have remained controversial for the following reasons. Gieren et al. (2005) showed that the Baade-Wesselink type distances of LMC Cepheids were a function of period if a conventional $``p"$ factor was used to convert observed radial velocities to pulsational velocities of the stars. Changing $``p"$ to remove this effect brings the LMC and Galaxy PL slopes into agreement within the uncertainties. As regards the ‘cluster’ distances, these partly depend, especially at longer periods, not on clusters but on stellar associations. The distances of these associations remain somewhat uncertain and their inclusion can affect any derived PL slope (as can be inferred from the early work of Feast & Walker 1987, their table 3). In view of these various uncertainties, one of our aims has been to derive the PL($W$) slope in our Galaxy.
Two PL($W$) relations are of particular importance for extragalactic applications.\
(1) The relation adopted by Freedman et al. (2001) in their HST key project on the Cepheid calibration of the Hubble Constant ($H_{0}$), $$M_{W} = -3.255 \log P + 2.45 (V - I) - 2.724 .$$ The $\log P$ coefficient was derived from OGLE LMC data (Udalski et al. 1999) and the colour coefficient from their adopted reddening law. The equation as given is applicable to local Galactic Cepheids. The zero-point is derived from their adopted LMC modulus and metallicity correction.\
(2) The relations for Galactic Cepheids used by Sandage et al. (2006 henceforth S2006)[^1] in their work on a Cepheid-based $H_{0}$ are equivalent to: $$M_{W} = -3.746 \log P + 2.563(V - I) - 2.213 .$$ The basis on which this equation was derived was discussed above.\
A third reddening-free relation is of relevance. This was derived (FB2007) from a “cleaned" selection of 581 LMC OGLE Cepheids and adopting the Freedman et al. reddening law $$\begin{aligned}
M_{W} = -3.29 (\pm 0.01)\log P + 2.45 (V - I) - \nonumber\\
(Mod -15.94 (\pm 0.01)) ,\end{aligned}$$ where $Mod$ is the distance modulus of the LMC and no metallicity correction has been applied.
After making corrections for reddening, Udalski et al. (1999) derived a true PLC relation for LMC Cepheids in the OGLE data base, which can be written as, $$\begin{aligned}
V_{0} = -3.25(\pm 0.02) \log P + 2.41 (\pm 0.03) (V-I) + \nonumber\\
15.88(\pm 0.02)).\end{aligned}$$ Comparison of equations 5 and 6 shows that the coefficients of the slopes and colour terms are nearly the same in the two equations. Thus, we are justified (at least in the LMC) in considering a reddening-free relation as also very close to a true PLC relation. This is important for two reasons. First, as discussed above, we expect a PLC relation to be very narrow. Secondly, because Cepheid overtone pulsators are in general bluer than those of the same fundamental period, they lie systematically above fundamental pulsators in a PL($V$) plot. However, they lie together with the fundamental pulsators in a PLC plot (see, e.g. Beaulieu et al. 1995). This is important if we wish to include overtone pulsators in a Cepheid calibration. We test this result in the case of Galactic Cepheids below.
In analysing the data we have proceeded in two ways:\
(1) For a limited number of Cepheids the percentage errors in the parallaxes are sufficiently small that individual values of $M_{W}$ can be directly combined to derive PL relations; Cepheids selected in this way require a Lutz-Kelker type bias correction (Lutz & Kelker 1973). We have scaled these corrections to be on the same system as that adopted by FB2007. We are primarily interested in using this subset of stars to obtain an estimate of $\alpha$ in equation 2.\
(2) We combine the main body of data using the method of reduced parallaxes (e.g. Feast 2002) and fixed values of $\alpha$ to obtain the zero-point, $\gamma$, in equation 2.
The slope $\alpha$ of the Galactic PL($W$) relation
===================================================
Table 2 lists the subset of Cepheids used in this section. The table gives, Hipparcos number, name, the parallax and its standard error from the combined Hipparcos and HST data, the absolute magnitude, $M_{W}$, (without Lutz-Kelker correction) and its standard error (= $2.17\sigma_{\pi}/\pi$), from this parallax and the adopted photometry together with $\beta = 2.45$. Also given are the Lutz-Kelker corrections applicable in this case and the $\log$ of the fundamental period. Fig. 1 shows $M_{W}$ with Lutz-Kelker corrections applied plotted against $\log P$ and with our finally adopted relation ($\alpha =-3.29$, $\beta =2.45$ and $\gamma = -2.58$) shown. The residuals in this case are also listed in Table 2.
$\log P$ Res.
-------- -------------- ----------------- ---------- ------------------- ------- -------
11767 $\alpha$ UMi $7.72 \pm 0.12$ 0.754 $ -5.08 \pm 0.03$ 0.00 –0.02
13367 SU Cas $2.57 \pm 0.33$ 0.440 $ -4.18 \pm 0.28$ –0.13 –0.15
26069 $\beta$ Dor $3.26 \pm 0.14$ 0.993 $ -5.72 \pm 0.09$ –0.02 +0.13
30827 RT Aur $2.31 \pm 0.19$ 0.572 $ -4.35 \pm 0.18$ –0.06 +0.11
34088 $\zeta$ Gem $2.74 \pm 0.12$ 1.006 $ -5.98 \pm 0.10$ –0.02 –0.09
47854 $l$ Car $2.03 \pm 0.16$ 1.551 $ -7.70 \pm 0.17$ –0.05 –0.01
61136 BG Cru $2.23 \pm 0.30$ 0.678 $ -4.63 \pm 0.29$ –0.15 +0.18
87072 X Sgr $3.17 \pm 0.14$ 0.846 $ -5.22 \pm 0.10$ –0.02 +0.14
88567 W Sgr $2.30 \pm 0.19$ 0.880 $ -5.62 \pm 0.18$ –0.06 –0.15
89968 Y Sgr $2.13 \pm 0.29$ 0.761 $ -5.13 \pm 0.30$ –0.15 –0.05
93124 FF Aql $2.64 \pm 0.16$ 0.650 $ -4.66 \pm 0.13$ –0.03 +0.06
102949 T Vul $2.06 \pm 0.22$ 0.647 $ -4.43 \pm 0.23$ –0.09 +0.28
104185 DT Cyg $2.19 \pm 0.33$ 0.550 $ -4.15 \pm 0.33$ –0.18 +0.24
110991 $\delta$ Cep $3.71 \pm 0.12$ 0.730 $ -5.05 \pm 0.07$ –0.01 –0.07
Weighted least square fits to equation 2 were made to various subsets of the data and the slopes ($\alpha$) derived are listed in Table 3. This is in two parts: Table 3(a) adopts $\beta$ = 2.45 (as in Freedman et al. 2001 ) and Table 3(b) adopts $\beta$ = 2.523 (as in S2006) (see section 3).
No Sample $\alpha$ $\beta$
---- ---------------------------------- --------------------- ---------
(a)
1 10 HST stars + HST phot. $-3.335 \pm 0.172$ 2.45
2 10 HST stars + New phot. $-3.473 \pm 0.183$ 2.45
3 10 stars HST + Hipp $-3.328 \pm 0.188$ 2.45
4 \(3) + Polaris (11 stars) $-3.285 \pm 0.169$ 2.45
5 weight $>10$ (13 stars) $-3.273 \pm 0.155$ 2.45
6 weight $>8$ (14 stars) $-3.288 \pm 0.151$ 2.45
7 \(6) omitting $l$ Car (13 stars) $-3.265 \pm 0.230 $ 2.45
8 LMC (OGLE) $-3.29 \pm 0.01 $ 2.45
9 Freedman (adopted) $-3.26 $ 2.45
(b)
10 10 HST stars + New phot. $-3.502 \pm 0.182$ 2.523
11 10 stars HST + Hipp $-3.357 \pm 0.186$ 2.523
12 \(11) + Polaris (11 stars) $-3.330 \pm 0.165$ 2.523
13 weight $>$ 10 (13 stars) $-3.315 \pm 0.152$ 2.523
14 weight $>$ 8 (14 stars $-3.330 \pm 0.149$ 2.523
15 Sandage (adopted) $-3.75 $ 2.523
: Determinations of the slope, $\alpha$, in equation 2.
From Table 3(a) we draw the following conclusions:\
1. Slight changes in the adopted photometry affect the value of the slope. However, this is not the main source of uncertainty.\
2. Adding the overtone Polaris at its fundamental period does not change the slope appreciably (see also the next section).\
3. Leaving out $l$ Car does not affect the slope appreciably. This is important since $l$ Car is by far the longest period star in this sample and therefore, when included, has a major effect on the slope determined.\
4. Our best determinations (solutions 6 and 7 of Table 3) give values of $\alpha$ close to that determined for LMC stars from the OGLE data (as shown in the table). There is still significant uncertainty in our value of the Galactic slope. However, within the uncertainties it agrees with that determined for the LMC.\
From Table 3(b) we draw the following conclusions:\
1. Using the “Sandage" colour coefficient ($\beta$) we get slopes which are not significantly different from those in Table 3(a).\
2. Our slopes, especially the higher weight ones, are distinctly different from that adopted by S2006, and in view of the uncertainties surrounding this latter result (see section 3) we suggest it should be replaced by our best value. Nevertheless, in the next section we give zero-points derived using a value of $\alpha = -3.75$.\
The main conclusion of this section is that within current uncertainties the value of $\alpha$ is the same in the Galaxy as in the LMC and our final results will be based on this assumption.
The zero-point $\gamma$ of the PL($W$) relation
===============================================
No. Stars $\alpha$ $\beta$ $\gamma$ notes
----- ------- ---------- --------- -------------------- -------------------
(a)
1 14 –3.255 2.45 $-2.606 \pm 0.022$
2 10 –3.255 2.45 $-2.568 \pm 0.036$ HST result
3 14 –3.29 2.24 $-2.579 \pm 0.022$
4. 14 –2.75 2.523 $-2.264 \pm 0.028$
(b)
5 240 –3.255 2.45 $-2.604 \pm 0.030$
6 240 –3.29 2.45 $-2.576 \pm 0.030$
7 239 –3.29 2.45 $-2.558 \pm 0.044$ \(6) no Polaris
8 213 –3.29 2.45 $-2.563 \pm 0.046$ \(6) no overtones
9 240 –3.75 2.523 $-2.263 \pm 0.030$
: The zero-point, $\gamma$, of equation 2 with fixed $\alpha$ and $\beta$.
Table 4 gives results of the determinations of the zero-point, $\gamma$ in equation 2 using fixed values of $\alpha$ and $\beta$. In Table 4(a) are the values of $\gamma$ obtained directly from the 14 stars in Table 2 with Lutz-Kelker corrections applied. In Table 4(b) we give the values of $\gamma$ derived by the method of reduced parallaxes (e.g. Feast 2002) to our bulk sample. Points to note are:\
1. Values of $\gamma$ in Table 4(a) agree closely with the corresponding values in Table 4(b).\
2. In Table 4(b), leaving out the high weight overtone pulsator, Polaris, makes no significant difference. Nor does leaving out all the known overtone pulsators. This is consistent with the discussion in section 3 which noted that the reddening-free relation in the LMC was very closely the same as a true PLC relation.\
3. In carrying out the reduced parallax solutions we have assumed that the uncertainties are dominated by the errors in the parallaxes. That is, we have neglected the second term in equation 3 of Feast (2002). However, if we supposed that $\sigma_{m_{0}}$ = $\sigma_{M_{0}}$ = 0.14, which we believe would be a gross overestimate, then the value of $\gamma$ in Table 4(b) (solution 6) would only change from –2.576 to –2.554.
We adopt –2.58 (solution 6, Table 4) as the value for $\gamma$ to use with $\alpha = -3.29$ and $\beta = 2.45$ in equation 2.
The Hubble Constant
===================
An important use of Cepheids is as a basis for the determination of the distances of galaxies and from that the estimation of a value of $H_{0}$. This can then be compared with the value derived in other ways (e.g. from the microwave background) which require the adoption of a general cosmological model; thus providing a test of the model. Sandage and his co-workers have recently completed a major programme of reanalysing HST data of Cepheids in galaxies in which supernovae have been observed (see S2006, Saha et al. 2006 and earlier papers in the series). In their summary paper they use the Cepheid data to derive distance moduli to 10 normal SNIa. From these they derive the maximum SNIa brightness (as defined by them). They then use this as a zero-point for a determination of $H_{0}$. We have redetermined the distance moduli of these galaxies using equation 2 with our adopted coefficients from section 5 viz. $\alpha = -3.29$, $\beta = +2.45$, and $\gamma = -2.58$. We use the same selection of Cepheids as used by S2006 and adopt the corrected apparent SN magnitudes in Table 2 of that paper. We derive SNIa absolute magnitudes equivalent to those in Table 3 of S2006 and like them derive weighted means. Table 5 gives the weighted mean absolute magnitudes of S2006 and the corresponding values of $H_{0}$ which this implies in their work.
--------------------------- ------------------- ------------------- ------------------- -----------------------
$M_{B}$ $M_{V}$ $M_{I}$
$H_{0}$ (B) $H_{0}$ (V) $H_{0}$ (I) Adopted $H_{0}$
S2006 $-19.49 \pm 0.04$ $-19.46 \pm 0.04$ $-19.22 \pm 0.05$
$62.4 \pm 1.2$ $62.5 \pm 1.2$ $62.1 \pm 1.4$ $62.3 \pm 1.3$ (int.)
Revised (no metal cor.) $-19.26 \pm 0.05$ $-19.22 \pm 0.05$ $-18.98 \pm 0.07$
$69.4 \pm 1.6$ $69.8 \pm 1.6$ $69.4 \pm 2.3$ $69.5$
Revised (with metal cor.) $-19.26$ $-19.22$ $-18.97$
$69.4$ $69.8$ $69.7$ $69.6$
--------------------------- ------------------- ------------------- ------------------- -----------------------
Table 5 also contains the equivalent weighted mean absolute magnitudes derived from our estimates of the moduli. Evidently the difference between our absolute magnitude and theirs implies a change in $H_{0}$ and this is given in the table (in units of $\rm km\,s^{-1} \,Mpc^{-1}$). The table contains the results of two estimates we have made for the moduli. In one case we have applied no metallicity correction. In the other we have applied a metallicity correction based on the “Sakai" values of \[O/H\] in table 1 of S2006. These abundances are in the $T_{e}$ system and Macri et al. (2006) find from their work on NGC4258 that in this system a PL($W$) relation such as we have used requires a correction of –0.49 mag/dex. There is considerable uncertainties in the size of the required metallicity correction. Fortunately, as Table 5 shows, the result we obtain is quite insensitive to the correction used. This is due to the mean metallicity of the S2006 galaxies being close to that of the local Cepheids. Only a large, non-linear metallicity correction would change this conclusion.
Thus our best value of $H_{0}$ based on our PL($W$) relation but with all other data and assumptions as in S2006 is 69.6. S2006 obtained $62.3 \pm 5$. Our improved Cepheid results would in principle reduce the uncertainty to $\sim \pm 2$, but to be conservative we keep it the same. Our revised value ($70\pm 5$) is clearly compatible with the value of $73 \pm 3$ found from the WMAP data and a $\Lambda$CDM model (Spergel et al. 2006).
There are a large number of other determinations of $H_{0}$, some of them depending on a Cepheid scale. The most widely quoted is that of Freedman et al. (2001) who obtained $H_{0} = 72 \pm 8$. Since that paper was published there has been much work on the refinement of the basic HST data, on galaxy metallicities and on the various large scale distance indicators used by these workers. However, the fact that the values of $\alpha$ and $\beta$ in the PL($W$) relation they use are close to ours, and that the mean metallicity of their sample, weighted according to the contribution of an indicator to $H_{0}$ is close to solar, means that a satisfactory estimate of the effect of our work on theirs can be made by comparing PL($W$) zero-points. Table 4 solution 5 shows that with their $\alpha$ and $\beta$ we find $\gamma = -2.604$ whereas they used $-2.724$ at the metallicity of Galactic Cepheids (see equation 3 above). The difference (0.12 mag) implies an increase of their $H_{0}$ to $76 \pm
8$, where to be conservative we keep the error the same, though the discussion of Freedman et al. together with our new results would in principle reduce this to $\sim \pm 6$. Again this revised $H_{0}$ is quite compatible with the WMAP result.
The distance modulus of the LMC using $VI$ photometry
=====================================================
Combining the LMC PL($W$) relation (equation 5) with our derived value of $\gamma$ ($-2.58$) gives directly the true modulus of the LMC uncorrected for metallicity effects. We thus find a modulus of $18.52\pm
0.03$. Adopting the results of Andrievsky et al. (2002) and Sakai et al. (2004), as discussed by S2006 the LMC Cepheids are metal deficient by $\Delta
[O/H] = 0.26$ on the “$T_{e}"$ abundance scale. As already noted Macri et al. (2006) found a Cepheid metallicity effect, applicable to our PL($W$) results, of $-0.49 (\pm 0.15)$ mag/dex. Applying this leads to a metallicity corrected LMC modulus of $18.39 \pm 0.05$.
The main uncertainty in this result is due to the uncertainty in the metallicity correction to our PL($W$) relation (note that corrections at other wavelengths would not necessarily be the same). It has even been recently suggested that the effect may be negligible (Rizzi et al. 2007). This would be somewhat remarkable since it has been long known (Gascoigne & Kron 1965 and much further work) that the intrinsic colours of LMC Cepheids differed from those of Galactic Cepheids of the same period. This was shown by Laney & Stobie (1986) to be due to a combination of changes in atmospheric blanketing together with a real shift of the instability strip in temperature. Fortunately the metallicity correction problem is not important for the work on $H_{0}$ discussed in the previous section. The use of the LMC modulus, however determined, will remain an uncertain basis for an extragalactic scale based on Cepheids pending further work on the metallicity effect.
The Cepheid-based distance of the Maser-host galaxy NGC4258
===========================================================
NGC4258 is of special interest because a distance has been determined (Herrnstein et al. 1999) based on the motions of $H_{2}O$ masers around the central black hole. Macri et al. (2006) have obtained HST photometry of Cepheids in this galaxy. Here we concentrate on Cepheids in their inner field which has a metallicity close to solar and therefore is immune to the problem of metallicity corrections when combined with our Galactic calibration. Some discussion of the Cepheid-based distance of this galaxy was made in connection with the HST Cepheid parallax work (FB2007).
We first consider the coefficient $\alpha$ of equation 2. Taking the 69 Cepheids which pass the selection criteria of Macri et al.[^2] We find the following:\
1. For our adopted value of $\beta = 2.45$ we find $\alpha = -3.18 \pm 0.13$.\
2. For the value of $\beta = 2.523$ favoured by S2006 we find $\alpha = -3.19 \pm 0.13$.\
These values are less than $1 \sigma$ from our adopted slope of –3.29, and $4.3 \sigma$ different from that favoured by S2006 for Galactic Cepheids (–3.75). We take this as further evidence that the slope of the PL($W$) relation in the LMC applies also to Cepheids of approximately Galactic composition. Macri et al. reach a similar conclusion by a different method.
Adopting $\alpha = -3.29$, $\beta = 2.45$ and $\gamma = 2.58$, from section 5, for equation 2 we find from the data of Macri et al. (2006) for their inner region, a distance modulus of $29.22 \pm 0.03$ for NGC4258. The standard error takes into account the internal scatter in the NGC4258 Cepheid data and the uncertainty in the adopted $\gamma$. The currently available maser-based distance modulus is $29.29 \pm 0.15$ (Herrnstein et al. 1999) and is thus in good agreement with the parallax-based Cepheid modulus. The referee has suggested that the Cepheid modulus may be slightly underestimated due to the possible effects of blending on the NGC4258 Cepheids and an even closer agreement with the maser distance might be obtained if this could be taken into account. Improvements in the maser-based modulus are expected (Macri et al. 2006, Argon et al. 2007) and these should allow a more stringent comparison with the Cepheids.
Methods using $JK$ photometry
=============================
In the present section our aim is to establish zero points for PL and PLC relations in the near infrared. The data used were outlined in section 2 and listed in Table A1. In the case of the important overtone pulsator, Polaris, the 2MASS observation is heavily saturated and thus has a very large uncertainty and there appears to be no other ground-based near infrared photometry on a system which can be reliably transformed to the SAAO system. There is, however, extensive DIRBE data (Smith et al. 2004). This has been transformed to the SAAO system as follows. There are eleven SAAO $JHKL$ standard stars in the DIRBE catalogue with $J <1.51$ and $K<
1.00$ and no indication of confusion in the DIRBE flags. Using these as calibrators, together with Procyon which has $J=-0.40$ and $-0.65$ on the SAAO system (transformed from Glass (1974) and Bouchet et al. (1991)) leads to $J=0.98$ and $K=0.60$ for Polaris.
In view of the $VI$ results of the present paper and those at $VI$ and $K$ in FB2007, we confine ourselves to determining the zero-points of PL and PLC relations of forms established in the LMC. The most extensive data set in $J$ and $K$ for LMC Cepheids, based on multiple observations, is that of Persson et al. (2004). We adopt their PL and PLC relations. Transformed (Carter 1990) onto the SAAO system these are: $$M_{K} = -3.258 \log P + \gamma_{1}$$ and $$M_{K} = -3.457 \log P + 1.894(J-K)_{0} + \gamma_{2} .$$ We have carried through the calculations with reddening corrections according to both the reddening law derived by Laney & Stobie (1993) and that of Cardelli, Clayton & Mathis (1989) and taking values of $E(B-V)$ from either Fernie et al. (1995) or Tammann et al. (2003) as outlined in section 2. The differences in the results obtained were all very small and much smaller than the standard errors of the quantities sought, so we list only one of them (Tammann et al. reddenings, Laney & Stobie reddening law).
In the case of a PL relation we expect the overtones, at their fundamental periods, to be brighter than fundamental pulsators of the same period, because of a temperature difference, and this is clear from the PL($K$) relations in the LMC by, e.g. Groenewegen (2000). We must therefore treat fundamental pulsators and overtones (at their fundamental periods) separately in discussing equation 7. Note that because of the relation between fundamental and first overtone periods (equation 1) the PL relation when transformed from the fundamental to the observed, overtone, periods will have a somewhat different slope, as found by Groenewegen. Using the method of reduced parallaxes as outlined earlier we obtain the results listed in Table 6 where Polaris is treated separately. Using the whole sample or only those with well covered light curves makes no significant difference to the results.
sample stars $\gamma_{1}$
------------------------- ------- ------------------
Fundamentals 220 $-2.40\pm 0.05$
Overtones (not Polaris) 26 $-2.33 \pm 0.15$
Polaris 1 $-2.50 \pm 0.03$
: PL($K$) zero-point, $\gamma_{1}$, determinations.
The difference in $\gamma_{1}$ between Polaris (at its fundamental period) and the mean of the fundamental pulsators ($-0.07 \pm 0.06$ mag) is not significant. However, it in fact agrees closely with the difference between overtones (at their fundamental periods) and fundamental pulsators expected (–0.08 mag) at the period of Polaris from the LMC data of Groenewegen (2000).
Results for the PLC zero-point are given in Table 7.
sample stars $\gamma_{2}$
------------------------- ------- ------------------
All except Polaris 246 $-3.01 \pm 0.05$
Fundamentals 220 $-3.02 \pm 0.05$
Overtones (not Polaris) 26 $-2.85 \pm 0.15$
Polaris 1 $-3.07 \pm 0.03$
: Determinations of the PLC($J,K$) zero-point, $\gamma_{2}$.
We expect for the PLC relation that overtones at their fundamental periods should follow the same relation as fundamental pulsators. The table shows that Polaris is $+0.05 \pm 0.06$ mag fainter than the mean of the fundamental pulsators, not significantly different from zero.
The LMC modulus from $J$ and $K$
================================
The LMC relations of Persson et al. (2004) converted to the SAAO system using the relations of Carter (1990) are: $$K_{0} = -3.258 \log P + 16.048$$ and $$K_{0} = -3.457 \log P + 1.894(J-K)_{0} + 15.402.$$ Also Groenewegen (2000) obtained for LMC overtone pulsators in the 2MASS system[^3]: $$K_{0} = -3.381 \log P_{1} + 15.533.$$ The zero-point, $\gamma_{1}$, for fundamental pulsators in Table 6 together with equation 9 gives an LMC modulus of $18.45\pm 0.05$. Polaris, at its observed (overtone) period with $K=0.58$ (the value from section 9 converted to the 2MASS system using Carpenter (2001)) and equation 11, gives $18.49\pm 0.04$ (taking into account the uncertainty in Groenewegen’s result. A straight mean of these two values ($18.47 \pm 0.03$) is our best estimated of the PL($K$) modulus of the LMC uncorrected for abundance effects. This is in agreement with $18.45\pm 0.04$ derived by FB2007.
A weighted mean of the last three entries in Table 7 leads to $\gamma_{2} = -3.05\pm 0.02$. Together with equation 10 and an estimate of its uncertainty leads to a infrared PLC modulus of $18.45\pm 0.04$ again without any metallicity correction.
These values may be compared with those found from $VI$ in section 7 which were $18.52\pm 0.03$ uncorrected for metallicity effects and $18.39\pm 0.05$ with a metallicity correction from Macri et al. (2006). These results suggest that any metallicity corrections to the infrared relations will be small.
Conclusions
===========
Our main conclusions based on the combined revised Hipparcos and HST data are the following;\
1. The coefficient of the $\log P$ term in a reddening free $V,I$ relation is found to be the same, within the uncertainties, in our Galaxy and in the LMC.\
2. This result is supported by an analysis of the data of Macri et al. (2006) for Cepheids in the inner region of NGC4258.\
3. Our reddening-free $V,I$ relation applied to the Cepheids in the inner region of NGC4258 leads to a modulus of $29.22 \pm 0.03$. in agreement (but of higher accuracy than) the maser-based distance ($29.29 \pm 0.15$)\
4. Our revised Cepheid $V,I$ calibration leads to a revision of the Cepheid-based distances to the 10 galaxies on which S2006 base their SNIa calibration and Hubble constant. We revise their results from $H_{0} = 62$ to $\rm 70 \pm 5\, km\,s^{-1}\,Mpc^{-1}$ This result is immune to metallicity effects on the Cepheid scale as long as these are linear. The Freeman et al. (2001) result is revised from 72 to $H_{0} = 76 \pm 8\,\rm {km\,s^{-1}\,Mpc^{-1}}$. Both these results are consistent with the recent WMAP value ($H_{0} = 73 \pm 3 \, \rm {km\,s^{-1}\,Mpc^{-1}}$)\
5. The zero-points of Galactic PL($K$) and PLC($J,K$) are derived.\
6. Applying our various relations to the LMC we find the following distance moduli, uncorrected for metallicity effects: $18.52 \pm 0.03$ from a reddening-free $V,I$ relation; $18.47 \pm 0.03$ from a PL($K$) relation; $18.45 \pm 0.04$ from a PLC($J,K$) relation.\
7. Applying a metallicity correction derived by Macri et al. (2006) to our LMC modulus leads to a true (metallicity corrected) modulus of $18.39 \pm 0.05$
Acknowledgments {#acknowledgments .unnumbered}
===============
We are very grateful to Dr L. Berdnikov for placing the unpublished revision of his catalogue of Cepheid photometry at our disposal. Dr L. Macri and Dr Y. Ita, very kindly and promptly, answered a number of queries about their work. We thank the referee (Prof. W. Gieren) for his comments
Alcock C. et al., 1995, AJ, 109, 1653 Andrievsky S.M. et al., 2002, A&A, 381, 32 Antonello E., Poretti E., 1986, A&A, 169, 149 Antonello E., Poretti E., Reduzzi L., 1990, A&A, 236, 138 Argon A.L., Greenhill L.J., Reid M.J., Moran, J.M., Humphreys E.M.L., 2007, astro-ph/0701396 Beaulieu J.P. et al., 1995, A&A, 303, 137 Benedict G.F., MacArthur B.E., Feast M.W., Barnes T.G., Harrision T. E., Patterson R.J., Menzies J,W., Bean J.L., Freedman W.L., 2007, AJ, in press (astro-ph/0612465) (FB2007) Benedict G.F. et al., 2002, AJ, 124, 1695 (FB2002) Berdnikov L.N., 1992, Sov. Ast. Let., 18, 398 Berdnikov L.N., Dambis A.K., Voziakova M.A., 2000, A&A Sup, 143, 211 Bouchet P., Manfroid J., Schmider F.X., 1991, A&A Sup, 91, 409 Caldwell J.A.R., Coulson I.M., 1987, AJ, 93, 1090 Caldwell J.A.R., Coulson I.M., 1989, MNRAS, 240, 285 Cardelli J.A., Clayton G.C., Mathis J.S., 1989, ApJ, 345, 245 Carpenter J.M., 2001, AJ, 121, 2851 Carter B.S.C., 1990, MNRAS, 242, 1 ESA, 1997, The Hipparcos Catalogue, ESA SP-1200 Evans N.R., 1992a, ApJ, 384, 220 Evans N.R., 1992b, ApJ, 389, 657 Evans N.R., 1994, ApJ, 436, 273 Evans N.R., 1995, ApJ, 445, 393 Evans N.R., Carpenter K.G., Robinson R., Kienzle F., Dekas A.E., 2005, AJ, 130, 789 Feast M.W., 1998, Mem. Soc. Ast. It., 69, 31 Feast M.W., 1999, PASP, 111, 775 Feast M.W., 2002, MNRAS, 337, 1035 Feast M.W., Catchpole R.M., 1997, MNRAS, 286, L1 Feast M.W., Walker A.R., 1987, ARA&A, 25, 345 Fernie J.D., Beattie B., Evans N.R., Seager S., 1995, IBVS, 4148 (http://ddo.astro.utoronto.ca/cepheids.html) Freedman W.L. et al., 2001, ApJ, 553, 47 (WLF 2001) Gascoigne S.C.B., Kron G.E., 1965, MNRAS, 130, 333 Gieren W., Storm J., Barnes T.G., Fouqué P., Pietrzyński G., Kienzle F., 2005, ApJ, 627, 224 Glass I.S., 1974, Mon. Not. Ast. Soc. S.A. 33, 53 Gorynya N.A., Samus N.N., Sachkov M.E., Antipin S.V., Rastorgouev A.S., 2000, in: The Impact of Large-Scale Surveys on Pulsating Star Research, (eds.) Szabados, L, Kurtz D.W., ASP Conf. Ser. 203, 242 Groenewegen M.A.T., 1999, A&A Sup, 139, 245 Groenewegen M.A.T., 2000, A&A 363, 901, 2000 Groenewegen M.A.T., Oudmaijer R.D., 2000, A&A, 356, 849 Herrnstein et al., 1999, Nature, 400, 539 Ita Y. et al., 2004, MNRAS, 347, 720 Kienzle F., Moskalik P., Bersier D., Pont, F., 1999, A&A, 341, 818 Kienzle F, Pont F., Moskalik P., Bersier D., 2000, in: The Impact of Large-Scale Surveys on Pulsating Star Research, (eds.) Szabados L., Kurtz D.W., ASP conf. Ser. 203, 239 Kiss L.L., Vinkó J., 2000, MNRAS, 314, 420 Laney C.D., Stobie R.S., 1986, MNRAS, 222, 449 Laney C.D., Stobie R.S., 1992, A&A Sup, 93, 93 Laney C.D., Stobie R.S., 1993, MNRAS, 263, 921 Laney C.D., Stobie R.S., 1994, MNRAS, 266, 441 Leavitt H.S., 1908, Harvard Ann. 60, 87 Leavitt H.S., 1912, Harvard Circ. 173 Luck R.E., Kovtyukh V.V., Andrievsky S.M., 2001, A&A, 373, 589 Lutz T.E., Kelker D.H., 1973, PASP, 85, 573 Macri L.M., Stanek K.Z., Bersier D., Greenhill L.J., Reid M.J., 2006, ApJ, 652, 1133 (LM2006) Madore B.F., 1976, RGO Bulletins No. 182, 153 Martin W.L., Warren P.R., Feast M.W., 1979, MNRAS, 188, 139 Mantegazza L., Poretti E., 1992, A&A, 261, 137 Morgan S.M., 1995, AJ, 109, 1263 Ngeow C-C., Kanbur S.M., 2005, MNRAS, 360, 1033 Ngoew C-C., Kanbur S.M., Nikolaev S., Buonaccorsi J., Cook K.H., Welch D.L., 2005, MNRAS, 363, 831 Persson S.E., Madore B.F., Krzemiński W., Freedman W.L., Roth M., Murphy D.C., 2004, AJ, 128, 2239 Rizzi, L. et al., 2007, astro-ph/0701518 Sachkov M.E., 1997, Inf. Bull. Var. Stars. No.4522 Sachkov M.E., 2002, Sov. Ast. Let., 28, 589 Saha A., Thim F., Tammann G.A., Reindl B., Sandage A., 2006, ApJ Sup, 165, 108 Sakai S., Ferrarese L., Kennicutt R.C., Saha A., 2004, ApJ, 608, 42 Sandage A., Tammann G. A., Reindl B., 2004, A&A, 424,43 Sandage A., Tammann G.A., Saha A., Reindl B., Macchetto F.D., Panagia N., 2006, ApJ, 653, 843 Smith B.J., Price S.D., Baker R.I., 2004, ApJ Sup, 154, 673 Soszyński I., Gieren W., Pierrzyński, 2005, PASP, 117, 823 Spergel D.N. et al., 2006, astro-ph/0603449 Szabados L., 2003, IBVS, 5394 (http://www.konkoly.hu/CEP/ intro.html) Tammann G.A., Sandage A., Reindl B., 2003, A&A, 404, 423 Turner D.G., Berdnikov L.N., 2004, A&A, 423, 335 Udalski A. et al., 1999, Act. Ast., 49, 201 van Leeuwen F., 2005, A&A, 439, 791 van Leeuwen F., 2007, Hipparcos; The new reduction of the raw data, Springer, Berlin (in press) van Leeuwen F., Fantino E., 2005, A&A, 439, 791 Zakrzewski B., Og[ł]{}oza W., Moskalik P., 2000, Act. Ast., 50, 387
[^1]: This is the final paper of a series
[^2]: Of the 74 Cepheids which pass their stated selection criteria, the following stars were rejected by them as outliers: I-040434, kI-008361, I-144755, I-081614, I-I-009241 (L. Macri private communication).
[^3]: Groenewegen’s relation for LMC fundamental pulsators is in good agreement with that of Persson et al. (2004). This latter is in the LCO system which is very close to that of 2MASS. The PL slopes given by Ita et al. (2004) (LCO system) do not agree well with Groenewegen or Persson et al. This may be connected with the inclusion of LMC Cepheids with $\log P < 0.4$ in the Ita et al. sample (Y. Ita private communication)
| {
"pile_set_name": "ArXiv"
} |
---
abstract: |
We present a 21-cm emission-absorption study towards extragalactic sources in the Canadian Galactic Plane Survey (CGPS). We have analyzed [H ]{}spectra towards 437 sources with $S_{\nu} \ge$ 150 mJy, giving us a source density of 0.6 sources per square degree at arcminute resolution. We present the results of a first analysis of the [H ]{}temperatures, densities, and feature statistics. Particular emphasis is placed on 5 features with observed spin temperatures below 40 K.
We find most spin temperatures in the range from 40 K to 300 K. A simple [H ]{}two-component model constrains the bulk of the cold component to temperatures ($T_c$) between 40 K and 100 K. $T_c$ peaks in the Perseus arm region and clearly drops off with Galactocentric radius, R, beyond that. The [H ]{}density follows this trend, ranging from a local value of 0.4 cm$^{-3}$ to less than 0.1 cm$^{-3}$ at R = 20 kpc. We find that [H ]{}emission alone on average traces about 75% of the total [H ]{}column density, as compared to the total inferred by the emission and absorption. Comparing the neutral hydrogen absorption to CO emission no correlation is found in general, but all strong CO emission is accompanied by a visible [H ]{}spectral feature. Finally, the number of spectral [H ]{}absorption features per kpc drop off exponentially with increasing R.
author:
- 'S. Strasser'
- 'A.R. Taylor'
title: 1420 MHz Continuum Absorption toward Extragalactic Sources in the Galactic Plane
---
INTRODUCTION
============
In the last 30 years our knowledge of the interstellar neutral hydrogen (H ) in the Galaxy has expanded dramatically. Many of the discoveries came from emission-absorption studies towards continuum background sources at 1420 MHz. Perhaps the most important realization of the early surveys [e.g. @c65; @htc71; @rmlw72] was that the emission line-widths are clearly broader than the corresponding absorption, indicating that the emitting gas must be dominated by a much warmer phase than the absorbing gas. [@l75], [@dst78], [@mwkg82] analyzed the spin temperature ($T_s$), and found values from approximately 40 K up to several thousand Kelvin, peaking between 80 K and 150 K. Due to the probable mix of cold and warm gas, the physical interpretation of $T_s$ has remained difficult. [@bw92] applied a two-component model to absorption and emission data and found cold phase temperatures in agreement with the above spin temperatures. However two recent studies [@ht02; @ht02b; @d02], fitting individual features, find the cold component temperature peaking between 40 K and 60 K. [@ht02b] also finds the warm component ranging from roughly 1000 K to 9000 K. This puts ideas about the ISM existing in distinct stable temperature and density regimes [e.g. @mo77] into question. We also do not have a firm theoretical or observational basis on the lower limit (if such exists) of the cold component temperature, before most of the gas becomes molecular.
Addressing the variation with Galactocentric radius, [@kh88] state that the supernova-dominated McKee-Ostriker (MO) model [@mo77] might describe the inner Galaxy better, while the outer, more quiescent, Galaxy is better described by the FGH model [@fgh69]. [@dmgg02] and [@kjbd02] define the mean opacity, $<\kappa>$, as the optical depth line of the sight integral divided by the radial velocity change due to the rotation curve, for radial increments in the Galaxy. They found an decrease in $<\kappa>$ with increasing radius in the first and fourth quadrant respectively. Overall, however, our knowledge of the radial variation of the ISM phases is still quite limited. This is mainly because of the time requirements to observe a large number of sources over a wide area of the Galactic plane at the sensitivity needed to get a good source density. New surveys, as those used by [@dmgg02] and [@kjbd02], are starting to give us a much better picture of the [H ]{}states and the variations with Galactocentric radius.
We present a new emission-absorption study toward compact continuum background sources in the Canadian Galactic Plane Survey [CGPS, @t03]. We have applied an “on-off method” to extract the emission brightness temperature and optical depth toward the sources. Due to the high resolution and sensitivity of the CGPS we are able to study the radial variation of the [H ]{}temperature and density for a much higher source density in the Galactic Plane than ever before. Section 2 gives an overview of the CGPS and the on-off method used. In section 3 we present the spin temperatures and the results of a simple two-temperature model. Further we discuss the spatial variation of the [H ]{}column density, and the correlation with CO. We also present a brief analysis of absorption feature statistics. Section 4 summarizes our results and indicates some areas of future research.
OBSERVATIONS AND THE ON-OFF METHOD
==================================
The CGPS 21-cm dataset
----------------------
The Canadian Galactic Plane Survey [@t03] is a multi-wavelength study at 1 arc-minute resolution from Galactic longitude 74to 147, and latitude -3.6to 5.6. The observations for the 21 cm dataset were carried out with the Dominion Radio Astrophysical Observatory (DRAO) Synthesis Array, near Penticton, Canada [for a description, see @ldbetal00]. The array consists of seven 9-metre antennas. Four of the antennas are movable to yield a total of 12 configurations, or 144 baselines. For the survey each field was observed in each position for 12 hours to obtain full uv-sampling. Short spacing data were taken with the DRAO 26-m single dish telescope. At 21-cm the CGPS images have a spatial resolution of 1$\times$1cosec($\delta$). The instrumental velocity resolution of 1.319 km/s is sampled every 0.824 km/s, covering LSR velocities from approximately 40 km/s to -160 km/s. The RMS noise of the emission in a single velocity channel at the field centres is about 3 K in low emission regions, and up to 6 K in high emission regions [@gthd00]. The continuum data are taken in four 7.5 MHz bands (centred on 1406.9 MHz, 1413.8 MHz, 1427.4 MHz, and 1434.3 MHz). A technical matter of importance for our work is the fact that a slightly different uv-taper is applied to the (continuum subtracted) [H ]{}datacube than the continuum maps. We convolved the CGPS continuum maps to match the beam shapes exactly. For a detailed description of the data acquisition and reduction process see [@t03].
Compact Continuum Background Sources in the CGPS
------------------------------------------------
For this study we have identified 437 compact sources with a peak continuum brightness temperature higher than 30 K in the CGPS, corresponding to integrated flux densities higher than 150 mJy. The cut-off value of 30 K was chosen because the absorption signal to noise for sources below this temperature is almost always less than 1. This gives a source density of about 0.6 sources per square degree in the CGPS area. Figure \[fig:sloc\] shows the source locations. The sources provide a fairly uniform coverage of this section of the Galactic plane. Note that we do not exclude any sources that might be of Galactic origin. By inspecting the emission versus absorption velocity cut-offs we find about 15 potential Galactic sources.
The On-Off Method
-----------------
Past emission-absorption studies typically used one on-source position and several off-source positions distributed around the source to approximate the “unabsorbed” spectrum (e.g. a hexagonal pattern, see [@cst88]). For this work a different approach was taken, since we have a fully sampled, mosaiced dataset. First, the continuum sources were fitted with 2-D elliptical Gaussians. An average off-source emission brightness temperature, $T_{B}(v)$, was then extracted in a one arc-minute wide elliptical annulus, with its inner edge tracing the fitted ellipse in shape but scaled to be one FWHM from the centre. Most sources are unresolved, so the annuli are typically approximately one arc-minute (one beamwidth) away from the source centre. The RMS variations in $T_B(v)$ over the annulus give an estimate of the emission error. The off temperature is assumed to represent the emission on-source in the absence of continuum absorption. It is thus important to keep the separation between the on and off measurement directions as small as possible. At 1 kpc line of sight distance, one arc-minute corresponds to about 0.3 pc.
An average optical depth ($\tau$) was computed from the on-source spectrum using pixel values within one HWHM from the centre of the background source, and any pixels further away for which the emission temperature signal to noise was higher than 5. Under Local Thermal Equilibrium (LTE) the optical depth in any channel is then given by $$\label{eqn:tau}
\mathrm{exp}\left[-\tau(v)\right] = \left<1-\frac{T_{B} - T_{on}(x,y)}{T_{bg}(x,y)}\right>$$ where $T_{bg}$ is the continuum brightness temperature of the background source and $T_{on}(x,y)$ is the continuum subtracted profile at the pixel location ($x,y$). Typically about 10 pixels per channel are used in the average optical depth computation. The spin temperature is derived by, $$\label{eqn:Ts}
T_s(v) = \frac{T_{B}(v)}{1-e^{-\tau(v)}}$$ Note that for large optical depth, $T_{B}-T_{on}$ approaches $T_{bg}$. Due to noise in the data the optical depth can diverge. This situation was typically encountered for at least a few channels in any given spectrum. Since this corresponds to completely optically thick gas, we set $\mathrm{exp}(-\tau)$ and $T_s(v)$ equal to zero and $T_B(v)$ respectively in these cases.
To test our on-off method we inserted absorption from artificial background sources into the data cubes. The artificial continuum sources were Gaussian shaped to match the synthesized beam and had peak brightness temperatures ranging from 30 K to 300 K. The line emission brightness temperature in each channel was modified at the position of the artificial source in proportion to the background source intensity and the [H ]{}brightness in that channel at the source position. The resulting absorption spectra were extracted using the on-off method and compared to the known input profiles. The agreement between the two was consistently well within one sigma for both the spin temperatures and optical depth spectra. We also compared spectra common to this study and [@dkgh83]. In general the agreement is good, but in noisy channels they sometimes differ by several $\sigma$.
RESULTS AND INTERPRETATION
==========================
Figures \[fig:spec01\] and \[fig:spec02\] show two representative spectra, illustrating the quality of the data for two sources with very different continuum brightness temperatures. We do not show all the spectra in this paper due to their large number. All the spectral data are available online ([*[http://www.ras.ucalgary.ca/CGPS/products/]{}*]{}) or by request from S. Strasser. Some features are common to many spectra. We can almost always identify the local (10 to -20 km/s) and Perseus arms (around -25 to -50 km/s) as broad peaks in emission, and in many cases also the outer arm (between -80 and -110 km/s). In general the absorption spectra show more structure and narrower line widths than the emission spectra. However this difference is less pronounced than for past, lower angular resolution, studies. Some features follow the “$T_s - \tau$ relation” ($\log \tau$ decreases linearly with increasing $\log T_s$) quite nicely. This is attributed to the difference in absorption versus emission line widths, as the former is dominated by cool gas, and the latter by the warm component. In many cases however, the observed relation is more in the shape of a loop, where the spin temperature for a given optical depth on one side of the feature is systematically higher or lower than on the other side. Several recent studies confirm this relationship [see e.g. @dmgg02].
The Spin Temperatures
---------------------
The spin temperature ($T_s$) is the simplest temperature indicator we can compute. Under assumption of a single-component medium in equilibrium it is equal to the kinetic temperature of the gas [e.g. @kh88]. The distribution of spin temperatures is given in figure \[fig:Tshist\] for all channels with S/N$(T_{on-off}$) $>$ 5 (6318 channels in total). The median of the distribution is 120 K, and 90% of the values lie between 60 K and about 280 K. This is in rough agreement with previous studies [@dst78; @mwkg82; @pst82] and recent results for the southern Galactic plane [@dmgg02].
It is generally believed though that the neutral hydrogen has at least two components, the cold neutral medium (CNM) and the warm neutral medium (WNM). The spin temperature measures a harmonic mean of the CNM and WNM, weighted by the column density of each component. It therefore only provides us with an upper limit to the true cold component temperature and a lower limit to the warm component temperature. This is confirmed by more recent analyses that fit the CNM and WNM separately and find lower cold component [H ]{}temperatures than the associated spin temperatures. By fitting profiles with Gaussian functions [@ht02; @ht02b] find cold component temperatures in the range from about 10 K to 75 K. [@dmgg02] use a novel technique of fitting the emission and absorption and find a range for the cold temperature between 20 K and 110 K, much lower than their spin temperatures.
A 2-Component Model {#sec:2c}
-------------------
We have applied a two-component model to the data extracted from the CGPS. Our approach is similar to a model applied to the Galaxy and M31 by [@bw92]. We use the transfer equation for a case where two components contribute to the emission brightness temperature: $$\label{eqn:multi}
T_{B}(v) = T_w\left(1-e^{-\tau_w(v)q}\right) + T_c\left(1-e^{-\tau_c(v)}\right)e^{-\tau_w(v)q} +
T_w\left( 1-e^{-\tau_w(v)(1-q)}\right)e^{-\left[\tau_c(v) + \tau_w(v)q\right]}$$ Here we assume that the gas has a cold component at temperature $T_c$ and optical depth $\tau_c(v)$ and a warm component at temperature $T_w$ with optical depth $\tau_w(v)$. A fraction $q$ of the warm gas is in front of the cold gas at every velocity. The first term in equation \[eqn:multi\] represents WNM emission in front of the CNM. The second term accounts for the CNM emission, absorbed by the WNM in front. The third term represents the WNM behind the cold component, which is absorbed by both the CNM and WNM foregrounds. Since $\tau \propto N/T$, we further assume that the absorption due to the WNM is much smaller than that due to the CNM and that the WNM does not absorb the CNM. For $\tau_w(v) \ll \tau_c(v)$ and $\tau_w(v) \ll 1$ we then derive $$\label{eqn:2comp}
T_B(v) = T_0(v)e^{-\tau_c(v)} + T_{\infty}(v)\left(1 - e^{-\tau_c(v)}\right)$$ where $T_0 = T_w\tau_w(v)$ and $T_{\infty} = T_w\tau_w(v)q + T_c$ are the limits of $T_B$ as $\tau_c$ goes to zero and infinity respectively. Note that, even though these limiting temperatures are functions of velocity, we treat them as constants by assuming that $\tau_w(v)$ is approximately constant. We make this assumption because we fit the data for all sources at once, hence averaging out the velocity dependence.
In figure \[fig:2cfig01\] a scatter plot of $T_{B}(v)$ versus $1-e^{-\tau_c(v)}$ for all sources is displayed. On these axes $T_0$ and $T_{\infty}$ are the values of the $T_B(v)$ at $1-e^{\tau_c(v)}$ equal to 0 and 1 respectively. Every data point represents one velocity channel. Only points with $T_{on-off}$ S/N higher than 5 are plotted. The centre line is a least squares fit of equation \[eqn:2comp\] to the data, yielding $T_0$ = 24 K and $T_{\infty}$ = 88 K. We cannot explicitly solve for $T_w$, only the factor $T_{\infty} = T_w\tau_w$. Statistically, however, we expect an average value of $q$ = 0.5, and assuming this we can solve for the cold component temperature. This gives $T_c$ = 76 K. Varying $q$ gives an estimate of the uncertainty in the average cold component temperature. $T_c$ goes to 88 K and 64 K for $q$ equal to 0 and 1 respectively.
We interpret the scatter in the fit as being due to a range of $T_c$ and $T_0 = T_w\tau_w$. In figure \[fig:2cfig01\] we show visually selected upper and lower envelopes with the form of equation \[eqn:2comp\] (solid lines). $T_0$ ranges from 0 K to 60 K. Points on the lower envelope, which has $T_0=0$, thus represent regions where there is no warm component, and as we move toward the upper envelope, $T_c$ and $T_{\infty}$ increase. From the envelopes we find $T_c$ ranging between 50 K and $100 \pm 30$ K (the upper limit is for $q$ = 0.5 and the uncertainty comes from varying $q$ between 0 and 1).
There are a few scattered points below the lower envelope. We analyzed these points in more detail to see if they could be due to a very cold neutral hydrogen component (below 40 K). We find that several well defined (relatively unblended) features contribute most of the points. Table \[tab:coldfea\] summarizes these features. Selecting features with line centre $T_s$ below 40 K yields the same list. The optical depth and spin temperature line profiles for these features are shown in figure \[fig:coldbox\]. The last column in table \[tab:coldfea\] shows the cold temperatures obtained by fitting our two-component model to each feature region individually (for a $q$ = 0.5). Fitting individual features in this way is similar to the approach taken by [@dmss00], except that we fit two components instead of one, and the fitting is done by least squares rather than by eye. The cold temperatures are confirmed by the two-component model. Except for the feature towards (85.089, +3.620), which has a computed $T_c$ of only 4 K (range of 0 K for $q=1$ to 13 K for $q=0$), all of the features have a $T_c$ of approximately 20 K, with a typical range between 10 K for $q=1$ and 30 K for $q=0$. Note that the velocities of these features place all of them quite far in the outer Galaxy (unfortunately making it unlikely that they show up as [H ]{}self-absorption).
While the model envelopes (figure \[fig:2cfig01\]) fit the data reasonably well, we can clearly see that $T_{off}$ increases more rapidly at small $\tau_c$ and flattens out at a maximum value of approximately 120 K. This peak brightness temperature is a well known observational fact for spiral galaxies. We are probably seeing two effects here: First, because galactic rotation spreads out the gas in velocity, only a certain amount of gas ever piles up at the same velocity limiting the maximum brightness temperature [@dmgg02]. Second, clearly if the optical depth in a line becomes much greater than 1, it will start to saturate, which creates a non-linear dependence of $T_B$ on $e^{-\tau}$. These two effects would explain why a simple radiation transfer, as modeled, does not fit the upper envelope better. The dashed line in figure \[fig:2cfig01\] shows a fit to the upper envelope below the “break” with $T_0$ = 25 K and $T_{\infty}$ = 212 K. For $q$ = 0.5 this gives a cold component temperature, $T_c$ = 200 K (with a range from 188 K to 212 K for $q$ equal to 1 and 0 respectively). If saturation is a dominant effect at high optical depth, then this line reflects what the state of the unsaturated gas is. The much higher $T_c$ value obtained by this fit, compared to fitting the whole range of $\tau$, is accounted for by the fact that the warmer the gas gets, the harder it is for it to reach high optical depth.
[@bw92], applied a similar model to data for 87 sources covering a large range of Galactic latitude. They found best fit values of $T_c$ = 105 K and $T_0$ = 4 K, and envelope values from 54 K to 157 K for $T_c$ and from 2 K to 6 K for $T_0$. The inconsistency between these and our results might arise primarily from the typically higher latitude of sources used in their study.
Radial Variation of Physical Conditions
---------------------------------------
### Temperature
We examined the radial variation of the temperature by fitting the two-component model to bins in Galactocentric radius. We used the rotation curve from [@bb93] to compute the Galactocentric radius corresponding to a given velocity, assuming a rotation velocity of 220 km/s for the local standard of rest, and 8.5 kpc Galactic centre to Sun distance. To find the radial variation of our model parameters, $T_c$ and $T_0$, we binned the data in 1 kpc intervals out to 16 kpc and beyond in 2 kpc intervals to 20 kpc. The bin sizes were chosen so that every bin contains a large enough number of points to apply our model. We fitted the data for each bin in the same way as the whole dataset in section \[sec:2c\], finding a best fit value and an upper and lower envelope.
Figure \[fig:2cradvar\] shows the variation of $T_c$ and $T_0$ as a function of radius for $q$ = 0.5. The dots show the fitted values, while the lines above and below indicate the temperature range given by the envelopes. The average cold component temperature rises from a local value of about 65 K to 80 K around 14 kpc and then drops off to 45 K in the outer Galaxy. The uncertainty introduced by $q$ is a maximum of about 20 K for the inner bins, and drops to around 10 K in the outer bins (R $>$ 14 kpc). The lower envelope follows the same trend ranging from 40 K to 60 K, while the upper envelope depends only weakly on R ranging from 80 K to 100 K, with a minimum around 12 kpc and a peak at 14 kpc.
$T_0$ starts out at a maximum value of 36 K, and also has a local peak around 14 kpc before falling off to about 10 K at 20 kpc. These results indicate that warm gas is rare in the outer Galaxy. Of course the variation in $T_0$ could be due to a combination of changing warm temperature and opacity. Further, at 20 kpc a line of sight at latitude 5$^{\circ}$ is 1.8 kpc above the plane, and the Galactic warp could be as large as 4 kpc [@dl90]. Especially the negative latitude lines of sight can thus look at quite a different height of the disk locally as compared to the far outer Galaxy. This could affect our results (our spin temperatures, however, do not indicate any dependence on $z$).
### Volume Densities
We have used the absorption and emission spectra to study the density of the [H ]{}gas. For a gas at temperature $T_s= T_B\left(1-e^{-\tau}\right)$ the element of column density in a velocity channel of width $\Delta v$ is given by $$\label{eqn:N}
\Delta N(v) = \frac{C\tau(v)T_B(v)}{1-e^{-\tau}}\Delta v$$ and represents the number of atoms per cm$^2$ in the velocity range $v$ to $v + \Delta v$. For $v$ in km/s, the constant $C$ is equal to $1.83\times10^{18}$ cm$^{-2}$K$^{-1}$(km/s)$^{-1}$. A spatial density as a function of Galactocentric radius can be computed by dividing the line of sight into $R$ bins. The density, $n(R)$ in any bin is given by, $$\label{eqn:nR}
n(R) = \frac{\sum_{\Delta V}N(v)}{\Delta L}$$ where $\Delta L$ is the line of sight distance through the bin and $\Delta V$ is the corresponding velocity range computed from [@bb93]. Figure \[fig:NvsR\] shows the mean density as a function of Galactocentric radius for all the sources. A robust exponential least-squares fit of the form $y = Ae^{-Bx}$ to the distribution is shown, which characterizes the decrease quite well beyond 14 kpc, with values for $A = 6.04$ cm$^{-3}$ and $B = 0.29$ kpc$^{-1}$. We can see a local increase in the density around 12 kpc, due to the Perseus arm.
We also examine the variation of density with height above the plane, $z$. To derive a spatial density as a function of $z$, $n(z)$, we sum the data in a given $z$ bin as in equation \[eqn:nR\], but now $\Delta V$ refers to the velocity range contributing to the bin and $\Delta L$ is the line of sight range, as before. In figure \[fig:Nz\] we show the mean value of $n(v)$ as a function of $z$. The mean of the density drops off with scale heights of approximately 350 and 280 pc above and below the plane respectively (found from exponential fits to the positive and negative sides of the distribution). The difference is due to the Galactic warp, which shows up as a density enhancement clearly visible at positive $z$ in figure \[fig:Nz\]. The scale heights correspond to a combination of the warm and cold components (averaged over the CGPS region).
In the absence of information on absorption, [H ]{}mass is often inferred from emission line data. The question of the ability to trace [H ]{}gas by the emission alone is thus interesting to address. In deriving [H ]{}mass from emission the optically-thin case is assumed where equation \[eqn:N\] reduces to $$\label{eqn:Nem}
N_{em}(v) = C T_B(v)\Delta v,$$ where we have designated $N_{em}$ as the column density derived from emission alone. Figure \[fig:NvsN\] shows the emission column density, $N_{em}$, versus the column density $N_{HI}$ given by absorption using equation \[eqn:N\]. By combining equation \[eqn:N\] with equation \[eqn:Nem\] we can find a functional relation between $N_{HI}$ and $N_{em}$, $$\label{eqn:Nfit}
N_{HI} = -A \ln \left(1 - \frac{N_{em}}{A} \right)$$ where $A = C T_s \Delta v$. In reality many different spin temperatures contribute to each measured column density. However equation \[eqn:Nfit\] provides a convenient way to characterize the relationship between $N_{HI}$ and $N_{em}$ . By least squares fitting the parameter A, we find a value of $2.1 \times 10^{22}$ cm$^{-2}$; The corresponding fit is shown in figure \[fig:NvsN\] and well represents the data.
The average ratio of $N_{HI}/N_{em}$ is equal to 1.32 (or $N_{em}$ = 75% of $N_{HI}$). At the highest column densities $N_{em}$ measures only about 65% of the true column. [@dmgg02] find typical values for $N_{HI}$/$N_{em}$ of 1.4 to 1.6 in the inner Galaxy.[@dkgh83] derived a somewhat lower value of 1.2 for $N_{HI}/N_{em}$. The difference is probably due to our concentration of lines of sight very close to the Galactic plane, where the opacity is highest, so the difference between $N_{HI}$ and $N_{em}$ is the largest on average. [@ht02] clearly show this dependence of $N_{HI}/N_{em}$ on Galactic latitude in their figure 5, and their values close to the Galactic plane qualitatively agree with the results presented here.
Correlation with CO
-------------------
Another of the CGPS data products are the reprocessed images from the FCRAO Outer Galaxy Survey of the $^{12}$CO J = 1 - 0 rotational transition at 2.6 mm [@hbs98]. The survey has an angular resolution of 1.67 arc-minutes, and covers the same velocity range as the 21-cm dataset at the same spectral resolution. CO is often used as a tracer for molecular hydrogen, as it exists in a similar physical environment. The accuracy with which it traces $H_2$, however, is still a subject of frequent debate. While the presence of CO correlates with $H_2$ or very cold atomic hydrogen, the reverse often does not seem to be the case [e.g @rkh94; @mh96].
Recent studies [@gthd00; @ht02b; @kb01; @kd03] are finding clear evidence for [H ]{}below 30 K, where we expect most the hydrogen to be in molecular form . Our own 2-component analysis reveals several absorption features below 40 K. To check for the co-existence of [H ]{}absorbing gas and molecular material we compared the [H ]{}absorption spectra to the CO emission spectra towards the sources.
Figure \[fig:COcorr\] shows the CO emission brightness temperature versus the [H ]{}optical depth for the 134 sources which are in the FCRAO survey area (from Galactic longitude 102$^\circ$ to 143$^\circ$ and all CGPS latitudes). It is evident that overall there is no correlation between the CO emission and [H ]{}absorption. The CO noise level is $\sim 0.15$ K. There is a very clear lack of data points with CO emission brightness temperature above the noise level at low [H ]{}optical depth. This, and visual inspection of the spectra, show that in most cases where we do see CO emission, [H ]{}absorption is also present. We find about 60 clear matches between [H ]{}and CO features, and for approximately 20 of these, $e^{-\tau}$ goes to zero. Many such features are only a few channels wide both in CO and [H ]{}optical depth. Figure \[fig:COsamplespecs\] shows two sample features. The narrow feature in the left panel, illustrates a case in which $e^{-\tau}$ goes to zero. The panel on the right displays two broad [H ]{}absorption lines, one of which shows a clear correlation with CO emission.
Of the two cold features listed in table \[tab:coldfea\] for which we have CO data, one (112.107,+2.362) has associated CO emission. In general, however, there is no preference for low spin temperatures in [H ]{}channels with associated CO. As noted earlier though, if a cloud has multiple temperature components, the spin temperature is a harmonic mean, and therefore, even if the cold-component temperature is very low, the apparent spin temperature does not necessarily have to be low. So, while there is no overall correspondence between the CO and [H ]{}features, the association of over 60 features is interesting, and a good indication that at least some of the atomic hydrogen is at very low temperatures. We plan to pursue this by modeling the individual features and by obtaining more observations at higher spectral resolution and signal to noise to be able to better quantify our current results.
Feature Distribution and Properties
-----------------------------------
The spatial density of absorption peaks is a simple tracer of cold ISM structures. An absorption feature in general corresponds to one cloud, or more if several lines are blended. The feature density (number of features per kpc) is thus a lower limit to the cloud density. We have developed an automatic absorption feature identification routine that finds peaks in the $T_{on-off}$ spectra. Such an identification is never completely unambiguous, as noise spikes have to be separated from real features. The criteria applied were chosen to agree with a visual inspection of the spectra. Features have to be stronger than 3 sigma in $T_{on-off}$ to be identified. To determine the line of sight distance corresponding to a given radius interval we used the rotation curve from [@bb93]. Mathematically for any radial bin the feature density is given by, $$\mathrm{\#\,features\, /\, kpc} = \frac{\sum_{sources} \mathrm{\#\,features\,/\,kpc}}{N_{bin}}$$ where $N_{bin}$ is the total number of sources contributing to any given bin.
Figure \[fig:feaR\] shows the density of features as a function of Galactocentric radius for two optical depth cut-off levels. The solid histogram is for all features with $\tau > 0.5$, while the dashed histogram is for $\tau > 1.5$. The error bars indicate the standard deviation in the bin divided by the square-root of the number of sources contributing to each bin. We show an exponential least-squares fit to the distribution for $\tau > 0.5$ which is given by 19.7e$^{-0.36R}$ features/kpc, and which describes the distribution quite well. As the histogram for $\tau > 1.5$ illustrates, higher optical depth features are clearly concentrated in the local and Perseus arms and disappear entirely beyond 16 kpc. [@gd89] found the feature density for all peak values of $\tau$ to have an average value of about 0.42 $\pm$ 0.08 kpc$^{-1}$ from R = 0 to 4.25 kpc and 0.8 $\pm$ 0.1 from R = 4.25 kpc to 8.5 kpc which agrees with our results. The feature density thus seems to reach a peak around the solar neighbourhood and decreases both in the inner and outer Galaxy.
We have to keep in mind though that as the line of sight distance increases, a given change in velocity becomes a larger and larger change in line of sight distance due to the rotation curve. While 1.32 km/s (our spectral resolution) locally corresponds to about 200 to 300 pc (using the rotation curve to convert from velocity to distance), at -100 km/s the same velocity corresponds to scales from 200 pc to a few kpc (depending on longitude). Thus, if the dominant scale of structure in the [H ]{}is on the order of a 100 pc rather than several hundred parsecs to kiloparsecs, then this observed radial dependence could at least partially be an observational effect. As we look further and further away along the line of sight, we lose clouds in our identification due to the decreasing spatial resolution. In the outer Galaxy, an increase in the line of sight distance always corresponds to an increase in Galactocentric radius. Thus, since our data is largely in the outer Galaxy, the observed effect will give a drop in the number of features.
CONCLUSIONS
===========
We have presented a new emission-absorption survey of the Galactic Plane at 21-cm with an unprecedented source density of about 0.6 sources per square degree. In agreement with previous studies we find spin temperatures between 40 K and 300 K. Due to the difficulty in interpreting the spin temperature we apply a simple 2-component model to the data. We find $T_c$ and $T_w\tau_w$ ranging from approximately 50 K to 80 K, and 0 K to 60 K respectively. Both $T_c$ and $T_w\tau_w$ show a peak at about 14 kpc and a decrease in the outer Galaxy.
The derived density of neutral atomic hydrogen varies roughly as e$^{-0.3R}$, with Galactocentric radius (in kpc). However directly interpreting this as a drop in gas density with radius is complicated by the presence of the Galactic warp and the changing scale height visible in the $z$ profile of N. Comparing the column density traced by emission to the total column we find that on average the emission traces about 75% of the total [H ]{}mass, and at the highest optical depths only about 65%.
We looked for evidence of very cold hydrogen in the Galaxy in our two-component model results and the spin temperature. This yielded a list of 5 features, with $T_s < 40$ K and lying below the two-component model lower envelope. Further a comparison to CO shows that even though there is no overall correlation between CO emission and [H ]{}absorption, about half the spectra show at least one [H ]{}absorption feature with clearly associated CO emission.
This study is a first look at the [H ]{}absorption in the CGPS. We hope to continue this analysis by incorporating the Southern and the VLA Galactic Plane Surveys [@mgdghw01; @t03], giving us more information on the inner Galaxy, and the difference between the southern and northern parts of the Galactic plane. While the 2-component model applied here gives us some interesting results, it is clear that a more detailed feature by feature analysis could reveal much new information. This will also be pursued further.
We wish to thank J.M. Dickey, S.J. Gibson, and C. Heiles for helpful comments and discussions. Thanks also to J.C. Brown for her help with IDL and to L.A. Higgs for modifying the DRAO software for our purposes.
Brand, J., Blitz, L. 1993, A&A, 275, 67 Braun, R., Walterbros, R.A.M. 1992, , 386, 120 Clark, B.G. 1965, , 142, 1398 Colgan, S.W.J., Salpeter, E.E., Terzian, Y. 1988, , 328, 275 Dickey, J.M. 2002, in ASP Conf. Ser. 276, Seeing Through the Dust, ‘ ed. A.R. Taylor, T.L. Landecker & A.G. Willis (San Francisco:ASP), 248 Dickey, J.M., Lockman, F.J. 1990, , 28, 215 Dickey, J.M, Kulkarni, S.R., van Gorkum, J.H., Heiles, C.E. 1983, , 53, 591 Dickey, J.M., McClure-Griffiths, N.M., Gaensler, B.M., Green, A.J. 2003, , 585, 801 Dickey, J.M., Mebold, U., Stanimirovic, S., Staveley-Smith, L. 2000, , 536, 756 Dickey, J.M., Salpeter, E.E., Terzian, Y. 1978, , 36, 77 Field, G.B., Goldsmith, D.W., Habing, H.J. 1969, , 155, L149 Garwood, R.W., Dickey, J.M. 1989, , 338, 841 Gibson, S.J., Taylor, A.R., Higgs, L.A., Dewdney, P.E. 2000, , 540, 851 Heiles, C., Troland, T.H. 2003a, , 145, 329 Heiles, C., Troland, T.H. 2003b, , 586, 1067 Heyer, M.H., Brunt, C., Snell, R.L., Howe, J.E., Schloerb, F.P. Carpenter, J.M. 1998, , 115, 241 Hughes, M.P., Thompson, A.R., Colvin, R.S. 1971, , 23, 323 Kavars, D.W., Dickey, J.M., , accepted Knee, L.B.G., Brunt, C.M. 2001, , 412, 308 Kolpak, M.A., Jackson, J.M., Bania, T.M., Dickey, J.M. 2002, , 578, 868 Kulkarni, S.R., Heiles, C. 1988, in Galactic and Extragalactic Radio Astronomy, Verschuur, ed. G.L., Kellermann, K.I., Berlin: Springer, 117 Landecker, T.L., et al. 2000, , 145, 509 Lazareff, B. 1975, , 42, 25 McKee, C.F., Ostricker, J.P. 1977, , 218, 148 McClure-Griffiths, N.M., Green, A.J., Dickey, J.M., Gaensler, B.M., Haynes, R.F., Wieringa, M.H. 2001, , 551, 394 Mebold, U., Winnberg, A., Kalberla, P.M.W., Goss, W.M. 1982, , 115, 223 Meyerdierks, H., Heithausen, A. 1996, , 313, 929 Payne, H.E., Salpeter, E.E., Terzian, Y. 1982, , 48, 199 Radhakrishnan, V., Murray, J.D., Lockhart, P., Whittle, R.P.J. 1972, , 24, 15 Reach, W. T., Koo, B., Heiles, C. 1994, , 429, 672 Taylor, A.R. 1999, in ASP Conf. Ser. 168, New Perspectives on the Interstellar Medium, ed. A.R. Taylor, T.L. Landecker & G. Joncas (San Francisco:ASP), 3 Taylor A.R., et al. 2003, , 125, 3145 Taylor, A.R., Stil, J.M., Dickey, J.M., McClure-Griffiths, N.M., Martin, P.G., Rothwell, T., Lockman, F.J. 2002, in ASP Conf. Ser. 276, Seeing Through the Dust, ed. A.R. Taylor, T.L. Landecker & A.G. Willis (San Francisco:ASP), 68
[lcccccccc]{} 4C+35.49 & 76.366 & -2.674 & -74.02 & 36 & 3 & 0.71 & 18 & 20\
MG4 J203647+4654 & 85.089 & 3.620 & -103.70 & 18 & 2 & 1.94 & 15 & 4\
3C+434.1 & 94.112 & 1.223 & -81.47 & 31 & 2 & 1.01 & 20 & 20\
87GB 231107.2+625244 & 112.107 & 2.362 & -93.80 & 27 & 2 & 3.00 & 26 & 19\
87GB 020322.8+623159 & 131.459 & 1.157 & -97.93 & 32 & 2 & 0.89 & 19 & 22\
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We study the temporal asymmetry of over 600 bursts from the BATSE 3B catalog, encompassing a 200-fold range in peak flux. By comparing the rates of rise and fall of the flux near the highest burst peak, we find that about two-thirds of the bursts exhibit a preferred asymmetry in the sense that the flux rises more rapidly than it falls, confirming the conclusions of previous studies employing smaller databases. The statistical significance of the average time asymmetry of the sample is $>99.999$%; therefore, models that predict time symmetry of the burst profile are ruled out. We find no statistically significant correlation between burst temporal asymmetry and peak. This result is consistent with both cosmological and local interpretations of the gamma ray burst phenomenon.'
author:
- Bennett Link
- 'Richard I. Epstein'
title: STATISTICS OF GAMMA RAY BURST TEMPORAL ASYMMETRY
---
INTRODUCTION
============
The origin of gamma ray bursts (GRBs) remains a mystery since their detection by the [*Vela*]{} satellites in July 1969 ([@discovery]). The Burst and Transient Source Experiment (BATSE) on board the [*Compton Gamma Ray Observatory*]{} continues to detect GRBs distributed with striking isotropy on the sky, and with a dearth of faint events compared to that expected for a homogeneous distribution ([@isotropy1]; [@isotropy2]). With counterparts at other wavelengths yet to be identified, the distances to GRBs remain uncertain. The idea that GRBs originate at cosmological distances has emerged as a serious possibility, while the prospect that bursts originate within the halo of the Galaxy remains tenable. In the face of these uncertainties, the identification and interpretation of fundamental properties of GRB variability is needed to assess proposed models.
Cosmological GRB models account naturally for the observed isotropy while explaining the dearth of faint events as due to a modification of the observed distribution by the universal expansion. These models also predict a general time dilation of the more distant, fainter events. Recently, some authors have found indications of time dilation consistent with cosmic expansion (see, , [@dilation1]; [@dilation2]; [@dilation3]; [@dilation4]), but the evidence is not yet statistically compelling. Models in which GRBs originate near the Galaxy have the appeal of modest luminosity requirements compared to cosmological models, but need fine tuning to maintain consistency with the observed isotropy and number-flux relationship.
Some GRB models, such as those accounting for burst time structure as due soley to beams crossing our line of sight, predict time-symmetric light curves. Recent quantitative studies of the shapes of GRB light curves, however, have established that GRBs are time-asymmetric ([@LEP]; [@norris_asymmetry]; [@timeasymmetry]; [@mitrofanov]; [@fishman_asymmetry]), confirming ealier claims (see, , [@barat_asymmetry]). Link, Epstein, & Priedhorsky (1993) utilized a [*skewness function*]{}, similar to that used by Weisskopf (1978), and found that the majority of 20 bright GRBs selected from the first 48 detected by BATSE are time-asymmetric in the sense that the flux rises more rapidly than it falls. In a subsequent study, Nemiroff (1994) quantified the degree of time asymmetry by considering the ratio of the number of times where the counts in a given time bin are lower than in the previous bin to the number of the times the counts are higher. Nemiroff (1994) studied about 40 bright bursts, and confirmed with high statistical significance the result found by Link, Epstein, & Priedhorsky (1993). Mitrofanov (1994) constructed an average light curve from 260 bursts that showed a quickly rising flux followed by a slower decay. The results of these studies exclude models that predict time-symmetric GRB light curves.
Inasmuch as temporal asymmetry is independent of burst duration and intensity, it is a useful statistic to compare with other burst properties. The purpose of this paper is to study temporal asymmetry for a sample of bursts encompassing a large range in peak flux, and to test for a correlation between temporal asymmetry and peak flux. The discovery of a correlation would provide evidence for evolutionary or cosmological effects, though the lack of such a correlation would not necessarily constitute evidence against the cosmological interpretation. We find that approximately two-thirds of bursts spanning a 200-fold range in peak flux exhibit the same temporal asymmetry, with no statistically significant correlation between the asymmetry and the peak flux.
In §2 we describe the analysis techniques we use in studying burst temporal asymmetry and the attributes of the data set tested. In §3 we discuss our results, and in §4 we summarize our results and their implications.
ANALYSIS
========
As a measure of the shape of GRB light curves, we define a [*time-asymmetry*]{} parameter as the third-moment of the burst time profile: $$\ta\equiv {\langle (t - \langle t \rangle)^3 \rangle
\over
\langle (t - \langle t \rangle)^2 \rangle^{3/2}},$$ where $\langle\rangle$ denotes an average over the data sample performed as: $$\langle f(t) \rangle \equiv {\sum_i (c_i - c_{th}) f (t_i)
\over \sum_i (c_i - c_{th})}.$$ Here $c_i$ is the measured number of counts in the $i$th bin, $t_i$ is the time of the $i$th bin, and $c_{th}$ is a threshold count level. The time-asymmetry parameter is calculated for a contiguous data sample including the burst peak and nearby bins in which the counts exceed $c_{th}$. We define the threshold as $$c_{th} \equiv f (c_p - b) + b,$$ where $c_p$ is the peak count rate, $b$ is the background, and $f(<1)$ is a fraction that will be fixed for each data set. For a given $f$, this definition of the threshold ensures that $\ta$ is calculated to the same fraction of the peak flux, relative to the background, for each burst in the data sample. Larger values of $f$ emphasize the structure of the peak over that of the surrounding foothills. The normalization of $\ta$ has been chosen in such a way as to make it independent of burst amplitude, duration and background. For a time-symmetric burst peak, $\ta=0$, and $\ta>0$ ($<0$) for a burst whose flux rises (falls) more quickly than it falls (rises). For an infinitely fast rise followed by an exponential decay, $\ta=2$, independent of the time-constant of the decay.
RESULTS
=======
In Fig. 1 we show the burst time-asymmetry parameter $\ta$ as a function of peak flux for 631 bursts from the BATSE 3B catalog, selected as described below. Our sample contains faint bursts as well as bright ones, spanning a 200-fold range in peak flux. We use the BATSE PREB plus DISC data types at 64 ms time resolution, with the four energy channels 25-50 keV, 50-100 keV, 100-300 keV, and $>300$ keV combined to attain the best statistics. For each burst, $\ta$ was calculated for a contiguous sample of data containing the highest burst peak and for which each bin satisfies $c_i\ge c_{th}$. The only requirement for a burst to be tested is that the data sample satisfying $c_i\ge c_{th}$ contain at least three bins.
The error bars in Fig. 1 represent one-$\sigma$ deviations from the calculated time asymmetry due to photon counting statistics; they were obtained by randomizing the measured counts according to Poisson statistics and calculating the variance of the time-asymmetry parameter for many trials. A preponderance of positive time asymmetries is apparent in Fig. 1 for all peak fluxes; about two-thirds of the bursts have positive time asymmetry. In Table 1 we show the weighted average of the time-asymmetry parameter for different values of $f$. The small values of the standard deviation in Table 1 show that the positive time asymmetry is not an artifact of Poisson noise. To estimate the statistical significance of the positive time asymmetry, we consider the probability that the observed fraction of positive $\ta$ bursts occurs by chance. For example, for our largest sample in Table 1 containing 631 bursts, 68% of the bursts have positive temporal asymmetry. If $\ta$ is a random variable of zero mean, the probability of such a high percentage of $\ta>0$ bursts occurring by chance is $\lap 0.001$%.
While most bursts have positive time asymmetry, there are numerous counterexamples. Examples of bursts with positive and negative $\ta$ are shown in Fig. 2. Quite often the $\ta>0$ events have a simple structure of rapid rise followed by slow decline, while the $\ta<0$ events have multiple peaks, as illustrated in the figure. This result is consistent with the conclusion of Bhat (1994) that bursts with more complex structure are less frequent.
In Fig. 1 it appears that time asymmetry increases with peak flux. Is this apparent trend statistically significant? In Table 2 we show the number-averages of $\ta$ computed for the bright and dim halves of the sample. For our largest data set ($f=0.1$), $\ta_{\rm bright}/\ta_{\rm
dim}$ is $\sim 2$; for larger values of $f$, the ratio is larger. However, because noise has zero average time asymmetry, we expect fainter events, which have lower signal to noise than bright events, to exhibit $\ta$ values closer to zero. To estimate the significances of the $\ta_{\rm bright}/\ta_{\rm dim}$ values found, we studied how the $\ta$ values of the bright half of bursts change upon degrading them to fainter peak fluxes. With each burst in the bright half of the sample, we identified a peak flux selected at random from the dim half. Each bright burst was then degraded by reducing the counts in each bin by a factor of the peak count rate of the selected dim burst to the peak count rate of the bright burst, $c_{p,{\rm dim}}/c_{p,{\rm
bright}}$. To each bin in this simulated dim burst, we added a Poisson-deviate with a mean of the new number of counts. In this way we produced simulated dim bursts with the same intrinsic temporal asymmetry as the bright bursts. To estimate the significances of the ratios $\ta_{\rm bright}/\ta_{\rm dim}$ in Table 2, we produced numerous simulated data sets, calculated the ratio for each simulation, and determined the frequency with which the simulated $\ta_{\rm bright}/\ta_{\rm dim}$ exceeded the value from the real data set. The simulated value of $\ta_{\rm bright}/\ta_{\rm dim}$ exceeded the value measured from the real data set in 39% to 56% of the simulations, depending on $f$. We conclude that the apparent trend of $\ta$ with brightness in the BATSE 3B catalog is not statistically significant.[^1]
SUMMARY AND CONCLUSIONS
=======================
We have applied a simple measure of burst temporal asymmetry to a large, uniform sample of bright and dim bursts. About two-thirds of GRBs have time-asymmetric peaks in the sense that the flux rises more rapidly than it falls, confirming the results of previous analyses of bright bursts ([@LEP]; [@timeasymmetry]; [@mitrofanov]). We conservatively estimate the significance of the preferred time asymmetry at over 99.999%, thus excluding GRB models that predict time-symmetric light curves. We find that the preferred time asymmetry shows no significant dependence on peak flux. This result is consistent with both cosmological and local interpretations of the GRB phenomenon.
It is a pleasure to thank E. E. Fenimore and J. S. Bloom for providing us with data from the BATSE 3B catalog in a form convenient for this analysis. This work was performed under the auspices of the U.S. Department of Energy, and was supported in part by NASA EPSCoR grant \#291471.
[cccccc]{} 0.1 & 631 & 68% & 0.14 & 0.00041 & 99.9993 0.2 & 603 & 66% & 0.092 & 0.00048 & 99.99 0.5 & 463 & 62% & 0.086 & 0.0013 & 99.1 0.67& 350 & 59% & 0.12 & 0.0024 & 91
[ccc]{} 0.1 & 1.7 & 56% 0.2 & 2.0 & 39% 0.5 & 2.7 & $\sim 50$%
Barat, C., Hayles, R. I., Hurley, K., Niel, M., Vedrenne, G., Estulin, I. V., & Zenchenko, V. M. 1984 , 285, 791
Bhat, P. N., Fishman, G. J., Meegan, C. A., Wilson, R. B., & Paciesas, W. S. 1994, Proceedings of the 17th Texas Symposium on Relativistic Astrophysics, p. 197
Briggs, M. S., Paciesas, W. S., Pendleton, G. N., Meegan, C. A., Fishman, G. J., Horack, J. M., Brock, M., Kouveliotou, C., Hartman, D. H., & Hakkila, J. 1996, , March 1 issue.
Davis, S. P., Norris, J. P., Kouveliotou, C., Fishman, G. J., Meegan, C. A., Pacesias, W. C. 1994, AIP Conf. Proc. 307, [*Gamma Ray Bursts*]{}, ed. G. J. Fishman, J. J. Brainerd, & K. C. Hurley (New York: AIP), p. 182
Fenimore, E. E. & Bloom, J. S., , in press
Fishman, G. J. 1994, , 92, 229
Klebesadel, R., Strong, I. B., & Olson, R. A. 1973, , 182, L85
Link, B., Epstein, R. I., & Priedhorsky, W. C.1993, , 408, L81
Meegan, C. A., Fishman, G. J., & Wilson, R. B. 1992, Nature, 355, 143
Mitrofanov, I. G., Chernenko, A. M., Pozanenko, A. S., Paciesas, W. S., Kouveliotou, C., Meegan, C. A., Fishman, G. J., & Sagdeev, R. Z. 1994, AIP Conf. Proc. 307, [*Gamma Ray Bursts*]{}, ed. G. J. Fishman, J. J. Brainerd, & K. C. Hurley (New York: AIP), p. 187
Nemiroff, R. J. Norris, J. P., Kouveliotou, C., Fishman, G. J., Meegan, C. A., Paciesas, W. S. 1994, , 423, 432
Norris, J. P. 1994, , 424, 540
Norris, J. P. 1994, AIP Conf. Proc. 307, [*Gamma Ray Bursts*]{}, ed. G. J. Fishman, J. J. Brainerd, & K. C. Hurley (New York: AIP), p. 177
Norris, J. P., Davis, S. P., Kouveliotou, C., Fishman, G. J., Meegan, C. A., Wilson, R. B., & Paciesas, W. S. 1993, AIP Conf. Proc. 280, [*Compton Gamma Ray Observatory*]{}, ed. M. Friedlander, N. Gehrels, & D. J. Macomb, p. 959
Weisskopf, M. C., Sutherland, P. G., Katz, J. I., & Canizares, C. R. 1978, , 223, L17
[^1]: Computing weighted averages $\aveta$ in the bright and dim halves of the data set also shows no statistically significant dependence of $\ta$ on peak flux.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We establish the existence of an integer degree for the natural projection map from the space of parameterizations of asymptotically conical self-expanders to the space of parameterizations of the asymptotic cones when this map is proper. As an application we show that there is an open set in the space of cones in $\mathbb{R}^3$ for which each cone in the set has a strictly unstable self-expanding annuli asymptotic to it.'
address:
- 'Department of Mathematics, Johns Hopkins University, 3400 N. Charles Street, Baltimore, MD 21218'
- 'Department of Mathematics, University of Wisconsin-Madison, 480 Lincoln Drive, Madison, WI 53706'
author:
- Jacob Bernstein
- Lu Wang
title: 'An integer degree for asymptotically conical self-expanders'
---
Introduction
============
A *hypersurface*, i.e., a properly embedded codimension-one submanifold, $\Sigma\subset\mathbb{R}^{n+1}$, is a *self-expander* if $$\label{ExpanderEqn}
\mathbf{H}_\Sigma-\frac{\mathbf{x}^\perp}{2}=\mathbf{0}.$$ Here $$\mathbf{H}_\Sigma=\Delta_\Sigma\mathbf{x}=-H_\Sigma\mathbf{n}_\Sigma=-\mathrm{div}_\Sigma(\mathbf{n}_\Sigma)\mathbf{n}_\Sigma$$ is the mean curvature vector, $\mathbf{n}_\Sigma$ is the unit normal, and $\mathbf{x}^\perp$ is the normal component of the position vector. Self-expanders arise naturally in the study of mean curvature flow. Indeed, $\Sigma$ is a self-expander if and only if the family of homothetic hypersurfaces $$\left\{\Sigma_t\right\}_{t>0}=\left\{\sqrt{t}\, \Sigma\right\}_{t>0}$$ is a *mean curvature flow* (MCF), that is, a solution to the flow $$\left(\frac{\partial \mathbf{x}}{\partial t}\right)^\perp=\mathbf{H}_{\Sigma_t}.$$ Self-expanders are expected to model the behavior of a MCF as it emerges from a conical singularity [@AIC]. They are also expected to model the long time behavior of the flow [@EHAnn].
Given an integer $k\geq 2$ and $\alpha\in (0,1)$ let $\Gamma$ be a $C^{k,\alpha}_{*}$-asymptotically conical $C^{k,\alpha}$-hypersurface in $\mathbb{R}^{n+1}$ and let $\mathcal{L}(\Gamma)$ be the link of the asymptotic cone of $\Gamma$. For instance, if $\lim_{\rho\to 0^+} \rho \Gamma=\mathcal{C}$ in $C^{k, \alpha}_{loc} (\mathbb{R}^{n+1}\setminus\{\mathbf{0}\})$, where $\mathcal{C}$ is a cone, then $\Gamma$ is $C^{k,\alpha}_{*}$-asymptotically conical with asymptotic cone $\mathcal{C}$. For technical reasons, the actual definition is slightly weaker – see Section 3 of [@BWBanachManifold] for the details. We denote the space of $C^{k,\alpha}_{*}$-asymptotically conical $C^{k,\alpha}$-hypersurfaces in $\mathbb{R}^{n+1}$ by $\mathcal{ACH}^{k,\alpha}_n$.
In [@BWBanachManifold], the authors showed that the space $\mathcal{ACE}_n^{k,\alpha}(\Gamma)$ – see below – of asymptotically conical parameterizations of self-expanders modeled on $\Gamma$ (modulo reparameterizations fixing the parameterization of the asymptotic cone) possesses a natural Banach manifold structure modeled on $C^{k, \alpha}(\mathcal{L}(\Gamma); \mathbb{R}^{n+1})$. The authors further showed that the map $$\Pi\colon \mathcal{ACE}_n^{k,\alpha}(\Gamma)\to C^{k, \alpha}(\mathcal{L}(\Gamma); \mathbb{R}^{n+1})$$ given by $\Pi([\mathbf{f}])=\mathrm{tr}^1_{\infty}[\mathbf{f}]$ is smooth and Fredholm of index $0$. Here $\mathrm{tr}^1_\infty[\mathbf{f}]$ is the trace at infinity of $\mathbf{f}$ – see . As such, by work of Smale [@Smale], as long as $\Pi$ is proper it possesses a well-defined mod 2 degree – see [@BWProperness] for natural situations in which $\Pi$ is proper. In this paper, we show that as long as $\Pi$ is proper it has a well-defined *integer* degree. Namely, let $$\mathcal{V}_{\mathrm{emb}}^{k,\alpha}(\Gamma)= {\left\{\varphi\in C^{k,\alpha}(\mathcal{L}(\Gamma);\mathbb{R}^{n+1}) \colon \mbox{$\mathscr{E}^{\mathrm{H}}_1[\varphi]$ is an embedding}\right\}},$$ be the space of parameterizations of embedded cones. Here $\mathscr{E}^\mathrm{H}_1[\varphi]$ is the homogeneous degree-one extension of $\varphi$ – see . For any $\varphi\in \mathcal{V}_{\mathrm{emb}}^{k,\alpha}(\Gamma)$ a regular value of $\Pi$ so that $\Pi^{-1}(\varphi)$ is a finite set let $$\mathrm{deg}(\Pi, \varphi)= \sum_{[\mathbf{f}]\in \Pi^{-1}(\varphi)} (-1)^{\mathrm{ind}([\mathbf{f}])}$$ where $\mathrm{ind}([\mathbf{f}])$ is the *Morse index* of $[\mathbf{f}]$ – see Section \[IndexSec\]. We claim that in many situations the map $\mathrm{deg}$ depends only on the component of $\varphi$ in $\mathcal{V}_{\mathrm{emb}}^{k,\alpha}(\Gamma)$. More precisely,
\[MainThm\] Given $\Gamma\in\mathcal{ACH}^{k,\alpha}_n$, let $\mathcal{V}\subset \mathcal{V}_{\mathrm{emb}}^{k,\alpha}(\Gamma)$ be a connected open set and let $\mathcal{U}\subset \mathcal{ACE}_n^{k,\alpha}(\Gamma)$ be open and so $\Pi|_{\mathcal{U}}\colon \mathcal{U}\to \mathcal{V}$ is proper. Then for $\varphi\in\mathcal{V}$ which is a regular value of $\Pi|_{\mathcal{U}}$, $$\mathrm{deg}(\Pi|_{\mathcal{U}}, \varphi)=\sum_{[\mathbf{f}]\in \mathcal{U}\cap \Pi^{-1}(\varphi) }(-1)^{\mathrm{ind}([\mathbf{f}])}$$ is independent of $\varphi$ – i.e., $\Pi|_{\mathcal{U}}$ has a well-defined integer degree $\mathrm{deg}(\Pi|_\mathcal{U},\mathcal{V})$.
This is analogous to a result proved by White [@WhiteEI] for a large class of compact critical points of elliptic variational problems. While the argument in this article follows the outline laid out in [@WhiteEI], the details are quite different and are substantially more involved. This is due to technical issues introduced both by the non-compactness of the problem and, more seriously, by the fast growth of the weight.
As an application of Theorem \[MainThm\] we obtain the existence of many strictly unstable asymptotically conical self-expanders.
\[ApplicationThm\] Let $\Gamma\in\mathcal{ACH}^{k,\alpha}_2$ be an annulus. There is an open set $\mathcal{V}_0\subset\mathcal{V}_{\mathrm{emb}}^{k,\alpha}(\Gamma)$ so that for each $\varphi\in\mathcal{V}_0$ there is an element $[\mathbf{f}]\in\mathcal{ACE}^{k,\alpha}_2(\Gamma)$ with $\mathrm{tr}_\infty^1[\mathbf{f}]=\varphi$ and so that $\mathbf{f}(\Gamma)$ is a strictly unstable self-expander.
It is necessary to invoke the integer degree for the map $\Pi$, instead of the mod $2$ degree, in order to conclude that the self-expanders in Theorem \[ApplicationThm\] are strictly unstable.
In [@AIC], Angenent-Ilmanen-Chopp show that there is a critical value $\delta_*$ so that for $0<\delta<\delta_*$ there are no connected rotationally symmetric self-expanders asymptotic to the rotationally symmetric double cone $$\mathcal{C}_{\delta}={\left\{x_1^2+\cdots +x_n^2= \delta^2 x_{n+1}^2\right\}},$$ while for $\delta>\delta_*$ there is at least one connected rotationally symmetric self-expanders asymptotic to $\mathcal{C}_\delta$. In [@Helmensdorfer], Helmensdorfer further analyzed this problem and showed that, for $\delta>\delta_*$, there is a second connected rotationally symmetric self-expander asymptotic to $\mathcal{C}_\delta$. It is likely that the solution constructed by Angenent-Ilmanen-Chopp is strictly stable while the one found by Helmensdorfer is strictly unstable (and there is a unique weakly stable example asymptotic to $\mathcal{C}_{\delta_*}$), but, to our knowledge, a complete proof has not appeared in the literature.
Notation and background {#NotationSec}
=======================
Basic notions {#NotionSubsec}
-------------
Denote a (open) ball in $\mathbb{R}^n$ of radius $R$ and center $x$ by $B_R^n(x)$ and the closed ball by $\bar{B}^n_R(x)$. We often omit the superscript $n$ when its value is clear from context. We also omit the center when it is the origin.
For an open set $U\subset\mathbb{R}^{n+1}$, a *hypersurface in $U$*, $\Sigma$, is a smooth, properly embedded, codimension-one submanifold of $U$. We also consider hypersurfaces of lower regularity and given an integer $k\geq 2$ and $\alpha\in (0,1)$ we define a *$C^{k,\alpha}$-hypersurface in $U$* to be a properly embedded, codimension-one $C^{k,\alpha}$ submanifold of $U$. When needed, we distinguish between a point $p\in\Sigma$ and its *position vector* $\mathbf{x}(p)$.
Consider the hypersurface $\mathbb{S}^n\subset\mathbb{R}^{n+1}$, the unit $n$-sphere in $\mathbb{R}^{n+1}$. A *hypersurface in $\mathbb{S}^n$*, $\sigma$, is a closed, embedded, codimension-one smooth submanifold of $\mathbb{S}^n$ and *$C^{k,\alpha}$-hypersurfaces in $\mathbb{S}^n$* are defined likewise. Observe that $\sigma$ is a closed codimension-two submanifold of $\mathbb{R}^{n+1}$ and so we may associate to each point $p\in\sigma$ its position vector $\mathbf{x}(p)$. Clearly, $|\mathbf{x}(p)|=1$.
A *cone* is a set $\mathcal{C}\subset\mathbb{R}^{n+1}\setminus\{\mathbf{0}\}$ that is dilation invariant around the origin. That is, $\rho\mathcal{C}=\mathcal{C}$ for all $\rho>0$. The *link* of the cone is the set $\mathcal{L}[\mathcal{C}]=\mathcal{C}\cap\mathbb{S}^{n}$. The cone is *regular* if its link is a smooth hypersurface in $\mathbb{S}^{n}$ and *$C^{k,\alpha}$-regular* if its link is a $C^{k,\alpha}$-hypersurface in $\mathbb{S}^n$. For any hypersurface $\sigma\subset\mathbb{S}^n$ the *cone over $\sigma$*, $\mathcal{C}[\sigma]$, is the cone defined by $$\mathcal{C}[\sigma]=\left\{\rho p\colon p\in\sigma, \rho>0\right\}\subset\mathbb{R}^{n+1}\setminus\{\mathbf{0}\}.$$ Clearly, $\mathcal{L}[\mathcal{C}[\sigma]]=\sigma$.
Function spaces {#FunctionSubsec}
---------------
Let $\Sigma$ be a properly embedded, $C^{k,\alpha}$ submanifold of an open set $U\subset\mathbb{R}^{n+1}$. There is a natural Riemannian metric, $g_\Sigma$, on $\Sigma$ of class $C^{k-1,\alpha}$ induced from the Euclidean one. As we always take $k\geq 2$, the Christoffel symbols of this metric, in appropriate coordinates, are well-defined and of regularity $C^{k-2,\alpha}$. Let $\nabla_\Sigma$ be the covariant derivative on $\Sigma$. Denote by $d_\Sigma$ the geodesic distance on $\Sigma$ and by $B^\Sigma_R(p)$ the (open) geodesic ball in $\Sigma$ of radius $R$ and center $p\in\Sigma$. For $R$ small enough so that $B_{R}^\Sigma(p)$ is strictly geodesically convex and $q\in B^\Sigma_R(p)$, denote by $\tau^\Sigma_{p,q}$ the parallel transport along the unique minimizing geodesic in $B^\Sigma_R(p)$ from $p$ to $q$.
Throughout the rest of this subsection, let $\Omega$ be a domain in $\Sigma$, $l$ an integer in $[0,k]$, $\gamma\in (0,1)$ and $d\in\mathbb{R}$. Suppose $l+\gamma\leq k+\alpha$. We first consider the following norm for functions on $\Omega$: $$\Vert f\Vert_{l; \Omega}=\sum_{i=0}^l \sup_{\Omega} |\nabla_\Sigma^i f|.$$ We then let $$C^l(\Omega)=\left\{f\in C_{loc}^l(\Omega)\colon \Vert f\Vert_{l; \Omega}<\infty\right\}.$$ We next define the Hölder semi-norms for functions $f$ and tensor fields $T$ on $\Omega$: $$[f]_{\gamma; \Omega} =\sup_{\substack{p,q\in\Omega \\ q\in B^\Sigma_{\delta}(p)\setminus\{p\}}} \frac{|f(p)-f(q)|}{d_\Sigma(p,q)^\gamma}
\mbox{ and }
[T]_{\gamma; \Omega} =\sup_{\substack{p,q\in\Omega \\ q\in B^\Sigma_{\delta}(p)\setminus\{p\}}} \frac{|T(p)-(\tau^\Sigma_{p,q})^* T(q)|}{d_\Sigma(p,q)^\gamma},$$ where $\delta=\delta(\Sigma,\Omega)>0$ so that for all $p\in\Omega$, $B^\Sigma_\delta(p)$ is strictly geodesically convex. We further define the norm for functions on $\Omega$: $$\Vert f\Vert_{l, \gamma; \Omega}=\Vert f\Vert_{l; \Omega}+[\nabla_\Sigma^l f]_{\gamma; \Omega},$$ and let $$C^{l, \gamma}(\Omega)=\left\{f\in C_{loc}^{l, \gamma}(\Omega)\colon \Vert f\Vert_{l, \gamma; \Omega}<\infty\right\}.$$
We also define the following weighted norm for functions on $\Omega$: $$\Vert f\Vert_{l; \Omega}^{(d)}=\sum_{i=0}^l\sup_{p\in\Omega} \left(|\mathbf{x}(p)|+1\right)^{-d+i} |\nabla_\Sigma^i f(p)|.$$ We then let $$C^{l}_d(\Omega)=\left\{f\in C^l_{loc}(\Omega)\colon \Vert f\Vert_{l; \Omega}^{(d)}<\infty\right\}.$$ We further define the following weighted Hölder semi-norms for functions $f$ and tensor fields $T$ on $\Omega$: $$\begin{aligned}
[f]_{\gamma; \Omega}^{(d)} & =\sup_{\substack{p,q\in\Omega \\ q\in B^\Sigma_{\delta_p}(p)\setminus\{p\}}} \left((|\mathbf{x}(p)|+1)^{-d+\gamma}+(|\mathbf{x}(q)|+1)^{-d+\gamma}\right) \frac{|f(p)-f(q)|}{d_\Sigma(p,q)^\gamma}, \mbox{ and}, \\
[T]_{\gamma; \Omega}^{(d)} & =\sup_{\substack{p,q\in\Omega \\ q\in B^\Sigma_{\delta_p}(p)\setminus\{p\}}} \left((|\mathbf{x}(p)|+1)^{-d+\gamma}+(|\mathbf{x}(q)|+1)^{-d+\gamma}\right) \frac{|T(p)-(\tau^\Sigma_{p,q})^* T(q)|}{d_\Sigma(p,q)^\gamma},\end{aligned}$$ where $\eta=\eta(\Omega,\Sigma)\in \left(0,\frac{1}{4}\right)$ so that for any $p\in\Sigma$, letting $\delta_p=\eta (|\mathbf{x}(p)|+1)$, $B_{\delta_p}^\Sigma(p)$ is strictly geodesically convex. Next we define the norm for functions on $\Omega$: $$\Vert f\Vert_{l, \gamma; \Omega}^{(d)}=\Vert f\Vert_{l; \Omega}^{(d)}+[\nabla_\Sigma^l f]_{\gamma; \Omega}^{(d-l)},$$ and we let $$C^{l,\gamma}_d(\Omega)=\left\{f\in C^{l,\gamma}_{loc}(\Omega)\colon \Vert f\Vert_{l, \gamma; \Omega}^{(d)}<\infty\right\}.$$ We follow the convention that $C^{l,0}_{loc}=C^{l}_{loc}$, $C^{l,0}=C^l$ and $C^{l,0}_d=C^l_d$ and that $C^{0, \gamma}_{loc}=C^\gamma_{loc}$, $C^{0,\gamma}=C^\gamma$ and $C^{0,\gamma}_d=C^\gamma_d$. The notation for the corresponding norms is abbreviated in the same fashion.
Finally, we define the following weighted integral norm for functions on $\Omega$: $$\Vert f\Vert^{W, (d)}_{l;\Omega}=\left(\int_{\Omega} \sum_{i=0}^l |\nabla^i_\Sigma f|^2 e^{d|\mathbf{x}|^2}\, d\mathcal{H}^n\right)^{\frac{1}{2}},$$ and we then let $$W^{l}_{d}(\Omega)={\left\{f\in W^{l,2}_{loc}(\Omega)\colon\Vert f\Vert^{W, (d)}_{l;\Omega}<\infty\right\}}.$$
In all above definitions of various norms, we often omit the domain $\Omega$ when it is clear from context. These norms can be extended in a straightforward manner to vector-valued functions and tensor fields. It is a standard exercise to verify that these spaces equipped with the corresponding norms are Banach spaces.
Homogeneous functions and homogeneity at infinity {#HomoFuncSubsec}
-------------------------------------------------
Fix a $C^{k,\alpha}$-regular cone $\mathcal{C}$ with its link $\mathcal{L}$. By our definition, $\mathcal{C}$ is a $C^{k,\alpha}$-hypersurface in $\mathbb{R}^{n+1}\setminus\{\mathbf{0}\}$. For $R>0$, let $\mathcal{C}_R=\mathcal{C}\setminus\bar{B}_R$. There is an $\eta=\eta(\mathcal{L},R)>0$ so that for any $p\in\mathcal{C}_R$, $B^{\mathcal{C}}_{\delta_p}(p)$ is strictly geodesically convex, where $\delta_p=\eta(|\mathbf{x}(p)|+1)$. We also fix an integer $l\in [0,k]$ and $\gamma\in [0,1)$ with $l+\gamma\leq k+\alpha$.
A map $\mathbf{f}\in C^{l,\gamma}_{loc}(\mathcal{C}; \mathbb{R}^M)$ is *homogeneous of degree $d$* if $\mathbf{f}(\rho p)=\rho^d \mathbf{f}(p)$ for all $p\in\mathcal{C}$ and $\rho>0$. Given a map $\varphi\in C^{l,\gamma}(\mathcal{L}; \mathbb{R}^M)$ the *homogeneous extension of degree $d$* of $\varphi$ is the map $\mathscr{E}_d^{\mathrm{H}}[\varphi]\in C^{l,\gamma}_{loc}(\mathcal{C}; \mathbb{R}^M)$ defined by $$\label{HomoExtEqn}
\mathscr{E}_d^{\mathrm{H}}[\varphi](p)=|\mathbf{x}(p)|^d \varphi(|\mathbf{x}(p)|^{-1}p).$$ Conversely, given a homogeneous $\mathbb{R}^M$-valued map of degree $d$, $\mathbf{f}\in C^{l,\gamma}_{loc}(\mathcal{C}; \mathbb{R}^M)$, let $\varphi=\mathrm{tr}[\mathbf{f}]\in C^{l,\beta}(\mathcal{L}; \mathbb{R}^M)$, the *trace* of $\mathbf{f}$, be the restriction of $\mathbf{f}$ to $\mathcal{L}$. Clearly, $\mathbf{f}$ is the homogeneous extension of degree $d$ of $\varphi$.
A map $\mathbf{g}\in C^{l,\gamma}_{loc}(\mathcal{C}_R; \mathbb{R}^M)$ is *asymptotically homogeneous of degree $d$* if $$\lim_{\rho\to 0^+} \rho^d \mathbf{g}(\rho^{-1}p)=\mathbf{f}(p) \mbox{ in $C^{l,\gamma}_{loc}(\mathcal{C}; \mathbb{R}^M)$}$$ for some $\mathbf{f}\in C^{l,\gamma}_{loc}(\mathcal{C}; \mathbb{R}^M)$ that is homogeneous of degree $d$. For such a $\mathbf{g}$ we define the *trace at infinity* of $\mathbf{g}$ by $\mathrm{tr}^d_\infty[\mathbf{g}]=\mathrm{tr}[\mathbf{f}]$. We define $$C^{l,\gamma}_{d,\mathrm{H}}(\mathcal{C}_R; \mathbb{R}^M)=\left\{\mathbf{g}\in C^{l,\gamma}_d(\mathcal{C}_R; \mathbb{R}^M)\colon \mbox{$\mathbf{g}$ is asymptotically homogeneous of degree $d$}\right\}.$$ It is straightforward to verify that $C^{l,\gamma}_{d,\mathrm{H}}(\mathcal{C}_R; \mathbb{R}^M)$ is a closed subspace of $C^{l,\gamma}_d(\mathcal{C}_R; \mathbb{R}^M)$ and that $$\mathrm{tr}^d_\infty\colon C^{l,\gamma}_{d,\mathrm{H}}(\mathcal{C}_R; \mathbb{R}^M)\to C^{l,\gamma}(\mathcal{L}; \mathbb{R}^M)$$ is a bounded linear map. Finally, $\mathbf{x}|_{\mathcal{C}_R}\in C^{k,\alpha}_{1,\mathrm{H}}(\mathcal{C}_R; \mathbb{R}^{n+1})$ and $\mathrm{tr}_\infty^1[\mathbf{x}|_{\mathcal{C}_R}]=\mathbf{x}|_{\mathcal{L}}$.
Asymptotically conical hypersurfaces {#ACHSubsec}
------------------------------------
A $C^{k,\alpha}$-hypersurface, $\Sigma\subset\mathbb{R}^{n+1}$, is *$C^{k,\alpha}_{*}$-asymptotically conical* if there is a $C^{k,\alpha}$-regular cone, $\mathcal{C}\subset\mathbb{R}^{n+1}$, and a homogeneous (of degree $0$) transverse section, $\mathbf{V}$, on $\mathcal{C}$ such that $\Sigma$, outside some compact set, is given by the $\mathbf{V}$-graph of a function in $C^{k,\alpha}_1\cap C^{k}_{1,0}(\mathcal{C}_R)$ for some $R>1$. Here a transverse section is a regularized version of the unit normal – see Section 2.4 of [@BWBanachManifold] for the precise definition. Observe, that by the Arzelà-Ascoli theorem one has that, for every $l,\gamma$ with $l+\gamma<k+\alpha$, $$\lim_{\rho\to 0^+} \rho\Sigma = \mathcal{C} \mbox{ in } C^{l, \gamma}_{loc}(\mathbb{R}^{n+1}\setminus \{\mathbf{0}\}).$$ Clearly, the asymptotic cone, $\mathcal{C}$, is uniquely determined by $\Sigma$ and so we denote it by $\mathcal{C}(\Sigma)$. Let $\mathcal{L}(\Sigma)$ denote the link of $\mathcal{C}(\Sigma)$ and, for $R>0$, let $\mathcal{C}_R(\Sigma)=\mathcal{C}(\Sigma)\setminus \bar{B}_R$. Denote the space of $C^{k,\alpha}_{*}$-asymptotically conical $C^{k,\alpha}$-hypersurfaces in $\mathbb{R}^{n+1}$ by $\mathcal{ACH}^{k,\alpha}_n$.
Finally, let $K$ be a compact subset of $\Sigma$ and let $\Sigma^\prime=\Sigma\setminus K$. By definition, we may choose $K$ large enough so $\pi_{\mathbf{V}}$ – the projection of a neighborhood of $\mathcal{C}(\Sigma)$ along $\mathbf{V}$ – restricts to a $C^{k,\alpha}$ diffeomorphism of $\Sigma^\prime$ onto $\mathcal{C}_R(\Sigma)$. Denote its inverse by $\theta_{\mathbf{V}; \Sigma^\prime}$.
Traces at infinity {#TraceSubsec}
------------------
Fix an element $\Sigma\in\mathcal{ACH}_n^{k,\alpha}$. Let $l$ be an integer in $[0,k]$ and $\gamma\in [0,1)$ such that $l+\gamma<k+\alpha$. A map $\mathbf{f}\in C_{loc}^{l,\gamma}(\Sigma; \mathbb{R}^M)$ is *asymptotically homogeneous of degree $d$* if $\mathbf{f}\circ\theta_{\mathbf{V}; \Sigma^\prime}\in C^{l,\gamma}_{d,\mathrm{H}}(\mathcal{C}_R(\Sigma); \mathbb{R}^M)$ where $\mathbf{V}$ is a homogeneous transverse section on $\mathcal{C}(\Sigma)$ and $\Sigma^\prime, \theta_{\mathbf{V}; \Sigma^\prime}$ are introduced in the previous subsection. The *trace at infinity* of $\mathbf{f}$ is then $$\label{TraceEqn}
\mathrm{tr}_\infty^d[\mathbf{f}]=\mathrm{tr}_\infty^d[\mathbf{f}\circ\theta_{\mathbf{V}; \Sigma^\prime}] \in C^{l,\gamma}(\mathcal{L}(\Sigma); \mathbb{R}^M).$$ Whether $\mathbf{f}$ is asymptotically homogeneous of degree $d$ and the definition of $\mathrm{tr}_{\infty}^d$ are independent of the choice of homogeneous transverse sections on $\mathcal{C}(\Sigma)$. Clearly, $\mathbf{x}|_{\Sigma}$ is asymptotically homogeneous of degree $1$ and $\mathrm{tr}_\infty^{1}[\mathbf{x}|_\Sigma]=\mathbf{x}|_{\mathcal{L}(\Sigma)}$.
We next define the space $$C_{d,\mathrm{H}}^{l,\gamma}(\Sigma; \mathbb{R}^M)=\left\{\mathbf{f}\in C_{d}^{l,\gamma}(\Sigma; \mathbb{R}^M)\colon \mbox{$\mathbf{f}$ is asymptotically homogeneous of degree $d$}\right\}.$$ One can check that $C_{d,\mathrm{H}}^{l,\gamma}(\Sigma; \mathbb{R}^M)$ is a closed subspace of $C_d^{l,\gamma}(\Sigma; \mathbb{R}^M)$, and the map $$\mathrm{tr}_{\infty}^d\colon C_{d,\mathrm{H}}^{l,\gamma}(\Sigma; \mathbb{R}^M)\to C^{l,\gamma}(\mathcal{L}(\Sigma); \mathbb{R}^M)$$ is a bounded linear map. We further define the set $C^{l,\gamma}_{d,0}(\Sigma;\mathbb{R}^M)\subset C_{d,\mathrm{H}}^{l,\gamma}(\Sigma; \mathbb{R}^M)$ to be the kernel of $\mathrm{tr}_\infty^d$.
Asymptotically conical embeddings {#ACESubsec}
---------------------------------
Fix an element $\Gamma\in \mathcal{ACH}^{k,\alpha}_n$. We define the space of $C^{k,\alpha}_{*}$-asymptotically conical embeddings of $\Gamma$ into $\mathbb{R}^{n+1}$ to be $$\mathcal{ACH}^{k,\alpha}_n(\Gamma)=\left\{\mathbf{f}\in C_{1}^{k,\alpha}\cap C^k_{1,\mathrm{H}}(\Gamma; \mathbb{R}^{n+1})\colon\mbox{$\mathbf{f}$ and $\mathscr{E}_{1}^{\mathrm{H}}\circ\mathrm{tr}_\infty^1[\mathbf{f}]$ are embeddings}\right\}.$$ Clearly, $\mathcal{ACH}^{k,\alpha}_n(\Gamma)$ is an open set in the Banach space $C^{k,\alpha}_{1}\cap C^{k}_{1,\mathrm{H}}(\Gamma; \mathbb{R}^{n+1})$ with the $\Vert\cdot \Vert_{k,\alpha}^{(1)}$ norm. The hypotheses on $\mathbf{f}$, $\mathrm{tr}_\infty^1[\mathbf{f}]\in C^{k,\alpha}(\mathcal{L}(\Gamma);\mathbb{R}^{n+1})$ ensure $$\mathcal{C}[\mathbf{f}]=\mathscr{E}_{1}^{\mathrm{H}}\circ\mathrm{tr}_\infty^1[\mathbf{f}]\colon \mathcal{C}(\Gamma)\to \mathbb{R}^{n+1}\setminus\{\mathbf{0}\}$$ is a $C^{k,\alpha}$ embedding. As this map is homogeneous of degree one, it parameterizes the $C^{k,\alpha}$-regular cone $\mathcal{C}(\mathbf{f}(\Gamma))$ – see [@BWBanachManifold Proposition 3.3].
Finally, we introduce a natural equivalence relation on $\mathcal{ACH}_{n}^{k,\alpha}(\Gamma)$. First, say a $C^{k,\alpha}$ diffeomorphism $\phi\colon\Gamma\to \Gamma$ *fixes infinity* if $\mathbf{x}|_{\Gamma}\circ \phi\in \mathcal{ACH}^{k,\alpha}_n(\Gamma)$ and $$\mathrm{tr}^1_{\infty}[\mathbf{x}|_{\Gamma}\circ \phi]=\mathbf{x}|_{\mathcal{L}(\Gamma)}.$$ Two elements $\mathbf{f}, \mathbf{g}\in\mathcal{ACH}_{n}^{k,\alpha}(\Gamma)$ are equivalent, written $\mathbf{f}\sim\mathbf{g}$, provided there is a $C^{k,\alpha}$ diffeomorphism $\phi\colon\Gamma\to\Gamma$ that fixes infinity so that $\mathbf{f}\circ\phi=\mathbf{g}$. Given $\mathbf{f}\in \mathcal{ACH}_{n}^{k,\alpha}(\Gamma)$ let $[\mathbf{f}]$ be the equivalence class of $\mathbf{f}$. Following [@BWBanachManifold], we define the space $$\label{ACEDef}
\mathcal{ACE}_n^{k,\alpha}(\Gamma)={\left\{[\mathbf{f}]\colon \mathbf{f}\in \mathcal{ACH}^{k,\alpha}_n(\Gamma) \mbox{ and $\mathbf{f}(\Gamma)$ satisfies \eqref{ExpanderEqn}}\right\}}.$$
Various symmetric bilinear forms {#BilinearFormSubsec}
--------------------------------
Let $\Sigma\subset\mathbb{R}^{n+1}$ be a $C^2$-hypersurface and $\mathbf{v}$ a transverse section on $\Sigma$. Suppose $\mathbf{f}\colon\Sigma\to\mathbb{R}^{n+1}$ is a $C^2$ proper embedding. We set $\Lambda=\mathbf{f}(\Sigma)$ and $\mathbf{w}=\mathbf{v}\circ\mathbf{f}^{-1}$. We also assume $\mathbf{w}$ is transverse to $\Lambda$. If $g_{\mathbf{f}}$ is the pullback metric of the Euclidean one via $\mathbf{f}$, then we let $$\Omega_\mathbf{f}=\sqrt{\frac{\det g_{\mathbf{f}}}{\det g_\Sigma}}\, e^{\frac{|\mathbf{f}|^2-|\mathbf{x}|^2}{4}}$$ which is the ratio of the pullback weighted measure on $\Lambda$ and the one on $\Sigma$, and $$\Omega_{\mathbf{f},\mathbf{v}}=\left(\frac{\mathbf{v}\cdot (\mathbf{n}_\Lambda\circ\mathbf{f})}{\mathbf{v}\cdot\mathbf{n}_\Sigma}\right)^2 \Omega_\mathbf{f}.$$ Observe that $\Omega_{\mathbf{x}|_\Sigma}=\Omega_{\mathbf{x}|_\Sigma,\mathbf{v}}=1$.
First we introduce natural weighted inner products for functions on $\Sigma$. Suppose $u,v\in C_c^0 (\Sigma)$. Let $$B_\Sigma[u,v]=\int_{\Sigma} uv e^{\frac{|\mathbf{x}|^2}{4}} \, d\mathcal{H}^n$$ and $$B_{\Sigma,\mathbf{v}}[u,v]=B_\Sigma[(\mathbf{v}\cdot\mathbf{n}_\Sigma)u,(\mathbf{v}\cdot\mathbf{n}_\Sigma)v]=\int_\Sigma uv(\mathbf{v}\cdot\mathbf{n}_\Sigma)^2 e^{\frac{|\mathbf{x}|^2}{4}} \, d\mathcal{H}^n.$$ Likewise, let $$B_\mathbf{f}[u,v]=B_\Lambda\left[u\circ\mathbf{f}^{-1},v\circ\mathbf{f}^{-1}\right]=\int_\Sigma uv \Omega_\mathbf{f} e^{\frac{|\mathbf{x}|^2}{4}} \, d\mathcal{H}^n$$ and $$B_{\mathbf{f},\mathbf{v}}[u,v]=B_{\Lambda,\mathbf{w}} \left[u\circ\mathbf{f}^{-1},v\circ\mathbf{f}^{-1}\right]=\int_{\Sigma} uv(\mathbf{v}\cdot(\mathbf{n}_\Lambda\circ\mathbf{f}))^2 \Omega_\mathbf{f} e^{\frac{|\mathbf{x}|^2}{4}} \, d\mathcal{H}^n.$$ Thus, we have that $$B_{\mathbf{f}}[u,v]=B_\Sigma[\Omega_\mathbf{f} u, v] \mbox{ and } B_{\mathbf{f},\mathbf{v}}[u,v]=B_{\Sigma,\mathbf{v}}[\Omega_{\mathbf{f},\mathbf{v}} u, v].$$
Next, we introduce the following Dirichlet energy forms: for $u,v\in C^1_c(\Sigma)$, $$D_\Sigma[u,v]=\int_\Sigma \nabla_\Sigma u\cdot\nabla_\Sigma v \, e^{\frac{|\mathbf{x}|^2}{4}}\, d\mathcal{H}^n,$$ and $$D_{\Sigma,\mathbf{v}}[u,v]=D_\Sigma[(\mathbf{v}\cdot\mathbf{n}_\Sigma)u,(\mathbf{v}\cdot\mathbf{n}_\Sigma) v].$$ If $\Sigma$ is a $C^3$-hypersurface, then $$D_{\Sigma,\mathbf{v}}[u,v]=\int_\Sigma \left(\nabla_\Sigma u\cdot\nabla_\Sigma v+a_{\Sigma,\mathbf{v}}uv\right) (\mathbf{v}\cdot\mathbf{n}_\Sigma)^2 e^{\frac{|\mathbf{x}|^2}{4}} \, d\mathcal{H}^n$$ where $$a_{\Sigma,\mathbf{v}}=-(\mathbf{v}\cdot\mathbf{n}_\Sigma)^{-1} \left(\Delta_\Sigma+\frac{1}{2}\mathbf{x}\cdot\nabla_\Sigma\right) (\mathbf{v}\cdot\mathbf{n}_\Sigma).$$ Likewise, let $$D_\mathbf{f}[u,v]=D_{\Lambda}\left[u\circ\mathbf{f}^{-1},v\circ\mathbf{f}^{-1}\right]=\int_\Sigma (g_\mathbf{f}^{-1})^{ij}\nabla_i u\nabla_j v \, \Omega_\mathbf{f} e^{\frac{|\mathbf{x}|^2}{4}} \, d\mathcal{H}^n,$$ and $$D_{\mathbf{f},\mathbf{v}}[u,v]=D_{\Lambda,\mathbf{w}}\left[u\circ\mathbf{f}^{-1},v\circ\mathbf{f}^{-1}\right].$$ When $\Lambda$ is of class $C^3$, $$D_{\mathbf{f},\mathbf{v}}[u,v]=\int_\Sigma \left((g_\mathbf{f}^{-1})^{ij}\nabla_i u\nabla_j v+(a_{\Lambda,\mathbf{w}}\circ\mathbf{f}) uv\right) (\mathbf{v}\cdot(\mathbf{n}_\Lambda\circ\mathbf{f}))^2 \Omega_\mathbf{f} e^{\frac{|\mathbf{x}|^2}{4}} \, d\mathcal{H}^n.$$
Finally, we consider the Jacobi operator (also called the stability operator) on $\Sigma$, $$L_\Sigma=\Delta_\Sigma+\frac{1}{2}\mathbf{x}\cdot\nabla_\Sigma+|A_\Sigma|^2-\frac{1}{2}.$$ The operator $L_\Sigma$ is associated to the symmetric bilinear form $$Q_\Sigma[u,v]=\int_\Sigma \left(\nabla_\Sigma u\cdot\nabla_\Sigma v-|A_\Sigma|^2 uv+\frac{1}{2} uv\right) e^{\frac{|\mathbf{x}|^2}{4}} \, d\mathcal{H}^n$$ in the sense that for $u,v\in C^2_c(\Sigma)$, $$Q_\Sigma[u,v]=-B_\Sigma[u,L_\Sigma v].$$ In particular, $L_\Sigma$ is symmetric with respect to the inner product $B_\Sigma$.
We also consider the $\mathbf{v}$-Jacobi operator $$L_{\Sigma,\mathbf{v}}=(\mathbf{v}\cdot\mathbf{n}_\Sigma) L_\Sigma\left((\mathbf{v}\cdot\mathbf{n}_\Sigma)\times\right).$$ It will also be convenient to consider the related operator $$L_{\Sigma,\mathbf{v}}^\prime=(\mathbf{v}\cdot\mathbf{n}_\Sigma)^{-2} L_{\Sigma,\mathbf{v}}.$$ These operators are both associated to the symmetric bilinear form $$Q_{\Sigma,\mathbf{v}}[u,v] =Q_\Sigma[(\mathbf{v}\cdot\mathbf{n}_\Sigma)u,(\mathbf{v}\cdot\mathbf{n}_\Sigma)v].$$ When $\Sigma$ is of class $C^3$, $$Q_{\Sigma,\mathbf{v}}[u,v] = \int_\Sigma \left(\nabla_\Sigma u\cdot\nabla_\Sigma v+a^\prime_{\Sigma,\mathbf{v}}uv\right) (\mathbf{v}\cdot\mathbf{n}_\Sigma)^2 e^{\frac{|\mathbf{x}|^2}{4}}\, d\mathcal{H}^n$$ where $$a^\prime_{\Sigma,\mathbf{v}}=-(\mathbf{v}\cdot\mathbf{n}_\Sigma)^{-1} L_\Sigma(\mathbf{v}\cdot\mathbf{n}_\Sigma).$$ Specifically one has, for $u,v\in C^2_c(\Sigma)$, $$Q_{\Sigma,\mathbf{v}}[u,v]=-B_\Sigma[u,L_{\Sigma,\mathbf{v}} v]=-B_{\Sigma,\mathbf{v}}[u,L_{\Sigma,\mathbf{v}}^\prime v],$$ and so $L_{\Sigma,\mathbf{v}}$ and $L_{\Sigma,\mathbf{v}}^\prime$ are symmetric with respect to $B_\Sigma$ and $B_{\Sigma,\mathbf{v}}$, respectively.
Likewise, we consider the symmetric bilinear form $$\begin{aligned}
Q_{\mathbf{f}}[u,v] & = Q_{\Lambda}\left[u\circ\mathbf{f}^{-1},v\circ\mathbf{f}^{-1}\right] \\
& = \int_\Sigma \left((g_\mathbf{f}^{-1})^{ij}\nabla_i u\nabla_j v-(|A_\Lambda|^2\circ\mathbf{f})uv+\frac{1}{2}uv\right)\Omega_{\mathbf{f}} e^{\frac{|\mathbf{x}|^2}{4}} \, d\mathcal{H}^n.\end{aligned}$$ Let $$L_\mathbf{f}v=(L_\Lambda (v\circ\mathbf{f}^{-1}))\circ\mathbf{f}.$$ Observe, that for $u,v\in C^2_c(\Sigma)$, $$Q_{\mathbf{f}}[u,v]= -B_\mathbf{f}[u,L_\mathbf{f} v],$$ so $L_\mathbf{f}$ is symmetric with respect to $B_\mathbf{f}$. There also is a symmetric bilinear form $$Q_{\mathbf{f},\mathbf{v}}[u,v]= Q_{\Lambda,\mathbf{w}}\left[u\circ\mathbf{f}^{-1},v\circ\mathbf{f}^{-1}\right].$$ When $\Lambda$ is of class $C^3$, $$Q_{\mathbf{f},\mathbf{v}}[u,v] = \int_\Sigma \left((g_\mathbf{f}^{-1})^{ij}\nabla_iu\nabla_jv+(a_{\Lambda,\mathbf{w}}^\prime\circ\mathbf{f})uv\right)(\mathbf{v}\cdot(\mathbf{n}_\Lambda\circ\mathbf{f}))^2\Omega_\mathbf{f} e^{\frac{|\mathbf{x}|^2}{4}} \, d\mathcal{H}^n.$$ If we let $$L_{\mathbf{f},\mathbf{v}} v=(L_{\Lambda,\mathbf{w}}(v\circ\mathbf{f}^{-1}))\circ\mathbf{f} \mbox{ and } L_{\mathbf{f},\mathbf{v}}^\prime=(\mathbf{v}\cdot(\mathbf{n}_\Lambda\circ\mathbf{f}))^{-2} L_{\mathbf{f},\mathbf{v}},$$ then we observe that for $u,v\in C^2_c(\Sigma)$, $$Q_{\mathbf{f},\mathbf{v}}[u,v]=-B_\mathbf{f}[u,L_{\mathbf{f},\mathbf{v}} v]=-B_{\mathbf{f},\mathbf{v}}[u,L_{\mathbf{f},\mathbf{v}}^\prime v].$$ As such, $L_{\mathbf{f},\mathbf{v}}^\prime$ is symmetric with respect to $B_{\mathbf{f},\mathbf{v}}$.
We will always write the quadratic form associated to $B_\Sigma$ as $B_\Sigma[u]=B_\Sigma[u,u]$ and do so for the other bilinear forms.
First and second variations of the $E$-functional {#VariationSubsec}
-------------------------------------------------
Fix a $C^2$-hypersurface $\Sigma\subset\mathbb{R}^{n+1}$ and a transverse section $\mathbf{v}$ on $\Sigma$. Suppose $\mathbf{f}\colon\Sigma\to\mathbb{R}^{n+1}$ is a $C^2$ proper embedding and that $\mathbf{w}=\mathbf{v}\circ\mathbf{f}^{-1}$ is transverse to $\Lambda=\mathbf{f}(\Sigma)$. We also assume $u,v\in C^2_c(\Sigma)$. For $|s|,|t|$ sufficiently small, $$\mathbf{f}_{s,t}=\mathbf{f}+(su+tv)\mathbf{v}$$ is a proper embedding of $\Sigma$ into $\mathbb{R}^{n+1}$ and $\mathbf{v}\circ\mathbf{f}_{s,t}^{-1}$ is a transverse section on $\mathbf{f}_{s,t}(\Sigma)$.
We have by the first variation formula (see [@BWBanachManifold Proposition 4.1]), $${\frac{d}{ds}\vline}_{s=0} E[\mathbf{f}_{s,t}(\Sigma)]=-\int_{\Sigma} u\, \Xi_{\mathbf{f},\mathbf{v}}[tv] \Omega_{\mathbf{f}_{0,t}} e^{\frac{|\mathbf{x}|^2}{4}} \, d\mathcal{H}^n=-B_{\mathbf{f}_{0,t}}[\Xi_{\mathbf{f},\mathbf{v}}[tv],u]$$ where $$\Xi_{\mathbf{f},\mathbf{v}}[tv]=\mathbf{v}\cdot \left(\mathbf{H}-\frac{\mathbf{x}^\perp}{2}\right)[\mathbf{f}_{0,t}].$$ It will also be convenient to introduce $$\tilde{\Xi}_{\mathbf{f},\mathbf{v}}[tv]=\Omega_{\mathbf{f}_{0,t}} \Xi_{\mathbf{f},\mathbf{v}}[tv] \mbox{ and } \hat{\Xi}_{\mathbf{f},\mathbf{v}}[tv]=(\mathbf{v}\cdot\mathbf{n}_{\Sigma})^{-2}\Omega_{\mathbf{f}_{0,t}}\Xi_{\mathbf{f},\mathbf{v}}[tv].$$ We remark that the quantity $\tilde{\Xi}_{\mathbf{f},\mathbf{v}}$ is closer to what is considered in [@WhiteEI], however for various reasons we find $\hat{\Xi}_{\mathbf{f},\mathbf{v}}$ easier to work with. We also observe that $$B_{\mathbf{f}_{0,t}}[\Xi_{\mathbf{f},\mathbf{v}}[tv],u]=B_\Sigma[\tilde{\Xi}_{\mathbf{f},\mathbf{v}}[tv],u]=B_{\Sigma,\mathbf{v}}[\hat{\Xi}_{\mathbf{f},\mathbf{v}}[tv],u].$$ Thus, we have that $${\frac{d}{ds}\vline}_{s=0} E[\mathbf{f}_{s,t}(\Sigma)]=-B_\Sigma[\tilde{\Xi}_{\mathbf{f},\mathbf{v}}[tv],u]=-B_{\Sigma,\mathbf{v}}[\hat{\Xi}_{\mathbf{f},\mathbf{v}}[tv],u].$$ In particular, $\mathbf{f}_{0,t}(\Sigma)$ is a self-expander if and only if $$\Xi_{\mathbf{f},\mathbf{v}}[tv]=\tilde{\Xi}_{\mathbf{f},\mathbf{v}}[tv]=\hat{\Xi}_{\mathbf{f},\mathbf{v}}[tv]=0.$$
We next compute the second variation at critical points of $E$. If $\mathbf{f}(\Sigma)$ is a self-expander, then, by [@BWBanachManifold Proposition 4.2], we have $${\frac{d^2}{dtds}\vline}_{t=s=0} E[\mathbf{f}_{s,t}(\Sigma)]=Q_{\mathbf{f},\mathbf{v}}[u,v]=-B_{\Sigma,\mathbf{v}}[\Omega_{\mathbf{f},\mathbf{v}}L^\prime_{\mathbf{f},\mathbf{v}} v, u],$$ while differentiating the expression for the first variation formula directly gives, at critical points of $E$, $${\frac{d^2}{dtds}\vline}_{t=s=0} E[\mathbf{f}_{s,t}(\Sigma)]=-B_{\Sigma,\mathbf{v}}[D\hat{\Xi}_{\mathbf{f},\mathbf{v}}(0)v, u].$$ Hence, as long as $\mathbf{f}(\Sigma)$ is a self-expander and $\mathbf{v}\circ\mathbf{f}^{-1}$ is transverse to $\mathbf{f}(\Sigma)$, $$\label{2ndVarEqn}
D\hat{\Xi}_{\mathbf{f},\mathbf{v}}(0)=\Omega_{\mathbf{f},\mathbf{v}}L_{\mathbf{f},\mathbf{v}}^\prime$$ is symmetric with respect to $B_{\Sigma,\mathbf{v}}$.
Fast decay of eigenfunctions of $L_{\Sigma}$ and $L_{\Sigma,\mathbf{v}}^\prime$ {#DecayEigenfunSubsec}
-------------------------------------------------------------------------------
Let $\Sigma\in\mathcal{ACH}^{k,\alpha}_n$ be a self-expander. For $p\in\Sigma$ let $r(p)=|\mathbf{x}(p)|$. It follows from the expander equation, that there exist constants $R_0=R_0(\Sigma)$ and $C_0=C_0(\Sigma)$ so that on $E_{R_0}=\Sigma\setminus B_{R_0}$, $$\label{REstimatesEqn}
||\nabla_\Sigma r|-1|\leq C_0r^{-4} \leq \frac{1}{4}, \, |\nabla_\Sigma^2 r^2-2g_\Sigma|\leq C_0r^{-2} \leq\frac{1}{4},$$ and, $$\label{AEstimatesEqn}
|A_\Sigma|^2\leq C_0r^{-2} \leq\frac{1}{4} \mbox{ and } \left|H_{S_\rho}-\frac{n-1}{\rho}\right| \leq C_0 \rho^{-3} \leq \frac{1}{4} \rho^{-1} \mbox{ for $\rho>R_0$}.$$ Here $S_\rho=\Sigma\cap \partial B_\rho$ and $H_{S_\rho}$ is the mean curvature of $S_\rho$. Thus $(E_{R_0},g_\Sigma,r)$ is a weakly conical end in the sense of [@Bernstein Section 2].
We use results from [@Bernstein] to show the eigenfunctions of $L_\Sigma$ or $L_{\Sigma,\mathbf{v}}^\prime$ that below a certain threshold, actually decay extremely fast.
\[DecayEigenfunProp\] Suppose that $\Sigma\in\mathcal{ACH}^{k,\alpha}_n$ is a self-expander, $\mathbf{v}\in C^{k,\alpha}_0\cap C^{k}_{0,\mathrm{H}}(\Sigma;\mathbb{R}^{n+1})$ is a transverse section on $\Sigma$ so $\mathcal{C}[\mathbf{v}]=\mathscr{E}^{\mathrm{H}}_1\circ\mathrm{tr}_\infty^1[\mathbf{v}]$ is a transverse section on $\mathcal{C}(\Sigma)$, and $\mu\in\mathbb{R}$. If $u\in W^{1,2}_{loc}\cap W^0_\gamma(E_R)$ for some $\gamma>0$ and $R>R_0$, and satisfies $L_\Sigma u=\mu u$ or $L_{\Sigma,\mathbf{v}}^\prime u=\mu u$, then, for any $\beta<\frac{1}{2}$, $$\Vert e^{\beta |\mathbf{x}|^2} u\Vert_{k,\alpha; E_{4R}}<\infty.$$
Let $U=(\mathbf{v}\cdot\mathbf{n}_\Sigma) u$. Suppose $u\in W^{1,2}_{loc}\cap W^0_\gamma(E_R)$ and so is $U$. We first observe that $L_\Sigma, L_{\Sigma,\mathbf{v}}^\prime$ are uniformly elliptic operators with $C^{k-2,\alpha}_1$ coefficients (cf. (5.21) of [@BWBanachManifold]). If $L_\Sigma u=\mu u$ or $L_{\Sigma,\mathbf{v}}^\prime u=\mu u$ (i.e., $L_\Sigma U=\mu U$), then the standard elliptic regularity theory (see, e.g., [@GilbargTrudinger Chapters 6 and 8]) implies that $u\in C^{k,\alpha}_{loc}(E_{2R})$ and $u$ decays, in the pointwise sense, like $e^{-\frac{\gamma}{4}|\mathbf{x}|^2}$. Furthermore, it follows from [@Bernstein Theorems 9.1 and 7.2] that $u\in W_\beta^0(E_{2R})$ for all $\beta<1/2$. Hence, the claim follows from the local boundedness for weak solutions and the Schauder estimates (cf. [@GilbargTrudinger Theorems 8.7 and 6.2]).
A modification of the smooth dependence theorem from [@BWBanachManifold] {#ModifiedSmoothDependSec}
========================================================================
In this section we slightly modify the maps obtained in [@BWBanachManifold Theorem 7.1]. This modification brings the maps closer in line with those obtained in [@WhiteEI Theorem 3.2] and can be thought of as picking charts around each point of $\mathcal{ACE}^{k,\alpha}_n(\Gamma)$ different from the ones given in [@BWBanachManifold]. One important point of emphasis, our proof of the modified theorem uses [@BWBanachManifold Theorem 7.1] and it is not clear that one could prove the modified theorem directly. The issue has to do with the rapid growth of the weight, which can only be tamed using decay coming from [@BWBanachManifold Theorem 7.1].
Suppose that $\Sigma$ is a $C^2$-hypersurface in $\mathbb{R}^{n+1}$ and $\mathbf{v}$ is a transverse section on $\Sigma$. If $\mathbf{f}\colon\Sigma\to\mathbb{R}^{n+1}$ is a $C^2$ proper embedding, then we define $$\Xi_\mathbf{v}[\mathbf{f}]=\mathbf{v}\cdot\left(\mathbf{H}-\frac{\mathbf{x}^\perp}{2}\right)[\mathbf{f}] \mbox{ and } \hat{\Xi}_\mathbf{v}[\mathbf{f}]=(\mathbf{v}\cdot\mathbf{n}_\Sigma)^{-2}\Omega_{\mathbf{f}} \Xi_\mathbf{v}[\mathbf{f}].$$
\[ModifiedSmoothDependThm\] Let $\Sigma\in\mathcal{ACH}^{k,\alpha}_n$ be a self-expander, and let $\mathbf{v}\in C^{k,\alpha}_0\cap C^{k}_{0,\mathrm{H}}(\Sigma;\mathbb{R}^{n+1})$ be a transverse section on $\Sigma$ so that $\mathcal{C}[\mathbf{v}]$ is a transverse section on $\mathcal{C}(\Sigma)$ and that $$\Delta_\Sigma\mathbf{v}+\frac{1}{2}\mathbf{x}\cdot\nabla_\Sigma\mathbf{v}\in C^{k-2,\alpha}_{-2}(\Sigma;\mathbb{R}^{n+1}).$$ Let $$\mathcal{K}_\mathbf{v}={\left\{\kappa\in W^1_{\frac{1}{4}}(\Sigma)\colon L_{\Sigma,\mathbf{v}}^\prime \kappa=0\right\}}.$$ Then $\dim\mathcal{K}_\mathbf{v}<\infty$, and there exist smooth maps $$\begin{aligned}
\hat{F}_{\mathbf{v}}\colon \mathcal{U}_1\times\mathcal{U}_2 & \to \mathcal{ACH}_{n}^{k,\alpha}(\Sigma), \mbox{ and}, \\
\hat{G}_{\mathbf{v}}\colon \mathcal{U}_1\times\mathcal{U}_2 & \to \mathcal{K}_{\mathbf{v}},\end{aligned}$$ where $\mathcal{U}_1$ is some neighborhood of $\mathbf{x}|_{\mathcal{L}(\Sigma)}$ in $C^{k,\alpha}(\mathcal{L}(\Sigma); \mathbb{R}^{n+1})$ and $\mathcal{U}_2$ is some neighborhood of $0$ in $\mathcal{K}_{\mathbf{v}}$, such that the following hold:
1. \[TraceFItem\] For $(\varphi,\kappa)\in\mathcal{U}_1\times\mathcal{U}_2$, $\mathrm{tr}_\infty^1[\hat{F}_{\mathbf{v}}[\varphi,\kappa]]=\varphi$.
2. \[IdentityItem\] $\hat{F}_{\mathbf{v}}[\mathbf{x}|_{\mathcal{L}(\Sigma)},0]=\mathbf{x}|_\Sigma$.
3. \[EStationaryItem\] For $(\varphi,\kappa)\in\mathcal{U}_1\times\mathcal{U}_2$, $\hat{F}_{\mathbf{v}}[\varphi,\kappa]$ is $E$-stationary if and only if $\hat{G}_{\mathbf{v}}[\varphi,\kappa]=0$.
4. \[FDiffItem\] For $\kappa\in\mathcal{K}_{\mathbf{v}}$, $D_2 \hat{F}_{\mathbf{v}}(\mathbf{x}|_{\mathcal{L}(\Sigma)},0)\kappa=\kappa\mathbf{v}$.
5. \[SubmanifoldItem\] $\hat{G}_{\mathbf{v}}^{-1}(0)$ is a [smooth]{} submanifold of codimension equal to $\dim \mathcal{K}_{\mathbf{v}}$. It contains $\{0\}\times\mathcal{K}_{\mathbf{v}}$ in its tangent space at $(\varphi,0)$. Equivalently, $D_1 \hat{G}_{\mathbf{v}}(\mathbf{x}|_{\mathcal{L}(\Sigma)},0)$ is of rank equal to $\dim\mathcal{K}_{\mathbf{v}}$ and $D_2 \hat{G}_{\mathbf{v}}(\mathbf{x}|_{\mathcal{L}(\Sigma)},0)=0$.
6. \[ExpanderMeanCurvatureItem\] One has $$\hat{\Xi}_{ \mathbf{v}}[\hat{F}_{\mathbf{v}}[\varphi, \kappa]]=\hat{G}_{\mathbf{v}}[\varphi, \kappa]\in \mathcal{K}_{\mathbf{v}}.$$
Such a transverse section $\mathbf{v}$ always exists. For instance, outside a compact set of $\Sigma$, one may choose $\mathbf{v}$ to be $\mathbf{V}\circ\pi_{\mathbf{V}}|_\Sigma$ for $\mathbf{V}$ a homogeneous transverse section on $\mathcal{C}(\Sigma)$, and then extend $\mathbf{v}$ to be a transverse section on $\Sigma$. It is straightforward to verify that $\mathbf{v}$ satisfies the properties in Theorem \[ModifiedSmoothDependThm\].
The only difference with [@BWBanachManifold Theorem 7.1] is the weight terms that appear in Item . Namely, in that result one has $$\Xi_{ \mathbf{v}}[{F}_{\mathbf{v}}[\varphi, \kappa]]={G}_{\mathbf{v}}[\varphi, \kappa]\in\mathcal{K}_{\mathbf{v}}.$$
To prove Theorem \[ModifiedSmoothDependThm\] we will need several auxiliary propositions and lemmas. For $\beta,\mu\in\mathbb{R}$ we define $$\mathscr{L}_{\Sigma,\beta}^\mu=\Delta_\Sigma+2\beta\mathbf{x}\cdot\nabla_\Sigma-\mu.$$ Observe that, for $f,g\in C^{2}_c(\Sigma)$, $$\int_\Sigma f (\mathscr{L}_{\Sigma,\beta}^0 g) e^{\beta |\mathbf{x}|^2} \, d\mathcal{H}^n = -\int_{\Sigma} \nabla_\Sigma f\cdot\nabla_\Sigma g \, e^{\beta |\mathbf{x}|^2} \, d\mathcal{H}^n.$$ For brevity, we write $\mathscr{L}_\Sigma^\mu=\mathscr{L}_{\Sigma,\frac{1}{4}}^\mu$ and $\mathscr{L}_\Sigma=\mathscr{L}_{\Sigma,\frac{1}{4}}^{\frac{1}{2}}$.
\[ModifiedIsoProp\] Let $\Sigma\in\mathcal{ACH}^{k,\alpha}_n$ be a self-expander. For all $\beta\in \left[\frac{1}{4}, \frac{3}{8}\right]$, $$\mathscr{L}_{\Sigma}\colon W^2_\beta (\Sigma)\to W^0_\beta (\Sigma)$$ is an isomorphism.
First the Lax-Milgram theorem implies that given an $f\in W^0_\beta (\Sigma)$ there is a unique solution $u\in W^1_{\frac{1}{4}}(\Sigma)$ to $\mathscr{L}_\Sigma u=f$ in the distributional sense. Now suppose that $f$ is smooth and has compact support. As such, there is an $R^\prime\geq R_0(\Sigma)$ sufficiently large so that ${{\ensuremath{{\mathrm{spt}}\:}}}(f)\subset B_{R^\prime}$ and $\Sigma\setminus\bar{B}_{R^\prime}$ is a weakly conical end by the definition of $R_0$ given in Section \[DecayEigenfunSubsec\]. By standard elliptic regularity – see, e.g., Chapter 6 and 8 of [@GilbargTrudinger] – $\Sigma$ is indeed smooth and so is $u$. As $\mathscr{L}_{\Sigma} u=0$ in $\Sigma\backslash B_{R^\prime}$ it follows from Proposition \[DecayEigenfunProp\] that $$\label{WeightW2Eqn}
\int_{\Sigma} \left(|\nabla_\Sigma^2 u|^2+|\nabla_\Sigma u|^2 +u^2\right) e^{\frac{7}{16}|\mathbf{x}|^2} \, d\mathcal{H}^n<\infty.$$ In other words, $u\in W^{2}_{\frac{7}{16}}(\Sigma)$. As $\beta<\frac{7}{16}$, this fact will be used to justify integration by parts in several places.
Indeed, as $$\mathscr{L}_{\Sigma}^0 (u^2)=2 |\nabla_\Sigma u|^2 +2 u \mathscr{L}_{\Sigma}^0 u= 2 |\nabla_\Sigma u|^2+ u^2 +2u f,$$ integrating by parts, which is justified by , gives $$\begin{aligned}
0 &= \int_{\Sigma} \mathscr{L}_{\Sigma, \beta}^0(u^2) e^{\beta |\mathbf{x}|^2}\, d\mathcal{H}^n=\int_{\Sigma} \left( \mathscr{L}_{\Sigma}^0(u^2)+\left(2\beta-\frac{1}{2}\right) \mathbf{x}^\top \cdot \nabla_\Sigma u^2\right) e^{\beta |\mathbf{x}|^2} \, d\mathcal{H}^n\\
&= \int_{\Sigma} \left( 2 |\nabla_\Sigma u|^2+ u^2 +2u f+ \left( 4\beta-1\right) u \mathbf{x}^\top\cdot \nabla_\Sigma u \right) e^{\beta |\mathbf{x}|^2} \, d\mathcal{H}^n\\
&\geq \int_{\Sigma} \left( 2 |\nabla_\Sigma u|^2+ u^2 -\frac{1}{2} u^2 -2 f^2- \left( 4\beta-1\right) \left( |\nabla_\Sigma u|^2 +\frac{1}{4}|\mathbf{x}^\top|^2 u^2 \right) \right) e^{\beta |\mathbf{x}|^2} \, d\mathcal{H}^n,\end{aligned}$$ where the last inequality used the absorbing inequality and the fact that $\beta\geq \frac{1}{4}$. Hence, by of Proposition \[GradControlsL2Prop\], one obtains $$\int_{\Sigma} \left( 2 |\nabla_\Sigma u|^2+ \frac{1}{2} u^2 -2 f^2- 4\beta |\nabla_\Sigma u|^2 \right) e^{\beta |\mathbf{x}|^2} \, d\mathcal{H}^n \leq 0.$$ Hence, as $\beta\leq 3/8$, $$\int_{\Sigma} \left(\frac{1}{2} |\nabla_{\Sigma} u|^2 +\frac{1}{2}u^2 \right) e^{\beta|\mathbf{x}|^2} \, d\mathcal{H}^n\leq 2 \int_{\Sigma} f^2 e^{\beta|\mathbf{x}|^2}d\mathcal{H}^n.$$ That is, for $\beta\in \left[\frac{1}{4},\frac{3}{8}\right]$, $$\label{WeightW1Eqn}
\Vert u \Vert_{1}^{W, (\beta)} \leq 4\Vert f \Vert_0^{W, (\beta)}.$$
Next consider the Bochner type formula $$\begin{aligned}
\mathscr{L}_{\Sigma}^0 \left(\frac{1}{2} |\nabla_\Sigma u|^2\right)& =\nabla_{\Sigma} u \cdot \nabla_{\Sigma}(\mathscr{L}_{\Sigma}^0 u)+ |\nabla^2_\Sigma u|^2+\mathrm{Ric}_\Sigma (\nabla_\Sigma u, \nabla_\Sigma u)\\
& \quad -\frac{\mathbf{x}\cdot \mathbf{n}_\Sigma}{2} A_{\Sigma}(\nabla_{\Sigma} u, \nabla_{\Sigma} u)-\frac{1}{2}|\nabla_{\Sigma}u |^2 \\
&\geq \nabla_{\Sigma} u \cdot \nabla_{\Sigma}(\mathscr{L}_{\Sigma}^0 u)+ |\nabla^2_\Sigma u|^2-\left(\frac{1}{2}+K_{\Sigma}\right)|\nabla_{\Sigma}u |^2\end{aligned}$$ where the inequality uses the Gauss equation, the self-expander equation and $$\label{KSigmaDef}
K_{\Sigma}=\sup_{p\in\Sigma} (1+|\mathbf{x}(p)|^2)|A_{\Sigma}(p)|^2< \infty.$$ Integrating by parts (which can be justified by together with a standard cutoff argument) yields $$\begin{aligned}
\int_{\Sigma} |\nabla_\Sigma^2 u|^2 & e^{\beta |\mathbf{x}|^2} \, d\mathcal{H}^n \leq \int_{\Sigma} \left( \mathscr{L}_{\Sigma}^0 \left(\frac{1}{2} |\nabla_\Sigma u|^2\right)- \nabla_\Sigma u\cdot\nabla_\Sigma (\mathscr{L}_\Sigma^0 u ) \right) e^{\beta |\mathbf{x}|^2} \, d\mathcal{H}^n \\
& + \left(K_\Sigma+\frac{1}{2} \right) \int_{\Sigma} |\nabla_{\Sigma} u|^2 e^{\beta |\mathbf{x}|^2} \, d\mathcal{H}^n\\
& \leq \int_{\Sigma} \left( \frac{\left( 1-4\beta\right)}{2}\mathbf{x}^\top \cdot \nabla_\Sigma \left( \frac{1}{2} |\nabla_\Sigma u|^2 \right)+ ( \mathscr{L}_{\Sigma, \beta}^0 u) (\mathscr{L}_\Sigma^0 u ) \right) e^{\beta |\mathbf{x}|^2} \, d\mathcal{H}^n \\
& + \left(K_\Sigma+\frac{1}{2} \right) \int_{\Sigma} |\nabla_{\Sigma} u|^2 e^{\beta |\mathbf{x}|^2} \, d\mathcal{H}^n.\end{aligned}$$
One computes, $$\begin{aligned}
\left| \int_{\Sigma}\mathbf{x}^\top \cdot \nabla_\Sigma \left( \frac{1}{2} |\nabla_\Sigma u|^2 \right) e^{\beta |\mathbf{x}|^2} \, d\mathcal{H}^n\right| & = \left| \int_{\Sigma} \nabla^2_\Sigma u (\nabla_\Sigma u , \mathbf{x}^\top) e^{\beta |\mathbf{x}|^2} \, d\mathcal{H}^n\right|\\
&\leq \int_{\Sigma} |\nabla^2_\Sigma u| |\mathbf{x}^\top | |\nabla_\Sigma u|e^{\beta |\mathbf{x}|^2} \, d\mathcal{H}^n\\
&\leq \int_{\Sigma} \left( |\nabla^2_\Sigma u| ^2 +\frac{1}{4} |\mathbf{x}^\top |^2 |\nabla_\Sigma u|^2 \right)e^{\beta |\mathbf{x}|^2} \, d\mathcal{H}^n.\end{aligned}$$ Hence, by of Proposition \[GradControlsL2Prop\] applied to $|\nabla_\Sigma u|$ and the fact that $\frac{1}{4}\leq \beta\leq \frac{3}{8}$, $$\begin{aligned}
\left( \frac{ 1-4\beta}{2}\right) \int_{\Sigma}\mathbf{x}^\top \cdot \nabla_\Sigma \left(\frac{1}{2} |\nabla_\Sigma u|^2 \right) e^{\beta |\mathbf{x}|^2} \, d\mathcal{H}^n &\leq 2\beta \int_{\Sigma} |\nabla^2_\Sigma u| ^2 e^{\beta |\mathbf{x}|^2} \, d\mathcal{H}^n\\
&\leq \frac{3}{4}\int_{\Sigma} |\nabla^2_\Sigma u| ^2 e^{\beta |\mathbf{x}|^2} \, d\mathcal{H}^n.\end{aligned}$$ Similarly, as $\frac{1}{4}\leq \beta\leq \frac{3}{8}$, $$\begin{aligned}
\int_{\Sigma} ( \mathscr{L}_{\Sigma, \beta}^0 u) (\mathscr{L}_\Sigma^0 u ) & e^{\beta |\mathbf{x}|^2} d\mathcal{H}^n =\int_{\Sigma} \left( (2\beta-\frac{1}{2})\mathbf{x}^\top \cdot \nabla_\Sigma u + \frac{1}{2} u+f\right) \left(\frac{1}{2} u+f\right)e^{\beta |\mathbf{x}|^2} d\mathcal{H}^n \\
&\leq \int_{\Sigma}\left(\frac{1}{4} |\mathbf{x}^\top \cdot \nabla_\Sigma u| \left(\frac{1}{2}|u|+|f|\right) + \left(\frac{1}{2} u+f\right)^2\right) e^{\beta |\mathbf{x}|^2} \, d\mathcal{H}^n \\
&\leq \int_{\Sigma} \left(\frac{1}{128} |\mathbf{x}^\top \cdot \nabla_\Sigma u|^2+ 3\left(\frac{1}{2}|u|+|f|\right)^2 \right) e^{\beta |\mathbf{x}|^2} \, d\mathcal{H}^n\\
&\leq \frac{1}{8} \int_{\Sigma}|\nabla^2_\Sigma u| ^2 e^{\beta |\mathbf{x}|^2} \, d\mathcal{H}^n + 2\int_{\Sigma}\left( u^2 +4 f^2 \right) e^{\beta |\mathbf{x}|^2} \, d\mathcal{H}^n,\end{aligned}$$ where the last inequality uses of Proposition \[GradControlsL2Prop\].
Hence, we have arrived at $$\begin{aligned}
\frac{1}{8}\int_{\Sigma} |\nabla_\Sigma^2 u|^2 e^{\beta |\mathbf{x}|^2}d\mathcal{H}^n &\leq \int_{\Sigma} \left\{2 \left( u^2 +4 f^2\right) + \left(K_\Sigma+\frac{1}{2} \right) |\nabla_{\Sigma} u|^2\right\} e^{\beta |\mathbf{x}|^2} \, d\mathcal{H}^n.\end{aligned}$$ Combining this with the previous estimate gives $$\Vert u \Vert_{2}^{W,(\beta)} \leq C(K_\Sigma) \Vert f \Vert_{0}^{W,(\beta)}.$$
The proof is completed by observing that for a fixed $f\in W^{0}_{\beta}(\Sigma)$, by a standard cutoff and mollifier trick, there is a sequence of smooth, compactly supported functions on $\Sigma$, $f_j$, so that $f_j\to f$ in $W^{ 0}_{\beta}(\Sigma)$. Hence, if $u_j$ is the unique solution of $\mathscr{L}_\Sigma u_j=f_j$, then ${\left\{u_j\right\}}$ is a Cauchy sequence in $W^2_\beta(\Sigma)$. Hence, letting $j\to \infty$ one obtains a solution $u\in W^2_\beta(\Sigma)$ of $\mathscr{L}_\Sigma u=f$. The uniqueness of such solutions are ensured by the uniqueness of the solutions in $W^1_{\frac{1}{4}}(\Sigma)$.
We recall the Banach space $\mathcal{D}^{k,\alpha}(\Sigma)$ introduced in Section 5 of [@BWBanachManifold]: $$\mathcal{D}^{k,\alpha}(\Sigma)={\left\{f\in C^{k,\alpha}_1\cap C^{k-1,\alpha}_{0} \cap C^{k-2,\alpha}_{-1}(\Sigma)\colon \mathbf{x}\cdot\nabla_\Sigma f\in C^{k-2,\alpha}_{-1}(\Sigma)\right\}}$$ equipped with the norm $$\Vert f\Vert^*_{k,\alpha}=\Vert f\Vert^{(-1)}_{k-2,\alpha}+\sum_{i=k-1}^k \Vert\nabla_\Sigma^i f\Vert^{(1-k)}_{\alpha}+\Vert\mathbf{x}\cdot\nabla_\Sigma f\Vert^{(-1)}_{k-2,\alpha}.$$ We will adopt the convention that if $X_1$ and $X_2$ are Banach spaces, then $X_1\cap X_2$ is the Banach space with the norm $$\Vert f\Vert_{X_1\cap X_2}=\Vert f\Vert_{X_1}+\Vert f\Vert_{X_2}.$$
\[FredholmCor\] Let $\Sigma$ and $\mathbf{v}$ be as given in Theorem \[ModifiedSmoothDependThm\]. For $\beta\in \left[\frac{1}{4}, \frac{3}{8}\right]$ the maps $$L_{\Sigma}, L_{\Sigma, \mathbf{v}}^\prime\colon \mathcal{D}^{k, \alpha}\cap W^2_{\beta}(\Sigma)\to C^{k-2, \alpha}_{-1}\cap W^0_\beta (\Sigma)$$ are both Fredholm of index $0$.
By Proposition \[ModifiedIsoProp\] and [@BWBanachManifold Theorem 5.7] the map $$\mathscr{L}_\Sigma\colon \mathcal{D}^{k,\alpha}\cap W^2_\beta(\Sigma)\to C^{k-2,\alpha}_{-1}\cap W^0_\beta(\Sigma)$$ is an isomorphism. As $\Sigma$ is asymptotically conical, we have $|A_\Sigma|\in C^{k-2,\alpha}_{-1}(\Sigma)$. Thus the Poincaré inequality, of Proposition \[GradControlsL2Prop\], and the Arzelà-Ascoli theorem imply that $L_{\Sigma}$ and $\mathscr{L}_{\Sigma}$ differ by a compact operator. This proves that $L_{\Sigma}$ is Fredholm of index zero.
Observe that $L_{\Sigma,\mathbf{v}}^\prime=M_\mathbf{v}\circ L_{\Sigma,\mathbf{v}}$ where $M_{\mathbf{v}}[f]=(\mathbf{v}\cdot\mathbf{n}_\Sigma)^{-2} f$ is the operator given by multiplication by $(\mathbf{v}\cdot\mathbf{n}_\Sigma)^{-2}$. As $|\mathbf{v}\cdot\mathbf{n}_\Sigma|\in C^{k-1,\alpha}_0(\Sigma)$ and has a strictly positive lower bound, it is not hard to see that $M_\mathbf{v}$ is an isomorphism of $C^{k-2,\alpha}_{-1}\cap W_\beta^0(\Sigma)$. And the Poincaré inequality, of Proposition \[GradControlsL2Prop\], together with Lemma 5.10 and Proposition 5.11 of [@BWBanachManifold] implies that $L_{\Sigma,\mathbf{v}}$ and $\mathscr{L}_\Sigma$ differ by a compact operator from $\mathcal{D}^{k, \alpha}\cap W^2_{\beta}(\Sigma)$ to $C^{k-2, \alpha}_{-1}\cap W^0_\beta (\Sigma)$. Hence $L^\prime_{\Sigma,\mathbf{v}}$ is Fredholm of index $0$.
\[CokerLem\] Let $\Sigma$ and $\mathbf{v}$ be as given in Theorem \[ModifiedSmoothDependThm\]. If $\kappa \in \mathcal{K}_\mathbf{v}\setminus{\left\{0\right\}}$, then there are no solutions in $ W^2_{\frac{1}{4}} (\Sigma)$ to $L_{\Sigma,\mathbf{v}}^\prime u=\kappa$.
We argue by contradictions. If there were such a solution, then integration by parts would be valid by the fact that $u\in W^2_{\frac{1}{4}}(\Sigma)$ and Proposition \[DecayEigenfunProp\]. Thus, $$B_{\Sigma, \mathbf{v}}[\kappa, \kappa]=B_{\Sigma, \mathbf{v}}[\kappa, L_{\Sigma, \mathbf{v}}^\prime u]=-Q_{\Sigma, \mathbf{v}}[\kappa, u]=B_{\Sigma, \mathbf{v}}[L_{\Sigma, \mathbf{v}}^\prime \kappa, u]=0.$$ That is, $\kappa$ is identically zero.
For $\Sigma$ and $\mathbf{v}$ as given in Theorem \[ModifiedSmoothDependThm\], by Corollary \[FredholmCor\], $\dim \mathcal{K}_\mathbf{v}<\infty$. Let $\Pi_{\mathbf{v}}$ be the orthogonal projection to $\mathcal{K}_\mathbf{v}$ with respect to $B_{\Sigma,\mathbf{v}}$ and $\Pi^\perp_{\mathbf{v}}=\mathrm{Id}-\Pi_{\mathbf{v}}$. Recall the smooth map obtained in [@BWBanachManifold Theorem 7.1], $F_\mathbf{v}$ from a neighborhood of $(\mathbf{x}|_{\mathcal{L}(\Sigma)},0)$ in $C^{k,\alpha}(\mathcal{L}(\Sigma); \mathbb{R}^{n+1})\times\mathcal{K}_\mathbf{v}$ to a neighborhood of $\mathbf{x}|_\Sigma$ in $\mathcal{ACH}^{k,\alpha}_n(\Sigma)$ so that $$F_\mathbf{v}[\mathbf{x}|_{\mathcal{L}(\Sigma)},0]=\mathbf{x}|_\Sigma, \Xi_\mathbf{v}[F_\mathbf{v}[\varphi,\kappa]]\in\mathcal{K}_\mathbf{v} \mbox{ and } \mathscr{L}_\Sigma F_{\mathbf{v}}[\varphi,\kappa]\in C^{k-2,\alpha}_{-1}(\Sigma; \mathbb{R}^{n+1}).$$ Denote by $\mathcal{B}_R(f; X)$ the open ball in Banach space $X$ with radius $R$ and center $f$.
\[SmoothMapLem\] Let $\Sigma$ and $\mathbf{v}$ be as given in Theorem \[ModifiedSmoothDependThm\]. For $\beta\in \left(0,\frac{1}{2}\right)$ and sufficiently small $r_0>0$ the map $$\hat{\Theta}_\mathbf{v}\colon \mathcal{B}_{r_0}(\mathbf{x}|_{\mathcal{L}(\Sigma)}; C^{k,\alpha}(\mathcal{L}(\Sigma);\mathbb{R}^{n+1}))\times \mathcal{B}_{r_0}(0; \mathcal{D}^{k,\alpha}\cap W^2_\beta(\Sigma)) \to C^{k-2,\alpha}_{-1}\cap W^0_\beta(\Sigma)$$ given by $$\hat{\Theta}_\mathbf{v}[\varphi,u]=\Omega_{F_\mathbf{v}^\prime[\varphi,u]}^{-1}\Pi^\perp_{\mathbf{v}}\circ\hat{\Xi}_\mathbf{v}[F^\prime_\mathbf{v}[\varphi,u]] \mbox{ where $F_\mathbf{v}^\prime[\varphi,u]=F_\mathbf{v}[\varphi,0]+u\mathbf{v}$}$$ is a well-defined smooth map.
First, we write $$\begin{aligned}
\hat{\Theta}_\mathbf{v}[\varphi,u] & =(\mathbf{v}\cdot\mathbf{n}_\Sigma)^{-2} \left(\Xi_\mathbf{v}[F_\mathbf{v}^\prime[\varphi,u]]-\Xi_\mathbf{v}[F_\mathbf{v}[\varphi,0]]\right)+(\mathbf{v}\cdot\mathbf{n}_\Sigma)^{-2}\Xi_\mathbf{v}[F_\mathbf{v}[\varphi,0]] \\
& \quad -\Omega^{-1}_{F_\mathbf{v}^\prime[\varphi,u]} \Pi_{\mathbf{v}}\circ\hat{\Xi}_\mathbf{v}[F_\mathbf{v}^\prime[\varphi,u]].\end{aligned}$$ By Proposition \[DecayEigenfunProp\], $\mathcal{K}_{\mathbf{v}}\subset\mathcal{D}^{k,\alpha}\cap W^2_\beta(\Sigma)$. By [@BWBanachManifold Lemmas 7.2 and 7.3], the map $$(\varphi,u)\mapsto\Xi_\mathbf{v}[F_\mathbf{v}^\prime[\varphi,u]]-\Xi_\mathbf{v}[F_\mathbf{v}[\varphi,0]]$$ is a smooth map from a neighborhood of $(\mathbf{x}|_{\mathcal{L}(\Sigma)},0)$ in $C^{k,\alpha}(\mathcal{L}(\Sigma))\times (\mathcal{D}^{k,\alpha}\cap W^2_\beta(\Sigma))$ to $C^{k-2,\alpha}_{-1}\cap W^0_\beta(\Sigma)$. Similarly, one shows that for $r_0>0$ sufficiently small the maps $$\varphi\mapsto \Xi_\mathbf{v}[F_\mathbf{v}[\varphi,0]] \mbox{ and } (\varphi,u)\mapsto\Pi_{\mathbf{v}}\circ\hat{\Xi}_\mathbf{v}[F_\mathbf{v}^\prime[\varphi,u]]$$ are both smooth maps into $\mathcal{K}_\mathbf{v}$. Finally, we observe that, in view of Proposition \[DecayEigenfunProp\], the operator given by multiplication by $\Omega_{\mathbf{f}}^{-1}$ for $\mathbf{f}\in\mathcal{ACH}^{k,\alpha}_n(\Sigma)$ with $\mathrm{tr}^1_\infty[\mathbf{f}]$ sufficiently close to $\mathbf{x}|_{\mathcal{L}(\Sigma)}$ is a smooth map from $\mathcal{K}_\mathbf{v}$ to $C^{k-2,\alpha}_{-1}\cap W^0_\beta(\Sigma)$, and that the operator given by multiplication by $(\mathbf{v}\cdot\mathbf{n}_\Sigma)^{-2}$ is an isomorphism of $C^{k-2,\alpha}_{-1}\cap W^0_\beta(\Sigma)$. Therefore we have shown $\hat{\Theta}_\mathbf{v}$ is a smooth map.
We are now ready to prove the main result of this section.
First, Corollary \[FredholmCor\] gives $\dim \mathcal{K}_\mathbf{v}<\infty$, and Proposition \[DecayEigenfunProp\] gives $\mathcal{K}_\mathbf{v}\subset\mathcal{D}^{k,\alpha}\cap W^2_{\frac{3}{8}}(\Sigma)$. Let $$\begin{aligned}
\hat{\mathcal{K}}_{\mathbf{v}}^\perp & =\left\{f\in C^{k-2,\alpha}_{-1}\cap W^0_{\frac{3}{8}} (\Sigma)\colon B_{\Sigma, \mathbf{v}}[f, \kappa]=0 \mbox{ for all $\kappa\in\mathcal{K}_{\mathbf{v}}$}\right\}, \mbox{ and}, \\
\hat{\mathcal{K}}_{\mathbf{v},*}^\perp & =\left\{f\in\mathcal{D}^{k,\alpha}\cap W^2_{\frac{3}{8}}(\Sigma)\colon B_{\Sigma, \mathbf{v}}[f, \kappa]=0 \mbox{ for all $\kappa\in\mathcal{K}_{\mathbf{v}}$}\right\}.\end{aligned}$$ We consider the map $$\hat{\Theta}_{\mathbf{v}}^\prime\colon \mathcal{B}_{r_0} (\mathbf{x}|_{\mathcal{L}(\Sigma)}; C^{k,\alpha}(\mathcal{L}(\Sigma); \mathbb{R}^{n+1}))\times\mathcal{B}_{\frac{r_0}{2}}(0; \mathcal{K}_{\mathbf{v}})\times\mathcal{B}_{\frac{r_0}{2}} (0; \hat{\mathcal{K}}^\perp_{\mathbf{v},*}) \to \hat{\mathcal{K}}^\perp_{\mathbf{v}}$$ defined by $$\hat{\Theta}_{\mathbf{v}}^\prime [\varphi,\kappa,u]=\hat{\Theta}_\mathbf{v}[\varphi,\kappa+u].$$ By Lemma \[SmoothMapLem\], this is a well-defined smooth map.
By , one sees that $D_3\hat{\Theta}_{\mathbf{v}}^\prime (\mathbf{x}|_{\mathcal{L}(\Sigma)},0,0)$ is given by $L_{\Sigma,\mathbf{v}}'$ restricted to $\hat{\mathcal{K}}^\perp_{\mathbf{v},*}$. Thus, it follows from Corollary \[FredholmCor\] and Lemma \[CokerLem\] that $D_3\hat{\Theta}_{\mathbf{v}}^\prime (\mathbf{x}|_{\mathcal{L}(\Sigma)},0,0)$ is an isomorphism between $\hat{\mathcal{K}}^\perp_{\mathbf{v},*}$ and $\hat{\mathcal{K}}^\perp_{\mathbf{v}}$. Hence, by the implicit function theorem, there is a neighborhood $\mathcal{U}=\mathcal{U}_1\times\mathcal{U}_2$ of $(\mathbf{x}|_{\mathcal{L}(\Sigma)},0)$ in $C^{k,\alpha}(\mathcal{L}(\Sigma); \mathbb{R}^{n+1})\times\mathcal{K}_{\mathbf{v}}$ and a [smooth]{} map $\hat{F}_{\mathbf{v},*}\colon \mathcal{U}\to\hat{\mathcal{K}}^\perp_{\mathbf{v},*}$ that gives the unique solution in a neighborhood of $0\in\hat{\mathcal{K}}_{\mathbf{v},*}^\perp$ to $$\hat{\Theta}_{\mathbf{v}}^\prime[\varphi,\kappa,\hat{F}_{\mathbf{v},*}[\varphi,\kappa]]=0.$$
Now let $$\begin{aligned}
\hat{F}_{\mathbf{v}}[\varphi,\kappa] & =F_{\mathbf{v}}[\varphi, 0]+(\kappa+\hat{F}_{\mathbf{v},*}[\varphi,\kappa])\mathbf{v}, \mbox{ and}, \\
\hat{G}_{\mathbf{v}}[\varphi,\kappa] & =\Pi_{\mathbf{v}}\circ\hat{\Xi}_{ \mathbf{v}}[\hat{F}_{\mathbf{v}}[\varphi, \kappa]].\end{aligned}$$ Items , , and follow directly from the definitions of $\hat{F}_{\mathbf{v}}$ and $\hat{G}_{\mathbf{v}}$. Furthermore, Items and follow the exact same proofs as in [@BWBanachManifold Theorem 7.1]; for readers’ convenience we include them below.
To establish Item it suffices to show $D_2\hat{F}_{\mathbf{v},*}(\mathbf{x}|_{\mathcal{L}(\Sigma)},0)=0$. Take any $\kappa\in\mathcal{K}_\mathbf{v}$. Observe that for $|s|<\epsilon$, $$\label{ProjExpanderMCEqn}
\Pi^\perp_\mathbf{v}\circ\hat{\Xi}_\mathbf{v}[\mathbf{x}|_\Sigma+(s\kappa+\hat{F}_{\mathbf{v},*}[\mathbf{x}|_{\mathcal{L}(\Sigma)},s\kappa])\mathbf{v}]=0.$$ Differentiating at $s=0$, we use and the chain rule to get $$\Pi^\perp_\mathbf{v}\circ L_{\Sigma,\mathbf{v}}^\prime\circ D_2\hat{F}_{\mathbf{v},*}(\mathbf{x}|_{\mathcal{L}(\Sigma)},0)\kappa=0.$$ It follows from Lemma \[CokerLem\] and the definition of $\hat{F}_{\mathbf{v},*}$ that $$D_2\hat{F}_{\mathbf{v},*}(\mathbf{x}|_{\mathcal{L}(\Sigma)},0)\kappa\in\mathcal{K}_\mathbf{v}\cap\mathcal{K}_{\mathbf{v},*}^\perp={\left\{0\right\}},$$ proving the claim.
Next, Items and together with imply that $$D_2\hat{G}_\mathbf{v}(\mathbf{x}|_{\mathcal{L}(\Sigma)},0)\kappa=L^\prime_{\Sigma,\mathbf{v}}[\mathbf{v}\cdot D_2\hat{F}_{\mathbf{v}}(\mathbf{x}|_{\mathcal{L}(\Sigma)},0)\kappa]=L^\prime_{\Sigma,\mathbf{v}}\kappa=0.$$ Similarly, one uses the more general second variation formula, [@BWBanachManifold Proposition 4.2], to compute that, for $\zeta\in C^{k,\alpha}(\mathcal{L}(\Sigma))$, $$D_1\hat{G}_\mathbf{v}(\mathbf{x}|_{\mathcal{L}(\Sigma)},0)[\zeta\mathbf{V}]=(\mathbf{v}\cdot\mathbf{n}_\Sigma)^{-1}L_{\Sigma}[\mathbf{n}_\Sigma\cdot D_1\hat{F}_\mathbf{v}(\mathbf{x}|_{\mathcal{L}(\Sigma)},0)[\zeta\mathbf{V}]].$$ Thus, arguing as in [@BWBanachManifold Theorem 7.1], it follows that $D_1\hat{G}_{\mathbf{v}}(\mathbf{x}|_{\mathcal{L}(\Sigma)},0)$ is of rank equal to $\dim \mathcal{K}_\mathbf{v}$. Hence we have shown Item .
It is convenient to let $$\hat{\mathcal{F}}_{\mathbf{v}}[\varphi, \kappa]=\mathbf{v}\cdot \left(\hat{F}_{\mathbf{v}}[\varphi, \kappa]-\hat{F}_{\mathbf{v}}[\varphi, 0]\right).$$ With this notation $$\hat{F}_{\mathbf{v}}[\varphi, \kappa]=\hat{F}_{\mathbf{v}}[\varphi, 0]+ \hat{\mathcal{F}}_{\mathbf{v}}[\varphi, \kappa] \mathbf{v}.$$ It follows from the proof of Theorem \[ModifiedSmoothDependThm\], that $$\label{OrthogFcalEqn}
\hat{\mathcal{F}}_{\mathbf{v}}[\varphi, \kappa]-\kappa=\hat{F}_{\mathbf{v}, *}[\varphi, \kappa]-\hat{F}_{\mathbf{v},*}[\varphi,0]\in \hat{\mathcal{K}}_{\mathbf{v}, *}^\perp\subset \mathcal{D}^{k, \alpha}\cap W^2_{\frac{3}{8}}(\Sigma).$$ Moreover, $$\hat{\Xi}_{\hat{F}_{\mathbf{v}}[\varphi, 0], \mathbf{v}}[\hat{\mathcal{F}}_{\mathbf{v}}[\varphi, \kappa]]=\hat{\Xi}_{\hat{F}_{\mathbf{v}}[\varphi, \kappa], \mathbf{v}}[0]=\hat{G}_{\mathbf{v}}[\varphi, \kappa].$$ In particular, if $\hat{G}_\mathbf{v}[\varphi,\kappa]=0$, i.e., $\hat{F}_\mathbf{v}[\varphi,\kappa]$ is $E$-stationary, then $$\label{DiffIdentEqn}
\begin{split}
D_2 \hat{G}_{\mathbf{v}}(\varphi,\kappa) & =D\hat{\Xi}_{\hat{F}_{\mathbf{v}}[\varphi, \kappa],\mathbf{v}}(0)\circ D_2 \hat{\mathcal{F}}_{\mathbf{v}}(\varphi,\kappa) \\
& =\Omega_{\hat{F}_{\mathbf{v}}[\varphi, \kappa],\mathbf{v}}L_{\hat{F}_{\mathbf{v}}[\varphi, \kappa], \mathbf{v}}^\prime\circ D_2 \hat{\mathcal{F}}_{\mathbf{v}}(\varphi,\kappa).
\end{split}$$
Index and nullity {#IndexSec}
=================
We recall the notion of index and nullity for asymptotically conical self-expanders and relate these integers to certain other spectral invariants. As observed in Section \[VariationSubsec\], the expander equation is an elliptic variational problem and so, as is usual for such problems, we define the *(Morse) index* of an asymptotically conical self-expander, $\Sigma$, to be $$\mathrm{ind}(\Sigma)=\sup {\left\{\dim V\colon V\subset C^{2}_c(\Sigma) \mbox{ so that } Q_{\Sigma}[u]<0, \forall u\in V\backslash {\left\{0\right\}}\right\}}.$$ The *nullity* of $\Sigma$, $\mathrm{null}(\Sigma)$ is defined to be the dimension of the kernel of $L_\Sigma$, $$\mathcal{K}={\left\{\kappa\in W^{1}_{\frac{1}{4}}(\Sigma)\colon L_\Sigma\kappa=0\right\}}.$$ Observe, that if $\mathbf{v}\in C^{k,\alpha}_0\cap C^{k}_{0,\mathrm{H}}(\Sigma;\mathbb{R}^{n+1})$ is a transverse section on $\Sigma$ so that $\mathcal{C}[\mathbf{v}]$ is transverse to $\mathcal{C}(\Sigma)$, then $\dim\mathcal{K}_\mathbf{v}=\dim\mathcal{K}$. One readily checks that $B_{\Sigma}$ and $B_{\Sigma, \mathbf{v}}$ extend to $W^0_{\frac{1}{4}}(\Sigma)$, and $D_{\Sigma}$, $D_{\Sigma, \mathbf{v}}$ $Q_{\Sigma}$ and $Q_{\Sigma, \mathbf{v}}$ extend to $W_{\frac{1}{4}}^1(\Sigma)$.
\[SpectrumLem\] Let $\Sigma\in\mathcal{ACH}^{k,\alpha}_n$ be a self-expander, and let $\mathbf{v}\in C^{k,\alpha}_0\cap C^{k}_{0,\mathrm{H}}(\Sigma;\mathbb{R}^{n+1})$ be a transverse section on $\Sigma$ so that $\mathcal{C}[\mathbf{v}]$ is transverse to $\mathcal{C}(\Sigma)$. The operators $$-L_\Sigma, -L_{\Sigma,\mathbf{v}}^\prime \colon W^{2}_{\frac{1}{4}}(\Sigma)\to W^{0}_{\frac{1}{4}}(\Sigma)$$ are self-adjoint in $\left(W^0_{\frac{1}{4}}(\Sigma),B_\Sigma\right)$ and $\left(W^0_{\frac{1}{4}}(\Sigma),B_{\Sigma,\mathbf{v}}\right)$, respectively, and both have discrete spectrums.
First we observe that $W^2_{\frac{1}{4}}(\Sigma)$ is dense in $\left(W^0_{\frac{1}{4}}(\Sigma),B_\Sigma\right)$. If we choose $\gamma>0$ so $$\sup_{\Sigma}|A_\Sigma|^2 \leq \gamma,$$ then the Lax-Milgram theorem and Proposition \[ModifiedIsoProp\] imply, that given an $f\in W^0_{\frac{1}{4}}(\Sigma)$ there is a unique solution $u\in W^2_{\frac{1}{4}}(\Sigma)$ to $(-L_\Sigma+\gamma)u=f$. Moreover, $$\Vert u\Vert^{W,(\frac{1}{4})}_2 \leq C(\gamma) \Vert f\Vert_{0}^{W,(\frac{1}{4})}.$$ That is, one can define a bounded operator, $R_\gamma=(-L_\Sigma+\gamma)^{-1}$ in $\left(W^0_{\frac{1}{4}}(\Sigma), B_\Sigma\right)$. Moreover, $R_\gamma$ is self-adjoint because $-L_\Sigma$ is symmetric with respect to $B_\Sigma$. Now one writes $-L_\Sigma=R_\gamma^{-1}-\gamma$. It is easy to check (see, e.g., [@Grigoryan pages 107-108]) that $R_\gamma^{-1}$ is a self-adjoint operator in $\left(W^0_{\frac{1}{4}}(\Sigma), B_\Sigma\right)$ and so is $-L_\Sigma$.
To prove the spectrum of $-L_\Sigma$ is discrete, we follow the arguments in [@Grigoryan Theorem 10.20]. As the natural embedding of $W^2_{\frac{1}{4}}(\Sigma)$ into $W^0_{\frac{1}{4}}(\Sigma)$ is compact by the Poincaré inequality, of Proposition \[GradControlsL2Prop\], it follows that $R_\gamma$ is a compact operator. As the kernel of $R_\gamma$ is trivial, the Hilbert-Schmidt theorem implies that there is a countable orthonormal basis ${\left\{\psi_i\right\}}$ of $\left(W^0_{\frac{1}{4}}(\Sigma), B_\Sigma\right)$ that consists of the eigenfunctions $\psi_i$ of $R_\gamma$ with eigenvalues $\rho_i>0$ so $\rho_i\to 0$. Clearly, the $\psi_i$ are also eigenfunctions of $-L_\Sigma$ with eigenvalues $\mu_i=\rho_i^{-1}-\gamma$, and ${\left\{\mu_i\right\}}$ has no accumulating points and $\mu_i\to\infty$. Observe, that for $\mu\neq\mu_i$, if $f=\sum_{i} a_i \psi_i$, then $$(L_\Sigma-\mu)^{-1} f=\sum_{i} \frac{a_i}{\mu_i-\mu}\psi_i$$ and it is bounded as $\inf_i |\mu_i-\mu|>0$. Hence the entire spectrum of $L_\Sigma$ is given by ${\left\{\mu_i\right\}}$, implying the discreteness of the spectrum of $L_\Sigma$.
As long as one observes that $a_{\Sigma,\mathbf{v}}^\prime$ is uniformly bounded (cf. [@BWBanachManifold (5.21)]), the proof for $L_{\Sigma,\mathbf{v}}^\prime$ is essentially the same.
This allows us to relate the index to spectral properties of $L_{\Sigma}$ and $L_{\Sigma,\mathbf{v}}^\prime$. Indeed,
\[IndexProp\] Let $\Sigma\in\mathcal{ACH}^{k,\alpha}_n$ be a self-expander, and let $\mathbf{v}\in C^{k,\alpha}_0\cap C^{k}_{0,\mathrm{H}}(\Sigma;\mathbb{R}^{n+1})$ be a transverse section on $\Sigma$ so that $\mathcal{C}[\mathbf{v}]$ is transverse to $\mathcal{C}(\Sigma)$. Then one has that $\mathrm{ind}(\Sigma)$ is equal to:
1. The maximum dimension of subspaces of $W^1_\beta(\Sigma)$, $\beta\geq \frac{1}{4}$ on which ${Q}_{\Sigma}$ or ${Q}_{\Sigma,\mathbf{v}}$ is negative definite.
2. The number of negative eigenvalues (counted with multiplicities) of $-L_{\Sigma}$ or $-L_{\Sigma,\mathbf{v}}^\prime$ in $W^1_\beta(\Sigma)$ for any $\beta\in \left(0,\frac{1}{2}\right)$.
As $C^2_c(\Sigma)$ is dense in $W^1_{\frac{1}{4}}(\Sigma)$ and $|A_\Sigma|$ is uniformly bounded, it follows from the dominated convergence theorem that $\mathrm{ind}(\Sigma)$ is equal to the maximum dimension of subspaces of $W^1_{\frac{1}{4}}(\Sigma)$ on which $Q_\Sigma$ is negative definite. Observe, that the map $M_\mathbf{v}$ given by the multiplication by $(\mathbf{v}\cdot\mathbf{n}_\Sigma)^{-1}$ is an isomorphism of $W^1_{\frac{1}{4}}(\Sigma)$ and $Q_{\Sigma,\mathbf{v}}[M_\mathbf{v} u, M_\mathbf{v} v]=Q_{\Sigma}[u,v]$. Thus, $Q_\Sigma$ is negative definite on a subspace $V\subset W^1_{\frac{1}{4}}(\Sigma)$ if and only if $Q_{\Sigma,\mathbf{v}}$ is so on $M_\mathbf{v}(V)$. Hence the first claim follows from the trivial inclusions for $\beta\geq \frac{1}{4}$, $C^\infty_c(\Sigma)\subset W^1_\beta(\Sigma)\subset W^1_{\frac{1}{4}}(\Sigma)$.
To prove the second, we appeal to Lemma \[SpectrumLem\] and the min-max principle for self-adjoint operators. It follows that $\mathrm{ind}(\Sigma)$ is equal to the number of negative eigenvalues of $-L_\Sigma$ or $-L_{\Sigma,\mathbf{v}}^\prime$ in $W^1_{\frac{1}{4}}(\Sigma)$. Invoking Proposition \[DecayEigenfunProp\], we have that eigenfunctions of $-L_\Sigma$ or $-L_{\Sigma,\mathbf{v}}^\prime$ in $W^1_{\frac{1}{4}}(\Sigma)$ are in $W^1_{\beta}(\Sigma)$ for any $\beta\in\left(0,\frac{1}{2}\right)$ and vice versa. This proves the second claim.
We extend the definition of index and nullity to parameterizations in a natural way. For $\Sigma\in\mathcal{ACH}^{k,\alpha}_n$, if $\mathbf{f}\in\mathcal{ACH}^{k,\alpha}_n(\Sigma)$ is $E$-stationary, i.e., $\Lambda=\mathbf{f}(\Sigma)$ is a self-expander, then we define $$\mathrm{ind}(\mathbf{f})=\mathrm{ind}(\Lambda) \mbox{ and }\mathrm{null}(\mathbf{f})=\mathrm{null}(\Lambda),$$ and $\mathrm{ind}([\mathbf{f}])$ and $\mathrm{null}([\mathbf{f}])$ are defined likewise. Similarly, we have the following:
\[ParaSpectrumLem\] Let $\Sigma\in\mathcal{ACH}^{k,\alpha}_n$ and $\mathbf{v}\in C^{k,\alpha}_0\cap C^{k}_{0,\mathrm{H}}(\Sigma;\mathbb{R}^{n+1})$ a transverse section on $\Sigma$ so that $\mathcal{C}[\mathbf{v}]$ is transverse to $\mathcal{C}(\Sigma)$. If $\mathbf{f}\in\mathcal{ACH}^{k,\alpha}_n(\Sigma)$ is $E$-stationary, and $\mathbf{v}\circ\mathbf{f}^{-1}$ and $\mathcal{C}[\mathbf{v}]\circ\mathcal{C}[\mathbf{f}]^{-1}$ are transverse to, respectively, $\Lambda=\mathbf{f}(\Sigma)$ and $\mathcal{C}(\Lambda)$, then the operators $$-L_{\mathbf{f}}, -L_{\mathbf{f},\mathbf{v}}^\prime\colon \mathbf{f}^* W^2_{\frac{1}{4}}(\Lambda)\to \mathbf{f}^* W^0_{\frac{1}{4}}(\Lambda)$$ are self-adjoint in $\left( \mathbf{f}^* W^0_{\frac{1}{4}}(\Lambda), B_\mathbf{f}\right)$ and $\left( \mathbf{f}^* W^0_{\frac{1}{4}}(\Lambda), B_{\mathbf{f},\mathbf{v}}\right)$, respectively, and both have discrete spectrums.
\[ParaIndexProp\] Let $\Sigma\in\mathcal{ACH}^{k,\alpha}_n$ and $\mathbf{v}\in C^{k,\alpha}_0\cap C^{k}_{0,\mathrm{H}}(\Sigma;\mathbb{R}^{n+1})$ a transverse section on $\Sigma$ so that $\mathcal{C}[\mathbf{v}]$ is transverse to $\mathcal{C}(\Sigma)$. There is an $\epsilon_0>0$ sufficiently small, depending only on $K_\Sigma$ and $\inf_\Sigma|\mathbf{v}\cdot\mathbf{n}_\Sigma|$, so that if $\mathbf{f}\in \mathcal{B}_{\epsilon_0}(\mathbf{x}|_\Sigma; \mathcal{ACH}^{k,\alpha}_n(\Sigma))$ and is $E$-stationary, then $\mathrm{ind}(\mathbf{f})$ is equal to
1. The maximum dimension of subspaces of $C^2_c(\Sigma)$ on which ${Q}_{\mathbf{f}}$ is negative definite.
2. The maximum dimension of subspaces of $W^1_{\beta}(\Sigma)$, $\beta\geq \frac{3}{8}$ on which ${Q}_{\mathbf{f}}$ or ${Q}_{\mathbf{f},\mathbf{v}}$ is negative definite.
3. The number of negative eigenvalues (counted with multiplicities) of $-L_{\mathbf{f}}$ or $-L_{\mathbf{f},\mathbf{v}}^\prime$ in $W^1_{\beta}(\Sigma)$ for any $\beta\in \left[\frac{1}{8}, \frac{3}{8}\right]$.
Concentration estimates for eigenfunctions and stability of certain quadratic forms {#ConcentrationSec}
===================================================================================
The main goal of this section is to show that, for eigenfunctions of $-L_\Sigma$ with small eigenvalue, one can prove effective estimates on concentration of such eigenfunctions in a compact ball. Here $\Sigma$ is an asymptotically conical self-expander as in Section \[DecayEigenfunSubsec\] and $R_0=R_0(\Sigma)$ is chosen so and hold. We also let $K_\Sigma$ be given by . The estimates will depend only on $K_\Sigma, n$ and the measure of smallness of the eigenvalue. This will be crucial to showing that the quadratic forms from Section \[NotationSec\], are stable for perturbations of $\Sigma$, even those that change the asymptotic cone, when one entry is such an eigenfunction.
Let $S_\rho=\Sigma\cap\partial B_\rho$ and $$L(\rho)=\int_{S_{\rho}} u^2 |\nabla_\Sigma r| \, d\mathcal{H}^{n-1},$$ and for $\gamma\in {\mathbb R}$, $$L_\gamma(\rho)=e^{\gamma \rho^2}L(\rho).$$ Notice that, by the co-area formula, when $\rho\geq R_0$, $$\label{CoAreaEqn}
\frac{1}{2}\int_{\bar{E}_\rho} u^2 e^{\gamma r^2} d\mathcal{H}^{n} \leq \int_{\bar{E}_\rho} u^2 e^{\gamma r^2} |\nabla_\Sigma r|^2 d\mathcal{H}^{n}=\int_\rho^\infty L_\gamma(t) dt \leq \int_{\bar{E}_\rho} u^2 e^{\gamma r^2} d\mathcal{H}^{n}.$$
\[ConcentrationLem\] Let $\Sigma\in\mathcal{ACH}^{k,\alpha}_n$ be a self-expander. Fix an $\epsilon>0$ and $\mu\leq \frac{1}{4}$. If $u\in W^1_{\frac{1}{4}}(\Sigma)$ satisfies $$-L_{\Sigma} u=\mu u,$$ then there is an $R_1=R_1(K_\Sigma, n, \epsilon)\geq R_0$ so that $$\int_{\Sigma\backslash B_{R_1}} \left(|\nabla_\Sigma u|^2+u^2\right) e^{\frac{3}{8} |\mathbf{x}|^2} \, d\mathcal{H}^n\leq \epsilon \int_{\Sigma} u^2 e^{\frac{|\mathbf{x}|^2}{4}} \, d\mathcal{H}^n.$$
First observe that, by Proposition \[DecayEigenfunProp\], $u\in W^2_{\beta}(\Sigma)$ for all $\beta\in \left(0, \frac{1}{2}\right)$. This fact, with $\beta=\frac{7}{16}$, will justify several integrations by parts. The first variation formula, and give, for $\rho\geq R_0$ and $\gamma\leq \frac{1}{2}$, that $$\begin{aligned}
{L}'_\gamma(\rho)&\leq 2 e^{\gamma \rho^2} \int_{S_\rho} u \nabla_\Sigma u\cdot\frac{\nabla_\Sigma r}{|\nabla_\Sigma r|} \, d\mathcal{H}^{n-1}+\frac{n}{\rho} {L}_\gamma(\rho)+2\gamma \rho L_{\gamma}(\rho) \\
&=-2 e^{\left(\gamma-\frac{1}{4}\right) \rho^2 } \int_{\bar{E}_\rho}\left( |\nabla_\Sigma u|^2 +u\mathscr{L}_\Sigma^0 u\right) e^{\frac{r^2}{4}} \, d\mathcal{H}^{n} +\frac{n}{\rho} {L}_\gamma(\rho)+2\gamma \rho L_{\gamma}(\rho) \\
&= -2 e^{\left(\gamma-\frac{1}{4}\right) \rho^2 }\int_{\bar{E}_\rho}\left( |\nabla_\Sigma u|^2 -|A_\Sigma|^2 u^2 +\frac{1}{2} u^2 +u L_\Sigma u\right) e^{\frac{r^2}{4}} \, d\mathcal{H}^{n}\\
& +\frac{n}{\rho} {L}_\gamma(\rho)+2\gamma \rho L_{\gamma}(\rho) \\
&\leq -2(1-2\gamma) e^{\left(\gamma-\frac{1}{4}\right) \rho^2 }\int_{\bar{E}_\rho} |\nabla_\Sigma u|^2 e^{\frac{r^2}{4}} \, d\mathcal{H}^{n}+\frac{n}{\rho}{L}_\gamma(\rho)\\
&+2 e^{\left(\gamma-\frac{1}{4}\right) \rho^2 }\int_{\bar{E}_\rho} \left( |A_{\Sigma}|^2+\mu -\frac{1}{2}\right) u^2 e^{\frac{r^2}{4}} \, d\mathcal{H}^{n},\end{aligned}$$ where the last inequality follows from the Poincaré inequality, Proposition \[GradControlsBoundaryProp\]. As $\gamma\leq \frac{1}{2}$, $\mu\leq \frac{1}{4}$ and $|A_{\Sigma}|^2\leq\frac{1}{4} $ on $E_{R_0}$ this means that, for $\rho\geq R_0$, $$\label{hatLdecay}
\frac{d}{d\rho}\left(\rho^{-n} {L}_\gamma(\rho)\right) \leq -2(1-2\gamma)\rho^{-n} e^{\left(\gamma-\frac{1}{4}\right) \rho^2 }\int_{\bar{E}_\rho} |\nabla_\Sigma u|^2 e^{\frac{r^2}{4}} \, d\mathcal{H}^{n}\leq 0.$$ Hence, for $\rho\geq \rho_0\geq R_0$ and $\gamma\leq \frac{1}{2}$, $$\label{hatLdecayEst2}
\rho^{-n} L_{\gamma}(\rho)\leq \rho_0^{-n} L_{\gamma}(\rho_0) \mbox{ and } \frac{e^{\gamma\rho^2}}{\rho^n} \int_{S_\rho} u^2 \, d\mathcal{H}^{n-1} \leq \frac{2 e^{\gamma\rho_0^2}}{\rho_0^n} \int_{S_{\rho_0}} u^2 \, d\mathcal{H}^{n-1}.$$ A consequence of this, the co-area formula and the mean value theorem is that, for $\rho\geq \rho_0 +1\geq R_0+1$ and $\gamma<\frac{1}{2}$, $$\label{TraceEst}
\rho^{-n} e^{\gamma\rho^2} \int_{S_\rho} u^2 \, d\mathcal{H}^{n-1} \leq 2 \rho_0^{-n} \int_{\bar{E}_{\rho_0}} u^2 e^{\gamma r^2} \, d\mathcal{H}^{n} \leq 2 \rho_0^{-n} \int_{\Sigma} u^2 e^{\gamma r^2} \, d\mathcal{H}^{n} .$$ Furthermore, combining and one has for $\rho\geq \rho_1\geq R_0$ that $$\begin{aligned}
\int_{\bar{E}_\rho} u^2 e^{\frac{7}{16} {r^2}} \, d\mathcal{H}^n & \leq 2\int_{\rho}^\infty L_{7/16}(t) \, dt=2\int_{\rho}^\infty t^{-n} L_{1/2}(t) t^n e^{-\frac{t^2}{16}}\, dt \\
&\leq 2 \rho_{1}^{-n} L_{1/2}(\rho_1) \int_{\rho}^\infty t^n e^{-\frac{t^2}{16}} \, dt \\
&\leq C(n) \rho^{n-1} \rho_1^{-n} e^{-\frac{\rho^2}{16}+\frac{\rho^2_1}{2}} \int_{S_{\rho_1}} u^2 \, d\mathcal{H}^{n-1}.\end{aligned}$$ Combining this with for $\gamma=\frac{1}{4}$ and one has for $\rho\geq\rho_1\geq \rho_0+1\geq R_0+1$, $$\label{RefinedEst}
\begin{split}
\int_{\bar{E}_\rho} u^2 e^{\frac{7}{16} {r^2}} \, d\mathcal{H}^n &\leq 2 C(n) \rho^{n-1} \rho_0^{-n} e^{-\frac{\rho^2}{16} +\frac{\rho^2_1}{4}}\int_{\bar{E}_{\rho_0}} u^2 e^{\frac{r^2}{4}} \, d\mathcal{H}^n.
\end{split}$$ In particular, for any $\epsilon>0$, there is an $R=R(\epsilon, K_\Sigma,n) \geq R_0+1$ so that $$\int_{\bar{E}_{R}} u^2 e^{\frac{7}{16} r^2} \, d\mathcal{H}^n\leq \frac{\epsilon}{2} \int_{\Sigma} u^2 e^{\frac{r^2}{4}} \, d\mathcal{H}^n.$$
To complete the proof we observe that computing as above gives that, for $\rho\geq R_0$, $$\begin{aligned}
\frac{d}{d\rho} \left(\rho^{-n} L_{\frac{3}{8}}(\rho)\right) &\leq \frac{3}{4} \rho^{-n+1} L_{\frac{3}{8}}(\rho)+2\rho^{-n} e^{\frac{3}{8} \rho^2} \int_{S_\rho} u \nabla_\Sigma u\cdot\frac{\nabla_\Sigma r}{|\nabla_\Sigma r|} \, d\mathcal{H}^{n-1}\\
&= \frac{3}{4} \rho^{-n+1} e^{\frac{3}{8} \rho^2} L(\rho)-\frac{3}{2} \rho^{-n}e^{\frac{\rho^2}{8} } \int_{\bar{E}_{\rho}} \left(|\nabla_\Sigma u|^2 +u \mathscr{L}_{\Sigma}^0 u \right) e^{\frac{r^2}{4} } \, d\mathcal{H}^n
\\ & -\frac{1}{2}\rho^{-n} \int_{\bar{E}_{\rho}} \left( |\nabla_\Sigma u|^2 + u \mathscr{L}_\Sigma^0 u + \frac{1}{4} r u \partial_r u \right) e^{\frac{3}{8} r^2} \, d\mathcal{H}^n\\
&\leq -\frac{1}{2} \rho^{-n} \int_{\bar{E}_{\rho}} |\nabla_\Sigma u|^2 e^{\frac{3}{8} r^2} \, d\mathcal{H}^n-\frac{1}{8}\rho^{-n}\int_{\bar{E}_\rho} r u \partial_r u \, e^{\frac{3}{8} r^2} \, d\mathcal{H}^n\\
&-\frac{3}{2} \rho^{-n}e^{\frac{ \rho^2}{8}} \int_{\bar{E}_{\rho}} u (\mathscr{L}_{\Sigma}^0 u) e^{\frac{r^2}{4} } \, d\mathcal{H}^n-\frac{1}{2}\rho^{-n} \int_{\bar{E}_{\rho}} u (\mathscr{L}_\Sigma^0 u) e^{\frac{3}{8} r^2} \, d\mathcal{H}^n\end{aligned}$$ where the last inequality uses the Poincaré inequality, Proposition \[GradControlsBoundaryProp\] and the integration by parts is justified by Proposition \[DecayEigenfunProp\]. Hence, as $\mu\leq \frac{1}{4}$ and $|A_{\Sigma}|^2\leq \frac{1}{4}$, $$\begin{aligned}
\frac{d}{d\rho} \left(\rho^{-n} L_{\frac{3}{8}}(\rho)\right) &\leq -\frac{1}{2} \rho^{-n} \int_{\bar{E}_{\rho}} |\nabla_\Sigma u|^2 e^{\frac{3}{8} r^2} \, d\mathcal{H}^n-\frac{1}{8} \rho^{-n}\int_{\bar{E}_\rho} r u \partial_r u \, e^{\frac{3}{8} r^2} \, d\mathcal{H}^n\\
&+\frac{3}{2} \rho^{-n}e^{\frac{\rho^2}{8} } \int_{\bar{E}_{\rho}}\left( -u L_{\Sigma} u +|A_\Sigma|^2 u^2-\frac{1}{2}u^2\right)e^{\frac{r^2}{4} } \, d\mathcal{H}^n\\
&+\frac{1}{2}\rho^{-n} \int_{\bar{E}_{\rho}} \left( -u L_{\Sigma} u +|A_\Sigma|^2 u^2-\frac{1}{2}u^2\right) e^{\frac{3}{8} r^2} \, d\mathcal{H}^n\\
&\leq -\frac{1}{2} \rho^{-n} \int_{\bar{E}_{\rho}} |\nabla_\Sigma u|^2 e^{\frac{3}{8} r^2} \, d\mathcal{H}^n-\frac{1}{8}\rho^{-n}\int_{\bar{E}_\rho} r u \partial_r u \, e^{\frac{3}{8} r^2} \, d\mathcal{H}^n.\end{aligned}$$
Rearranging terms and using the absorbing inequality give $$\frac{1}{4} \rho^{-n} \int_{\bar{E}_{\rho}} |\nabla_\Sigma u|^2 e^{\frac{3}{8} r^2} \, d\mathcal{H}^n \leq -\frac{d}{d\rho} \left(\rho^{-n} L_{\frac{3}{8}}(\rho)\right)+\frac{1}{64} \rho^{-n}\int_{\bar{E}_\rho} r^2 u^2 e^{\frac{3}{8} r^2}\, d\mathcal{H}^n.$$ Pick $\tau>0$ so that $\rho^2 e^{-\frac{\rho^2}{16}}\leq 1$ for $\rho\geq \tau$. Hence, for $\rho_2\geq R+\tau +\rho_1 +1\geq R_0\geq 2$, if $2\rho_2\geq \rho\geq \rho_2$, this together with gives, $$\begin{aligned}
\frac{1}{4} (2\rho_2)^{-n} & \int_{\bar{E}_{2\rho_2}} |\nabla_\Sigma u|^2 e^{\frac{3}{8} r^2} \, d\mathcal{H}^n \leq \frac{1}{4} \rho^{-n} \int_{\bar{E}_{\rho}} |\nabla_\Sigma u|^2 e^{\frac{3}{8} r^2} \, d\mathcal{H}^n\\
&\leq-\frac{d}{d\rho} \left(\rho^{-n}L_{\frac{3}{8}}(\rho)\right)+\frac{1}{64} \rho^{-n} \int_{\bar{E}_\rho} u^2 e^{\frac{7}{16} r^2} \, d\mathcal{H}^n\\
&\leq -\frac{d}{d\rho} \left(\rho^{-n} L_{\frac{3}{8}}(\rho)\right)+C(n) \rho^{-1} \rho_0^{-n} e^{-\frac{\rho^2}{16} +\frac{\rho^2_1}{4}} \int_{\Sigma} u^2 e^{\frac{r^2}{4}} \, d\mathcal{H}^n\\
&\leq -\frac{d}{d\rho} \left(\rho^{-n} L_{\frac{3}{8}}(\rho)\right)+C(n) \rho_2^{-1} \rho_0^{-n} e^{-\frac{\rho^2_2}{16} +\frac{\rho^2_1}{4}} \int_{\Sigma} u^2 e^{\frac{r^2}{4}} \, d\mathcal{H}^n.\end{aligned}$$
Pick $R^{\prime}=R^{\prime}(n, K_\Sigma,\epsilon)\geq R+\tau+\rho_1+2$ so that if $\rho_2\geq R^{\prime}\geq 2$, then $$C(n) \rho_2^{-1} \rho_0^{-n} e^{-\frac{\rho_2^2}{16}+\frac{\rho^2_1}{4}}\leq \frac{1}{64} (2\rho_2)^{-n} \epsilon.$$ Hence, for such $\rho_2$ and $2\rho_2\geq \rho \geq \rho_2$, $$\frac{1}{4} (2\rho_2)^{-n} \int_{\bar{E}_{2\rho_2}} |\nabla_\Sigma u|^2 e^{\frac{3}{8} r^2} \, d\mathcal{H}^n\leq -\frac{d}{d\rho} \left(\rho^{-n}L_{\frac{3}{8}}(\rho)\right)+ \frac{\epsilon}{64} (2\rho_2)^{-n}\int_{\Sigma} u^2 e^{\frac{r^2}{4}} \, d\mathcal{H}^n.$$
Integrating the above inequality in $\rho$ over $[\rho_2+1, \rho_2+2]\subset [\rho_2, 2\rho_2]$ yields $$\frac{1}{4} (2\rho_2)^{-n} \int_{\bar{E}_{2\rho_2}} |\nabla_\Sigma u|^2 e^{\frac{3}{8} r^2} \, d\mathcal{H}^n\leq (\rho_2+1)^{-n} L_{\frac{3}{8}}(\rho_2+1) +\frac{\epsilon}{64} (2\rho_2)^{-n} \int_{\Sigma} u^2 e^{\frac{r^2}{4}} \, d\mathcal{H}^n.$$ Applying with $\rho=\rho_2+1$ and $\gamma=\frac{3}{8}$ gives, $$\frac{1}{4} (2\rho_2)^{-n} \int_{\bar{E}_{2\rho_2}} |\nabla_\Sigma u|^2 e^{\frac{3}{8} r^2} \, d\mathcal{H}^n\leq 2 \rho_2^{-n} \int_{\bar{E}_{\rho_2}} u^2 e^{\frac{3}{8}r^2} \, d\mathcal{H}^n +\frac{\epsilon}{64}(2\rho_2)^{-n} \int_{\Sigma} u^2 e^{\frac{r^2}{4}} \, d\mathcal{H}^n.$$ Applying , gives $$\begin{aligned}
\rho_2^{-n}\int_{\bar{E}_{\rho_2}} u^2 e^{\frac{3}{8}r^2} d\mathcal{H}^n & \leq 2 C(n) \rho_2^{-1} \rho_0^{-n} e^{-\frac{\rho_2^2}{16} +\frac{\rho^2_1}{4}}\int_{\Sigma}u^2 e^{\frac{r^2}{4}} \, d\mathcal{H}^n\\
&\leq \frac{\epsilon}{32}(2\rho_2)^{-n} \int_{\Sigma}u^2 e^{\frac{r^2}{4}} \, d\mathcal{H}^n,\end{aligned}$$ Therefore, $$\int_{\bar{E}_{2\rho_2}} |\nabla_\Sigma u|^2 e^{\frac{3}{8} r^2} \, d\mathcal{H}^n\leq \frac{\epsilon}{2} \int_{\Sigma}u^2 e^{\frac{r^2}{4}} \, d\mathcal{H}^n.$$ The result is thus proved by setting $R_1=\max{\left\{2R,R^\prime\right\}}$.
We next prove estimates on certain integrals that will be needed in studying perturbation results for the quadratic forms.
\[EigenEst\] Let $\Sigma\in\mathcal{ACH}^{k,\alpha}_n$ be a self-expander. Fix a $\mu\leq \frac{1}{4}$. Given $\delta>0$, there is an $\epsilon_1=\epsilon_1(K_{\Sigma}, n, \delta)<\frac{1}{16}$ and a $C_1=C_1(K_{\Sigma}, n)$ so that for $\mathbf{f}\in \mathcal{B}_{\epsilon_1}(\mathbf{x}|_\Sigma; \mathcal{ACH}^{k,\alpha}_n(\Sigma))$ and $u\in W^1_{\frac{1}{4}}(\Sigma)$ that satisfies $$-L_{\Sigma} u=\mu u,$$ one has $$\int_{\Sigma} \left(|\nabla_\Sigma u|^2+u^2\right)\left( \Omega_{\mathbf{f}} +\Omega_{\mathbf{f}}^{-1}\right) e^{\frac{|\mathbf{x}|^2}{4}} \, d\mathcal{H}^n \leq C^2_1 \int_{\Sigma} u^2 e^{\frac{|\mathbf{x}|^2}{4}} \, d\mathcal{H}^n$$ and $$\int_{\Sigma} \left(|\nabla_\Sigma u|^2+ u^2\right) (1-\Omega_{\mathbf{f}})^2(1+\Omega_{\mathbf{f}}^{-1}) e^{\frac{|\mathbf{x}|^2}{4}} \, d\mathcal{H}^n\leq \delta^2 \int_{\Sigma} u^2 e^{\frac{|\mathbf{x}|^2}{4}} \, d\mathcal{H}^n.$$
First observe that, by taking $\epsilon_1<\frac{1}{16}$ one has a constant $C=C(n,K_{\Sigma})>2+K_\Sigma$ so that if $\mathbf{f}\in \mathcal{B}_{\epsilon_1}(\mathbf{x}|_\Sigma; \mathcal{ACH}^{k,\alpha}_n(\Sigma))$, then $$C^{-1}e^{-\frac{|\mathbf{x}|^2}{32}}\leq \Omega_{\mathbf{f}} \leq C e^{\frac{|\mathbf{x}|^2}{32}}$$ and $$C^{-1} g_{\Sigma}^{-1}\leq g_{\mathbf{f}}^{-1}\leq C g_{\Sigma}^{-1}.$$ We observe that, by the hypotheses, $$Q_{\Sigma}[u]=-B_{\Sigma}[u, L_\Sigma u]=\mu B_\Sigma[u]\leq \frac{1}{4} B_\Sigma[u]$$ and hence $$D_{\Sigma}[u]\leq Q_{\Sigma}[u] +\left(K_\Sigma-\frac{1}{2}\right) B_{\Sigma}[u]\leq K_\Sigma B_\Sigma[u].$$
Let $\epsilon=\epsilon(K_\Sigma,n,\delta)>0$ be a small constant whose precise value will be chosen at different points in the argument. By Lemma \[ConcentrationLem\], there is an $R_1=R_1(K_\Sigma,n,\epsilon)>0$ so that $$\label{kappaEst}
\int_{\Sigma\backslash B_{R_1}} \left(|\nabla_\Sigma u|^2 +u^2\right) e^{\frac{3}{8}|\mathbf{x}|^2} \, d\mathcal{H}^n\leq \epsilon^2\int_{\Sigma} u^2 e^{\frac{|\mathbf{x}|^2}{4}} \, d\mathcal{H}^n=\epsilon^2 B_\Sigma[u].$$ With $\epsilon=1$, one has $$\begin{aligned}
&\int_{\Sigma} \left(|\nabla_\Sigma u|^2+u^2\right) \left( \Omega_{\mathbf{f}} +\Omega_{\mathbf{f}}^{-1}\right) e^{\frac{|\mathbf{x}|^2}{4}} \, d\mathcal{H}^n \leq 2C\int_{\Sigma} \left(|\nabla_\Sigma u|^2+u^2\right) e^{\frac{5}{16}|\mathbf{x}|^2} \, d\mathcal{H}^n\\
&\leq 2Ce^{\frac{R_1^2}{16}}\int_{\Sigma\cap \bar{B}_{R}} \left(|\nabla_\Sigma u|^2+u^2\right) e^{\frac{|\mathbf{x}|^2}{4}} \, d\mathcal{H}^n+ 2C \int_{\Sigma\backslash B_{R_1}} \left(|\nabla_\Sigma u|^2+u^2\right) e^{\frac{5}{16}|\mathbf{x}|^2} \, d\mathcal{H}^n\\
&\leq 2Ce^{\frac{R_1^2}{16}}\left( D_\Sigma[u]+ B_\Sigma[u]\right) + 2C B_{\Sigma}[u]\\
&\leq 2C\left(e^{\frac{R_1^2}{16}}(K_\Sigma+1)+1\right) B_\Sigma[u].\end{aligned}$$ The first inequality follows by taking $$C_1^2=2C\left(e^{\frac{R_1^2}{16}}(K_\Sigma+1)+1\right)$$ and observing that $R_1$ and $C$ depend only on $n$ and $K_\Sigma$.
We next observe that, up to increasing $C$ a bit we have $$(1-\Omega_{\mathbf{f}})^2(1+\Omega_{\mathbf{f}}^{-1})e^{\frac{|\mathbf{x}|^2}{4}} \leq 2 C^2 e^{ \frac{11}{32}|\mathbf{x}|^2}\leq2 C^2 e^{ \frac{3}{8}|\mathbf{x}|^2}\ .$$ In particular, by one has for a large enough $R_1=R_1(K_\Sigma,n,\epsilon^2)>0$, $$\int_{\Sigma\backslash B_{R_1}} \left(|\nabla_\Sigma u|^2+ u^2\right) (1-\Omega_{\mathbf{f}})^2 (1+\Omega_{\mathbf{f}}^{-1}) e^{\frac{|\mathbf{x}|^2}{4}} d\mathcal{H}^n\leq \epsilon^2 B_\Sigma[u]$$ where we will again fix $\epsilon=\epsilon(K_\Sigma, n, \delta)$ later. Notice that by taking $\epsilon_1=\epsilon_1(K_{\Sigma},n,\epsilon)$ sufficiently small one has that on $B_{R_1}$, $$(1-\Omega_{\mathbf{f}})^2 (1+\Omega_{\mathbf{f}}^{-1})\leq \epsilon^2.$$ Hence, $$\begin{aligned}
&\int_{\Sigma\cap \bar{B}_{R_1}} \left(|\nabla_\Sigma u|^2+ u^2\right) (1-\Omega_{\mathbf{f}})^2 (1+\Omega_{\mathbf{f}}^{-1}) e^{\frac{|\mathbf{x}|^2}{4}} \, d\mathcal{H}^n \\
& \leq \epsilon^2 \int_{\Sigma\cap \bar{B}_{R_1}} \left(|\nabla_\Sigma u|^2+ u^2\right) e^{\frac{|\mathbf{x}|^2}{4}} \, d\mathcal{H}^n \\
& \leq \epsilon^2 \left( D_{\Sigma}[u]+B_\Sigma[u]\right)\leq \epsilon^2 (K_\Sigma+1) B_{\Sigma}[u].\end{aligned}$$ As such, the second inequality follows by taking $
\epsilon^2= \frac{\delta^2}{2(K_\Sigma+1)}< \frac{\delta^2}{2}.
$
\[EstProp\] Let $\Sigma\in\mathcal{ACH}^{k,\alpha}_n$ be a self-expander. Fix a $\beta\in \left[\frac{1}{4}, \frac{3}{8}\right], \mu\leq \frac{1}{4}$ and $\delta\in\left(0, \frac{1}{2}\right)$. Let $u\in W^1_{\frac{1}{4}}(\Sigma)\setminus{\left\{0\right\}}$ satisfy $$-L_{\Sigma} u=\mu u.$$ There is an $\epsilon_1=\epsilon_1(K_\Sigma, n, \delta)>0$ so that if $\mathbf{f}\in \mathcal{B}_{\epsilon_1}(\mathbf{x}|_\Sigma;\mathcal{ACH}^{k,\alpha}_n(\Sigma))$, then, for any $v\in W^1_\beta(\Sigma)$, $$\left| B_{\mathbf{f}}[u,v]-B_{\Sigma}[u,v]\right|^2\leq \delta^2 B_{\Sigma} [u] B_{\Sigma}[v]$$ and $$\left| B_{\mathbf{f}}[u,v]-B_{\Sigma}[u,v]\right|^2\leq \delta^2 B_{\Sigma} [u] B_{\mathbf{f}}[v]\leq 2\delta^2 B_{\mathbf{f}}[u]B_{\mathbf{f}}[v].$$ In addition, one has $$\left| Q_{\mathbf{f}}[u,v]-Q_{\Sigma}[u,v]\right|^2\leq \delta^2 B_{\Sigma} [u] \left( D_{\Sigma}[v]+B_\Sigma[v]\right)$$ and $$\left| Q_{\mathbf{f}}[u,v]-Q_{\Sigma}[u,v]\right|^2\leq \delta^2 B_{\Sigma} [u] \left( D_{\mathbf{f}}[v]+B_{\mathbf{f}}[v]\right).$$
Let $\delta^\prime=\delta^\prime(K_\Sigma, n, \delta)$ be a small positive number which we will specify in the course of the proof. Using this $\delta^\prime$, pick $\epsilon_1=\epsilon_1(K_\Sigma, n, \delta^\prime)<\frac{1}{16}$ as in Lemma \[EigenEst\] for the given $u$. As long as $\delta^\prime\leq \delta$, one has $$\begin{aligned}
\left| B_{\mathbf{f}}[u,v]-B_\Sigma[u,v]\right|^2 &=\left| \int_{\Sigma} u v (\Omega_{\mathbf{f}}-1) e^{\frac{|\mathbf{x}|^2}{4}} \, d\mathcal{H}^n\right|^2\\
&\leq \int_{\Sigma} u^2 (\Omega_{\mathbf{f}}-1)^2 e^{\frac{|\mathbf{x}|^2}{4}} \, d\mathcal{H}^n\int_{\Sigma} v^2 e^{\frac{|\mathbf{x}|^2}{4}} \, d\mathcal{H}^n \\
&\leq (\delta')^2 B_\Sigma[u] B_\Sigma[v]\leq \delta^2 B_\Sigma[u] B_\Sigma[v]. \end{aligned}$$ Here the last inequality follows from Lemma \[EigenEst\]. Similarly, one has $$\begin{aligned}
\left| B_{\mathbf{f}}[u,v]-B_\Sigma[u,v]\right|^2 &=\left| \int_{\Sigma} u v (\Omega_{\mathbf{f}}-1) e^{\frac{|\mathbf{x}|^2}{4}} \, d\mathcal{H}^n\right|^2\\
&\leq \int_{\Sigma} u^2 \Omega_{\mathbf{f}}^{-1} (\Omega_{\mathbf{f}}-1)^2 e^{\frac{|\mathbf{x}|^2}{4}} \, d\mathcal{H}^n\int_{\Sigma} v^2 \Omega_{\mathbf{f}} e^{\frac{|\mathbf{x}|^2}{4}}\, d\mathcal{H}^n \\
&\leq (\delta')^2 B_\Sigma[u] B_\mathbf{f}[v]\leq \delta^2 B_\Sigma[u] B_\mathbf{f}[v].\end{aligned}$$ Here again the first part of the second inequality follows from Lemma \[EigenEst\]. The final inequality follows by observing that by taking $u=v$ what we have already shown gives $$\begin{aligned}
\left| B_{\mathbf{f}}[u]-B_{\Sigma}[u]\right|^2 &\leq \delta^2 B_{\mathbf{f}}[u] B_{\Sigma}[u] \leq \delta^2 B_{\mathbf{f}}[u]^2 + \delta^2 B_{\mathbf{f}}[u] \left( B_{\Sigma}[u] -B_{\mathbf{f}}[u]\right)\\
&\leq \frac{3}{2} \delta^2B_{\mathbf{f}}[u]^2 +\frac{1}{2} \left| B_{\mathbf{f}}[u]-B_{\Sigma}[u]\right|^2\end{aligned}$$ where the last inequality follows from the absorbing inequality. Hence, $$\left| B_{\mathbf{f}}[u]-B_{\Sigma}[u]\right|^2\leq 3 \delta^2 B_{\mathbf{f}}[u]^2\leq B_{\mathbf{f}}[u]^2$$ and so $$B_{\Sigma}[u]\leq B_{\mathbf{f}}[u] + B_{\mathbf{f}}[u]\leq 2 B_{\mathbf{f}}[u].$$ This verifies the second part of the inequality.
To prove the second pair of inequalities one first observes that, by shrinking $\epsilon_1$ and $\delta^\prime$ as needed, one can ensure that $$\frac{1}{2} g_\Sigma^{-1} \leq \left(1- \delta' \right) g_\Sigma^{-1} \leq g_{\mathbf{f}}^{-1}\leq \left(1+ \delta^\prime\right) g_\Sigma^{-1}\leq 2 g_\Sigma^{-1}$$ and $$-\delta^\prime+|A_{\Sigma}|^2\leq |A_{\mathbf{f}(\Sigma)}\circ \mathbf{f}|^2 \leq \delta^\prime+|A_{\Sigma}|^2\leq K_{\Sigma}+1.$$ As such, one computes, $$\begin{aligned}
& \left|D_{\mathbf{f}}[u,v]-D_\Sigma[u,v]\right| = \left|\int_{\Sigma} \left( \Omega_{\mathbf{f} }g_{\mathbf{f}}^{-1} -g_{\mathbf{f}}^{-1}+g_{\mathbf{f}}^{-1}-g_\Sigma^{-1}\right)^{ij}\nabla_i u \nabla_j v \, e^{\frac{|\mathbf{x}|^2}{4}} \, d\mathcal{H}^n \right|\\
&\leq \int_{\Sigma} \left|(g_{\mathbf{f}}^{-1})^{ij} \nabla_i u \nabla_j v \right| \left| \Omega_{\mathbf{f}}-1\right| e^{\frac{|\mathbf{x}|^2}{4}} \, d\mathcal{H}^n +\int_{\Sigma} \left| \left(g_{\mathbf{f}}^{-1} -g_{\Sigma}^{-1}\right)^{ij} \nabla_i u \nabla_j v \right| e^{\frac{|\mathbf{x}|^2}{4}} \, d\mathcal{H}^n\\
&\leq 2\int_{\Sigma} |\nabla_{\Sigma} u| |\nabla_\Sigma v| \left| \Omega_{\mathbf{f}}-1\right| e^{\frac{|\mathbf{x}|^2}{4}} \, d\mathcal{H}^n+\delta^\prime\int_{\Sigma} \left| \nabla_\Sigma u \right| \left|\nabla_\Sigma v \right| e^{\frac{|\mathbf{x}|^2}{4}} \, d\mathcal{H}^n\\
&\leq 2 D_{\Sigma}[v]^{\frac{1}{2}} \left(\int_{\Sigma}|\nabla_\Sigma u|^2 (\Omega_{\mathbf{f}}-1)^2 e^{\frac{|\mathbf{x}|^2}{4}} \, d\mathcal{H}^n\right)^{\frac{1}{2}}+\delta^\prime D_\Sigma[v]^{\frac{1}{2}} D_\Sigma[u]^{\frac{1}{2}}\\
&\leq \delta^\prime \left(2+K_\Sigma^{\frac{1}{2}}\right) D_{\Sigma}[v]^{\frac{1}{2}}B_\Sigma[u]^{\frac{1}{2}}. \end{aligned}$$ Similarly, one has $$\begin{aligned}
\left|D_\mathbf{f}[u,v]-D_\Sigma[u,v]\right| & \leq 2D_\mathbf{f}[v]^{\frac{1}{2}} \left(\int_\Sigma |\nabla_\Sigma u|^2 (\Omega_\mathbf{f}-1)^2\Omega_\mathbf{f}^{-1} e^{\frac{|\mathbf{x}|^2}{4}} \, d\mathcal{H}^n\right)^{\frac{1}{2}}\\
&+2\delta^\prime D_\mathbf{f}[v]^{\frac{1}{2}} \left(\int_\Sigma |\nabla_\Sigma u|^2 \Omega^{-1}_\mathbf{f} e^{\frac{|\mathbf{x}|^2}{4}} \, d\mathcal{H}^n\right)^{\frac{1}{2}} \\
& \leq 2\delta^\prime (C_1+1)D_\mathbf{f}[v]^{\frac{1}{2}} B_\Sigma[u]^{\frac{1}{2}}.\end{aligned}$$ Hence, by taking $$\delta^\prime\leq \frac{\delta}{8}\min{\left\{\left(2+K_\Sigma^{1/2}\right)^{-1}, (C_1+1)^{-1}\right\}},$$ one has $$|D_{\mathbf{f}}[u,v]-D_\Sigma[u,v]| \leq \frac{\delta}{4} B_\Sigma[u]^{\frac{1}{2}} \min{\left\{D_{\Sigma}[v]^{\frac{1}{2}},D_\mathbf{f}[v]^{\frac{1}{2}}\right\}}.$$
To complete the proof of the third inequality we observe that $$\begin{aligned}
\left| Q_{\mathbf{f}}[u,v]-Q_{\Sigma}[u,v]\right| &\leq \left|D_\mathbf{f}[u,v]-D_{\Sigma}[u,v]\right| +\frac{1}{2} \left| B_\mathbf{f}[u,v]-B_{\Sigma}[u,v]\right|\\
&+\left| \int_{\Sigma} uv \left(|A_{\mathbf{f}(\Sigma)}\circ \mathbf{f}|^2 \Omega_{\mathbf{f}} - |A_{\Sigma}|^2\right) e^{\frac{|\mathbf{x}|^2}{4}} d\mathcal{H}^n\right|\\
& \leq \left|D_\mathbf{f}[u,v]-D_{\Sigma}[u,v]\right| +\frac{1}{2} \left| B_\mathbf{f}[u,v]-B_{\Sigma}[u,v]\right|\\
&+(K_\Sigma+1) \int_{\Sigma} \left|uv(\Omega_{\mathbf{f}}-1)\right| e^{\frac{|\mathbf{x}|^2}{4}} d\mathcal{H}^n+ \delta' \int_{\Sigma}|uv| e^{\frac{|\mathbf{x}|^2}{4}} \, d\mathcal{H}^n.\end{aligned}$$ Combining the previous estimates with the Cauchy-Schwarz inequality and Lemma \[EigenEst\] gives $$\left| Q_{\mathbf{f}}[u,v]-Q_{\Sigma}[u,v]\right| \leq \frac{\delta}{4} B_\Sigma[u]^{\frac{1}{2}} D_{\Sigma}[v]^{\frac{1}{2}} +\left(\frac{\delta^\prime}{2}+(K_{\Sigma}+2) \delta^\prime\right) B_\Sigma[u]^{\frac{1}{2}} B_\Sigma[v]^{\frac{1}{2}}.$$ Taking $\delta^\prime\leq \frac{\delta}{4(K_{\Sigma}+2)} \leq \frac{\delta}{4}$ one has $$\begin{aligned}
\left| Q_{\mathbf{f}}[u,v]-Q_{\Sigma}[u,v]\right| &\leq \frac{\delta}{2} B_\Sigma[u]^{\frac{1}{2}} \left(D_{\Sigma}[v]^{\frac{1}{2}} + B_\Sigma[v]^{\frac{1}{2}}\right)\\
&\leq \delta B_\Sigma[u]^{\frac{1}{2}} \left(D_{\Sigma}[v] + B_\Sigma[v]\right)^{\frac{1}{2}},\end{aligned}$$ which proves the third inequality.
The fourth inequality is proved in a similar fashion. Namely, one estimates as previous: $$\begin{aligned}
\left|Q_\mathbf{f}[u,v]-Q_\Sigma[u,v]\right| & \leq \left|D_\mathbf{f}[u,v]-D_{\Sigma}[u,v]\right| +\frac{1}{2} \left| B_\mathbf{f}[u,v]-B_{\Sigma}[u,v]\right| \\
& +(K_\Sigma+1) B_\mathbf{f}[v]^{\frac{1}{2}} \left(\int_\Sigma u^2 (\Omega_\mathbf{f}-1)^2\Omega_\mathbf{f}^{-1} e^{\frac{|\mathbf{x}|^2}{4}} \, d\mathcal{H}^n\right)^{\frac{1}{2}} \\
& + \delta^\prime B_\mathbf{f}[v]^{\frac{1}{2}} \left( \int_\Sigma u^2 \Omega_\mathbf{f}^{-1} e^{\frac{|\mathbf{x}|^2}{4}} \, d\mathcal{H}^n\right)^{\frac{1}{2}} \\
& \leq \frac{\delta}{4} B_\Sigma[u]^{\frac{1}{2}} D_\mathbf{f}[v]^{\frac{1}{2}} +\left(\frac{\delta^\prime}{2}+(K_\Sigma+1+C_1^{1/2})\delta^\prime\right) B_\Sigma[u]^{\frac{1}{2}} B_\mathbf{f}[v]^{\frac{1}{2}}.\end{aligned}$$ Hence, taking $\delta^\prime \leq \frac{\delta}{4\left(K_\Sigma+1+C_1^{1/2}\right)}\leq\frac{\delta}{4}$ one has the fourth inequality by the Cauchy-Schwarz inequality.
\[EigenSpacePerturbCor\] Let $\Sigma\in\mathcal{ACH}^{k,\alpha}_n$ be a self-expander, and let $\mathbf{v}\in C^{k,\alpha}_0\cap C^{k}_{0,\mathrm{H}}(\Sigma;\mathbb{R}^{n+1})$ be a transverse section on $\Sigma$ so that $\mathcal{C}[\mathbf{v}]$ is transverse to $\mathcal{C}(\Sigma)$. Let $$C_{\Sigma,\mathbf{v}}=\sup_{\Sigma} |\mathbf{v}\cdot\mathbf{n}_\Sigma|^{-1}+\Vert\mathbf{v}\Vert_{1}.$$ Fix a $\beta\in \left[\frac{1}{4}, \frac{3}{8}\right], \mu\leq \frac{1}{4}$ and $\delta\in \left(0,\frac{1}{2}\right)$. Suppose $u\in W^1_{\frac{1}{4}}(\Sigma)\setminus{\left\{0\right\}}$ satisfies $$-L_{\Sigma, \mathbf{v}}^\prime u=\mu u.$$ There is an $\epsilon_2=\epsilon_2(K_\Sigma, C_{\Sigma,\mathbf{v}}, n, \delta)>0$ so that if $\mathbf{f}\in \mathcal{B}_{\epsilon_2}(\mathbf{x}|_\Sigma; \mathcal{ACH}^{k,\alpha}_n(\Sigma))$, then, for any $v\in W^1_{\beta}(\Sigma)$, $$\left|B_{\mathbf{f}, \mathbf{v}}[u,v]-B_{\Sigma, \mathbf{v}}[u,v]\right|^2\leq \delta^2 B_{\Sigma, \mathbf{v}} [u] B_{\Sigma, \mathbf{v}}[v]$$ and $$\left|B_{\mathbf{f}, \mathbf{v}}[u,v]-B_{\Sigma, \mathbf{v}}[u,v]\right|^2\leq \delta^2 B_{\Sigma, \mathbf{v}} [u] B_{\mathbf{f}, \mathbf{v}}[v]\leq 2\delta^2 B_{\mathbf{f},\mathbf{v}}[u]B_{\mathbf{f},\mathbf{v}}[v].$$ Likewise, $$\left| Q_{\mathbf{f}, \mathbf{v}}[u,v]-Q_{\Sigma, \mathbf{v}}[u,v]\right|^2\leq \delta^2 B_{\Sigma, \mathbf{v}} [u] \left(D_{\Sigma,\mathbf{v}}[v]+B_{\Sigma, \mathbf{v}}[v]\right)$$ and $$\left| Q_{\mathbf{f}, \mathbf{v}}[u,v]-Q_{\Sigma, \mathbf{v}}[u,v]\right|^2\leq \delta^2 B_{\Sigma, \mathbf{v}} [u] \left(D_{\mathbf{f},\mathbf{v}}[v]+B_{\mathbf{f}, \mathbf{v}}[v]\right).$$
We first observe that if $-L_{\Sigma,\mathbf{v}}^\prime u=\mu u$, then $U=(\mathbf{v}\cdot\mathbf{n}_\Sigma) u$ satisfies $$-L_\Sigma U=\mu U.$$ If $V=(\mathbf{v}\cdot\mathbf{n}_\Sigma) v$, then $$B_{\Sigma,\mathbf{v}}[u,v]=B_\Sigma[U,V], D_{\Sigma,\mathbf{v}}[u,v]=D_\Sigma[U,V] \mbox{ and } Q_{\Sigma,\mathbf{v}}[u,v]=Q_\Sigma[U,V].$$ Let $\Lambda=\mathbf{f}(\Sigma)$. There is a $C=C(K_\Sigma,C_{\Sigma,\mathbf{v}},n)$ so that $$\Vert\mathbf{v}\cdot(\mathbf{n}_\Lambda\circ\mathbf{f})-\mathbf{v}\cdot\mathbf{n}_\Sigma\Vert_1\leq C\epsilon_2.$$ Thus, by choosing $\epsilon_2$ sufficiently small, we have that $$\frac{1}{4}B_\mathbf{f}[V]\leq B_{\mathbf{f},\mathbf{v}}[v] \leq 4 B_\mathbf{f}[V]$$ and $$\frac{1}{4}\left(B_{\mathbf{f}}[V]+D_\mathbf{f}[V]\right) \leq B_{\mathbf{f},\mathbf{v}}[v]+D_{\mathbf{f},\mathbf{v}}[v] \leq 4\left(B_{\mathbf{f}}[V]+D_\mathbf{f}[V]\right).$$ And there is a $C^\prime$ depending on $K_\Sigma,C_{\Sigma,\mathbf{v}}$ and $n$ so that $$\label{EigenPerturbEqn}
\left|B_{\mathbf{f},\mathbf{v}}[u,v]-B_\mathbf{f}[U,V]\right| \leq C^\prime\epsilon_2 B_\mathbf{f}[|U|,|V|]$$ and $$\begin{aligned}
\left|Q_{\mathbf{f},\mathbf{v}}[u,v]-Q_\mathbf{f}[U,V]\right| \leq C^\prime\epsilon_2 & \left(B_\mathbf{f}[|U|,|V|]+B_\mathbf{f}[|\nabla_\Sigma U|,|V|]\right. \\
& \left.+B_\mathbf{f}[|U|,|\nabla_\Sigma V|]+B_\mathbf{f}[|\nabla_\Sigma U|,|\nabla_\Sigma V|]\right).\end{aligned}$$
By the Cauchy-Schwarz inequality and Lemma \[EigenEst\], for $\epsilon_2$ sufficiently small $$B_\mathbf{f}[|U|,|V|] \leq B_\Sigma[U(\Omega_\mathbf{f}-1)]^{\frac{1}{2}} B_\Sigma [V]^{\frac{1}{2}}+B_\Sigma[U]^{\frac{1}{2}} B_\Sigma[V]^{\frac{1}{2}} \leq 2B_\Sigma[U]^{\frac{1}{2}} B_\Sigma[V]^{\frac{1}{2}}.$$ Thus, by choosing $\epsilon_2$ sufficiently small $$\left|B_{\mathbf{f},\mathbf{v}}[u,v]-B_\mathbf{f}[U,V]\right| \leq \frac{\delta}{2} B_\Sigma[U]^{\frac{1}{2}} B_\Sigma[V]^{\frac{1}{2}}.$$ Hence, by Proposition \[EstProp\] and possibly shrinking $\epsilon_2$, $$\begin{aligned}
\left| B_{\mathbf{f},\mathbf{v}}[u,v]-B_{\Sigma,\mathbf{v}}[u,v]\right| & \leq \left|B_{\mathbf{f},\mathbf{v}}[u,v]-B_\mathbf{f}[U,V]\right|+\left| B_\mathbf{f}[U,V]-B_\Sigma[U,V]\right| \\
& \leq \delta B_\Sigma[U]^{\frac{1}{2}} B_\Sigma [V]^{\frac{1}{2}}=\delta B_{\Sigma,\mathbf{v}}[u]^{\frac{1}{2}} B_{\Sigma,\mathbf{v}}[v]^{\frac{1}{2}},\end{aligned}$$ proving the first inequality.
To prove the second inequalities, one again uses the Cauchy-Schwarz inequality and Lemma \[EigenEst\] to compute, $$\begin{aligned}
B_\mathbf{f}[|U|,|V|] & \leq B_\Sigma\left[U(\Omega_\mathbf{f}-1)\Omega_\mathbf{f}^{-\frac{1}{2}}\right]^{\frac{1}{2}} B_\mathbf{f}[V]^{\frac{1}{2}}+B_\Sigma \left[U\Omega_\mathbf{f}^{-\frac{1}{2}}\right]^{\frac{1}{2}} B_\mathbf{f}[V]^{\frac{1}{2}} \\
& \leq (1+C_1)B_\Sigma[U]^{\frac{1}{2}} B_\mathbf{f}[V]^{\frac{1}{2}}.\end{aligned}$$ Thus, by choosing $\epsilon_2$ sufficiently small gives $$\left|B_{\mathbf{f},\mathbf{v}}[u,v]-B_\mathbf{f}[U,V]\right| \leq \frac{\delta}{4} B_\Sigma[U]^{\frac{1}{2}} B_\mathbf{f}[V]^{\frac{1}{2}}.$$ Hence, invoking Proposition \[EstProp\] and shrinking $\epsilon_2$ (if needed) one has $$\left| B_{\mathbf{f},\mathbf{v}}[u,v]-B_{\Sigma,\mathbf{v}}[u,v]\right| \leq \frac{\delta}{2} B_\Sigma[U]^{\frac{1}{2}} B_\mathbf{f}[V]^{\frac{1}{2}} \leq \delta B_{\Sigma,\mathbf{v}}[u]^{\frac{1}{2}} B_{\mathbf{f},\mathbf{v}}[v]^{\frac{1}{2}}.$$ Now taking $u=v$ it follows that $$\begin{aligned}
\left|B_{\mathbf{f},\mathbf{v}}[u]-B_{\Sigma,\mathbf{v}}[u]\right|^2 & \leq \delta^2 B_{\Sigma,\mathbf{v}}[u]B_{\mathbf{f},\mathbf{v}}[u]\leq \delta^2 B_{\mathbf{f},\mathbf{v}}[u]^2+\delta^2 B_{\mathbf{f},\mathbf{v}}[u]\left|B_{\mathbf{f},\mathbf{v}}[u]-B_{\Sigma,\mathbf{v}}[u]\right| \\
& \leq \frac{3}{2} \delta^2 B_{\mathbf{f},\mathbf{v}}[u]^2+\frac{1}{2} \left|B_{\mathbf{f},\mathbf{v}}[u]-B_{\Sigma,\mathbf{v}}[u]\right|^2.\end{aligned}$$ Thus, as $\delta<1/2$ we have $$\left|B_{\mathbf{f},\mathbf{v}}[u]-B_{\Sigma,\mathbf{v}}[u]\right|^2 \leq 3\delta^2 B_{\mathbf{f},\mathbf{v}}[u]^2 \leq B_{\mathbf{f},\mathbf{v}}[u]^2$$ and so $$B_{\Sigma,\mathbf{v}}[u] \leq B_{\mathbf{f},\mathbf{v}}[u]+\left|B_{\mathbf{f},\mathbf{v}}[u]-B_{\Sigma,\mathbf{v}}[u]\right| \leq 2B_{\mathbf{f},\mathbf{v}}[u].$$ This completes the proof of the second inequalities.
The other pair of inequalities are proved in a similar fashion.
\[PerturbCor\] Let $\Sigma\in\mathcal{ACH}^{k,\alpha}_n$ be a self-expander, and let $\mathbf{v}\in C^{k,\alpha}_0\cap C^{k}_{0,\mathrm{H}}(\Sigma;\mathbb{R}^{n+1})$ be a transverse section on $\Sigma$ so that $\mathcal{C}[\mathbf{v}]$ is transverse to $\mathcal{C}(\Sigma)$. Let $$C_{\Sigma,\mathbf{v}}=\sup_{\Sigma}|\mathbf{v}\cdot\mathbf{n}_\Sigma|^{-1}+\Vert\mathbf{v}\Vert_1.$$ Fix a $\beta\in \left[\frac{1}{4}, \frac{3}{8}\right]$, $\mu\leq \frac{1}{4}$ and $\delta\in \left(0,\frac{1}{2}\right)$. Let $V$ be a subspace of $W^1_{\frac{1}{4}}(\Sigma)$ spanned by $N$ linearly independent eigenfunctions of $-L_{\Sigma,\mathbf{v}}^\prime$ with eigenvalues less than or equal to $\mu$. There is an $\epsilon_3=\epsilon_3(K_\Sigma, C_{\Sigma,\mathbf{v}},n, \delta, N)>0$ so that if $\mathbf{f}\in \mathcal{B}_{\epsilon_3}(\mathbf{x}|_\Sigma;\mathcal{ACH}^{k,\alpha}_n(\Sigma))$, then, for any $u\in V\setminus{\left\{0\right\}}$ and $v\in W^1_\beta(\Sigma)$, $$\left|B_{\mathbf{f}, \mathbf{v}}[u,v]-B_{\Sigma, \mathbf{v}}[u,v]\right|^2\leq \delta^2 B_{\Sigma, \mathbf{v}} [u] B_{\Sigma, \mathbf{v}}[v]$$ and $$\left|B_{\mathbf{f}, \mathbf{v}}[u,v]-B_{\Sigma, \mathbf{v}}[u,v]\right|^2\leq \delta^2 B_{\Sigma, \mathbf{v}}[u]B_{\mathbf{f}, \mathbf{v}} [v]\leq 2\delta^2 B_{\mathbf{f}, \mathbf{v}} [u]B_{\mathbf{f}, \mathbf{v}} [v].$$ Likewise, $$\left| Q_{\mathbf{f}, \mathbf{v}}[u,v]-Q_{\Sigma, \mathbf{v}}[u,v]\right|^2\leq \delta^2 B_{\Sigma, \mathbf{v}} [u] \left(D_{\Sigma,\mathbf{v}}[v]+B_{\Sigma, \mathbf{v}}[v]\right)$$ and $$\left| Q_{\mathbf{f}, \mathbf{v}}[u,v]-Q_{\Sigma, \mathbf{v}}[u,v]\right|^2\leq \delta^2 B_{\Sigma, \mathbf{v}} [u] \left(D_{\mathbf{f},\mathbf{v}}[v]+B_{\mathbf{f}, \mathbf{v}}[v]\right).$$
Pick a $B_{\Sigma, \mathbf{v}}$-orthonormal basis, $u_1, \ldots, u_N$, of $V$ of eigenfunctions of $-L_{\Sigma, \mathbf{v}}^\prime$. This is possible as $-L_{\Sigma, \mathbf{v}}^\prime$ is self-adjoint with respect to $B_{\Sigma, \mathbf{v}}$. In particular, $Q_{\Sigma, \mathbf{v}}[u_i,u_j]=\mu_i \delta_{ij}$ where $\mu_i$ is the eigenvalue of $u_i$. That is, this orthonormal basis also diagonalizes $Q_{\Sigma, \mathbf{v}}$. Pick $\epsilon_3$ so for any $u_i$ one has $$\left|B_{\mathbf{f}, \mathbf{v}}[u_i,v]-B_{\Sigma, \mathbf{v}}[u_i,v]\right|^2\leq \frac{\delta^2}{N} B_{\Sigma, \mathbf{v}}[u_i] B_{\Sigma, \mathbf{v}}[v].$$
Write $u=\sum_{i=1}^N a_i u_i$. One has, by the Cauchy-Schwarz inequality, $$\begin{aligned}
\left|B_{\mathbf{f}, \mathbf{v}}[u,v]-B_{\Sigma, \mathbf{v}}[u,v]\right|^2 &= \left| \sum_{i=1}^N a_i \left(B_{\mathbf{f}, \mathbf{v}}[u_i,v]- B_{\Sigma, \mathbf{v}}[u_i,v]\right)\right|^2\\
&\leq N \sum_{i=1}^N a_i^2 \left|B_{\mathbf{f}, \mathbf{v}}[u_i,v]-B_{\Sigma, \mathbf{v}}[u_i,v]\right|^2\\
&\leq N \sum_{i=1}^N a_i^2 \frac{\delta^2}{N} B_{\Sigma, \mathbf{v}}[u_i] B_{\Sigma, \mathbf{v}}[v]=\delta^2 B_{\Sigma, \mathbf{v}}[u] B_{\Sigma, \mathbf{v}}[v]\end{aligned}$$ where the last equality used that ${\left\{u_i\right\}}$ is an orthonormal basis.
The proof of remaining inequalities proceed in a similar fashion.
Stability of bilinear form {#StableFormSec}
==========================
In this section we show the stability of the positive and negative spaces of the bilinear form $Q_{\mathbf{f},\mathbf{v}}$ as one perturbs $\mathbf{f}$ around $\mathbf{x}|_{\Sigma}$ in $\mathcal{ACH}^{k,\alpha}_n(\Sigma)$ for $\Sigma$ an asymptotically conical self-expander. This is a non-trivial result as such a perturbation can radically change the weight that appears. In what follows we will use the (easily checked) fact that if $\mathbf{f}\in \mathcal{B}_{\frac{1}{8}}(\mathbf{x}|_\Sigma; \mathcal{ACH}^{k,\alpha}_n(\Sigma))$, then for all $u\in W^1_{\frac{3}{8}}(\Sigma)$ one has $$Q_{\mathbf{f}, \mathbf{v}}[u]+D_{\mathbf{f}, \mathbf{v}}[u]+ B_{\mathbf{f}, \mathbf{v}}[u]<\infty.$$ This explains the choice of weighted space that will appear below.
We define $$\mathcal{E}_{\Sigma, \mathbf{v}}^\prime[\mu]=\mathrm{span} {\left\{u\in W^1_{\frac{1}{4}}(\Sigma)\colon -L_{\Sigma, \mathbf{v}}^\prime u = \mu^\prime u \mbox{ for some $\mu^\prime\leq \mu$}\right\}}$$ to be the subspace of $W^1_{\frac{1}{4}}(\Sigma)$ spanned by eigenfunctions of $-L_{\Sigma, \mathbf{v}}^\prime$ with eigenvalues less than or equal to $\mu$. We also let $$\mathcal{E}_{\Sigma, \mathbf{v}}^\prime(\mu)=\mathrm{span} {\left\{u\in W^1_{\frac{1}{4}}(\Sigma)\colon -L_{\Sigma, \mathbf{v}}^\prime u = \mu^\prime u \mbox{ for some $\mu^\prime< \mu$}\right\}}$$ be the space spanned by eigenfunctions with eigenvalues strictly less than $\mu$. Let $\mathcal{N}_{\Sigma, \mathbf{v}}^\prime=\mathcal{E}_{\Sigma, \mathbf{v}}^\prime(0)$ be the space spanned by the eigenfunctions with negative eigenvalues and let $$\mathcal{K}_{\Sigma, \mathbf{v}}^\prime=\mathcal{E}_{\Sigma, \mathbf{v}}^\prime [0]\setminus \mathcal{E}_{\Sigma, \mathbf{v}}^\prime(0)=\mathcal{K}_{\mathbf{v}}$$ be the space of elements in $W^1_{\frac{1}{4}}(\Sigma)$ in the kernel of $L_{\Sigma, \mathbf{v}}^\prime$. Observe that, by Proposition \[DecayEigenfunProp\], for any $\mu$, $$\mathcal{E}_{\Sigma, \mathbf{v}}^\prime(\mu)\subset \mathcal{E}_{\Sigma, \mathbf{v}}^\prime[\mu]\subset W^1_{\frac{3}{8}}(\Sigma).$$ and so the same is true of $\mathcal{N}_{\Sigma, \mathbf{v}}^\prime$ and $\mathcal{K}_{\Sigma, \mathbf{v}}^\prime$. Using this fact, let $$\mathcal{P}_{\Sigma, \mathbf{v}}^\prime={\left\{f\in W^1_{\frac{3}{8}}(\Sigma)\colon B_{\Sigma, \mathbf{v}}[f, u]=0, \forall u\in \mathcal{E}_{\Sigma, \mathbf{v}}^\prime[0]\right\}}.$$ By construction and the fact that $-L_{\Sigma, \mathbf{v}}^\prime$ is self-adjoint with respect to $B_{\Sigma, \mathbf{v}}$ and has a discrete spectrum, $$W^1_{\frac{3}{8}}(\Sigma)=\mathcal{N}_{\Sigma, \mathbf{v}}^\prime\oplus\mathcal{K}_{\Sigma, \mathbf{v}}^\prime\oplus\mathcal{P}_{\Sigma, \mathbf{v}}^\prime$$ and these three subspaces are orthogonal with respect to $B_{\Sigma, \mathbf{v}}$. Moreover, for all $u\in \mathcal{N}_{\Sigma, \mathbf{v}}^\prime$, $$Q_{\Sigma, \mathbf{v}}[u]\leq \mu_{\Sigma, \mathbf{v}}^- B_{\Sigma, \mathbf{v}}[u]\leq 0,$$ and for all $u\in\mathcal{P}_{\Sigma, \mathbf{v}}^\prime$, $$Q_{\Sigma, \mathbf{v}}[u]\geq \mu_{\Sigma, \mathbf{v}}^+ B_{\Sigma, \mathbf{v}}[u]\geq 0.$$ Here $ \mu_{\Sigma, \mathbf{v}}^- <0< \mu_{\Sigma, \mathbf{v}}^+$ are, respectively, the largest negative and smallest positive eigenvalues of $-L_{\Sigma, \mathbf{v}}^\prime$. Likewise, for $u\in \mathcal{K}_{\Sigma,\mathbf{v}}^\prime$ and $v\in W^1_{\frac{3}{8}}(\Sigma)$, $Q_{\Sigma, \mathbf{v}}[u, v]=0$.
First, a simple consequence of the inverse function theorem in Hölder spaces.
\[InverseLem\] Fix $\Sigma\in\mathcal{ACH}^{k,\alpha}_n$ and $\epsilon>0$. There is an $\eta=\eta(\Sigma,\epsilon)>0$ so that if $\mathbf{f}\in \mathcal{B}_{\eta}(\mathbf{x}|_\Sigma;\mathcal{ACH}^{k,\alpha}_n(\Sigma))$, then $$\mathbf{f}^{-1} \in \mathcal{B}_{\epsilon}(\mathbf{x}|_{\mathbf{f}(\Sigma)};\mathcal{ACH}^{k,\alpha}_n(\mathbf{f}(\Sigma))).$$
Let $\Lambda=\mathbf{f}(\Sigma)$. For $\eta$ sufficiently small, $\mathbf{f}\in \mathcal{ACH}^{k,\alpha}_n(\Sigma)$ and so $\Lambda\in\mathcal{ACH}^{k,\alpha}_n$. We first observe that as long as $\eta\leq \frac{1}{8}$, then by the implicit function theorem in $C^{k,\alpha}_1$ $$\Vert \mathbf{f}^{-1}\Vert_{k,\alpha}^{(1)}\leq C(n)$$ Observe that $$\mathbf{f}^{-1}-\mathbf{x}|_{\Gamma}=\mathbf{x}|_{\Sigma} \circ \mathbf{f}^{-1}-\mathbf{f}\circ \mathbf{f}^{-1},$$ and so by [@BWBanachManifold Proposition 3.1 (4)] $$\Vert \mathbf{f}^{-1}-\mathbf{x}|_{\Lambda}\Vert_{k,\alpha}^{(1)} \leq C^\prime\Vert \mathbf{f}-\mathbf{x}|_{\Sigma}\Vert_{k,\alpha}^{(1)}$$ for some $C^\prime=C^\prime(\Sigma,C(n))$. The claim follows immediately.
We now show the stability of the positive and negative spaces for the quadratic form $Q_{\mathbf{f}, \mathbf{v}}$ for $\mathbf{f}$ a small perturbation of $\mathbf{x}|_{\Sigma}$. The main challenge in proving this result is the fact that the weight is allowed to change a large amount near infinity and we cannot directly use the estimates of Section \[ConcentrationSec\] on the positive part.
\[QuadFormProp\] Let $\Sigma\in\mathcal{ACH}^{k,\alpha}_n$ be a self-expander, and let $\mathbf{v}\in C^{k,\alpha}_0\cap C^{k}_{0,\mathrm{H}}(\Sigma;\mathbb{R}^{n+1})$ a transverse section on $\Sigma$ so that $\mathcal{C}[\mathbf{v}]$ is transverse to $\mathcal{C}(\Sigma)$. There is an $\epsilon^\prime=\epsilon^\prime(\Sigma,\mathbf{v})$ so that if $\mathbf{f}\in \mathcal{B}_{\epsilon^\prime}(\mathbf{x}|_\Sigma;\mathcal{ACH}^{k,\alpha}_n(\Sigma))$ is $E$-stationary (i.e., $\mathbf{f}(\Sigma)$ is a self-expander), then $$Q_{\mathbf{f}, \mathbf{v}}[u]< 0,$$ for any $u\in \mathcal{N}_{\Sigma, \mathbf{v}}^\prime\setminus{\left\{0\right\}}$ and $$Q_{\mathbf{f}, \mathbf{v}}[u] >0,$$ for any $u\in \mathcal{P}_{\Sigma,\mathbf{v}}^\prime\setminus{\left\{0\right\}}$.
First observe that if $u\in \mathcal{N}_{\Sigma,\mathbf{v}}^\prime\setminus{\left\{0\right\}}$, then we have $$Q_{\Sigma, \mathbf{v}}[u]<\frac{1}{2} \mu_{\Sigma, \mathbf{v}}^-B_{\Sigma, \mathbf{v}}[u]<0.$$ Moreover, as $$Q_{\Sigma,\mathbf{v}}[u]\geq D_{\Sigma,\mathbf{v}}[u]-K_\Sigma B_{\Sigma,\mathbf{v}}[u]$$ one has $$D_{\Sigma,\mathbf{v}}[u]<K_\Sigma B_{\Sigma,\mathbf{v}}[u].$$ Thus, by Corollary \[PerturbCor\] with $\delta= -\frac{\mu_{\Sigma, \mathbf{v}}^-}{4\sqrt{K_\Sigma+1}}>0$ one has an $\epsilon_3$ given in the corollary so that if $\epsilon<\epsilon_3$, then $$\left| Q_{\mathbf{f}, \mathbf{v}}[u]-Q_{\Sigma, \mathbf{v}}[u]\right| \leq \delta \sqrt{K_\Sigma+1}\, B_{\Sigma, \mathbf{v}}[u].$$ Hence, $$\begin{aligned}
Q_{\mathbf{f},\mathbf{v}}[u] & \leq\delta \sqrt{K_\Sigma+1}\,B_{\Sigma, \mathbf{v}}[u] +Q_{\Sigma, \mathbf{v}}[u] \\
& <\left(\frac{1}{2} \mu_{\Sigma, \mathbf{v}}^-+\delta \sqrt{K_\Sigma+1}\right)B_{\Sigma, \mathbf{v}}[u]= \frac{1}{4} \mu_{\Sigma, \mathbf{v}}^-B_{\Sigma, \mathbf{v}}[u]<0.\end{aligned}$$ This proves the first part of the proposition.
For the second part, first observe that, by definition, for all $u\in \mathcal{E}_{\Sigma, \mathbf{v}}^\prime[0]=\mathcal{N}_{\Sigma,\mathbf{v}}^\prime\oplus \mathcal{K}_{\Sigma,\mathbf{v}}^\prime$ one has $Q_{\Sigma,\mathbf{v}}[u]\leq 0$. Hence, up to shrinking $\epsilon_3$, Corollary \[PerturbCor\] gives, as above, that $$Q_{\mathbf{f},\mathbf{v}}[u]\leq \frac{1}{4} \mu^+ B_{\mathbf{f},\mathbf{v}}[u],$$ where $$\mu^+=\frac{1}{4}\min {\left\{\mu_{\Sigma, \mathbf{v}}^+,\frac{1}{4}\right\}}>0.$$ We now argue by contradiction. That is, suppose there were a $v_0\in \mathcal{P}^\prime_{\Sigma, \mathbf{v}}\setminus{\left\{0\right\}}$ with $Q_{\mathbf{f},\mathbf{v}}[v_0]\leq 0$. Observe that, by definition, $v_0$ is orthogonal (with respect to $B_{\Sigma, \mathbf{v}}$) to $\mathcal{E}_{\Sigma, \mathbf{v}}^\prime[0]$. Up to further shrinking $\epsilon_3$, we may assume that for all $\mathbf{f}\in \mathcal{B}_{\epsilon_3}(\mathbf{x}|_\Sigma;\mathcal{ACH}^{k,\alpha}_n(\Sigma))$ if $\Lambda=\mathbf{f}(\Sigma)$, then $$K_{\Lambda}\leq K_\Sigma+1.$$ Moreover, as $$0\geq Q_{\mathbf{f}, \mathbf{v}}[v_0]\geq D_{\mathbf{f},\mathbf{v}}[v_0]-K_\Lambda B_{\mathbf{f}, \mathbf{v}}[v_0]$$ one has $$D_{\mathbf{f},\mathbf{v}}[v_0]\leq \left(K_\Sigma+1\right) B_{\mathbf{f}, \mathbf{v}}[v_0].$$ Now let $$V=\mathrm{span}{\left\{v_0\right\}}\oplus\mathcal{E}_{\Sigma, \mathbf{v}}^\prime[0]=\mathrm{span}{\left\{v_0\right\}}\oplus \mathcal{N}_{\Sigma,\mathbf{v}}^\prime\oplus\mathcal{K}_{\Sigma,\mathbf{v}}^\prime.$$ The choice of $v_0$ ensures that $$\dim V= 1+ \dim \mathcal{E}_{\Sigma, \mathbf{v}}^\prime [0]=1+ \dim \mathcal{N}_{\Sigma,\mathbf{v}}^\prime+\dim \mathcal{K}_{\Sigma,\mathbf{v}}^\prime=1+N.$$
We claim that, up to shrinking $\epsilon_3$, for any $v\in V$ one has $$Q_{\mathbf{f}, \mathbf{v}}[v]\leq \mu^+ B_{\mathbf{f},\mathbf{v}}[v].$$ Indeed, write $$v=c_0 v_0+ u$$ where $c_0\in {\mathbb R}$ and $u\in \mathcal{E}_{\Sigma, \mathbf{v}}^\prime[0]$. Using Corollary \[PerturbCor\] and the fact that $Q_{\Sigma, \mathbf{v}}[v_0,u]=0$, one computes, $$\begin{aligned}
|Q_{\mathbf{f}, \mathbf{v}}[v_0, u]|^2 & \leq \delta^2 \left(D_{\mathbf{f}, \mathbf{v}}[v_0]+B_{\mathbf{f}, \mathbf{v}}[v_0]\right) B_{\Sigma, \mathbf{v}}[u] \\
& \leq 2\delta^2 \left( K_{\Sigma}+2\right) B_{\mathbf{f}, \mathbf{v}}[v_0] B_{\mathbf{f}, \mathbf{v}}[u].\end{aligned}$$ Hence, picking $\delta$ (and hence $\epsilon_3$) small enough so $\delta\leq \frac{\mu^+}{16\sqrt{K_{\Sigma}+2}}$ yields $$\begin{aligned}
Q_{\mathbf{f}, \mathbf{v}}[v] &=c_0^2 Q_{\mathbf{f}, \mathbf{v}}[v_0]+2c_0 Q_{\mathbf{f}, \mathbf{v}}[v_0, u]+Q_{\mathbf{f}, \mathbf{v}}[u] \\
& \leq \frac{1}{4} |c_0| \mu^+ B_{\mathbf{f}, \mathbf{v}}[v_0]^{1/2} B_{\mathbf{f}, \mathbf{v}}[u]^{1/2}+\frac{1}{4} \mu^+ B_{\mathbf{f},\mathbf{v}}[u]\\
&\leq \frac{1}{2}\mu^+ \left( B_{\mathbf{f}, \mathbf{v}}[c_0 v_0]+ B_{\mathbf{f},\mathbf{v}}[u]\right)
\\& = \frac{1}{2}\mu^+ \left( B_{\mathbf{f}, \mathbf{v}}[c_0 v_0+u]-2B_{\mathbf{f}, \mathbf{v}}[c_0v_0, u]\right).\end{aligned}$$ As $B_{\Sigma, \mathbf{v}}[v_0,u]=0$, up to shrinking $\delta$ some more one has $$\begin{aligned}
|B_{\mathbf{f},\mathbf{v}}[c_0 v_0,u]| &\leq \frac{1}{2} B_{\mathbf{f},\mathbf{v}}[c_0 v_0]^{\frac{1}{2}} B_{\mathbf{f},\mathbf{v}}[u]^{\frac{1}{2}}\leq \frac{1}{4} \left( B_{\mathbf{f},\mathbf{v}}[c_0 v_0]+ B_{\mathbf{f},\mathbf{v}}[u]\right)\\
& \leq \frac{1}{4} \left( B_{\mathbf{f},\mathbf{v}}[c_0 v_0+u]+2|B_{\mathbf{f},\mathbf{v}}[c_0 v_0, u]|\right). \end{aligned}$$ That is, $$|B_{\mathbf{f},\mathbf{v}}[c_0 v_0,u]| \leq \frac{1}{2} B_{\mathbf{f},\mathbf{v}}[c_0 v_0+u],$$ and so $$Q_{\mathbf{f}, \mathbf{v}}[v]\leq \mu^+ B_{\mathbf{f}, \mathbf{v}}[c_0 v_0+u]=\mu^+ B_{\mathbf{f}, \mathbf{v}}[v],$$ verifying the claim.
Hence, there is an $N+1$ dimensional subspace, $V\subset W^1_{\frac{3}{8}}(\Sigma)$, so that, for all $v\in V$, $$Q_{\mathbf{f}, \mathbf{v}}[v]\leq \mu^+B_{\mathbf{f}, \mathbf{v}}[v].$$ By Lemma \[ParaSpectrumLem\], $-L_{\mathbf{f},\mathbf{v}}^\prime$ is self-adjoint with respect to $B_{\mathbf{f},\mathbf{v}}$ and has a discrete spectrum. As such, it follows from the min-max characterization of eigenvalues that the $(N+1)$-st eigenvalue (counted with multiplicities), $\mu_{\mathbf{f},\mathbf{v}}^{N+1}$, of $-L_{\mathbf{f},\mathbf{v}}'$ satisfies $\mu_{\mathbf{f},\mathbf{v}}^{N+1}\leq \mu^+$. Now let $$\mathcal{E}_{\mathbf{f},\mathbf{v}}^\prime[\mu_{\mathbf{f},\mathbf{v}}^{N+1}]=\mathrm{span} {\left\{u\in W^1_{\frac{3}{8}}(\Sigma)\colon -L_{\mathbf{f},\mathbf{v}}^\prime u=\mu^\prime u \mbox{ for some $\mu^\prime\leq \mu^{N+1}_{\mathbf{f},\mathbf{v}}$}\right\}},$$ and choose $W\subset \mathcal{E}_{\mathbf{f},\mathbf{v}}^\prime [\mu_{\mathbf{f},\mathbf{v}}^{N+1}]$ spanned by $(N+1)$ linearly independent eigenfunctions.
Using $K_{\Lambda}$ and $\delta=\frac{\mu^+}{4(K_\Lambda+1)}$ let $\epsilon_3^\prime$ be given by Corollary \[PerturbCor\] with $\Lambda$ in place of $\Sigma$. Observe $\epsilon_3^\prime=\epsilon_3^\prime(K_\Lambda, \mu^+, n, N+1)$ depends only on $\Sigma$. By Lemma \[InverseLem\], we may shrink $\epsilon_3$, in a manner depending only on $\Sigma$, so that $\mathbf{f}^{-1}\in \mathcal{B}_{\epsilon_3^\prime}(\mathbf{x}|_\Lambda;\mathcal{ACH}^{k,\alpha}_n(\Lambda))$. As such, we may apply Corollary \[PerturbCor\], with $\Lambda$ in place of $\Sigma$, to $w\in W$ to see that $$\begin{aligned}
Q_{\Sigma, \mathbf{v}}[w]&=Q_{\mathbf{f}^{-1}, \mathbf{v}\circ \mathbf{f}^{-1}} [w\circ \mathbf{f}^{-1}]-Q_{\mathbf{f}, \mathbf{v} }[w]+Q_{\mathbf{f}, \mathbf{v} }[w]\\
&= Q_{\mathbf{f}^{-1}, \mathbf{v}\circ \mathbf{f}^{-1}} [w\circ \mathbf{f}^{-1}]-Q_{\Lambda, \mathbf{v}\circ \mathbf{f}^{-1}}[w\circ \mathbf{f}^{-1}]+Q_{\mathbf{f}, \mathbf{v} }[w]\\
&\leq \frac{1}{2} \mu^+ B_{\Lambda, \mathbf{v}\circ \mathbf{f}^{-1}}[w\circ \mathbf{f}^{-1}]+Q_{\mathbf{f}, \mathbf{v}}[w]\\
&=\frac{1}{2}\mu^+ B_{\mathbf{f},\mathbf{v}}[w]+Q_{\mathbf{f}, \mathbf{v}}[w]
\leq \frac{3}{2} \mu^+ B_{\mathbf{f}, \mathbf{v}}[w].\end{aligned}$$ Similarly, applying Corollary \[PerturbCor\], with $\Lambda$ in place of $\Sigma$, gives $$B_{\mathbf{f}, \mathbf{v}}[w]= B_{\Lambda, \mathbf{v}\circ\mathbf{f}^{-1}}[w\circ\mathbf{f}^{-1}]\leq 2 B_{\mathbf{f}^{-1}, \mathbf{v}\circ\mathbf{f}^{-1}}[w\circ\mathbf{f}^{-1}]=2 B_{\Sigma, \mathbf{v}}[w].$$ Hence, as $\mu^+> 0$, $$Q_{\Sigma, \mathbf{v}}[w]\leq 3 \mu^+B_{\Sigma, \mathbf{v}}[w].$$ Recall, by Lemma \[SpectrumLem\], $-L_{\Sigma,\mathbf{v}}^\prime$ is self-adjoint with respect to $B_{\Sigma,\mathbf{v}}$ and has a discrete spectrum. As such, by the min-max characterization of eigenvalues, one has that $\mu_{\Sigma,\mathbf{v}}^{N+1}$, the $(N+1)$-st eigenvalue of $-L_{\Sigma, \mathbf{v}}^\prime$ satisfies $\mu_{\Sigma,\mathbf{v}}^{N+1}\leq 3 \mu^+\leq \frac{3}{4}\mu_{\Sigma, \mathbf{v}}^+$. However, as $N$ is the dimension of $\mathcal{E}_{\Sigma, \mathbf{v}}^\prime[0]$, one must have $\mu_{\Sigma,\mathbf{v}}^{N+1}=\mu_{\Sigma,\mathbf{v}}^+>\frac{3}{4}\mu_{\Sigma, \mathbf{v}}^+>0$. This contradiction proves the second part of the proposition and completes the arguments.
Ultimately, we will show an extension of [@WhiteEI Proposition 4] for asymptotically conical self-expanders. However, before stating the theorem we introduce some notation inspired by [@WhiteVM]. Fix a self-expander $\Sigma\in \mathcal{ACH}^{k,\alpha}_n$ and a transverse section on $\Sigma$, $\mathbf{v}\in C^{k,\alpha}_0\cap C^{k}_{0,\mathrm{H}}(\Sigma;\mathbb{R}^{n+1})$ so that $\mathcal{C}[\mathbf{v}]$ is transverse to $\mathcal{C}(\Sigma)$ and $\mathscr{L}_\Sigma^0\mathbf{v}\in C^{k-2,\alpha}_{-2}(\Sigma;\mathbb{R}^{n+1})$. Let $\hat{G}_{\mathbf{v}}$ and $\hat{F}_{\mathbf{v}}$ be given in Theorem \[ModifiedSmoothDependThm\], and $(\varphi, \kappa)$ is in a small neighborhood of $(\mathbf{x}|_{\mathcal{L}(\Sigma)},0)$ in $C^{k,\alpha}(\mathcal{L}(\Sigma); {\mathbb R}^{n+1})\times \mathcal{K}_{\mathbf{v}}$. Formally, define the following function (which is infinite valued) $$\hat{g}_{\mathbf{v}}[\varphi, \kappa]=\int_{\hat{F}_{\mathbf{v}}[\varphi,\kappa](\Sigma)} e^{\frac{|\mathbf{x}|^2}{4}} \, d\mathcal{H}^n=E[\hat{F}_{\mathbf{v}}[\varphi,\kappa]].$$ Differentiating (formally) in the second variable gives the following well-defined function $D_2 \hat{g}_{\mathbf{v}}$ on $\mathcal{K}_{\mathbf{v}}$: $$\begin{aligned}
D_2 \hat{g}_{\mathbf{v}}(\varphi, \kappa)\kappa_1&=-B_{\Sigma, \mathbf{v}}[\hat{\Xi}_{\hat{F}_{\mathbf{v}}[\varphi, \kappa],\mathbf{v}}[0], D_2\hat{\mathcal{F}}_\mathbf{v}(\varphi, \kappa)\kappa_1]\\
&=-B_{\Sigma, \mathbf{v}}[\hat{G}_{\mathbf{v}}[\varphi, \kappa], D_2\hat{\mathcal{F}}_\mathbf{v}(\varphi, \kappa)\kappa_1]\\
&=-B_{\Sigma, \mathbf{v}}[\hat{G}_{\mathbf{v}}[\varphi, \kappa], \kappa_1]\end{aligned}$$ where the last equality follows from the fact that $\hat{G}_{\mathbf{v}}[\phi, \kappa]\in \mathcal{K}_{\mathbf{v}}$ and . An immediate consequence of this is that $$D_2 \hat{g}_{\mathbf{v}}(\varphi, \kappa)=0 \iff \hat{G}_{\mathbf{v}}[\varphi, \kappa]=0.$$ That is, $\hat{F}_{\mathbf{v}}[\varphi, \kappa]$ is $E$-stationary if and only if $D_2 \hat{g}_{\mathbf{v}}(\varphi, \kappa)=0$.
Finally, by differentiating again at a $(\varphi, \kappa)$ where $\hat{G}_{\mathbf{v}}[\varphi, \kappa]=0$, that is where $\hat{F}_{\mathbf{v}}[\varphi, \kappa]$ is $E$-stationary, one has the bilinear form on $\mathcal{K}_{\mathbf{v}}$ given by $$\begin{aligned}
(D_{22}^2 \hat{g}_{\mathbf{v}}(\varphi, \kappa))[\kappa_1, \kappa_2]&=-B_{\Sigma, \mathbf{v}}[D\hat{\Xi}_{\hat{F}_{\mathbf{v}}[\varphi, \kappa]}(0)\circ D_2\hat{\mathcal{F}}_{\mathbf{v}}(\varphi,\kappa)\kappa_1, D_2\hat{\mathcal{F}}_{\mathbf{v}}(\varphi,\kappa)\kappa_2]\\
&=-B_{\Sigma, \mathbf{v}}[\Omega_{\hat{F}_{\mathbf{v}}[\varphi, \kappa],\mathbf{v}}L_{\hat{F}_{\mathbf{v}}[\varphi, \kappa], \mathbf{v}}^\prime\circ D_2\hat{\mathcal{F}}_{\mathbf{v}}(\varphi,\kappa)\kappa_1, D_2\hat{\mathcal{F}}_{\mathbf{v}}(\varphi,\kappa)\kappa_2]\\
&=Q_{\hat{F}_{\mathbf{v}}[\varphi, \kappa],\mathbf{v}}[D_2\hat{\mathcal{F}}_{\mathbf{v}}(\varphi, \kappa)\kappa_1,D_2\hat{\mathcal{F}}_{\mathbf{v}}(\varphi, \kappa)\kappa_2].\end{aligned}$$
\[MainTechThm\] Let $\Sigma\in \mathcal{ACH}^{k,\alpha}_n$ be a self-expander, and let $\mathbf{v}$ be a transverse section on $\Sigma$ so that $\mathcal{C}[\mathbf{v}]$ is transverse to $\mathcal{C}(\Sigma)$ and $\mathscr{L}_\Sigma^0\mathbf{v}\in C^{k-2,\alpha}_{-2}(\Sigma;\mathbb{R}^{n+1})$. There is an $\bar{\epsilon}=\bar{\epsilon}(\Sigma, \mathbf{v})>0$ so that if $(\varphi, \kappa)\in C^{k,\alpha}(\mathcal{L}(\Sigma); {\mathbb R}^{n+1}) \times \mathcal{K}_{\mathbf{v}}$ satisfies $\hat{G}_{\mathbf{v}}[\varphi, \kappa]=0$ and $\Vert \varphi -\mathbf{x}|_{\mathcal{L}(\Sigma)}\Vert_{k, \alpha}+\Vert \kappa \Vert_{\mathcal{D}^{k,\alpha}\cap W^2_{3/8}} <\bar{\epsilon}$, then
1. $\mathrm{ind} (\hat{F}_{\mathbf{v}}[\varphi, \kappa])= \mathrm{ind}(\Sigma)+ \mathrm{ind}(D_{22}^2 \hat{g}_{\mathbf{v}}(\varphi,\kappa))$.
2. $\mathrm{null}(\hat{F}_{\mathbf{v}}[\varphi, \kappa])= \mathrm{null}(D_{22}^2 \hat{g}_{\mathbf{v}}(\varphi,\kappa))$.
Pick $\epsilon^\prime>0$ as in Proposition \[QuadFormProp\] and choose $\bar{\epsilon}$ so $\hat{F}_{\mathbf{v}}[\varphi, \kappa]\in \mathcal{B}_{\epsilon^\prime}(\mathbf{x}|_\Sigma; \mathcal{ACH}^{k,\alpha}_n(\Sigma))$. For convenience, write $\mathbf{f}=\hat{F}_{\mathbf{v}}[\varphi, \kappa]$. Note that $\mathcal{K}_{\mathbf{v}}=\mathcal{K}_{\Sigma, \mathbf{v}}^\prime$. Set $$\tilde{\mathcal{K}}={\left\{D_2 \hat{\mathcal{F}}_{\mathbf{v}}(\varphi, \kappa) \kappa^\prime\colon \kappa^\prime\in \mathcal{K}_{\mathbf{v}}\right\}}=\mathrm{Im}(D_2 \hat{\mathcal{F}}_{\mathbf{v}}(\varphi, \kappa)).$$ By definition, $$D_2\hat{\mathcal{F}}_{\mathbf{v}}(\varphi,\kappa)\kappa^\prime=\kappa^\prime+ D_2\hat{F}_{\mathbf{v},*}(\varphi, \kappa)\kappa^\prime,$$ and so implies $$\tilde{\mathcal{K}}\subset \mathcal{K}_{\mathbf{v}}+\hat{\mathcal{K}}_{\mathbf{v}, *}^\perp\subset \mathcal{D}^{k,\alpha}\cap W^{2}_{\frac{3}{8}}(\Sigma).$$ Moreover, as $$D_2 \hat{G}_{\mathbf{v}}(\varphi, \kappa)\colon \mathcal{K}_{\mathbf{v}}\to \mathcal{K}_{\mathbf{v}}$$ it follows from that, setting $J^\prime=-\Omega_{\mathbf{f}, \mathbf{v}} L_{\mathbf{f}, \mathbf{v}}^\prime$, one has $$D_2 \hat{G}_{\mathbf{v}}(\varphi, \kappa)=-J^\prime \circ D_2 \hat{\mathcal{F}}_{\mathbf{v}}(\varphi, \kappa)$$ and so $J^\prime(\tilde{\mathcal{K}})\subset \mathcal{K}_{\mathbf{v}}=K_{\Sigma, \mathbf{v}}^\prime$. Furthermore, as $D_2\hat{F}_{\mathbf{v},*}(\varphi, \kappa)$ has image in $\hat{\mathcal{K}}_{\mathbf{v}, *}^\perp$, $\ker(D_2\hat{\mathcal{F}}_{\mathbf{v}})={\left\{0\right\}}$ and so $$l=\dim\tilde{\mathcal{K}}=\dim\mathcal{K}_{\mathbf{v}}=\mathrm{null}(\Sigma).$$ Finally, $$Q_{\mathbf{f}, \mathbf{v}}[u, v]=B_{\Sigma, \mathbf{v}}[u, -\Omega_{\mathbf{f}, \mathbf{v}}L_{\mathbf{f}, \mathbf{v}}^\prime v]= B_{\Sigma, \mathbf{v}}[u, J^\prime v]= B_{\Sigma, \mathbf{v}}[J^\prime u, v].$$
Now let $\tilde{Q}_{\mathbf{f}, \mathbf{v}}\colon \tilde{\mathcal{K}}\times \tilde{\mathcal{K}}\to {\mathbb R}$ be the bilinear form defined, for $\tilde{\kappa}_1, \tilde{\kappa}_2\in \tilde{\mathcal{K}}$ by $$\tilde{Q}_{\mathbf{f}, \mathbf{v}}[\tilde{\kappa}_1, \tilde{\kappa}_2]=Q_{\mathbf{f}, \mathbf{v}}[\tilde{\kappa}_1, \tilde{\kappa}_2].$$ Clearly, for $\kappa_1, \kappa_2\in\mathcal{K}_{\mathbf{v}}$, $$(D_{22}^2\hat{g}_{\mathbf{v}}(\varphi, \kappa))[\kappa_1, \kappa_2]=\tilde{Q}_{\mathbf{f}, \mathbf{v}}[D_2 \hat{\mathcal{F}}_{\mathbf{v}}(\varphi, \kappa)\kappa_1, D_2 \hat{\mathcal{F}}_{\mathbf{v}}(\varphi, \kappa)\kappa_2].$$ As the map $\kappa^\prime\mapsto D_2 \hat{\mathcal{F}}_{\mathbf{v}}(\varphi, \kappa)\kappa^\prime$ is injective, one has $$\mathrm{null}(\tilde{Q}_{\mathbf{f}, \mathbf{v}})=\mathrm{null}(D_{22}^2\hat{g}_{\mathbf{v}}(\varphi, \kappa))$$ and $$\mathrm{ind}(\tilde{Q}_{\mathbf{f}, \mathbf{v}})=\mathrm{ind}(D_{22}^2\hat{g}_{\mathbf{v}}(\varphi, \kappa)).$$
We are now ready to complete the proof. First, consider the decomposition of $W^{1}_{\frac{3}{8}}(\Sigma)$ into $$W^1_{\frac{3}{8}}(\Sigma)=\mathcal{N}_{\Sigma, \mathbf{v}}^\prime\oplus \mathcal{K}_{\Sigma, \mathbf{v}}^\prime\oplus \mathcal{P}_{\Sigma, \mathbf{v}}^\prime.$$ By Proposition \[QuadFormProp\], for $\epsilon^\prime$ sufficiently small, one has $Q_{\mathbf{f}, \mathbf{v}}$ positive definite on $\mathcal{P}_{\Sigma, \mathbf{v}}^\prime$ and negative definite on $\mathcal{N}_{\Sigma, \mathbf{v}}^\prime$. A consequence of this is that $$W^1_{\frac{3}{8}}(\Sigma)=\mathcal{N}_{\Sigma, \mathbf{v}}^\prime\oplus\tilde{\mathcal{K}}\oplus \mathcal{P}_{\Sigma, \mathbf{v}}^\prime.$$ Indeed, consider the projection map $$\pi\colon W^1_{\frac{3}{8}}(\Sigma)\to W^1_{\frac{3}{8}}(\Sigma)/(\mathcal{N}_{\Sigma, \mathbf{v}}^\prime+\mathcal{P}_{\Sigma, \mathbf{v}}^\prime)=\pi( \mathcal{K}_{\Sigma, \mathbf{v}}^\prime).$$ The orthogonal decomposition of $W^{1}_{\frac{3}{8}}(\Sigma)$ implies $\pi( \mathcal{K}_{\Sigma, \mathbf{v}}^\prime) \simeq {\mathbb R}^l$ where $l=\dim \mathcal{K}_{\Sigma, \mathbf{v}}^\prime=\dim \mathcal{K}_{\mathbf{v}}$. Hence, it suffices to show $\pi|_{\tilde{\mathcal{K}}}$ is surjective. As $\dim \tilde{\mathcal{K}}=l$ this is equivalent to showing this map is injective. If $\tilde{\kappa}\in \ker \pi|_{\tilde{\mathcal{K}}}$, then $\tilde{\kappa}=\tilde{p}+\tilde{n}$ for $\tilde{p}\in \mathcal{P}_{\Sigma, \mathbf{v}}^\prime$ and $\tilde{n}\in \mathcal{N}_{\Sigma, \mathbf{v}}^\prime$. However, $$Q_{\mathbf{f}, \mathbf{v}}[\tilde{n},\tilde{n}]+Q_{\mathbf{f}, \mathbf{v}}[\tilde{p}, \tilde{n}]=Q_{\mathbf{f}, \mathbf{v}}[\tilde{\kappa}, \tilde{n}]=B_{\Sigma, \mathbf{v}}[J^\prime\tilde{\kappa}, \tilde{n}]=0$$ where the last equality follows from the fact that $J^\prime\tilde{\kappa}\in \mathcal{K}_{\mathbf{v}}=\mathcal{K}_{\Sigma,\mathbf{v}}^\prime$ which is orthogonal with respect to $B_{\Sigma, \mathbf{v}}$ to $\mathcal{N}_{\Sigma, \mathbf{v}}^\prime$. Similar reasoning shows $$Q_{\mathbf{f}, \mathbf{v}}[\tilde{p}, \tilde{p}]+ Q_{\mathbf{f}, \mathbf{v}}[\tilde{p}, \tilde{n}]= Q_{\mathbf{f}, \mathbf{v}}[\tilde{p}, \tilde{\kappa}]=B_{\Sigma, \mathbf{v}}[\tilde{p}, J^\prime\tilde{\kappa}]=0.$$ Subtracting the first equality from the second gives $$Q_{\mathbf{f}, \mathbf{v}}[\tilde{p}, \tilde{p}]-Q_{\mathbf{f}, \mathbf{v}}[\tilde{n}, \tilde{n}]=0.$$ As $Q_{\mathbf{f}, \mathbf{v}}$ is positive definite on $\mathcal{P}_{\Sigma, \mathbf{v}}^\prime$ and negative definite on $\mathcal{N}_{\Sigma, \mathbf{v}}^\prime$ one must have $\tilde{p}=\tilde{n}=0$. That is, $\tilde{\kappa}=0$, which proves the claim.
Now decompose $\tilde{\mathcal{K}}$ into $$\tilde{\mathcal{K}}=\mathcal{N}^\prime\oplus \mathcal{K}^\prime\oplus\mathcal{P}^\prime$$ where each subspace $\mathcal{N}^\prime, \mathcal{K}^\prime, \mathcal{P}^\prime$ is orthogonal (with respect to $B_{\Sigma, \mathbf{v}}$) and $\tilde{Q}_{\mathbf{f}, \mathbf{v}}$ is, respectively, negative definite, zero and positive definite on $\mathcal{N}^\prime, \mathcal{K}^\prime$ and $\mathcal{P}^\prime$. Now consider $p\in \mathcal{P}^\prime_{\Sigma, \mathbf{v}}$ and $p^\prime\in \mathcal{P}^\prime$ one has $$\begin{aligned}
Q_{\mathbf{f}, \mathbf{v}}[p+p^\prime, p+p^\prime] &= Q_{\mathbf{f}, \mathbf{v}}[p,p]+2Q_{\mathbf{f},\mathbf{v}}[p,p^\prime]+Q_{\mathbf{f},\mathbf{v}}[p^\prime,p^\prime] \\
&=Q_{\mathbf{f},\mathbf{v}}[p,p]+2B_{\Sigma,\mathbf{v}}[p,J^\prime p^\prime]+Q_{\mathbf{f},\mathbf{v}}[p^\prime,p^\prime]\\
&=Q_{\mathbf{f}, \mathbf{v}}[p,p]+\tilde{Q}_{\mathbf{f}, \mathbf{v}}[p^\prime, p^\prime].\end{aligned}$$ Here the last equality uses the fact that $J^\prime p^\prime\in \mathcal{K}_{\mathbf{v}}=\mathcal{K}_{\Sigma, \mathbf{v}}'$ is orthogonal to $p$ and that $J^\prime$ is self-adjoint with respect to $B_{\Sigma, \mathbf{v}}$. By Proposition \[QuadFormProp\] and the fact that $p^\prime\in\mathcal{N}$ one has for $p+p^\prime\neq 0$ that $$Q_{\mathbf{f}, \mathbf{v}}[p+p^\prime, p+p^\prime]>0$$ That is, $Q_{\mathbf{f}, \mathbf{v}}$ is positive definite on $\mathcal{P}_{\Sigma, \mathbf{v}}\oplus\mathcal{P}^\prime$. Similar reasoning shows that $Q_{\mathbf{f}, \mathbf{v}}$ is negative definite on $\mathcal{N}_{\Sigma, \mathbf{v}}\oplus\mathcal{N}^\prime$, negative semi-definite on $\mathcal{N}_{\Sigma, \mathbf{v}}^\prime\oplus\mathcal{N}^\prime\oplus\mathcal{K}^\prime$ and positive semi-definite on $\mathcal{P}_{\Sigma, \mathbf{v}}\oplus\mathcal{P}^\prime\oplus\mathcal{K}^\prime$. As a consequence, $$\begin{aligned}
\mathrm{ind}(\hat{F}_{\mathbf{v}}[\varphi, \kappa])&= \mathrm{ind}(\mathbf{f})=\mathrm{ind}(Q_{\mathbf{f}, \mathbf{v}})\\
&= \dim (\mathcal{N}_{\Sigma, \mathbf{v}}^\prime\oplus\mathcal{N}^\prime)=\dim\mathcal{N}_{\Sigma, \mathbf{v}}^\prime+\dim\mathcal{N}^\prime\\
&= \mathrm{ind}(\Sigma)+\mathrm{ind}(\tilde{Q}_{\mathbf{f}, \mathbf{v}})=\mathrm{ind}(\Sigma)+\mathrm{ind}(D^2_{22}\hat{g}_{\mathbf{v}}(\varphi,\kappa)).\end{aligned}$$ One further concludes that, for $u\in W^1_{\frac{3}{8}}(\Sigma)$, $J^\prime u=0$ if and only if $u\in \mathcal{K}^\prime$ and so $$\mathrm{null}(\hat{F}_{\mathbf{v}}[\varphi, \kappa])=\dim\mathcal{K}^\prime=\mathrm{null}(D^2_{22}\hat{g}_{\mathbf{v}}(\varphi,\kappa)).$$ This completes the proof of the result.
Proof of Theorem \[MainThm\] {#MainThmSec}
============================
We now include a proof of theorem \[MainThm\]. We adapt the proof from [@WhiteEI], using Theorem \[MainTechThm\] in place of [@WhiteEI Proposition 4]. The details are included for the sake of the reader.
First of all, for $\Gamma\in\mathcal{ACH}^{k,\alpha}_n$ and $l\geq 1$, let $$\mathcal{S}_l={\left\{[\mathbf{f}]\in \mathcal{ACE}_n^{k,\alpha}(\Gamma)\colon \mathrm{null}(\mathbf{f})\geq l\right\}}.$$ We record some technical facts proved in [@WhiteEI] that readily adapt to the setting of this paper.
\[BaireLem\] For $\Gamma\in\mathcal{ACH}^{k,\alpha}_n$ the following is true:
1. $\Pi(\mathcal{S}_1)$ is of first Baire category in $C^{k,\alpha}(\mathcal{L}(\Gamma); {\mathbb R}^{n+1})$.
2. For each $\varphi_0, \varphi_1\in \mathcal{V}^{k,\alpha}_{\mathrm{emb}}(\Gamma)$ a generic, in the sense of Baire category, curve $\Phi\in C^2([0,1]; \mathcal{V}_{\mathrm{emb}}^{k,\alpha}(\Gamma))$ joining $\varphi_0$ to $\varphi_1$ will be embedded and transverse to $\Pi$.
This is proved in [@WhiteEI Theorem 5.1] using only properties of the spaces and of $\Pi$ that are the same as in our setting.
\[ComposeLem\] Let $\Sigma\in\mathcal{ACH}^{k,\alpha}_n$ be a self-expander, and let $\mathbf{f}\in\mathcal{ACH}^{k,\alpha}_n(\Sigma)$ satisfy $\mathscr{L}_\Sigma\mathbf{f}\in C^{k-2,\alpha}_{-1}(\Sigma;\mathbb{R}^{n+1})$. Let $\Lambda=\mathbf{f}(\Sigma)$. Then the following is true:
1. $\mathscr{L}_\Lambda\mathbf{f}^{-1}\in C^{k-2,\alpha}_{-1}(\Lambda;\mathbb{R}^{n+1})$.
2. If $u\in C^{k,\alpha}_{0}(\Sigma)$ satisfies $\mathscr{L}^0_\Sigma u\in C^{k-2,\alpha}_{-2}(\Sigma)$, then $\mathscr{L}_{\Lambda}^0(u\circ\mathbf{f}^{-1}) \in C^{k-2,\alpha}_{-2}(\Lambda).
$
For $\mathbf{f}\in\mathcal{ACH}^{k,\alpha}_n(\Sigma)$, by [@BWBanachManifold Proposition 3.3], $\Lambda\in\mathcal{ACH}^{k,\alpha}_n$ and $\mathbf{f}^{-1}\in\mathcal{ACH}^{k,\alpha}_n(\Lambda)$. Consequently, $\Delta_\Lambda\mathbf{f}^{-1}\in C^{k-2,\alpha}_{-1}(\Lambda;\mathbb{R}^{n+1})$. To prove the first claim, it suffices to show $\mathbf{x}\cdot\nabla_\Lambda\mathbf{f}^{-1}-\mathbf{f}^{-1}\in C^{k-2,\alpha}_{-1}(\Lambda;\mathbb{R}^{n+1})$. As $\Sigma\in\mathcal{ACH}^{k,\alpha}_n$ is a self-expander, one has $\mathscr{L}_\Sigma\mathbf{x}=0$ and so $\mathbf{x}\cdot\nabla_\Sigma\mathbf{x}-\mathbf{x}\in C^{k-2,\alpha}_{-1}(\Sigma;\mathbb{R}^{n+1})$. Substituting $\mathbf{x}|_\Sigma=\mathbf{f}^{-1}\circ\mathbf{f}$ and applying the chain rule give $$\mathbf{x}\cdot\nabla_\Sigma (\mathbf{f}^{-1}\circ\mathbf{f})-\mathbf{f}^{-1}\circ\mathbf{f}=(\nabla_\Lambda\mathbf{f}^{-1}\circ\mathbf{f})\cdot(\mathbf{x}\cdot\nabla_\Sigma\mathbf{f}-\mathbf{f})+(\mathbf{x}\cdot\nabla_\Lambda\mathbf{f}^{-1}-\mathbf{f}^{-1})\circ\mathbf{f}.$$ As $\mathbf{f}\in C^{k,\alpha}_1(\Sigma;\mathbb{R}^{n+1})$ and $\mathscr{L}_\Sigma\mathbf{f}\in C^{k-2,\alpha}_{-1}(\Sigma;\mathbb{R}^{n+1})$, the claim follows by invoking [@BWBanachManifold Propositions 3.1 and 3.3].
If $u\in C^{k,\alpha}_0(\Sigma)$ and $\mathbf{f}\in\mathcal{ACH}^{k,\alpha}_n(\Sigma)$, then, by [@BWBanachManifold Proposition 3.3], $u\circ\mathbf{f}^{-1}\in C^{k,\alpha}_0(\Lambda)$ and so $\Delta_\Lambda (u\circ\mathbf{f}^{-1})\in C^{k-2,\alpha}_{-2}(\Lambda)$. One further computes, $$\mathbf{x}\cdot\nabla_\Lambda(u\circ\mathbf{f}^{-1})=(\nabla_\Sigma u\circ\mathbf{f}^{-1})\cdot(\mathbf{x}\cdot\nabla_\Lambda\mathbf{f}^{-1}-\mathbf{f}^{-1})+(\mathbf{x}\cdot\nabla_\Sigma u)\circ\mathbf{f}^{-1},$$ which, by what we have shown and our hypotheses on $u$, is an element of $C^{k-2,\alpha}_{-2}(\Lambda)$. Thus, the second claim follows easily from this.
\[OrientationLem\] Let $\Sigma\in\mathcal{ACH}^{k,\alpha}_n$ be a self-expander, and let $\mathbf{v}\in C^{k,\alpha}_0\cap C^{k}_{0,\mathrm{H}}(\Sigma;\mathbb{R}^{n+1})$ be a transverse section on $\Sigma$ so that $\mathcal{C}[\mathbf{v}]$ is transverse to $\mathcal{C}(\Sigma)$ and $\mathscr{L}_\Sigma^0\mathbf{v}\in C^{k-2,\alpha}_{-2}(\Sigma;\mathbb{R}^{n+1})$. Suppose that $\Pi^{-1}(\mathbf{x}|_{\mathcal{L}(\Sigma)})\cap\mathcal{W}$ is a connected one-dimensional submanifold of $\mathcal{ACE}^{k,\alpha}_n(\Sigma)$ for $\mathcal{W}$ a neighborhood of $[\mathbf{x}|_\Sigma]\in\mathcal{ACE}^{k,\alpha}_n(\Sigma)$ and that $\kappa\in\mathcal{K}_\mathbf{v}$ with $B_{\Sigma,\mathbf{v}}[\kappa]=1$. Given $\varphi\in C^{k,\alpha}(\mathcal{L}(\Sigma);\mathbb{R}^{n+1})$ there is an $\epsilon=\epsilon(\Sigma,\mathbf{v},\varphi)>0$ so that the following holds:
1. If $|s|<\epsilon$, then $\Lambda_s=\hat{F}_{\mathbf{v}}[\mathbf{x}|_{\mathcal{L}(\Sigma)},s\kappa](\Sigma)\in\mathcal{ACH}^{k,\alpha}_n$ is a self-expander with $\mathcal{L}(\Lambda_s)=\mathcal{L}(\Sigma)$ and $\mathbf{w}_s=\mathbf{v}\circ \hat{F}_{\mathbf{v}}[\mathbf{x}|_{\mathcal{L}(\Sigma)},s\kappa]^{-1}$ is a transverse section to $\Lambda_s$. Moreover, setting $$\kappa_s=\left( \frac{d}{ds}\hat{\mathcal{F}}_\mathbf{v}[\mathbf{x}|_{\mathcal{L}(\Sigma)},s\kappa]\right)\circ\hat{F}_\mathbf{v}[\mathbf{x}|_{\mathcal{L}(\Sigma)},s\kappa]^{-1}\in \mathcal{K}_{\mathbf{w}_s},$$ $B_{\Sigma,\mathbf{v}}[D_1\hat{G}_\mathbf{v}(\mathbf{x}|_{\mathcal{L}(\Sigma)},0)\varphi,\kappa]>0$ implies $B_{\Lambda_s,\mathbf{w}_s}[D_1\hat{G}_{\mathbf{w}_s}(\mathbf{x}|_{\mathcal{L}(\Lambda_s)},0)\varphi, \kappa_s]>0$.
2. For $\mathbf{v}^\prime\in C^{k,\alpha}_0\cap C^{k}_{0,\mathrm{H}}(\Sigma;\mathbb{R}^{n+1})$ satisfying that $\Vert\mathbf{v}^\prime-\mathbf{v}\Vert^{(0)}_{k,\alpha}<\epsilon$ and $\mathscr{L}_\Sigma^0\mathbf{v}^\prime\in C^{k-2,\alpha}_{-2}(\Sigma;\mathbb{R}^{n+1})$, setting $$\kappa^\prime=\frac{(\mathbf{v}\cdot\mathbf{n}_\Sigma)\kappa}{\mathbf{v}^\prime\cdot\mathbf{n}_\Sigma} \in \mathcal{K}_{\mathbf{v}^\prime},$$ if $B_{\Sigma,\mathbf{v}}[D_1\hat{G}_\mathbf{v}(\mathbf{x}|_{\mathcal{L}(\Sigma)},0)\varphi,\kappa]>0$, then $B_{\Sigma,\mathbf{v}}[D_1\hat{G}_{\mathbf{v}^\prime}(\mathbf{x}|_{\mathcal{L}(\Sigma)},0)\varphi,\kappa^\prime]>0$.
Here, $\hat{F}_\mathbf{v}$, $\hat{G}_\mathbf{v}$, $\hat{G}_{\mathbf{v}^\prime}$, $\hat{G}_{\mathbf{w}_s}$, $\mathcal{K}_\mathbf{v}$, $\mathcal{K}_{\mathbf{v}^\prime}$, $\mathcal{K}_{\mathbf{w}_s}$ and $\hat{\mathcal{F}}_{\mathbf{v}}$ are given in Section \[ModifiedSmoothDependSec\].
First, for simplicity, denote by $\mathbf{f}_s=\hat{F}_\mathbf{v}[\mathbf{x}|_{\mathcal{L}(\Sigma)},s\kappa]$. Observe that by our constructions, for $\epsilon$ sufficiently small $\mathbf{f}_s\in\mathcal{ACH}^{k,\alpha}_n(\Sigma)$ and $\mathscr{L}_\Sigma \mathbf{f}_s\in C^{k-2,\alpha}_{-1}(\Sigma;\mathbb{R}^{n+1})$. Thus, by Lemma \[ComposeLem\], for $\epsilon$ sufficiently small $\hat{F}_{\mathbf{w}_s}$ and $\hat{G}_{\mathbf{w}_s}$ are well defined.
Using definitions of the maps from Theorem \[ModifiedSmoothDependThm\] and the rapid decay of the various terms one has $$B_{\Sigma,\mathbf{v}}[D_1\hat{G}_\mathbf{v}(\mathbf{x}|_{\mathcal{L}(\Sigma)},s\kappa)\varphi,\kappa_s\circ \mathbf{f}_s]=B_{\Sigma, \mathbf{v}}\left[{\frac{d}{dt}\vline}_{t=0}\hat{\Xi}_{\mathbf{v}}[\hat{F}_{\mathbf{v}}[\mathbf{x}|_{\mathcal{L}(\Sigma)}+t\varphi ,s\kappa]], \kappa_s\circ \mathbf{f}_s\right].$$ Arguing as in Section \[VariationSubsec\], the fact that the $\mathbf{f}_s$ are stationary for the functional $E$ and the second variation formula for $E$ give $$\begin{aligned}
{\frac{d}{dt}\vline}_{t=0} & \hat{\Xi}_{\mathbf{v}}[\hat{F}_{\mathbf{v}}[\mathbf{x}|_{\mathcal{L}(\Sigma)}+t\varphi ,s\kappa]]\\
&= {\frac{d}{dt}\vline}_{t=0}\hat{\Xi}_{\mathbf{f}_s, \mathbf{v}}[ t (\mathbf{v}\cdot (\mathbf{n}_{\Lambda_s}\circ \mathbf{f}_s))^{-1} D_1 \hat{F}_{\mathbf{v}}(\mathbf{x}|_{\mathcal{L}(\Sigma)} ,s\kappa)\varphi\cdot (\mathbf{n}_{\Lambda_s}\circ \mathbf{f}_s) ]\\
&= D\hat{\Xi}_{\mathbf{f}_s, \mathbf{v}}(0) ( (\mathbf{v}\cdot (\mathbf{n}_{\Lambda_s}\circ \mathbf{f}_s))^{-1} D_1 \hat{F}_{\mathbf{v}}(\mathbf{x}|_{\mathcal{L}(\Sigma)} ,s\kappa)\varphi\cdot (\mathbf{n}_{\Lambda_s}\circ \mathbf{f}_s))\\
&= \Omega_{\mathbf{f}_s, \mathbf{v}} L_{\mathbf{f}_s, \mathbf{v}}^\prime( (\mathbf{v}\cdot (\mathbf{n}_{\Lambda_s}\circ \mathbf{f}_s))^{-1} D_1 \hat{F}_{\mathbf{v}}(\mathbf{x}|_{\mathcal{L}(\Sigma)} ,s\kappa)\varphi\cdot (\mathbf{n}_{\Lambda_s}\circ \mathbf{f}_s))\\
&=\Omega_{\mathbf{f}_s, \mathbf{v}}(\mathbf{v}\cdot (\mathbf{n}_{\Lambda_s}\circ \mathbf{f}_s))^{-1} L_{\mathbf{f}_s}( D_1 \hat{F}_{\mathbf{v}}(\mathbf{x}|_{\mathcal{L}(\Sigma)} ,s\kappa)\varphi\cdot (\mathbf{n}_{\Lambda_s}\circ \mathbf{f}_s)).\end{aligned}$$ Hence, $$\begin{aligned}
& B_{\Sigma,\mathbf{v}}[D_1\hat{G}_\mathbf{v}(\mathbf{x}|_{\mathcal{L}(\Sigma)},s\kappa)\varphi,\kappa_s\circ \mathbf{f}_s] \\
&= B_{\Sigma, \mathbf{v}}[ \Omega_{\mathbf{f}_s, \mathbf{v}} ( \mathbf{v}\cdot (\mathbf{n}_{\Lambda_s}\circ \mathbf{f}_s) )^{-1} L_{\mathbf{f}_s}( D_1 \hat{F}_{\mathbf{v}}(\mathbf{x}|_{\mathcal{L}(\Sigma)} ,s\kappa)\varphi\cdot (\mathbf{n}_{\Lambda_s}\circ \mathbf{f}_s)), \kappa_s\circ \mathbf{f}_s]\\
& =B_{\Sigma}[ \Omega_{\mathbf{f}_s} ( \mathbf{v}\cdot (\mathbf{n}_{\Lambda_s}\circ \mathbf{f}_s) ) L_{\mathbf{f}_s}(D_1 \hat{F}_{\mathbf{v}}(\mathbf{x}|_{\mathcal{L}(\Sigma)} ,s\kappa)\varphi\cdot (\mathbf{n}_{\Lambda_s}\circ \mathbf{f}_s)), \kappa_s\circ \mathbf{f}_s]\\
&=B_{\mathbf{f}_s} [ L_{\mathbf{f}_s}(D_1 \hat{F}_{\mathbf{v}}(\mathbf{x}|_{\mathcal{L}(\Sigma)} ,s\kappa)\varphi \cdot (\mathbf{n}_{\Lambda_s}\circ \mathbf{f}_s)), ( \mathbf{v}\cdot ( \mathbf{n}_{\Lambda_s}\circ \mathbf{f}_s) ) \kappa_s\circ \mathbf{f}_s]\\
&= B_{\Lambda_s} [( \mathbf{w}_s\circ \mathbf{n}_{\Lambda_s} ) \kappa_s , L_{\Lambda_s}((D_1 \hat{F}_{\mathbf{v}}(\mathbf{x}|_{\mathcal{L}(\Sigma)} ,s\kappa)\varphi\circ \mathbf{f}_s^{-1}) \cdot \mathbf{n}_{\Lambda_s}]\\
& = \int_{\Lambda_s} (\mathbf{w}_s\cdot\mathbf{n}_{\Lambda_s})\kappa_s L_{\Lambda_s}\left((D_1\hat{F}_{\mathbf{v}}(\mathbf{x}|_{\mathcal{L}(\Sigma)},s\kappa)\varphi\circ \mathbf{f}_s^{-1})\cdot\mathbf{n}_{\Lambda_s}\right) e^{\frac{|\mathbf{x}|^2}{4}} \, d\mathcal{H}^n.\end{aligned}$$ Furthermore, the above integral can be simplified by using integration by parts (which is justified by Proposition 6.2 of [@BWBanachManifold]) to $$\int_{\mathcal{L}(\Lambda_s)} \mathrm{tr}_\infty^*[(\mathbf{w}_s\cdot\mathbf{n}_{\Lambda_s})\kappa_s] \mathrm{tr}^1_\infty[(D_1\hat{F}_\mathbf{v}(\mathbf{x}|_{\mathcal{L}(\Sigma)},s\kappa)\varphi\circ\mathbf{f}_s^{-1})\cdot\mathbf{n}_{\Lambda_s}]\, d\mathcal{H}^{n-1}$$ where $\mathrm{tr}_\infty^*$ is given in Section 6 of [@BWBanachManifold]. By Item of Theorem \[ModifiedSmoothDependThm\] and the linearity of $\mathrm{tr}_\infty^1$, we have that $\mathcal{L}(\Lambda_s)=\mathcal{L}(\Sigma)$, $\mathrm{tr}_\infty^1[\mathbf{f}_s]=\mathbf{x}|_{\mathcal{L}(\Sigma)}$ and $$\mathrm{tr}_\infty^1[(D_1\hat{F}_{\mathbf{v}}(\mathbf{x}|_{\mathcal{L}(\Sigma)},s\kappa)\varphi\circ\mathbf{f}_s^{-1})\cdot\mathbf{n}_{\Lambda_s}]=\varphi\cdot\mathbf{n}_{\mathcal{L}(\Lambda_s)}.$$ Thus, it follows that $$B_{\Sigma,\mathbf{v}}[D_1\hat{G}_\mathbf{v}(\mathbf{x}|_{\mathcal{L}(\Sigma)},s\kappa)\varphi,\kappa_s\circ \mathbf{f}_s]=\int_{\mathcal{L}(\Lambda_s)} \mathrm{tr}_\infty^*[(\mathbf{w}_s\cdot\mathbf{n}_{\Lambda_s})\kappa_s] (\varphi\cdot\mathbf{n}_{\mathcal{L}(\Lambda_s)}) \, d\mathcal{H}^{n-1}.$$ In particular, evaluating the identity at $s=0$ gives $$B_{\Sigma,\mathbf{v}}[D_1\hat{G}_\mathbf{v}(\mathbf{x}|_{\mathcal{L}(\Sigma)},0)\varphi,\kappa]=\int_{\mathcal{L}(\Sigma)} \mathrm{tr}_\infty^*[(\mathbf{v}\cdot\mathbf{n}_{\Sigma})\kappa] (\varphi\cdot\mathbf{n}_{\mathcal{L}(\Sigma)}) \, d\mathcal{H}^{n-1}.$$ Notice that, in the derivation of the above identity, we only use the fact that $\Sigma$ is an asymptotically conical self-expander embedded in a family of asymptotically conical self-expanders with the same asymptotic cone, that $\mathbf{v}$ is a transverse section on $\Sigma$ satisfying the hypotheses of the lemma and that $\kappa\in \mathcal{K}_{\mathbf{v}}$. Thus, the same identity holds true with $\Sigma$ replaced by $\Lambda_{s}$, $\mathbf{v}$ by $\mathbf{w}_{s}$ and $\kappa$ by $\kappa_s$, that is, $$B_{\Lambda_s,\mathbf{w}_s}[D_1\hat{G}_{\mathbf{w}_s}(\mathbf{x}|_{\mathcal{L}(\Lambda_s)},0)\varphi,\kappa_s]=\int_{\mathcal{L}(\Lambda_s)} \mathrm{tr}_\infty^*[(\mathbf{w}_s\cdot\mathbf{n}_{\Lambda_s})\kappa_s] (\varphi\cdot\mathbf{n}_{\mathcal{L}(\Lambda_s)}) \, d\mathcal{H}^{n-1}.$$ Hence, $$B_{\Lambda_{s},\mathbf{w}_{s}}[D_1\hat{G}_{\mathbf{w}_{s}}(\mathbf{x}|_{\mathcal{L}(\Lambda_{s})},0)\varphi,\kappa_{s}]=B_{\Sigma,\mathbf{v}}[D_1\hat{G}_\mathbf{v}(\mathbf{x}|_{\mathcal{L}(\Sigma)},s\kappa)\varphi,\kappa_{s}\circ \mathbf{f}_{s}].$$ Therefore, after possibly shrinking $\epsilon$, the first claim follows from this and the fact that $$s\mapsto B_{\Sigma,\mathbf{v}}[D_1\hat{G}_\mathbf{v}(\mathbf{x}|_{\mathcal{L}(\Sigma)},s\kappa)\varphi,\kappa_s\circ \mathbf{f}_s]$$ is smooth near $s=0$.
Finally, by what we have shown, $$B_{\Sigma,\mathbf{v}}[D_1\hat{G}_\mathbf{v}(\mathbf{x}|_{\mathcal{L}(\Sigma)},0)\varphi,\kappa]=B_{\Sigma,\mathbf{v}^\prime}[D_1\hat{G}_{\mathbf{v}^\prime}(\mathbf{x}|_{\mathcal{L}(\Sigma)},0)\varphi,\kappa^\prime],$$ proving the second claim.
Given two points $\varphi_0, \varphi_1 \in \mathcal{V}\backslash\Pi(\mathcal{S}_1)$ let $C=\Phi([0,1]])$ be the image of a generic curve in $\mathcal{V}$ connecting $\varphi_0$ to $\varphi_1$ as given by Lemma \[BaireLem\]. As $C$ is embedded and transverse to $\Pi$, $C^\prime=\Pi^{-1}(C)\cap \mathcal{U}$ is a (not necessarily connected) one-dimensional submanifold with boundary. Moreover, $$\partial C^\prime=\Pi^{-1}(\partial C)\cap \mathcal{U}=\Pi^{-1}({\left\{\varphi_0, \varphi_1\right\}})\cap \mathcal{U}.$$ Notice that $C^\prime$ is compact as $\Pi|_{\mathcal{U}}$ is proper.
Orient $C$ so $\varphi_0$ is the initial point. We now orient $C^\prime$ as follows:
1. If $[\mathbf{f}]\in C^\prime\backslash \mathcal{S}_1$, then $\Pi$ is a local diffeormorphism at $[\mathbf{f}]$ (by [@BWBanachManifold Theorem 1.1] or Theorem \[ModifiedSmoothDependThm\]). That is, there is an open neighborhood $\mathcal{U}^\prime\subset \mathcal{U}$ of $[\mathbf{f}]$ so that $\Pi|_{\mathcal{U}^\prime}$ is a diffeomorphism onto its image. Notice $\Pi(C^\prime\cap \mathcal{U}^\prime)\subset C$. Let $I^\prime=\mathcal{U}^\prime\cap C^\prime$ and orient $I^\prime$ so $\Pi|_{I^\prime}$ is orientation preserving if $[\mathbf{f}]$ has even Morse index and $\Pi|_{I^\prime}$ is orientation-reversing if $[\mathbf{f}]$ has odd Morse index.
2. If $[\mathbf{f}]\in C^\prime\cap \mathcal{S}_1$, then first observe that as $C$ is transverse to $\Pi$, $[\mathbf{f}]\in \mathcal{S}_1\backslash \mathcal{S}_2$. In this case a neighborhood, $I^\prime$, of $[\mathbf{f}]$ embeds in $\tilde{C}\times {\mathbb R}$ as $\hat{G}_{\mathbf{v}}^{-1}(0)$ (with $[\mathbf{f}]$ corresponding to $(\mathbf{x}|_{\mathcal{L}(\mathbf{f}(\Gamma))},0)$) where $$\tilde{C}={\left\{\mathrm{tr}_\infty^1[\mathscr{E}^\mathrm{H}_1[\varphi]\circ\mathcal{C}[\mathbf{f}]^{-1}]\colon \varphi\in C\right\}}$$ and $$D_{1} \hat{G}_{\mathbf{v}}(\mathbf{x}|_{\mathcal{L}(\mathbf{f}(\Gamma))},0)\neq 0.$$ The orientation of $\tilde{C}$ and the standard one of ${\mathbb R}$ determine an orientation on $\tilde{C}\times {\mathbb R}$. If $\mathrm{ind}(\mathbf{f})$ is even, orient $I^\prime=\hat{G}_{\mathbf{v}}^{-1}(0)$ so it is the boundary of $\hat{G}_{\mathbf{v}}>0$ while if $\mathrm{ind}(\mathbf{f})$ is odd, give $I^\prime$ the reverse orientation.
If $[\mathbf{f}]\in C^\prime\cap\mathcal{S}_1$ can be approximated by a sequence of regular points on $C^\prime$, then, by Theorem \[MainTechThm\], these two conventions are consistent. If $\Pi^{-1}(\mathrm{tr}_\infty^1[\mathbf{f}])$ has a component containing $[\mathbf{f}]$ that has a nonempty open subset of $C^\prime\cap\mathcal{S}_1$, then Lemma \[OrientationLem\] ensures that the second convention orients this component independent of choices. Thus, the above convention orients $C^\prime$ and the degree $\mathrm{deg}(\Pi|_\mathcal{U}, \mathcal{V})$ is just the usual degree of the map $\Pi|_{C^\prime}\colon C^\prime\to C$.
Proof of Theorem \[ApplicationThm\] {#ApplicationSec}
===================================
Finally, we prove Theorem \[ApplicationThm\]. First, we use the shrinking rotationally symmetric torus described by Angenent [@Angenent] to prove the following non-existence result for connected asymptotically conical self-expanders whose asymptotic cone is a “narrow" double cone.
\[NonExistLem\] There is a $\delta_0=\delta_0(n)>0$ so that: If $\mathcal{C}$ is a $C^{k,\alpha}$-regular cone in ${\mathbb R}^{n+1}$ with both $\mathcal{C}\cap {\left\{x_{n+1}>0\right\}}$ and $\mathcal{C}\cap {\left\{x_{n+1}<0\right\}}$ non-empty and $\mathcal{C}\subset {\left\{x_1^2+\ldots +x_n^2\leq \delta_0^2 x_{n+1}^2\right\}}$, then any self-expander, $\Sigma\in \mathcal{ACH}_n^{k,\alpha}$ with $\mathcal{C}(\Sigma)=\mathcal{C}$ is disconnected.
Let $\Lambda$ to the rotationally symmetric self-shrinker given by Angenent in [@Angenent] chosen so the axis of symmetry is $\ell_{n+1}$, the $x_{n+1}$-axis. That is, $\Lambda$ is of the form $$\Lambda= {\left\{f(x_1^2+\ldots+ x_n^2, x_{n+1})=0\right\}}$$ for some function $f\colon{\mathbb R}^2\to {\mathbb R}$. The construction of $\Lambda$ ensures there are $0<R_-<\sqrt{2n}< R_+$ so that $$\Lambda \subset \bar{B}_{R_+}\backslash B_{R_-} \mbox{ and } \Lambda\cap {\left\{x_{n+1}=0\right\}}=\left(\partial B_{R_+} \cup \partial B_{R_-}\right)\cap {\left\{x_{n+1}=0\right\}}.$$ As $\Lambda$ is disjoint from $\ell_{n+1}$, there is a $\delta_0=\delta_0(n)>0$ so that $$\Lambda \cap \left({\left\{x_1^2+\ldots +x_n^2\leq \delta^2_0 x_{n+1}^2\right\}}\cup B_{R_-}\right)=\emptyset.$$ In particular, for the given cone $\mathcal{C}\cap \Lambda=\emptyset$.
Consider the mean curvature flows $\mathcal{M}={\left\{\sqrt{t}\, \Sigma\right\}}_{t>0}$ and $\tilde{\mathcal{M}}={\left\{\sqrt{1-t}\, \Lambda\right\}}_{t<1}$ of $\Sigma$ and of $\Lambda$, respectively. As $\mathcal{C}(\Sigma)\cap \Lambda=\emptyset$ and $\Lambda$ is compact, the parabolic maximum principle implies that $$\sqrt{t}\, \Sigma\cap \sqrt{1-t}\, \Lambda=\emptyset$$ for all $t\in (0,1)$.
If $\Sigma$ is connected, then $\Sigma_0=\Sigma\cap {\left\{x_{n+1}=0\right\}}$ is non-empty as ${\left\{x_{n+1}>0\right\}}\cap \Sigma$ and ${\left\{x_{n+1}<0\right\}}\cap \Sigma$ are both non-empty by the hypotheses on $\mathcal{C}=\mathcal{C}(\Sigma)$. Moreover, as $\mathcal{C}(\Sigma)\cap {\left\{x_{n+1}=0\right\}}={\left\{\mathbf{0}\right\}}$, $\Sigma_0$ is compact. In particular, for $t$ sufficiently small $\sqrt{t}\, \Sigma_0\subset B_{\sqrt{1-t}\, R_-}$. Hence, there is a $t_*\in (0,1)$ so $\sqrt{t_*}\, \Sigma_0\cap \partial B_{\sqrt{1-t_*} \, R_-}\neq \emptyset$. The definition of $R_-$ means $\sqrt{t_*}\, \Sigma_0\cap \partial B_{\sqrt{1-t_*} \, R_-}\subset \sqrt{1-t_*} \, \Lambda$. As this contradicts the disjointness of the flows, one must have that $\Sigma$ is not connected.
We next use a construction of Angenent-Ilmanen-Chopp [@AIC] of connected rotationally symmetric self-expanders and a variational argument sketched by Ilmanen [@IlmanenNotes] and carried out by Ding [@Ding] in order to show existence of a connected self-expander asymptotic to a “wide" double cone.
\[ExistLem\] Fix $2\leq n \leq 6$. There is a $\delta_1=\delta_1(n)>0$ so that: If $\mathcal{C}$ is a $C^{k,\alpha}$-regular cone in ${\mathbb R}^{n+1}$ with $\mathcal{C}\subset {\left\{x_1^2+\ldots +x_n^2\geq \delta^2_1 x_{n+1}^2\right\}}$ and so that $\mathcal{L}[\mathcal{C}]=\mathcal{L}_+\cup \mathcal{L}_-$ where $\mathcal{L}_\pm$ are connected, $\mathcal{L}_+ \subset {\left\{ x_{n+1}>0\right\}}$ and $\mathcal{L}_-\subset {\left\{x_{n+1}<0\right\}}$ and $0\neq [\mathcal{L}_-]=[\mathcal{L}_+]\in H^{n-1}({\mathbb R}^{n+1}\backslash \ell_{n+1}; \mathbb{Z}_2)$, where $\ell_{n+1}$ is the $x_{n+1}$-axis, then there is a connected asymptotically conical self-expander, $\Sigma\in \mathcal{ACH}_n^{k,\alpha}$ with $\mathcal{C}(\Sigma)=\mathcal{C}$. Moreover, this $\Sigma$ may be chosen to be (weakly) stable.
By the construction of Angenent-Ilmanen-Chopp [@AIC], there is a $\delta>\delta_0>0$ so that there is a connected rotationally symmetric self-expander $\Lambda \in \mathcal{ACH}^{k,\alpha}_n$ with $$\mathcal{C}(\Lambda)={\left\{x_1^2+\ldots+x_n^2=\delta^2 x_{n+1}^2\right\}}.$$ This self-expander has the additional property that $\Lambda\cap \ell_{n+1}=\emptyset$, that is, it is disjoint from the $x_{n+1}$-axis. As such we may choose $\Omega$ to be the connected component of ${\mathbb R}^{n+1}\backslash \Lambda$ that is disjoint from $\ell_{n+1}$.
Now let $\delta_1=2\delta>0$. This choice ensures that there is an $R_0>0$ so for $R>R_0$, $R\mathcal{L}[\mathcal{C}]\subset \Omega\cap \partial B_R$. Using $\Lambda$ and $\partial B_R$ as barriers, one uses the direct method to find a smooth self-expander $\Sigma_R\subset \Omega\cap B_R$ with $\partial \Sigma_R= R\mathcal{L}[\mathcal{C}]$ – here we use the hypotheses that $[\mathcal{L}[\mathcal{C}]]=[\mathcal{L}_+]+[\mathcal{L}_-]=2[\mathcal{L}_+]=0$ and so $\mathcal{L}[\mathcal{C}]$ is null-homologous in $H^{n-1}({\mathbb R}^{n+1}\backslash \ell_{n+1}; \mathbb{Z}_2)$. As $\mathcal{L}_\pm$ are connected and there are no closed self-expanders, $\Sigma_R$ is connected.
Finally, by suitably modifying the argument of [@IlmanenNotes] (see also [@Ding Theorem 6.3] for full details) one verifies that, up to passing to a subsequence, one has $\lim_{R_i\to \infty} \Sigma_{R_i}=\Sigma$ with $\Sigma\in \mathcal{ACH}^{k,\alpha}_n$ and $\mathcal{C}(\Sigma)=\mathcal{C}$. Clearly, $\Sigma$ is connected and (weakly) stable.
For the sake of completeness, we include a brief explanation of how the argument of Ding in [@Ding Theorem 6.3] should be modified. The main issue is that Ding uses barriers to show that the limit $\Sigma$ is $C^0$ close to the cone $\mathcal{C}$ and one might worry that these are not consistent with the barriers used to construct $\Sigma_R$. To that end, consider the following family of functions depending on parameters $\mathbf{v}\in \mathbb{S}^n$, $\eta>0$ and $h\geq 0$ : $$f_{\mathbf{v}, \eta, h}(\mathbf{x})=2n+ |\mathbf{x}|^2-\left(1+\eta^{-2}\right) (\mathbf{x}\cdot \mathbf{v})^2+h.$$ We observe that the connected closed set $$E_{\mathbf{v}, \eta, h}={\left\{\mathbf{x}\in\mathbb{R}^{n+1}\colon f_{\mathbf{v}, \eta, h}(\mathbf{x})\leq 0\mbox{ and } \mathbf{x}\cdot \mathbf{v}\geq 0\right\}}$$ has boundary asymptotic to the connected, rotationally symmetric cone, $$\mathcal{C}_{\mathbf{v}, \eta}={\left\{|\mathbf{x}|^2= \left(1+\eta^{-2}\right)(\mathbf{x}\cdot \mathbf{v})^2, \mathbf{x}\cdot \mathbf{v}\geq 0\right\}},$$ that lies in the half-space $\mathbf{x\cdot \mathbf{v}\geq 0}$ and has axis parallel to $\mathbf{v}$ and cone angle $\eta^{-1}$. Furthermore, $E_{\mathbf{v}, \eta, h}$ lies entirely in the component of ${\mathbb R}^{n+1}\backslash \mathcal{C}_{\mathbf{v}, \eta}$ that contains $\mathbf{v}$. Observe that, by construction, $f_{\mathbf{v}, \eta,h}>0$ on ${\left\{|\mathbf{x}\cdot \mathbf{v}|<\eta (2n+h)\right\}}$ and so $$E_{\mathbf{v}, \eta, h}\cap {\left\{\mathbf{x}\cdot \mathbf{v}<\eta (2n+h)\right\}}=\emptyset.$$ A straightforward computation gives that on any self-expander, $\Sigma$, $$\mathscr{L}_{\Sigma}^0 f_{\mathbf{v}, \eta, h} -f_{\mathbf{v}, \eta, h} =-2\left(\eta^{-2}+1\right)|\nabla_\Sigma (\mathbf{v}\cdot\mathbf{x})|^2 -h\leq 0.$$
One readily checks that for each $p\in \mathcal{L}(C)$ and choice of normal $\mathbf{N}$ to $\mathcal{L}[\mathcal{C}]$ at $p$, there are $\eta>0$, $h>0$ and $\mathbf{v}_\pm\in \mathbb{S}^{n}$ so that $E_p=E_{\mathbf{v}_\pm, \eta, h}$ satisfies $E_p\subset \Omega \backslash \mathcal{C}$, $E_p\subset H_p$ where $H_p$ is the component of ${\left\{x_{n+1}\neq 0\right\}}$ containing $p$ and $\mathcal{L}[\mathcal{C}_{\mathbf{v_\pm}, \eta}]$ lies on the $\pm \mathbf{N}$ side of $\mathcal{L}[\mathcal{C}]$ and touches $\mathcal{L}[\mathcal{C}]$ only at $p$. By construction, $f_{\mathbf{v}_\pm, \eta, h}>0$ on the component of $\partial \Sigma_R$ that lies in $H_p$. Moreover, $f_{\mathbf{v}_\pm, \eta,h}>0$ on ${\left\{x_{n+1}=0\right\}}$. Hence, by the strong maximum principle, $f_{\mathbf{v}_\pm, \eta, h}>0$ on $\Sigma_R\cap H_p$. That is, $\Sigma_R\cap E_p=\emptyset$. As $$\left(\mathscr{L}_\Sigma^0-1\right)(|\mathbf{x}|^2+2n)=0$$ the strong maximum principle and compactness of $\mathcal{L}[\mathcal{C}]$ imply that, there is an $R^\prime>0$ so that $B_{R^\prime}\cap E_p\neq\emptyset$ for all $p\in\mathcal{L}[\mathcal{C}]$ and that $\Sigma_R\setminus B_{R^\prime}$, $R>R^\prime$, has exactly two components, one of which contains $R\mathcal{L}_+$ and the other contains $R\mathcal{L}_-$. Thus, by construction and further enlarging $R^\prime$, we have that, for $p\in\Sigma_R\setminus B_{R^\prime}$, $\mathrm{dist}(p,\mathcal{C})=O(|\mathbf{x}(p)|^{-1})$. Hence one obtains from this that $\Sigma$ is $C^0$ asymptotic to $\mathcal{C}$. The argument that $\mathcal{C}(\Sigma)=\mathcal{C}$ then proceeds exactly as in [@Ding Theorem 6.3].
We need to following refinement of a result of Fong-McGrath [@FongMcGrath].
\[SymmetryLem\] Let $\Sigma\in \mathcal{ACH}^{k,\alpha}_n$ be a (weakly) stable self-expander. If $\mathcal{C}(\Sigma)$ is symmetric with respect to rotation around the $x_{n+1}$-axis, $\ell_{n+1}$, then so is $\Sigma$.
Let $\mathbf{R}$ be a non-trivial rotational Killing vector field in ${\mathbb R}^{n+1}$ that is a symmetry of $\mathcal{C}(\Sigma)$. Consider $f=\mathbf{R}\cdot \mathbf{n}_{\Sigma}$. One readily checks that $f\in C^{k-1,\alpha}_{1, 0}\cap C^{\infty}_{loc}(\Sigma)$ and $L_{\Sigma} f =0$ (see, e.g., [@FongMcGrath Lemma 2.3]). As $\Sigma$ has a weakly conical end (cf. ), it follows from [@Bernstein Theorem 9.1] that $f\in W_{\frac{1}{4}}^{1}(\Sigma)$.
In particular, either $f$ identically vanishes and so $\mathbf{R}$ is a symmetry of $\Sigma$ or $f$ is an eigenfunction of $-L_{\Sigma}$ with eigenvalue $0$. As $\Sigma$ is (weakly) stable this latter situation would imply that $f$ is the lowest eigenfunction of $-L_\Sigma$ and so $f$ cannot change sign. However, as $\ell_{n+1}\cap \mathcal{C}(\Sigma)={\left\{\mathbf{0}\right\}}$, there must be a $p\in \Sigma$ that satisfies ${\mathrm{dist}}(p, \ell_{n+1})={\mathrm{dist}}(\Sigma, \ell_{n+1})$. Clearly, $f(p)=0$ and so one is in the situation that $f$ identically vanishes. As $\mathbf{R}$ was arbitrary, one deduces that $\Sigma$ is also rotationally symmetric about $\ell_{n+1}$.
By [@BWProperness], the map $\Pi$, for surfaces of a fixed topological type in ${\mathbb R}^3$, is proper. Lemma \[NonExistLem\] implies the degree of $\Pi$ is zero for annuli and so the existence of unstable solutions will follow from Lemma \[ExistLem\].
Let $$\Gamma_0= {\left\{x_1^2+x_2^2=x_3^2+1\right\}}\subset {\mathbb R}^3.$$ Observe that $\Gamma_0\in \mathcal{ACH}_2^{k,\alpha}$ and is a topological annulus. Let $\mathcal{V}$ be the component of $\mathcal{V}^{k,\alpha}_{\mathrm{emb}}(\Gamma_0)$ that contains $\mathbf{x}|_{\mathcal{L}(\Gamma_0)}$. Let $\mathcal{U}=\Pi^{-1}(\mathcal{V})$ where $$\Pi\colon \mathcal{ACE}_{2}^{k,\alpha}(\Gamma_0)\to C^{k,\alpha}(\mathcal{L}(\Gamma_0); {\mathbb R}^3)$$ is the natural projection map from [@BWBanachManifold]. By [@BWProperness], $\Pi_0=\Pi|_{\mathcal{U}}\colon \mathcal{U}\to \mathcal{V}$ is a proper map. Hence, by Theorem \[MainThm\], $\Pi_0$ possesses an integer degree $\mathrm{deg}(\Pi|_\mathcal{U}, \mathcal{V})$.
Write $\varphi=(\varphi_1, \varphi_2, \varphi_3)$. For $\lambda>0$, let $$\varphi_\lambda= (\varphi_1, \varphi_2, \lambda \varphi_3).$$ One readily checks that $\varphi_{\lambda}\in \mathcal{V}$ for all $\lambda > 0$. Moreover, one has $$\mathcal{C}_\lambda =\mathcal{C}[\varphi_\lambda(\mathcal{L}(\Gamma_0))]={\left\{ x_1^2+x_2^2=\lambda^{-2} x_3^2\right\}}.$$ On the one hand, for $\lambda_0=2 \delta_0^{-1}$ where $\delta_0$ is given by Lemma \[NonExistLem\], there is no connected self-expander with asymptotic cone $\mathcal{C}_{\lambda_0}$. As $\Gamma_0$ is connected, one has $\Pi^{-1}(\varphi_{\lambda_0})=\emptyset$ and so $\mathrm{deg}(\Pi|_\mathcal{U}, \mathcal{V})=0$. On the other hand, by the construction of Angenent-Ilmanen-Chopp [@AIC], there is a $\lambda_1\in (0, \lambda_0)$, so that there is a connected rotationally symmetric expander, $\Lambda$, asymptotic to $\mathcal{C}_{\lambda_1}$. That is, $\Pi^{-1}(\varphi_{\lambda_1})\neq \emptyset$. Clearly, we may assume this $\lambda_1$ satisfies $\lambda_1^{-1}>\delta_1(2)$ where $\delta_1(2)$ is given by Lemma \[ExistLem\] for $n=2$.
To conclude the proof we need to find a $\varphi\in \mathcal{V}$ so that $\varphi$ is a regular value of $\Pi$ and so $\Pi^{-1}(\varphi)$ is non-empty. To that end, first let $\Omega$ be the unique component of ${\mathbb R}^3 \backslash \mathcal{C}(\Lambda)$ that is disjoint from the $x_3$-axis. Next, observe that, as the set of regular values of $\Pi$ is generic (by [@BWBanachManifold]), there is a sequence, $\varphi_i\in \mathcal{V}$ of regular values of $\Pi$ with the property that $\varphi_i\to \varphi$ in $C^{k,\alpha}(\mathcal{L}(\Gamma_0); {\mathbb R}^3)$ and $\varphi_i(\mathcal{L}(\Gamma_0)) \subset \Omega$.
It follows from Lemma \[ExistLem\] that there is a sequence of (weakly) stable connected self-expanders $\Sigma_i\in \mathcal{ACH}^{k,\alpha}_2$ with $\mathcal{L}(\Sigma_i)= \varphi_i(\mathcal{L}(\Gamma_0))$. By the compactness results of [@BWProperness], up to passing to a subsequence and relabeling, the $\Sigma_i\to \Sigma$ in the smooth topology where $\Sigma$ is a (weakly) stable, connected, self-expander with $\mathcal{C}(\Sigma)=\mathcal{C}_{\lambda_1}$. By Lemma \[SymmetryLem\], as $\Sigma$ is (weakly) stable and $\mathcal{C}(\Sigma)$ is rotationally symmetric around the $x_3$-axis, $\Sigma$ is also rotationally symmetric. As $\Sigma$ is also connected this means that $\Sigma$ is an annulus and so, by the nature of the convergence, up to throwing out a finite number of terms, the $\Sigma_i$ are also annuli for all $i$.
As a consequence, there are $\mathbf{f}_i\in \mathcal{ACH}_2^{k,\alpha}(\Gamma_0)$ with $\mathbf{f}_i(\Gamma_0)=\Sigma_i$ and $\mathrm{tr}_\infty^1 [\mathbf{f}_i]=\varphi_i$. In particular, $[\mathbf{f}_i]\in \Pi^{-1}_0(\varphi_i)$ are stable elements. As the degree of $\Pi_0$ is $0$, this means that there must be at least one unstable element $[\mathbf{f}_i^\prime]\in \Pi^{-1}_0(\varphi_i)$. As $\Pi$ is a local diffeomorphism, at each regular point, the claim is proved by fixing some $i_0\geq 1$ large and letting the open set be some small neighborhood, in $C^{k,\alpha}$, of $\mathcal{C}(\Sigma_{i_0})$.
Poincaré inequality {#PoincareIneqnSec}
===================
\[GradControlsL2Prop\] Let $\Sigma\in \mathcal{ACH}_{n}^{k,\alpha}$ be a self-expander. For $\beta \geq \frac{1}{4}$ and $f\in W^{1}_{\beta}(\Sigma)$, $$\label{PI1eqn}
\int_{\Sigma} \left(4 n +|\mathbf{x}|^2\right)f^2 e^{\beta |\mathbf{x}|^2}\, d\mathcal{H}^n \leq 16 \int_{\Sigma}|\nabla_\Sigma f|^2 e^{\beta |\mathbf{x}|^2}\, d\mathcal{H}^n.$$ Likewise $$\label{PI2eqn}
(4\beta-1) \int_{\Sigma}|\mathbf{x}^\top|^2 f^2 e^{\beta |\mathbf{x}|^2} \, d\mathcal{H}^n \leq 4 \int_{\Sigma}|\nabla_\Sigma f|^2 e^{\beta |\mathbf{x}|^2}\, d\mathcal{H}^n.$$
First observe that as $\Sigma$ is a self-expander one has that $$\mathscr{L}_{\Sigma}^0 \left( |\mathbf{x}|^2+2n\right)=|\mathbf{x}|^2+2n.$$ Now suppose that $f$ has compact support. Integrating by parts gives $$\begin{aligned}
&\int_\Sigma \left(2 n +|\mathbf{x}|^2\right)f^2 e^{\beta |\mathbf{x}|^2} \, d\mathcal{H}^n = \int_\Sigma \left(\mathscr{L}_\Sigma^0 \left(2 n +|\mathbf{x}|^2\right) \right) f^2 e^{\beta |\mathbf{x}|^2}\, d\mathcal{H}^n \\
&= - \int_\Sigma \nabla_\Sigma \left(2 n +|\mathbf{x}|^2\right) \cdot \nabla_\Sigma \left( e^{\left( \beta-\frac{1}{4}\right)|\mathbf{x}|^2} f^2 \right) e^{\frac{|\mathbf{x}|^2}{4}}\, d\mathcal{H}^n \\
&=-\int_{\Sigma} \left( (4\beta-1) |\mathbf{x}^\top|^2 f^2 + 4 \phi \mathbf{x}^\top \cdot \nabla_\Sigma f \right) e^{\beta |\mathbf{x}|^2} \, d\mathcal{H}^n \\
&= - (4\beta-1) \int_{\Sigma} |\mathbf{x}^\top|^2 f^2 e^{\beta |\mathbf{x}|^2} \, d\mathcal{H}^n -4 \int_{\Sigma} f \mathbf{x}^\top \cdot \nabla_\Sigma f \, e^{\beta |\mathbf{x}|^2}\, d\mathcal{H}^n.\end{aligned}$$ The absorbing inequality and the hypothesis that $\beta\geq \frac{1}{4}$ give $$\int_\Sigma \left(2 n +|\mathbf{x}|^2\right) f^2 e^{\beta |\mathbf{x}|^2}\, d\mathcal{H}^n \leq \int_{\Sigma}\left( \frac{1}{2} |\mathbf{x}^\top|^2 f^2 + 8| \nabla_\Sigma f|^2 \right) e^{\beta |\mathbf{x}|^2}\, d\mathcal{H}^n.$$ Rearranging this gives the first inequality. The second inequality follows from the same computation but using a different version of the absorbing inequality.
To see that these estimates hold for all $f\in W^{1}_{\beta}(\Sigma)$ one uses the compactly supported cutoff $\psi_R\in C^\infty({\mathbb R}^{n+1})$ with the property that $0\leq \psi_R\leq 1$, $\psi_R=1$ in $B_R$, ${{\ensuremath{{\mathrm{spt}}\:}}}\psi_R\subset B_{2R}$ and $|\nabla \psi_R|\leq 2$. The estimates hold for $f_R=\psi_R f$ and hence, by the dominated convergence theorem, hold also for $f$.
\[GradControlsBoundaryProp\] Let $\Sigma\in \mathcal{ACH}_{n}^{k,\alpha}$ be a self-expander. Suppose $R_0\geq 1$ is such that on $\bar{E}_{R_0}$ the estimates hold with constant $C_0$. There is an $R_1=R_1(C)\geq R_0$ so that: For $\beta \geq \frac{1}{4}$, $f \in W^{1}_{\beta}(\Sigma)$ and $R>R_1$ one has the estimate $$R e^{\beta R^2} \int_{S_R} f^2\, d\mathcal{H}^{n-1} \leq 2 \int_{\bar{E}_R} |\nabla_\Sigma f|^2 e^{\beta |\mathbf{x}|^2} \, d\mathcal{H}^n.$$
Suppose first that $f$ has compact support. Let $$\mathbf{Z}= \frac{|\mathbf{x}| \mathbf{x}^\top}{|\mathbf{x}^\top|}=\frac{\frac{1}{2}\nabla_\Sigma |\mathbf{x}|^2}{|\nabla_\Sigma |\mathbf{x}||}.$$ By , $|\nabla_\Sigma |\mathbf{x}||\geq \frac{1}{2}$ so this vector field is well-defined. Using , one sees that, by taking $R_1$ sufficiently large, on $\bar{E}_{R_1}$, one has $$|\mathbf{x}^\top| \geq |\mathbf{x}|-\frac{C}{|\mathbf{x}|^3}\geq |\mathbf{x}| -\frac{1}{|\mathbf{x}|} \mbox{ and } \mathrm{div}_\Sigma \mathbf{Z} \geq n-\frac{1}{2} \geq \frac{1}{2}.$$ Hence, on $\bar{E}_R$, $$\begin{aligned}
\mathrm{div}_\Sigma \left( f^2 e^{\beta |\mathbf{x}|^2} \mathbf{Z} \right) &= \left( f^2 \mathrm{div}_\Sigma \mathbf{Z}+ 2 f \nabla_\Sigma f \cdot \mathbf{Z} + 2\beta f^2\mathbf{x}^\top \cdot \mathbf{Z}\right) e^{\beta|\mathbf{x}|^2} \\
& \geq \left(\frac{1}{2} f^2+ 2 f \nabla_\Sigma f \cdot \mathbf{Z} + \frac{1}{2} f^2 |\mathbf{Z}|\left(|\mathbf{x}|-|\mathbf{x}|^{-1}\right) \right) e^{\beta|\mathbf{x}|^2} \\
&\geq \left( \frac{1}{2} f^2-2 |\nabla_\Sigma f |^2 - \frac{1}{2} f^2 |\mathbf{Z}|^2 + \frac{1}{2} f^2 |\mathbf{Z}| \left(|\mathbf{x}|-|\mathbf{x}|^{-1}\right)\right) e^{\beta|\mathbf{x}|^2}\end{aligned}$$ where the last estimate uses the absorbing inequality. As $|\mathbf{Z}|=|\mathbf{x}|$ one has $$\mathrm{div}_\Sigma \left( f^2 e^{\beta |\mathbf{x}|^2} \mathbf{Z} \right) \geq -2 | \nabla_\Sigma f |^2 e^{\beta|\mathbf{x}|^2}.$$ Integrating over $\bar{E}_R$ and appealing to the divergence theorem give $$\begin{aligned}
R e^{\beta R^2} \int_{S_R} f^2 \, d\mathcal{H}^{n-1} &= -\int_{S_R} f^2 \mathbf{Z}\cdot \frac{-\mathbf{x}^\top}{|\mathbf{x}^\top|} e^{\beta|\mathbf{x}|^2} \, d\mathcal{H}^{n-1}\\
&=\int_{\bar{E}_R} -\mathrm{div}_\Sigma \left( f^2 e^{\beta |\mathbf{x}|^2} \mathbf{Z} \right) \, d\mathcal{H}^n \leq 2 \int_{\bar{E}_R} |\nabla_\Sigma f |^2 e^{\beta|\mathbf{x}|^2} \, d\mathcal{H}^n,\end{aligned}$$ which proves the claim for $f$ of compact support. For general $f\in W^{1}_\beta(\Sigma)$ one uses cutoffs and the dominated convergence theorem.
[99]{}
S.B. Angenent, *Shrinking doughnuts*, Nonlinear Diffusion Equations and Their Equilibrium States, 3 (Gregynog, 1989), 21–38, Progr. Nonlinear Differential Equations Appl., 7, Birkhäuser Boston, Boston, MA, 1992.
S.B. Angenent, T. Ilmanen and D.L. Chopp, *A computed example of non-uniqueness of mean curvature flow in ${\mathbb R}^3$*, Commun. in Partial Differential Equations 20 (1995), no. 11-12, 1937–1958.
J. Bernstein, *Asymptotic structure of almost eigenfunctions of drift Laplacians on conical ends*, Preprint. Available at <https://arxiv.org/abs/1708.07085>.
J. Bernstein and L. Wang, *The space of asymptotically conical self-expanders of mean curvature flow*, Preprint. Available at <https://arxiv.org/abs/1712.04366>.
J. Bernstein and L. Wang, *Smooth compactness for spaces of asymptotically conical self-expanders of mean curvature flow*, Preprint. Available at <https://arxiv.org/abs/1804.09076>.
Q. Ding, *Minimal cones and self-expanding solutions for mean curvature flows*, Preprint. Available at <http://arxiv.org/abs/1503.02612v2>.
K. Ecker and G. Huisken, *Mean curvature evolution of entire graphs*, Ann. of Math. (2) 130 (1989), no. 3, 453–471.
F. Fong and P. McGrath, *Rotational symmetry of asymptotically conical mean curvature flow self-expanders*, Comm. Anal. Geom., to appear. Available at <https://arxiv.org/abs/1609.02105>.
D. Gilbarg and N.S. Trudinger, Elliptic partial differential equations of second order, Reprint of the 1998 Edition, Classics in Mathematics, Springer-Verlag, Berlin, 2001.
A. Grigor’yan, Heat kernel and analysis on manifolds, AMS/IP Studies in Advanced Mathematics, 47, American Mathematical Society, Providence, RI; International Press, Boston, MA, 2009.
S. Helmensdorfer, *A model for the behaviour of fluid droplets based on mean curvature flow*, Preprint. Available at <https://arxiv.org/abs/1102.3546>.
T. Ilmanen, Lectures on mean curvature flow and related equations, Unpublished Notes. Available at <https://people.math.ethz.ch/~ilmanen/papers/notes.ps>.
S. Smale, *An infinite dimensional version of Sard’s theorem*, Amer. J. Math. 87 (1965), 861–866.
B. White, *The space of $m$-dimensional surfaces that are stationary for a parametric elliptic functional*, Indiana Univ. Math. J. 36 (1987), no. 3, 567–602.
B. White, *The space of minimal submanifolds for varying Riemannian metrics*, Indiana Univ. Math. J. 40 (1991), no. 1, 161–200.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We present the results of our study of astrometric stability of 200-in Hale (Mt. Palomar) and 10-m Keck II (Mauna Kea) telescopes, both with Adaptive Optics (AO) facilities. A group of nearby visual binaries and multiples was observed in near infrared, relative separations and position angles measured. We have also checked the influence of some systematic effects (e.g. atmospherical refraction, varying plate scale factor) on result and precision of astrometric measurements. We conclude that in visual binaries astrometrical observations it is possible to achieve much better precision than 1 miliarcsecond \[$mas$\], which in many cases allows detection of the astrometrical signal produced by planetary-mass object.'
address: 'Nicolaus Copernicus Astronomical Center, Rabiańska 8, 87-100 Toruń, Poland'
author:
- 'Krzysztof He[ł]{}miniak'
- Maciej Konacki
title: |
Precise Astrometry of Visual Binaries with Adaptive Optics.\
A way for finding exoplanets?
---
Introduction
============
Astrometry is thought to be the most promising method of exoplanets detection in the future. On the contrary to radial velocities measurements (RV), astrometry is almost independent on stellar physics, e.g. phenomena like the activity or pulsations, rather than the distance to the object. Todays interferometers are able to achieve precision at the level of mili- or microarcseconds, which is good enough to attempt exoplanet research.
Nevertheless, the same level is possible to achieve in small fields by CCD imaging with adaptive optics (AO) systems. To do that, one must subtract and correct some systematical effects in order to obtain a gaussian distribution of the measurements. In such case, the precision improves like $N^{-0.5}$, where $N$ is a number of single measurements. The goal of our studies was to check if a gaussian statistic can really be achieved with two top-class telescopes with AO systems: the 200-in Hale telescope + PHARO camera (Mt. Palomar) and 10-m Keck II telescope + NIRC2 (Mauna Kea), both working in infrared (IR). With these facilities we were imaging 17 objects, majority of which were M-type dwarf binaries, located closer than 20 $pc$ from the Sun.
Observations
============
Most of the observations were made with Hale telescope and [*Palomar High-Angular Resolution Observer*]{} (PHARO). The camera was using three wide-bandpass filters – K, K’ and Ks – and two narrow-band, centered on Br$\gamma$ and FeII lines (Hayward [@hay01]). PHARO’s imaging mode was used with 39.91 and (mostly) 25.10 $mas/pix$ scale. With this telescope we made about 30,000 images of 12 binaries/multiples – these are: GJ 195, GJ 352, GJ 458, GJ 507, GJ 661, GJ 767, GJ 860, GJ 873, GJ 9071 and MWC 1080 – and two fields in open clusters, i.e. NGC 1039 and NGC 6871. Observations were made between October 2001 and November 2002, during 11 nights in groups of 2 to 4. Objects were observed in $dithering$ mode, which means small shifts of the telescope’s position in order to subtract a variable IR background and reduce the influence of the chip’s distortion.
Three other multiples were observed with Keck II + [*Near Infra-Red Camera 2*]{} (NIRC2). These were 56 Per, GJ 300 and GJ 596. Main components were saturated, so the astrometry could be done only for double secondaries[^1]. Keck II was used 4th March 2002 in imaging mode with 9.942 and 39.686 $mas/pix$ scale, J, K’ and K-cont (narrow-band) filters, also with [ *dithering*]{} and, additionally, field rotation.
The method
==========
The closure of companions allows one to observe visual binaries in a smaller field, where atmospheric distortions are weakly affecting the measurements, star’s image is better sampled, and the measurements errors in pixels transfer to smaller errors in arcseconds. But, instead of parallax and proper motion, one have to take into account the orbital motion. Note also that in case of a positive detection of a 3-rd body, relative astrometry does not give the information around which particular star of the system the body is orbiting. There are also some systematic effects, which, in every astrometric research, need to be corrected in order to achieve gaussian statistics (white noise) of the measurements. For checking if the white-noise-behavior is present, we used Allan variance ($\sigma_A$), which for a random dispersed values shows -1 slope on a log-log plot of $\sigma_A$ vs. so-called [*lag*]{}[^2].
AO correction
-------------
Proper AO correction is a crucial factor. Our example of GJ 352 (separation $\simeq
350\, mas$) shows how the quality of measurements is dependent on it. For only 10 properly AO-corrected frames, $rms$ of the separation $\rho$ was 4.2 $mas$ while for 62 others (in some of them the components were not resolved), $rms$ reached 19.8 $mas$.
PHARO chip geometry
-------------------
Depending on the position of the binary on the chip, we were obtaining various values of $\rho$, what obviously reflects variations of the plate-scale around the chip. We adopted a model of distortion with a linear dependency of the plate-scale on the $x$ and $y$ coordinates of the chip, similar to the one found by Metchev & Hillenbrand ([@met04]). After this correction, the measurements truly followed the white-noise-behavior (Fig. 1).
We have also found a small rotation of PHARO, between August and November 2002, by a value of $0.641 \pm 0.013\, deg$. This is also in agreement with results obtained by Metchev & Hillenbrand ([@met04]).
Atmospheric refraction
----------------------
Precision bellow 1 $mas$ requires atmospheric refraction to be carefully taken into account. It’s zenithal distance ($z$) dependency causes that observed separation of two stars is smaller than the true one (due to stars’ different $z$’s). In our research we used a semi-full approach, described by Hełminiak ([@hel08]), which is suitable for IR.
Precision and detection limits
==============================
Assuming $3\sigma$ detection limit and transferring Pravdo & Shaklan’s ([@pra96]) Equation 8 (Sect. 6, p. 270), we get a relation: $$a M_P [AU \cdot M_{Jup}] = 1562.5 \frac{\sigma_{\rho} M_S}{\pi} \left[ \frac{mas \cdot M_{\odot}}{mas} \right]$$ where $a$ is a semi-major axis of a planet, $M_P$ is it’s mass, $\sigma_{\rho}$ is the precision of separation measurements, $M_S$ is the star’s mass and $\pi$ is it’s parallax. $3\sigma_{\rho}$ is used instead of the astrometric signal $\Theta$. This relation was used to calculate, what kind of planets can be found in a particular system. For stars from our sample with known mass and distance, best precision achieved and it’s corresponding limits are given in Table 1. The values show that with Hale and Keck II telescopes it is possible to find Jupiter- or even Saturn-mass bodies around given stars.
For several systems, observed in all epochs, we also calculated a level of a long-term astrometric precision. We fitted 2-nd order polynomials, and also obtained $rms$’ at the level of 100 $\mu as$ or sometimes better[^3].
Conclusions
===========
Todays telescopes with AO systems allows us to perform astrometric measurements with precision well bellow 1 $mas$. This refers to a single-epoch observations, as well as to a long-term stability. For many cases such a precision means an ability to detect exoplanets around nearby stars.
{width="0.95\columnwidth"}
---------- ------------------ -------- --------------- ------------------ --------------- ------------------ ------
Star Lowest Dist. Mass A Limit for A Mass B Limit for B Tel.
(GJ No.) $\sigma\, [mas]$ $[pc]$ $[M_{\odot}]$ $[AU \cdot M_J]$ $[M_{\odot}]$ $[AU \cdot M_J]$
195 0.12 13.89 0.53 1.38 0.19 0.50 Hale
352 1.11 10.53 0.44 8.04 0.41 7.49 Hale
458 0.28 15.32 0.40 2.68 0.37 2.48 Hale
507 0.33 13.16 0.46 3.12 0.37 2.51 Hale
569B 0.11 9.81 0.071 0.116 0.054 0.088 Keck
661 0.038 6.32 0.379 0.16 0.34 0.15 Hale
767 0.09 13.35 0.44 0.83 0.40 0.75 Hale
860 0.048 4.01 0.34 0.10 0.27 0.09 Hale
873 0.57 5.05 0.36 1.62 unknown unknown Hale
9071 0.20 13.89 0.53 2.22 0.49 2.05 Hale
---------- ------------------ -------- --------------- ------------------ --------------- ------------------ ------
: Smallest $\sigma_{\rho}$ and corresponding detection limits for researched stars.
Hayward, T. L. 2001, PASP ,113, 105 Hełminiak, K. G. 2008, arXiv:0805.3369v2 \[astro-ph\] Lane, B. F. and Muterspaugh, M. W. 2004, ApJ, 601, 1129 Metchev, S. A. and Hillenbrand, L. A. 2004, ApJ, 617, 1330 Pravdo, S. H. and Shaklan, S. B. 1996, ApJ, 465, 264
[^1]: Two ”secondary” stars of GJ 300 are actually field stars. The same is for GJ 873.
[^2]: For more details see Pravdo & Shaklan ([@pra96]) or Lane & Muterspaugh ([@lan04]).
[^3]: Measurements, achieved precisions, limits and fittings for all systems can be found at http://www.ncac.torun.pl/$\sim$xysiek/source/mgr.pdf Ch.6, pp.94-105 (in polish, uncorrected for distortion!).
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'This article proposes a collaborative control framework for an autonomous aerial swarm tasked with the surveillance of a convex region of interest. Each Mobile Aerial Agent (MAA) is equipped with a Pan-Tilt-Zoom (PTZ) camera of conical FOV and suffers from sensor-induced positional uncertainty. By utilizing a Voronoi-free tessellation strategy and a gradient scheme, the heterogeneous swarm self-organizes in a distributed manner to monotonically achieve optimal collective visual coverage of the region of interest, both in terms of quality and total area. Simulation studies are offered to investigate the effectiveness of the suggested scheme.'
author:
-
-
-
-
bibliography:
- 'Bibliography/SP\_bibliography.bib'
- 'Bibliography/SP\_conferences.bib'
- 'Bibliography/SP\_journals.bib'
nocite: '[@Bousias_ECC2019]'
title: '**Distributed surveillance by a swarm of UAVs operating under positional uncertainty**'
---
Cooperating Robots, Swarms, Multi-Robot Systems, Area Coverage
Introduction
============
Mobile robot teams have several potential applications, one of the frequently studied ones being area coverage problems. Coverage problems can be broadly categorized as static or sweep coverage. In static or blanket coverage [@Abbasi_Aut2017; @Pierson_IJRR2017] the objective of the mobile agents is a static configuration at which some performance criterion is optimized. In dynamic or sweep coverage problems [@Palacios_IEEETACR2016; @Franco_EJC2015a] the performance criterion is time–varying, resulting in the agents moving constantly. Other ways to categorize coverage problems are based on the properties of the region of interest [@Kantaros_Automatica15; @Alitappeh_SC2016], of the dynamic model of the mobile agents [@Sharifi_IEEETCST2015; @Luna_CDC2010] or on the type of their onboard sensors [@Stergiopoulos_ICRA14; @Arslan_ICRA2016]. The most common approach to coverage problems is geometric optimization [@Stergiopoulos_IEEETAC15] with other proposed approaches being event–triggered control [@Nowzari_Automatica2012], game theory [@Ramaswamy_ACC2016], annealing algorithms [@Kwok_IEEETCST2011] and model predictive control [@Mavrommati_IEEETR2018].
An inherent characteristic of most positioning systems is the uncertainty in their measurements. Some proposed solutions applied to mobile robots are probabilistic methods [@Habibi_IEEETAC2016], safe trajectory planning [@Davis_IEEERAL2016] or the use of Voronoi–like partitioning schemes [@Papatheodorou_IJARS2018; @Tzes_LCSS2018]. In this article the positioning uncertainty model is similar to the one used in [@Papatheodorou_IJARS2018; @Tzes_LCSS2018] but the approach followed differs. Instead of employing a Voronoi–like space partitioning, the positioning uncertainty is incorporated in the agents’ sensing patterns and the sensed space is partitioned using a Voronoi–free technique similar to [@Papatheodorou_RAS2017].
Aerial agents are a popular platform for area coverage tasks due to their high mobility and versatility. The case of cameras with up to 3 translational and 3 rotational degrees of freedom has been examined in [@Schwager_IEEE2011] and an algorithm for information exchange has also been developed. In this work however the cameras are not allowed to zoom and the cameras’ localization is precise. Moreover, regions covered by multiple agents contribute more to the objective, thus favoring overlapping between the agents’ sensed regions. Previous works have examined downwards facing cameras [@Papatheodorou_RAS2017; @Papatheodorou_CDC2017] and although positioning uncertainty has been successfully incorporated in these control schemes [@Tzes_LCSS2018], it was done using a Voronoi–like partitioning which is not easy to generalize in the case of pan-tilt-zoom cameras. Pan-tilt-zoom camera networks have been examined using Voronoi–like diagrams in [@Arslan_ICRA2018] although in that work the cameras were stationary instead of being affixed on mobile agents. In the present work the MAAs have 3 translational degrees of freedom and are equipped with pan-tilt-zoom cameras. The MAAs’ planar positioning uncertainty is taken into account by using a Voronoi–free partitioning scheme. Additionally, regions sensed by multiple agents do not contribute more to the objective, thus favoring separation of the MAAs’ sensing patterns. It should be noted that what the best approach concerning the overlapping of sensing patterns depends entirely on the intended use–case and the one used in this article can not be considered strictly better or worse than the one used in [@Schwager_IEEE2011].
Problem Statement
=================
We assume a compact convex region $\Omega \in \mathbb{R}^2$ to be placed under surveillance by a swarm of $n$ MAAs, each positioned at $X_i=\left[x_i,y_i,z_i\right]^T, i\in I_n$ where $I_n=\{1, \dots, n\}$. We define the vector $q_i = \left[x_i,y_i\right]^T \in \Omega$ denoting the projection of each MAA on the plane. Each MAA can fly within a predefined altitude range, thus $z_i \in \left[ z_i^{\min},z_i^{\max} \right], i\in I_n$ with $z_i^{\min}
\leq z_i^{\max}$. These altitude constraints ensure the safe operation of the MAA by avoiding collisions with ground obstacles as well as ensuring they remain within communication range of their base stations.
In addition, each MAA comes equipped with an onboard visual sensor capable of pan and tilt movements. Moreover, the sensor has a conical field of view and is able to alter its zoom. We denote the sensor’s pan and tilt angles $h_i$ and $\theta_i$ respectively, while its zoom level is represented by the angle of of the cone of vision which is denoted $2 \delta_i$ with $\delta_i \in
\left[\delta_i^{\min} , \delta_i^{\max} \right]$ and $\delta_i^{\min} \leq
\delta_i^{\max} < \frac{\pi}{2}$.
Given the conical field of view of the sensors, its intersection with the plane will be a conic section which we call the sensing pattern. The sensing pattern is the region of the plane an MAA is able to cover. More specifically, it is a circle for $h_i = 0$, an ellipse for $0 < \left|h_i\right| < \frac{\pi}{2} -
\delta_i$, a parabola for $\frac{\pi}{2} - \delta_i \leq \left|h_i\right| \leq
\frac{\pi}{2} + \delta_i$ and a hyperbola for $\frac{\pi}{2} + \delta_i <
\left|h_i\right|$. In the sequel we will examine only the case where $\left|h_i\right| < \frac{\pi}{2} - \delta_i$, i.e. circular and elliptical sensing patterns, in order to always have the sensing pattern bounded by a curve. In order to have static boundaries for the tilt angle we will constrain it inside the interval $\left(-h_i^{\max}, h_i^{\max} \right) \subseteq
\left(-\frac{\pi}{2} + \delta_i, \frac{\pi}{2} - \delta_i \right)$ where $h_i^{\max} = \frac{\pi}{2} - \delta_i^{\max}$.
We define the center of the sensing footprint $q_{c,i} \in \Omega$ and denote the semi-major and semi minor axis of the elliptic sensing pattern $a_i$ and $b_i$ respectively. The unit vector $w_i \in \mathbb{R}^2$ indicates the orientation of the sensing pattern and is parallel to the semi-major axis of the ellipse if the sensing pattern is elliptical. These result in each MAA’s sensing pattern being $$\label{eq:sensing}
{C_i^s} (X_i,h_i,\theta_i,\delta_i) = \mathbf{R}\left(\theta_i\right) C_i^b
+ q_{i,c}, \quad i \in I_n,$$ where $\mathbf{R}$ is the ${2\times2}$ rotation matrix, $||.||$ is the Euclidean metric and $$\begin{aligned}
C_i^b &= \left\{ q \in \mathbb{R}^2 \colon \left\Vert \left[ \begin{matrix}
\frac{1}{a_i} & 0 \\ 0 & \frac{1}{b_i} \end{matrix} \right] {q} \right\Vert
\leq 1 \right\}, \\
a_i &= \frac{z_i}{2} \left[ \tan(h_i + \delta_i) - \tan(h_i - \delta_i)
\right], \\
b_i &= {z_i} \tan(\delta_i) \sqrt{1 + \left[ \frac{ \tan(h_i + \delta_i) +
\tan(h_i - \delta_i)}{2} \right]^2}, \\
{q_{i,c}} &= {q_i} + w_i \frac{z_i}{2} \left[ \tan(h_i + \delta_i) +
\tan(h_i - \delta_i) \right], \\
w_i &= \begin{bmatrix} \cos(\theta_i) & \sin(\theta_i) \end{bmatrix}^T.\end{aligned}$$ The pan angle $\theta_i$ only affects the sensing pattern’s orientation, while the tilt angle $h_i$ affects the eccentricity of the elliptical sensing pattern. It can be shown that if $h_i = 0$ then $a_i = b_i$ and $C_i^s$ degenerates into a circle with $q_{c,i} = q_i$.
For the sake of simplicity, instead of using a complete dynamic model for the MAAs such as quadrotor dynamics, a single integrator kineamtic model is used instead. The MAAs are approximated by point masses able to move in $\mathbb{R}^3$. It is assumed that the visual sensors’ pan and tilt angles and zoom can be controlled by onboard servos, thus their states are decoupled from those of the MAA. Therefore the kinematic model of each MAA is $$\begin{aligned}
\label{eq:kin_q}
\dot{q_i} &= u_{i,q}, \quad q_i \in \Omega, & u_{i,q} \in \mathbb{R}^2, \\
\label{eq:kin_z}
\dot{z_i} &= u_{i,z}, \quad z_i \in \left[ z_i^{\min}, z_i^{\max} \right],
& u_{i,z} \in \mathbb{R}, \\
\label{eq:kin_theta}
\dot{\theta_i} &= u_{i,\theta}, \quad \theta_i \in \mathbb{R}, & u_{i,\theta}
\in \mathbb{R}, \\
\label{eq:kin_h}
\dot{h_i} &= u_{i,h}, \quad h_i \in \left( -h_i^{\max}, h_i^{\max} \right),
& u_{i,h} \in \mathbb{R}, \\
\label{eq:kin_delta}
\dot{\delta_i} &= u_{i,\delta}, \quad \delta_i \in \left[\delta_i^{\min},
\delta_i^{\max} \right], & u_{i,\delta} \in \mathbb{R}.\end{aligned}$$
The projection on the ground $q_i \in \Omega$ of each agent’s position is assumed to be known with a degree of uncertainty, whereas each MAA’s altitude $z_i$, sensor pan angle $\theta_i$, tilt angle $h_i$ and view angle $\delta_i$ are known with certainty. Given an upper bound $r_i$ for the positioning uncertainty of each MAA, its footprint $q_i$ may reside anywhere with a disk called the positioning uncertainty region. The positioning uncertainty region, denoted $C_i^u$, is defined as $$\label{eq:uncertainty}
{C_i^u} (q_i, r_i) = \Big\{ q \in \Omega : \left\Vert q-q_i \right\Vert
\leq r_i \Big\}, \quad i \in I_n.$$
Given the positioning uncertainty of each MAA, we also define the guaranteed sensed region $C_i^{gs} \subseteq C_i^s \subset \mathbb{R}^2$ as the region the MAA is guaranteed to cover for all its possible positions within $C_i^u$. The guaranteed sensed region is then defined as $$\begin{aligned}
\nonumber
{C_i^{gs}} (X_i,h_i,\theta_i,\delta_i,r_i) &\stackrel{\triangle}{=} \left\{
\underset{q_i \in C_i^u}{\bigcap} {C_i^s} \right\} \\
\label{eq:gsensing}
&= \mathbf{R}\left(\theta_i\right)C_i^{b_{gs}} + q_{i,c}, \quad i \in I_n,\end{aligned}$$ where $$\begin{aligned}
C_i^{b_{gs}}= \left\{ q \in \Omega \colon \left\Vert \left[ \begin{matrix}
\frac{1}{a_i-r_i} & 0 \\ 0 & \frac{1}{b_i-r_i} \end{matrix} \right] {q}
\right\Vert \leq 1 \right\}.\end{aligned}$$ Since the positioning uncertainty region $C_i^u$ and sensed region $C_i^s$ are circular and elliptical respectively, the guaranteed sensed region $C_i^{gs}$ is also an ellipse. If $C_i^s$ is a disk due to the tilt angle $h_i$ being $0$ then $C_i^{gs}$ will also be a disk. If $r_i=0$, i.e. the position of the MAA’s footprint is known precisely, then $C_i^{gs} = C_i^s$. As $r_i \to
\min(a_i,b_i)$ then $ C_i^gs$ approaches a line segment. For $r_i >
\min(a_i,b_i)$ we get that $C_i^{gs}=\emptyset$. Figure \[fig:problem\_statement\] illustrates the visual coverage concept described in this section. The agent’s guaranteed sensing pattern is shown filled in red for both $h_i = 0$ and $h_i \in \left( 0, \frac{\pi}{2} - \delta_i \right)$.
![Visual coverage concept.[]{data-label="fig:problem_statement"}](config3.eps){width="42.00000%"}
Due to the nature of visual sensors, objects further away from the sensor appear with lower quality than ones near the sensor. We model the coverage quality using a quality function $f_i \colon \left[z_i^{\min}, +\infty\right]
\rightarrow \left[0, 1\right]$, with $0$ and $1$ corresponding to the lowest and the highest possible quality respectively. In order to keep the control scheme simple, it is assumed that the coverage quality is uniform throughout each MAA’s sensed pattern. As the MAA’s altitude $z_i$ increases, the visual coverage quality of its sensed region decreases. The same is true while the sensor’s tilt angle $h_i$ increases, resulting in the center $q_{i,c}$ of the sensed pattern $C_i^s$ moving further away from the MAA. Similarly, zooming out, i.e. increasing $\delta_i$, also leads in a decrease in quality. However, the sensor’s pan angle $\theta_i$ as well as the agent’s footprint $q_i$ have no effect on the coverage quality. Except from being a decreasing function of $z_i$, $h_i$ and $\delta_i$, $f_i$ must also be first–order differentiable with respect to $z_i$, $h_i$ and $\delta_i$ within $C_i^{gs}$. This property is needed for computing the control law and will become apparent in the sequel.
Although any function having the previously mentioned properties could be chosen as the coverage quality function $f_i$, the following one was chosen arbitrarily
$$\label{eq:quality}
f_i (z_i,h_i,\delta_i) =
\begin{cases}
\frac{p\left(z_i, \ z_i^{\min}, \ z_i^{\max}\right)}{3} +
\frac{p\left(h_i, \ 0, \ h_i^{\max}\right)}{3} &\\
\quad + \frac{p\left(\delta_i, \ \delta_i^{\min}, \
\delta_i^{\max}\right)}{3}, &\ q \in C_i^{gs} \\
0, &\ q\notin C_i^{gs}
\end{cases},$$
where $$p\left(x, \ x^{\min}, \ x^{\max}\right) = \frac{\left(\left(x -
x^{\min}\right)^2 - \left(x^{\max} -
x^{\min}\right)^2\right)^2}{\left(x^{\max} - x^{\min}\right)^4}.$$ For the function $p$ it holds that $p\left(x^{\min}, \ x^{\min}, \
x^{\max}\right) = 1$ and that $p\left(x^{\max}, \ x^{\min}, \ x^{\max}\right) =
0$. Consequently $f_i (z_i^{\min}, 0, \delta_i^{\min}) = 1$ and $\underset{h_i
\to h_i^{\max}}{\lim} f_i (z_i^{\max}, h_i, \delta_i^{\max}) = 0$. The function $f_i$ is also dependent on the agent’s altitude, tilt and zoom constraints. It should be noted that this choice of quality function is not unique and that different quality functions result in different quality–coverage trade–offs.
Additionally, each point $q\in \Omega$ can be assigned an importance weight through a space density function $\phi \colon \Omega \rightarrow \mathbb{R}_+$ which expresses the a priori information regarding the importance of certain regions of $\Omega$. We define the following joint coverage-quality objective $$\label{eq:objective_initial}
\mathcal{H} \stackrel{\triangle}{=} \int_{\Omega}{ \ \max_{\substack{i \in
I_n}} f_i (z_i,h_i,\delta_i) \ \phi\left(q\right) \mathrm{d}q}.$$ This function accounts for both the area covered by the agents and the coverage quality over that area, while also taking into account the importance of points as encoded by $\phi(q)$. The goal of the MAA team is to maximize this objective. To that extent, a suitable partitioning scheme will be employed in order to distribute the computation of $\mathcal{H}$ among the agents. Then a gradient–based control law will be designed to lead the MAA team to a locally optimal configuration with respect to $\mathcal{H}$.
Area Partitioning Strategy
==========================
The most common choice of partitioning scheme for area coverage problems is the Voronoi diagram and similar diagrams inspired by it. Voronoi–like diagrams that can take into account the positioning uncertainty of mobile agents have been proposed in the past [@Papatheodorou_IJARS2018; @Tzes_LCSS2018]. However in this work a partitioning of just the sensed space is utilized, similarly to [@Papatheodorou_CDC2017]. This partitioning scheme assigns a region of responsibility (cell) to each agent based on guaranteed sensed regions $C_i^{gs}$ and the coverage quality over them. Each MAA is assigned a cell $W_i$ as follows $$\label{eq:partitioning}
W_i \stackrel{\triangle}{=} \left\{ q \in \Omega \colon \ f_i > f_j, \ i
\neq j \right\}, i \in I_n.$$ However the union these cells does not comprise a complete tessellation of the total guaranteed sensed region $\underset{i \in I_n}{\bigcup} C_i^{gs}$. This is due to the fact that regions sensed by multiple agents with the same coverage quality are left unassigned. These so called common regions still contribute towards the objective $\mathcal{H}$ so they must be taken into account. The set of agents with the same coverage quality $f^l$ and overlapping guaranteed sensed regions is $$\begin{aligned}
\mathcal{I}_l = &\left\{ i,j \in I_n, i \neq j \colon C_i^{gs} \cap
C_j^{gs} \neq \emptyset
\right.\\
&\quad\left.\wedge f_i = f_j = f^l \right\}, l \in I_L.\end{aligned}$$ The common regions are then computed as $$\begin{aligned}
\label{eq:partitioning_common}
W_c^l = \left\{ \exists i,j \in \mathcal{I}_l, i \neq j:q\in C_i^{gs} \cap
C_j^{gs} \right\},\quad l \in I_L.\end{aligned}$$
By utilizing the partitioning strategy , , coverage–quality objective can be written as $$\label{eq:objective}
\mathcal{H} = \sum_{\substack{i\in I_n}}
\int_{\substack{W_i}}{f_i \ \phi(q) \
\mathrm{d}q} + \sum_{l=1}^{L} \int_{\substack{W_c^l}}{ f^l
\ \phi(q) \ \mathrm{d}q}$$
We define the neighbors $N_i$ of an agent $i$ as $$\label{eq:neighbors}
N_i \stackrel{\triangle}{=} \left\{j \in I_n \setminus i \colon
C_i^{gs} \cap C_j^{gs} \neq \emptyset \right\}.$$ The neighbors of an agent $i$ are essentially the agents that affect the cell $W_i$ of agent $i$, thus they are the agents $i$ must be able to exchange information with.
By allowing the MAAs’ cameras to tilt it is possible that the sensed regions of two distant MAAs overlap. Since the partitioning scheme is based on the sensed regions, these MAAs should be able to communicate. However this might not always be practical given their distance. An algorithm for propagating state information in a mobile agent network has been proposed in [@Schwager_IEEE2011]. By utilizing this algorithm MAAs are able to exchange information with their neighbors in multiple hops instead of communicating directly.
Since the partitioning scheme , only partitions the guaranteed sensed region $\underset{i \in I_n}{\bigcup} C_i^{gs}$, a portion of $\Omega$ is left unpartitioned. This region is called the neutral region, is denoted $\mathcal{O}$ and can be computed as $$\begin{aligned}
%\nonumber
\label{eq:neutral}
\mathcal{O} &= \Omega \setminus \left\{ \underset{i \in I_n}{\bigcup}
{W_i} \cup \underset{i \in I_l}\bigcup{W_c^l} \right\}= \Omega \setminus \underset{i \in I_n}{\bigcup} C_i^{gs}.
\end{aligned}$$
Due to the fact that the coverage quality $f_i$ is constant throughout the guaranteed sensed region $C_i^{gs}$, the resulting cells $W_i$ are bounded by elliptical arcs of $C_i^{gs}$. Moreover, this partitioning scheme may result in some cells being non–convex, empty or consisting of multiple disjoint regions. However all of these cases are handled properly by the designed control law without the need for extensions.
Collaborative Control Development
=================================
Having defined the partitioning scheme , which allows distributing the computation of the objective $\mathcal{H}$ among the agents, what remains is the derivation of the gradient–based control law.
Given a team of MAAs with kinematics described by , , , , , sensing performance and positioning uncertainty , the following control law guarantees monotonic increase of the coverage–quality objective along the MAAs trajectories.
$$\begin{aligned}
\label{eq:control_q}
u_{i,q} &= K_{i,q} \Biggl[ \underset{\partial W_i \cap \partial
\mathcal{O}} {\int} {u_i^i \ n_i \ f_i \ \phi(q) \ \mathrm{d}q} \qquad\qquad\qquad\qquad\qquad \nonumber \\
&\quad\quad +\sum_{\substack{j \in I_n \\ j \neq i}} \underset{\partial W_i \cap
\partial W_j}{\int}{u_i^i \ n_i \left( f_i -f_j \right) \ \phi(q) \
\mathrm{d}q} \Biggr], \\
\label{eq:control_z}
u_{i,z} &= { K_{i,z} \Biggl[ \underset{\partial W_i \cap \partial
\mathcal{O}}{\int} {v_i^i \ n_i \ f_i \ \phi(q) \ \mathrm{d}q} + {
\underset{W_i}{\int}{ \frac{\partial f_i} {\partial z_i} \phi(q) \ \mathrm{d}q}} } \quad\quad \nonumber \\
&\quad\quad +\sum_{\substack{j \in I_n \\ j \neq i}} \underset{\partial W_i \cap
\partial W_j}{\int}{v_i^i \ n_i \left(f_i- f_j \right) \ \phi(q) \
\mathrm{d}q} \Biggr], \\
\label{eq:control_theta}
u_{i,\theta} &= K_{i,\theta} \Biggl[ \underset{\partial W_i \cap
\partial \mathcal{O}}{\int} {\tau_i^i \ n_i \ f_i \ \phi(q) \ \mathrm{d}q} \qquad\qquad\qquad\qquad\qquad \nonumber \\
&\quad\quad + \sum_{\substack{j \in I_n \\ j \neq i}} \underset{\partial W_i \cap
\partial W_j}{\int}{\tau_i^i \ n_i \left( f_i- f_j \right) \ \phi(q) \
\mathrm{d}q} \Biggr], \\
\label{eq:control_h}
u_{i,h} &={ K_{i,h} \Biggl[ \underset{\partial W_i \cap \partial
\mathcal{O}}{\int} {\sigma_i^i \ n_i \ f_i \ \phi(q) \ \mathrm{d}q} +
\underset{W_i}{\int}{ \frac{\partial f_i} {\partial h_i} \ \phi(q) \ \mathrm{d}q} } \quad\quad \nonumber \\
&\quad\quad {+ \sum_{\substack{j \in I_n \\ j \neq i}} \underset{\partial W_i
\cap \partial W_j}{\int}{\sigma_i^i \ n_i \left( f_i- f_j \right) \
\phi(q) \ \mathrm{d}q} } \Biggr], \\
\label{eq:control_delta}
u_{i,\delta} &={ K_{i,\delta} \Biggl[ \underset{\partial W_i \cap
\partial \mathcal{O}}{\int} {\mu_i^i \ n_i \ f_i \ \phi(q) \ \mathrm{d}q}
+\underset{W_i}{\int}{ \frac{\partial f_i} {\partial \delta_i} \ \phi(q) \: \mathrm{d}q} } \quad\quad \nonumber \\
&\quad\quad {+ \sum_{\substack{j \in I_n \\ j \neq i}} \underset{\partial W_i
\cap \partial W_j}{\int}{\mu_i^i \ n_i \left( f_i- f_j \right) \
\phi(q) \ \mathrm{d}q} } \Biggr],
\end{aligned}$$
where $K_{i,q},K_{i,z},K_{i,\theta},K_{i,h},K_{i,\delta}$ are positive constants, $n_i$ the outward pointing unit normal vector on $W_i$ and $u_i^i,v_i^i,\tau_i^i,\sigma_i^i,\mu_i^i$ the Jacobian matrices $$\begin{aligned}
u_j^i &\stackrel{\triangle}{=} \dfrac{\partial q}{\partial q_i},\quad q \in \partial W_j,i,j\in I_n \\
v_j^i &\stackrel{\triangle}{=} \dfrac{\partial q}{\partial z_i},\quad q \in \partial W_j,i,j\in I_n \\
\tau_j^i &\stackrel{\triangle}{=} \dfrac{\partial q}{\partial \theta_i},\quad q \in \partial W_j,i,j\in I_n \\
\sigma_j^i &\stackrel{\triangle}{=} \dfrac{\partial q}{\partial h_i},\quad q \in \partial W_j,i,j\in I_n \\
\mu_j^i &\stackrel{\triangle}{=} \dfrac{\partial q}{\partial \delta_i},\quad q \in \partial W_j,i,j\in I_n
\end{aligned}$$
In order to guarantee monotonic increase of $\mathcal{H}$, its time derivative is evaluated as $$\dfrac{\partial \mathcal{H}} {\partial t} = \dfrac{\partial \mathcal{H}} {\partial q_i} \dot{q_i} + \dfrac{\partial \mathcal{H}} {\partial z_i} \dot{z_i}+ \dfrac{\partial \mathcal{H}} {\partial \theta_i} \dot{\theta_i}+ \dfrac{\partial \mathcal{H}} {\partial h_i} \dot{h_i}+ \dfrac{\partial \mathcal{H}} {\partial \delta_i} \dot{\delta_i}$$ By selecting the following control inputs $$\begin{aligned}
u_{i,q} &= K_{i,q} \dfrac{\partial \mathcal{H}} {\partial q_i}, u_{i,z} = K_{i,z} \dfrac{\partial \mathcal{H}} {\partial z_i} , \nonumber\\
u_{i,\theta} &= K_{i,\theta} \dfrac{\partial \mathcal{H}} {\partial \theta_i} , u_{i,h} = K_{i,h} \dfrac{\partial \mathcal{H}} {\partial h_i} ,u_{i,\delta} = K_{i,\delta} \dfrac{\partial \mathcal{H}} {\partial\delta_i}, \end{aligned}$$
we guarantee, given the MAAs’ dynamics, that $\dfrac{\partial \mathcal{H}}
{\partial t} $ is non-negative since $$\begin{aligned}
&\dfrac{\partial \mathcal{H}} {\partial t} = \Bigg[ K_{i,q} \left( \dfrac{\partial \mathcal{H}} {\partial q_i} \right)^2 +K_{i,z} \left( \dfrac{\partial \mathcal{H}} {\partial z_i} \right)^2 + K_{i,\theta} \left( \dfrac{\partial \mathcal{H}} {\partial \theta_i} \right)^2 \qquad\qquad \qquad\qquad \nonumber\\
&+ K_{i,h} \left( \dfrac{\partial \mathcal{H}} {\partial h_i}+ \right)^2
+K_{i,\delta} \left( \dfrac{\partial \mathcal{H}} {\partial \delta_i} \right)^2
\Bigg] \geq 0, \nonumber\end{aligned}$$ where $K_{i,q},K_{i,z},K_{i,\theta},K_{i,h},K_{i,\delta}$ are positive constants, thus, ensuring that the coverage–quality criterion increases in a monotonic manner.
The partial derivative $\frac{\partial \mathcal{H}}{\partial q_i}$ is
$$\dfrac{\partial \mathcal{H}}{\partial q_i} = \dfrac{\partial}{\partial
q_i} \left\{ \sum_{\substack{i\in I_n}} \underset{W_i}{\int}{ f_i \
\phi(q) \ \mathrm{d}q} + \sum_{l=1}^{L} \underset{W_c^l}{\int}{ f^l \ \phi(q) \
\mathrm{d}q} \right\}.$$
By applying the Leibniz integral rule and since $\frac{\partial f_i
(z_i,h_i,\delta_i)}{\partial q_i}=\frac{\partial f_j(z_j,h_j,\delta_j)
}{\partial q_i}=0$ the previous equation yields $$\begin{aligned}
\dfrac{\partial \mathcal{H}} {\partial q_i} &= \underset{\partial
W_i}{\int} {u_i^i \ n_i \ f_i \ \phi(q) \ \mathrm{d}q} \\
&\quad + \sum_{\substack{j \in I_n \\ j \neq i}} \underset{\partial W_i \cap
\partial W_j}{\int}{u_j^i \ n_j \ f_j \phi(q) \ \mathrm{d}q}.\end{aligned}$$ We use a boundary decomposition of $\partial W_i$ into disjoint sets similarly to [@Papatheodorou_CDC2017] $$\begin{aligned}
\partial W_i &= \Bigg\{ \left\{ \partial W_i \cap \partial \Omega \right\} \cup \left\{ \partial W_i \cap \partial \mathcal{O} \right\} \cap \qquad\qquad\qquad \nonumber \\
&\qquad \Bigg\{ \bigcup_{\substack{i \neq j}}{\partial W_i \cap \partial W_j} \Bigg\} \cap \Bigg\{ \bigcup_{l=1}^{L}{\partial W_i \cap \partial W_c^l} \Bigg\} \Bigg\}\end{aligned}$$ and assuming a static region of interest $\Omega$. In addition, since $\partial
W_i\cup \partial W_c^l$ are subsets of some sensed region boundary $\partial
C_j$, independent of the state of node $i$, at $q \in \bigg\{ \partial \Omega
\cap \partial W_i \bigg\}$ and $q \in \bigg\{ \partial W_i \cup \partial
W_c^l \bigg\}$, the Jacobian matrix is $u_i^i=\bf0_{2\times2}$ resulting in the final expression for $\dfrac{\partial \mathcal{H}} {\partial q_i}$ $$\begin{aligned}
\dfrac{\partial \mathcal{H}} {\partial q_i} &= \underset{\partial W_i \cap
\partial \mathcal{O}}{\int} {u_i^i \ n_i f_i \ \phi(q) \ \mathrm{d}q} \\
&\qquad + \sum_{\substack{j \in I_n \\ j \neq i}} \underset{\partial W_i \cap
\partial W_j}{\int}{u_i^i \ n_i \left[f_i - f_j \right] \ \phi(q) \
\mathrm{d}q}.\end{aligned}$$
Through a similar procedure and given that $\frac{\partial f_j}{\partial
z_i}=\frac{\partial f_j}{\partial \theta_i}=\frac{\partial f_i}{\partial
\theta_i}=\frac{\partial f_j}{\partial h_i}=\frac{\partial f_j}{\partial
\delta_i}=0$ we obtain the rest of the control laws.
Simulation Studies
==================
Simulation studies were conducted in order to evaluate the efficiency of the proposed control strategy. For consistency, the region of interest $\Omega$ was selected to be the same as in [@Tzes_LCSS2018]. The space density function was assumed to be $\phi \left(q\right)=1,\forall q \in \Omega$, assigning equal importance to all points inside the region of interest. The camera state limits were $h_i^{\max} = 50^\circ$, $\delta_i^{\min} = 15^\circ$ and $\delta_i^{\max}
= 35^\circ, \ \forall i \in I_n$. The MAAs’ cells are shown filled in grey with solid black boundaries while the boundaries of guaranteed sensed regions are shown in dashed red lines in Figures 2 (a) and (b) and Figures 4 (a) and (b). The MAAs’ trajectories are shown as blue lines Figures 2 (b) and 4 (b).
Case study I
------------
This simulation examines the case of a team of 3 MAAs and it serves to highlight the fact that there exists a configuration with respect to the agents’ altitude $z_i$, camera tilt angle $h_i$ and zoom $\delta_i$ that is globally optimal. The MAAs altitude constraints were set to $ z_i^{min}=0.3$ and $z_i^{max}=3.8 \ \forall i \in I_3$. The initial and final configurations of the swarm are shown in Figures 2 (a) and 2 (b) respectively. We observe from Figure 2 (b) that the MAAs guaranteed sensed regions do not overlap and are completely contained inside $\Omega$. This indicates that the MAA team has reached a globally optimal configuration. It should be noted that due to the fact that no closed–form expression exists for the arc length of an ellipse, it is impossible to analytically compute the altitude, tilt angle and zoom that result in this configuration. This simulation study was also conducted with agents equipped with downward facing cameras unable to pan, tilt or zoom as in [@Papatheodorou_RAS2017]. Figure 3 shows the evolution of the coverage–quality objective $\mathcal{H}$ over time for both the pan-tilt-zoom and downwards facing cameras in solid black and dashed red respectively. It is observed that by allowing the MAAs’ cameras to pan, tilt and zoom, it is possible to achieve significantly higher coverage performance. Finally, it is observed that the monotonic increase of $\mathcal{H}$ has been achieved, confirming that the control design and implementation is correct.
[0.5]{} ![Simulation I[]{data-label="fig:sim1"}](simulations/case_study1/initial_config.eps "fig:")
[0.5]{} ![Simulation I[]{data-label="fig:sim1"}](simulations/case_study1/final_config_no_ptz.eps "fig:")
[0.5]{} ![Simulation I[]{data-label="fig:sim1"}](simulations/case_study1/final_config_ptz.eps "fig:")
![Case study I: Coverage–quality objective evolution.[]{data-label="fig:sim1_H"}](simulations/case_study1/H_compare.eps)
Case study II
-------------
A team of 6 MAAs is simulated in this case study. The MAAs altitude constraints were set to $z_i^{min}=0.3$ and $z_i^{max}=1.8 \ \forall i \in I_6$. The initial and final configurations of the MAA team are shown in Figures 4 (a) and 4 (b) respectively. Due to the greater number of agents in this simulation there is overlapping between their guaranteed sensed regions and the MAA team has not reached a globally optimal configuration with respect to $\mathcal{H}$. However the MAAs do reach a locally optimal configuration as was expected. This simulation study was also repeated with cameras unable to pan, tilt and zoom. Figure 5 shows the evolution of the coverage–quality objective $\mathcal{H}$ over time for both the pan-tilt-zoom and downwards facing cameras in solid black and dashed red respectively. Once again the benefits of using pan-tilt-zoom cameras become apparent and it is once again observed that $\mathcal{H}$ does indeed increase monotonically.
[0.45]{} ![Simulation II: Initial () and final configurations without PTZ-cameras () and with PTZ-cameras ().[]{data-label="fig:sim2"}](simulations/case_study2/initial_config.eps "fig:")
[0.45]{} ![Simulation II: Initial () and final configurations without PTZ-cameras () and with PTZ-cameras ().[]{data-label="fig:sim2"}](simulations/case_study2/final_config_no_ptz.eps "fig:")
![Simulation II: Initial () and final configurations without PTZ-cameras () and with PTZ-cameras ().[]{data-label="fig:sim2"}](simulations/case_study2/final_config_ptz.eps)
![Simulation II: Coverage–quality objective.[]{data-label="fig:sim2_H"}](simulations/case_study2/H_compare.eps)
Conclusions
===========
In this article, the efficiency collaborative visual aerial coverage by a swarm of MAAs, equipped with cameras able to pan, tilt and zoom while operating under positional uncertainty, has been examined. Regarding the partitioning, a Voronoi-free strategy has been utilized upon which a gradient-based control law was derived guaranteeing monotonic increase of the coverage quality criterion. Simulation studies were offered to evaluate the efficiency of the PTZ extended configuration in contrast with the downwards, not zoom-enabled cameras.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'New methods for $D$-decomposition analysis are presented. They are based on topology of real algebraic varieties and computational real algebraic geometry. The estimate of number of root invariant regions for polynomial parametric families of polynomial and matrices is given. For the case of two parametric family more sharp estimate is proven. Theoretic results are supported by various numerical simulations that show higher precision of presented methods with respect to traditional ones. The presented methods are inherently global and could be applied for studying $D$-decomposition for the space of parameters as a whole instead of some prescribed regions. For symbolic computations the Maple v.14 software and its package RegularChains are used.'
address: 'Ya.Z. Tsypkin Laboratory “Adaptive and Robust Control Systems”, V.A.Trapesnikov Institute of Control Sciences of Russian Academy of Sciences, Moscow, Russian Federation;Chair of Applied Mathematics and Computer Modelling, Department of Automation and Computer Science, I.M. Gubkin Russian State University of Oil and Gas, Moscow, Russian Federation.'
author:
- 'Oleg O. Vasil’ev'
title: 'Counting and computing regions of $D$-decomposition: algebro-geometric approach.'
---
Introduction
============
In this paper linear differential equation with constant coefficients $$\dot{x}(t)=Ax(t),$$ where $A$ is a complex(real) $n\times n$ matrix, will be called linear continuous system.
Similarly, linear discrete system is a finite-difference equation $$x(t+1)=Ax(t),$$ where $A$ is a complex(real) $n\times n$ matrix.
The problem of finding stability domains of a linear control system depending on a vector parameter is a classical one. It can be formulated as such: Let us consider a polynomial $P(s,k)$ with complex(for example, real) coefficients. polynomially depending on real parameters $k=(k_1,\ldots,k_l)\in {\mathbb K}.$ Complex parameters could be considered as $a+jb,$ where $a, b$ are real.
This polynomial is usually considered as a characteristic polynomial of a linear system. Vector $k$ understood as indeterminacy or parameters of a regulator.
We need to find such regions $U\subseteq {\mathbb K}$ that for all $k\in U$ $P(s,k)$ will be continuous-time(resp. discrete-time) stable.
Polynomial is called continuous-time stable(Hurwitz) if all of its roots have negative real part. Polynomial is called discrete-time stable(Schur) if all of it’s roots is have modulus lesser than $1$.
Continuous-time(res. discrete-time) system is stable if and only if its characteristic polynomial is a Hurwitz(resp. Schur) one.
It is obvious that the problems of finding regions of hurwitzness or schurness of a polynomial is equivalent. Clearly, polynomial $P(z)$ is Schur if and only if $(s-1)^{\deg P}P(\frac{s+1}{s-1})$ is Hurwitz. Therefore we can consider only continuous-time systems, and we will do that in any case, unless otherwise specified.
There are two cases which are of the greater interest for control theoretist, namely the case of characteristic polynomial of SISO system with affine dependency on parameters and the case of family of matrices affinely depending on parameters, for example, in the case of construction of a linear output feedback. In the latter case polynomial $P(s,k),$ will be $\det (sI -(A_0+\sum_{i=1}^lk_iA_i)).$
One of a methods of studying stability regions is a $D$-decomposition approach based on study of decomposition of parameter space into connected regions with equal numbers of stable and unstable roots(root-invariant regions or regions of $D$-decomposition). It is clear that the border of these regions is a semialgebraic hypersurface in $K$ parametrized by equalities $\Re(P(j\omega,k))=0, \Im (P(j\omega,k))=0$ and, possibly, a hypersurface with an equation $|a_{\deg P}(k)|=0,$ where $a_{\deg P}$ - is a leading term of $P.$ $\Re$ and $\Im$ here denote real and imaginary part of an expression.
The main idea of this approach goes back to I. Vishnegradsky[@Vish1876], but first clear formulation and detailed study of a method have been done by Yu.I. Neimark [@Nei1948], [@Nei1949]. The contemporary state-of-the-art could be obtained from paper [@GP2006] and review [@GPT2008].
The main idea of our paper is an application of various methods of computational algebra and real algebraic geometry to the study of $D$-decomposition- both computational(finding sample points from all root invariant regions) and theoretical - upper bounds to the number of regions.
The famous Tarski-Seidenberg theorem ([@Tar1951; @Sei1954]) tells us that the problem of description of all root-invariant regions, as well as stable ones is algorithmically solvable. Nevertheless, effective algorithms of study of real polynomial systems of equations have been included in general purpose computer algebra systems only recently. For example, Maple package $RegularChains$ was included in Maple only in 2005 and it is still rapidly developing. The history of those algorithms itself goes back to discovery of cylindric algebraic decomposition by G.E. Collins in 1975 [@Col1975] and partial cylindric algebraic decomposition by G.E.Collins and H.Hong in 1991[@CH1991]. Monograph [@BPC2006] is a standart reference book on the subject.
Moreover, applications of quantifier elimination algorithms to control theory goes back to the seminal paper by B.D.O. Anderson, N.K. Bose, E.I.Jury published in 1975 [@ABJ1975], where an application of quantifier elimination to the linear output feedback stabilisation problem have been studied. They used Routh-Hurwitz theory, multivariate resultants and classic Tarski-Seidenberg quantifier elimination technique to check existence or non-existence of stable output feedback. However, there was no effective computational techniques, software and sufficiently powerful machines at the time, so the interest to this topic have risen only in the midde of 1990’s[@DYA1997; @Neu1997; @Jir1996]. The application of real algebraic geometry to the study of the stable points set and the Nyquist map is stated as an unsolved problem in monograph by E.A. Jonckheere [@Jon1997] The approach used in this papers is quite similar to the Anderson-one, but authors have used not an original Tarski-Seidenberg algorithms but cylindrical algebraic decomposition implemented at that time in the QEPCAD software by H.Hong. His collaborative work with R. Liska and S.Steinberg [@HLS1996] gave rise to applications of partial cylindric algebraic decomposition algorithms to stability problems for PDE.
Later, H.Anai and S.Hara et al. in series of papers have introduced the new methods of robust control synthesis with quantifier elimination based on so-called Sturm-Habicht sequence, Symbolic-Numeric CAD etc (see [@AH2006],[@IYAY2009] and references therein), they have done computations with these methods in computer algebra system Risa/Asir. They have also designed a Maple package SyNRAC, where some of their algorithms was implemented.
Some similar considerations in the context of Nyquist criterium could be found in papers by N.P. Ke ([@Ke2000] and references therein). The paper [@SXZ2011] is devoted to the applications of computational real algebraic geometry to an asymptotic stability analysis of nonlinear systems. A.D. Bruno et al. [@BBV2012] introduces use of resultants for study of $D$-decomposition border.
This paper is organised as follows. In section 2 we obtain some upper bounds to the number of root invariant regions in the parameter space in two-dimensional and general cases. Using both Groebner bases and classical resultant-based elimination technique we obtain a variety containing projection of intersection of hypersurfaces $\Re(P(j\omega,k))=0, \Im (P(j\omega,k))=0$ onto ${\mathbb K}.$ Then we estimate the number of regions of its complement using classical Harnack and Bezout theorems in two-dimensional case and Warren inequality in the general case. These results are far-going generalisation of upper bounds from [@Gry2004; @GP2006].
In section 3 we give the description of Maple-based algorithm of studying $D$-decomposition and concentrate on some particular examples from [@GP2006], which illustrate main features of our approach. We use a combination of Groebner basis-based or resultant-based elimination and partial cylindric algebraic decomposition. Specifically, for the continuous-time system first appeared in [@QBRDYD1995], Example 2 and studied in [@GP2006], Example 15, we have found one new root-invariant region. This example tell us about one very important feature of our methodology - its inherent globality. We automatically obtain the structure of $D$-decomposition not only for some predefined small region, but for parameter space as a whole.
Estimates for the number of root invariant regions in the parameter space
=========================================================================
Let us consider a polynomial $P(s,k)$ with complex(for example, real) coefficients. polynomially depending on real parameters $k=(k_1,\ldots,k_l)\in {\mathbb K}.$ Complex parameters could be considered as $a+jb,$ where $a, b$ are real.
This polynomial is usually considered as a characteristic polynomial of a linear system. Vector $k$ understood as indeterminacy or parameters of a regulator.
We need to find such regions $U\subseteq {\mathbb K}$ that for all $k\in U$ $P(s,k)$ will be continuous-time(resp. discrete-time) stable. The case of discrete-time stability is equivalent to continous-time one up to linear fractional transformation. Therefore we can consider only the continous-time case unlessotherwise is specified.
The $D$-decomposition approach is based on study of decomposition of parameter space into connected regions with equal numbers of stable and unstable roots(root-invariant regions or regions of $D$-decomposition).
It is clear that the border of these regions is a semialgebraic hypersurface in $K$ parametrized by equalities $\Re(P(j\omega,k))=0, \Im (P(j\omega,k))=0$ and, possibly, a hypersurface with an equation $|a_{\deg P}(k)|=0,$ where $a_{\deg P}$ - is a leading term of $P.$ $\Re$ and $\Im$ here denote real and imaginary part of an expression.
In the following theorems there are obtained upper bounds for a number of root invariant regions.
\[th:1\] Let $P(s,k_1,\ldots,k_n)$ be a complex(in particular, may be real) polynomial of degree $t$ on $s$ and of degree $d$ on all $k_i$ together. Suppose that $\Re( P(j\omega,k_1,\ldots,k_n))$ and $\Im( P(j\omega,k_1,\ldots,k_n)),$ do not have any common divisors non-trivially depending on $\omega.$
Then the number of regions of $D$-decomposition is no greater than $$6(4td+4d)^n,$$
The border of $D$-decomposition in the continuous-time case is a semialgebraic set in the parameter space $p_1,\ldots,p_n$. It is defined parametrically by the pair of equations $\Re(P(j\omega,k_1,\ldots,k_n))=0$, $\Im(P(j\omega,k_1,\ldots,k_n))=0,$ and an equation $a_t(k_1,\ldots,k_n)=0.$ $a_t$ is a coefficient of $s^t$ in $P(s,k_1,\ldots,k_n)$([@PS2002] §4.1.2).
This semialgebraic set could be continued to a minimal algebraic variety containing it. Because $\Re(P(j\omega,k_1,\ldots,k_n), \Im(P(j\omega,k_1,\ldots,k_n))$ are coprime over ${\mathbb C}[\omega]$ its dimension will stay the same, therefore the number of connected components of complement will not be lesser.
Consider a minimal algebraic variety $X$ containing the set parametrized by $\Re(P(j\omega,k_1,\ldots,k_n))=0, \Im(P(j\omega,k_1,\ldots,k_n))=0.$ Let $g$ be a Groebner basis of an ideal generated by parametrization polynomials relative to some $\omega$-eliminating order. Then by an elimination theorem ([@KOL2000], Ch.3 §1, Theorem 2; [@KOL2000], Ch.3 §1, Ex.5) $X$ is defined by elements of $g$ that do not contain $\omega.$
By Proposition 1 in [@KOL2000], Ch.3 §6 the resultant $r$ of $\Re(P(j\omega,k_1,\ldots,k_n))$, $\Im(P(j\omega,k_1,\ldots,k_n))$ in variable $\omega$ defines an algebraic set containing the border of $D$-decomposition.
This set could be bigger than the set defined by a basis. But because of coprimeness of $\Re(P(j\omega,k_1,\ldots,k_n)), \Im(P(j\omega,k_1,\ldots,k_n))$ up to $\omega$ resultant $r$ is a non-zero polynomial. Therefore the set of zeros of $r$ divides ${\mathbb K}$ in a number of parts that is no greater than the number of parts generated by zeros of the Groebner basis.
In order to find that number we need to estimate the degree of $h=ra_t$(or $h=r(\Re^2(a_t)+\Im^2(a_t))$) It can be easily shown that it is no greater that $2td+2d.$ Finally, we need to bound the number of regions. Using Theorem 2 [@War1968] we get $6(2 \deg h)^n.$
The bound for continuous-time is the same as for discrete-time one because linear fractional transformation does not change the degree of polynomials.
It must be noted that the main result of [@BPC2009] could signify that the multiplier $6$ in formulation of the theorem above could be replaced by $2.$
In $2$-dimensional case we can obtain better bound, even sharp in some sense. To prove an upper bound for the number of root invariant regions we have to prove the following lemma:
\[lem:1\] Let $X$ be the plane real (possibly singular and reducible) affine algebraic curve of degree $n.$
Then its complement in ${\mathbb R}^2$ consists of no more than $\frac{n^2+n+2}{2}$ connected components.
This bound is sharp and it is reached on union of lines in general position.
If $X$ is a real irreducible nonsingular projective curve of degree $n$ then, by Harnack inequality, ([@Gud1974] §3, [@JLC1931] Ch.4 Th. 11) it divides plane to no more than $\frac{(n-1)(n-2)}{2}+2$ components and $\frac{(n-1)(n-2)}{2}+1$ of them correspond to the ovals of $X.$
If $X$ is singular, then by First Harnack Theorem ([@JLC1931], Ch.4, Th.12) real irreducible projective curve has no more than $\frac{(n-1)(n-2)}{2}-\sum k_i(k_i-1)+1$ branches. Here $k_i$ are the multiplicities of singular points.
Note that every self-intersection or self-tangency of $l$ fragments of ovals of the curve in a singular point gives birth to no more than $2l$ new connected regions in a small neiborghood of a point. All other types of singularities do not lead to an increase of number of regions.
Therefore if the mutiplicity of a singular point has increased by $l,$ then the maximal possible number of ovals is reduced by $(k_i+l)(k_i+l-1).$ Hence the number of connected components of a complement to $X$ increases by no more than $2l-(k_i+l)(k_i+l-1),$ but if $k_i, l \geq 1, $ then $2l-(k_i+l)(k_i+l-1)\leq 0.$ Thus singular irreducible projective curve divides the projective plane into no more than $\frac{(n-1)(n-2)}{2}+2$ components.
Let $X$ be a reducible curve. Denote by the symbols $m\vdash n$ such a fact that $m=(m_1,\ldots,m_{|m|})$ is a partition of $n$ to the degrees of irreducible components of $X.$
Let $k=|\{i|m_i=1\}|$, $l=|\{i|m_i>1\}|.$
We can estimate the number of regions $Reg$ as $$\begin{array}{l}
Reg\leq \max_{m\vdash n}(\sum_{m_i\in m}\frac{(m_i-1)(m_i-2)+2}{2}+\sum_{i\neq j}m_im_j+n-k+1)=\\=\frac{n^2-n+2}{2}+\max_{m\vdash n}(2|m|-k)=\frac{n^2-n+2}{2}+\max_{m\vdash n}(2l+k)\leq\\\leq\frac{n^2-n+2}{2}+\max_{2l+k\leq n, 0\leq l,0\leq k}(2l+k)=\frac{n^2+n+2}{2}.
\end{array}$$
This estimate is obviously sharp. It is sufficient to take a reducible curve that consists from $n$ lines in general position(e.g. without parallels or more than double intersections).
\[locstr\] 
The only thing we have to prove is the following inequality: $$Reg\leq \max_{m\vdash n}(\sum_{m_i\in m}\frac{(m_i-1)(m_i-2)+2}{2}+\sum_{i\neq j}m_im_j+n-k+1)$$ The first term in the left part of inequality is trivially derived from previous considerations. The term $n-k$ corresponds to the transition from the projective case to affine one. Actually, there are no more than $n$ intersections between $X$ and the infinite line. $k$ of them are intersections between linear components and the infinite line, but this intersections do not add any new connected components.
We can deduce from Bezout theorem that there are no more than $m_im_j$ intersections between irreducible components of degrees $m_i$, $m_j.$ Therefore we need to prove that every such intersection creates no more connected components than its multiplicity.
Let us consider sufficiently small neighborhoods of points of two intersecting and non-intersecting irreducible components(see Figure 1). Those irreducible components are locally isotopic to an intersection of lines(one line if non-singular), therefore we need to consider only the case of lines , which is shown at the figure 1.
Now we need globalize the local structure. Let us consider the following numeration of regions.
Suppose that we have $p$ irreducible components. Let us denote by $N(i)$ a number of regions that the irreducible component $i$ adds.
Let us divide the naturals into $p+1$ parts. - from $2$ to $2+N(1)-1$, from $2+N(1)$ to $2+N(1)+N(2)-1$, from $N=2+\sum_{i=1}^pN(i)$ to $+\infty.$ We can break the plane into the finite union of neighborhoods such that each of them contains no more than one point of intersection of components.
We can enumerate connected regions appearing in those neighborhoods.
Let us take an arbitrary region that is not internal for any oval of an irreducible component. It will be the first region.
Then we take an arbitrary neighborhood. If there are a point of intersection of irreducible components in it then we bypass that point counterclockwise and enumerate all appearing non-numbered regions using the following rule:
1. If the component $i$ that borders current region from counterclockwise-side have appeared before and all numbers between $2+\sum_{w=0}^{i-1}N(w)$ and $2+\sum_{w=0}^{i}N(w) -1$ already correspond to some region, then let us numerate it by the first free number, that is greater than $N.$
2. If, otherwise, there are free number from $2+\sum_{w=0}^{i-1}N(w)$ to $2+\sum_{w=0}^{i}N(w) -1$ then let us numerate our region by first free number from this range.
Then we use the same rule to enumerate all non-enumerated regions around points of self-intersections of irreducible components.
Finally, all remaining regions will be enumerated upside-down and if this will not be possible - from left to right.
We have $e$ enumerated regions of the plane.
Let us call a (possibly non-compact) region on the plane with border containing more than one irreducible component a curvilinear polygon. Points(possibly infinite or even formal- i.e. directions in which border component changes) of intersections between two distinct irreducible components will be called vertices of a polygon. Parts of irreducible components lying betwóen vertices will be called edges.
Let $R_1\ldots R_z$ be a curvilinear polygon. Every irreducible component form no more than half of all edges of $R_1\ldots R_z.$
We have established a correspondence between the set of regions and segment of naturals $[1,\ldots, e+1).$ Consider it as a part of real line. Break every interval of form $[i,i+1)$ into parts corresponding to an irreducible components that appear in its border. The length of every subinterval in this partition is proportional to the number of edges formed by every component. Therefore there will be no element of the partition of length greater than $\frac{1}{2}-\epsilon$ and lesser than $1.$
Consider a neighborhood of any point of intersection between components. Let us bypass that point in any direction. During the bypass every component would appear to be a border for no more than $2j$ regions, where $j$ is the intersection multiplicity of that component in the point. From the other side the contribution of this component to the partition of interval $[1,\ldots, e+1)$ in the point is no greater than $2j\frac{1}{2}=j.$ This proves the lemma.
\[th:2\] Let $P(s,r,p)$ be a polynomial of degree $t$ with complex(in particular, possibly, real) coefficients.polynomially depending from two real parameters $r$, $p$ or one complex parameter of a form $r+jp.$ Denote by $d$ its maximal degree as polynomial on variables $r$ and $p$ together.
Suppose that $\Re( P(j\omega,r,p))$ and $\Im( P(j\omega,r,p)),$ as complex polynomials does not have any common divisors that nontrivially depends on $\omega.$
Then the number of regions of $D$-decomposition is no greater than $$\frac{q^2+q+2}{2},$$ with $q=2td+2d.$
Follows from Lemma \[lem:1\] and the proof of previous theorem.
This bound could not be generalised onto $n$-dimensional case, because of lack of any sharp Harnack-like bounds for higher-dimensional real algebraic varieties. Moreover, we can easily show that the case of hyperplanes in general position even in $3$-dimensional case do not give a maximal number of regions.
Namely, Bihan[@Bih2003] has shown that the maximal number of connected components of real projective algebraic surface of degree $q$ is equivalent as $q\to\infty$ to $dq^3,$ where $d\in [\frac{13}{36},\frac{5}{12}].$ Moreover [@Vir1998], from Comessatti-Petrovsky-Oleinik and Smith-Thom inequalities could be obtained an upper bound $\frac{5}{12}q^3-\frac{3}{2}q^2+\frac{25}{12}q$.
But $q$ planes could divide the space only in $\frac{q^3+5q+6}{6}$ parts.
Now we can write bounds for the case of polynomial matrix family.
Let $A(r,p)$ be a $t\times t$-matrix, which entries are complex(possibly real) polynomials on real parameters $r$,$p$(may be on one complex parameter $r+jp$) of degree no greater than $d$
Suppose that $\Re (\chi_A(j\omega,r,p))$ and $\Im (\chi_A(j\omega,r,p)),$ as complex polynomials does not have any common divisors that nontrivially depends on $\omega.$
Then the number of regions of $D$-decomposition is no greater than: $$\frac{q^2+q+2}{2},$$ where $q=2t^2d$ in continuous case and $q=2t^2d+2td$ in discrete one.
Let $A(k_1,\ldots,k_n)$ be a $t\times t$-matrix, which entries are complex(possibly real) polynomials on real parameters $k_1\ldots k_n$ of degree no greater than $d$.
Suppose that $\Re( \chi_A(j\omega,k_1,\ldots,k_n))$ and $\Im (\chi_A(j\omega,k_1,\ldots,k_n)),$ as complex polynomials does not have any common divisors that nontrivially depends on $\omega.$
Then the number of regions of $D$-decomposition in continuous case is no greater than: $6(4t^2d)^{n}, $ and no greater than $6(4t^2d+4td)^n$ in discrete one.
$D$-decompositions for 2-parametric families: computational study
=================================================================
In order to study 2-dimensional $D$-decompositions with Maple 14 we have used the following scheme.
1. Convertion of family of matrices to the family of polynomials $P(s,r,p)$, if needed.
2. If we think about our family as a discrete-time object then we convert it to continuous-time using transformation $P(s)\mapsto (s-1)^{\deg P}P(\frac{s+1}{s-1}).$
3. Computing an equation defining most of irreducible components of an algebraic curve, containing border of $D$-decomposition using command $EliminationIdeal$ for an ideal generated by $\Re( P(j\omega,r,p))$ and $\Im( P(j\omega,r,p))$ or eliminating $\omega$ using resultants.
4. if $P(s,r,p)$ has a real leading term then we add an irreducible component generated by its leading coefficient $a_{\deg P}(r,p)$ else we add $\Re^2(a_{\deg P}(r,p))+\Im^2(a_{\deg P}(r,p)).$ We get an algebraic curve containing the border of $D$-decomposition.
5. Using $PartialCylindricalAlgebraicDecomposition,$ we obtain a point cloud that contains at least one point from every region of $D$-decomposition.
6. We count number of stable roots for every point from a cloud and generate the list of stable points.
Let us proceed to examples.
Consider a discrete-time object $s^6+(r+jp)s^5+\frac{3}{2}.$ Converting it to continuous time we obtain: $p(s,r,p)=(s+1)^6+(r+jp)(s+1)^5(s-1)+\frac{3}{2}(s-1)^6.$
$D$-decomposition of family $z^n+(r+jp)z^{n-1}+\alpha,$ with $\alpha>1$ has $(n-1)^2+1$ regions [@GP2006]. In [@GP2006], there is also given an estimate for number of root invariant regions for families of a form $a(s)+(r+jp)b(s),$ namely $(n-1)^2+2.$ Polynomial parametrization of border of $D$-decomposition is: $$\begin{array}{l}
-r-4ps^5+4ps-\frac{75}{2}s^2+5rs^4+5rs^2-rs^6+\frac{5}{2}+\frac{75}{2}s^4-\frac{5}{2}s^6=0\\
-4rs-ps^6-3s^5+5ps^2-p-3s+5ps^4+4rs^5+10s^3=0
\end{array}$$
Explicit equation of border of $D$-decomposition has degree $10:$ $$\begin{array}{l}
9216p^{10}+46080p^8r^2+92160p^6r^4+92160p^4r^6+46080p^2r^8+9216r^{10}-\\-94464p^8-377856p^6r^2-566784p^4r^4-377856p^2r^6-94464r^8+301440p^6+\\+683136p^4r^2+1051776p^2r^4+276864r^6-309600p^4-619200p^2r^2-309600r^4+\\+122500p^2+122500r^2-15625=0
\end{array}$$ By the Lemma \[lem:1\], if degree of a border is $10,$ $D$-decomposition contains no more than $56$ regions. There are $26$ regions in the example considered.
There are $56$ regions if and only if the border of $D$-decomposition is an arrangement of $10$ lines without parallels and triple intersections.
At the figure 2 the $D$-decomposition and the point cloud are shown. There are no stability domains.
\[4.6\] 
The other example [@GP2006] is a discrete-time object $z^6+(r+jp)z^5+\frac{3}{20}.$
It is an example of one possible fenomenon appearing in a given way of studying $D$-decomposition. Namely, an equation $\Re^2(a_6(r,p))+\Im^2(a_6(r,p))=(r+\frac{23}{20})^2=0$ has nonzero number of roots but it does not separate root invariant regions – there are the same number of stable ad unstable roots from every side of the line defined by the equation.
The border of $D$-decomposition(without extraneous components) is given by an euation of degree $10:$ $$\begin{array}{l}
9,216\cdot10^{13}p^{10}+4,608\cdot10^{14}p^8r^2+9,216\cdot10^{14}p^6r^4+9,216\cdot10^{15}p^4r^6+\\+4,608\cdot10^{14}p^2r^8+9,216\cdot10^{13}r^{10}+8,1792\cdot10^{13}p^8+3,27168\cdot10^{14}p^6r^2+\\+4,90752\cdot10^{14}p^4r^4+3,27168\cdot10^{14}p^2r^6+8,1792\cdot10^{13}r^8+\\+1,30276416\cdot10^{15}p^6-1,821010752\cdot10^{16}p^4r^2+1,865389248\cdot10^{15}p^2r^4-\\-1,15483584\cdot10^{15}r^6+6,82460784\cdot10^{11}p^4+1,364921568\cdot10^{14}p^2r^2+\\+6,82460784\cdot10^{13}r^4+4,160322828658\cdot10^{15}p^2+4,160322828658\cdot10^{15}r^2-\\-3,573226485213841\cdot10^{15}=0.\end{array}$$
At the figure 3 the border of $D$-decomposition(including extraneous line) is shown.
\[4.7\] 
The other example shows us how a minimal algebraic curve containing the border of $D$-decomposition may differ from the border of $D$-decomposition itself.
Let us consider an output feedback problem of the form $K=\begin{pmatrix}r&p\\-p&r\end{pmatrix}$ for discrete-time system defined by matrices: $$A=\begin{pmatrix}0,4753&0,7579&7,9939\\-0,0415&0,8905&0,7579\\-0,0758&-0,0415&0,4753\end{pmatrix}$$
$$B=\begin{pmatrix}0,0801&0,0430\\-0,0015&0,0948\\-0,043&-0,0015\end{pmatrix},\qquad C=\begin{pmatrix}1&0&0\\0&1&0\end{pmatrix}$$
A minimal algebraic curve containing the border of $D$-decomposition consists from three intersecting ovals.
\[14\] 
But the border of $D$-decomposition is not an algebraic curve, because it does not contain part of the medium oval lying in the smallest one.
\[14.1\] 
The next example is an output feedback of a form $K=\begin{pmatrix}-r&p\\p&r\end{pmatrix}$ for a co ntinuous-time systems defined by matrices $$\begin{array}{l}
A=\begin{pmatrix}79&20&-30&-20\\-41&-12&17&13\\167&40&-60&-38\\33,5&9&-14,5&-11\end{pmatrix},\\\\ B=\begin{pmatrix}0,219&0,9346\\0,047&0,3835\\0,6789&0,5194\\0,6793&0,831\end{pmatrix}, C=\begin{pmatrix}0,0346&0,5297&0,0077&0,0668\\0,0535&0,6711&0,3834&0,4175\end{pmatrix}.
\end{array}$$
\[15\] 
\[15.1\] 
This system has been considered in [@QBRDYD1995], Example 2 and [@GP2006], Example 15. We have found a new connected region of $D$-decomposition with one stable root. This region isaway from all other components, its border corresponds to $\omega >20$.
This fact illustrates globality of our method. We don’t need to prescribe any parameter ranges, but we do automatically get an answer for a whole space.
Here we give an example of polynomial family with complex coefficient for which bound from Lemma \[lem:1\] is sharp.
Namely, let $$\begin{array}{l}
P(s,r,p)=(4-22j)s^2+((30-20j)r-(145+8j)p+68)s+\\+((21-8j)r^2-(34+13j)p^2+(40+54j)rp+(2-26j)r+(-34+19j)p-(8+6j))
\end{array}$$ Then the border of $D$-decomposition could be parametrised as: $$\begin{array}{l}
4s^2+(20r+8p-4)s-21r^2+40rp+2r-21p^2-34p-8=0\\
-22s^2-145ps+68s+30rs-54rp-8r^2+19p-13p^2-26r-6=0
\end{array}$$ An explicit equation of border of $D$-decomposition could be written as: $$\begin{array}{l}
1008000p^4-3303960p^3r+1782100p^2r^2+978760pr^3-627300r^4+62160p^3+\\+1366648p^2r+1219928pr^2-1200320r^3-1679328p^2+630352pr-145904r^2-\\-309120p+310272r+64512
\end{array}$$ The decomposition of plane into $11$ regions is shown on figure.
\[exp2\] 
Conclusions
===========
We obtain a new technique of studying $D$-decomposition using real algebraic geometry. Method of obtaining an explicit equation for a border of $D$-decomposition is given as well as method of constructing point clouds from all root invariant regions.
An estimates for a number of root invariant regions are given in $n$-dimensional and in $2$-dimensional cases. The last bound is close to the sharp one.
This technique could be applied in construction of low order regulators as well as in robust stability analysis.
[45]{} A bridge between robust control and computational algebra. A robust control synthesis by a special quantifier elimination //Constructive Algebra and Systems Theory B.Hanzon, M.Hazewinkel(eds.), Royal Netherlands Academy of Arts and Sciences, Amsterdam, Netherlands, 2006, P. 23-34 Output feedback stabilization and related problems – solution via decision methods.//IEEE Trans. Automat. Control Vol.20(1975), P. 53-65. Algorithms in real algebraic geometry. Âerlin, Heidelberg: Springer-Verlag, 2006. An asymptotically tight bound on the number of connected components of realizable sign conditions.// Combinatorica, Vol.29(2009) P. 523-546. Asymptotic behaviour of Betti numbers of real algebraic surfaces//Comment. Math. Helv. 2003. Vol.78 P. 227-244 Stability sets of multiparameter Hamiltonian systems //Journal of Applied Mathematics and Mechanics Vol.76(2012) P. 56-92 Quantifier elimination in the elementary theory of real closed fields by cylindrical algebraic decomposition//Lecture Notes in Comput. Sci. Vol.33(1975), P. 134–183, Berlin, Springer-Verlag. Partial cylindrical algebraic decomposition for quantifier elimination.// J. Symbolic Comput. Vol.12(1991), P. 299–328. A Treatise on Algebraic Plane Curves. Oxford: Oxford University Press, 1931.(Reprinted by Courier Dover Publications, 2004). Ideals, Varieties and Algorithms. An Introduction to Computational Algebraic Geometry and Commutative Algebra, Second Edition. New York: Springer, 1998. Application of Quantifier Elimination Theory to Robust Multi-objective Feedback Design//J. Symbolic Computation, Vol.24(1997) P. 153-159 The $D$-Decomposition Theory// Automation and Remote Control 2004. Vol.65, ¹12, P.:1872-1884. Stability regions in the parameter space: $D$-decomposition revisited.//Automatica 2006. Vol.42. P.: 13-26 $D$-decomposition technique state-of-the-art// Automation and Remote Control 2008. Vol.69, ¹12, P.:1991-2026. The topology of real projective algebraic varieties //Russian Mathematical Surveys, Vol.29(1974), ¹4, P. 1-79. Testing stability by quantifier elimination//J. Symbolic Computation Vol.24(1997) P. 161-187. An Effective Implementation of a Symbolic-Numeric Cylindrical Algebraic Decomposition for Quantifier Elimination//Proceedings of the 2009 conference on Symbolic numeric computation, 3-5 of July 2009, Kyoto, Japan P. 55-64 Algebraic methods for modeling and design in control. Linkoping Studies in Science and Technologies, Likoping University, Linkoping, Sweden, 1996. Algebraic and differential topology of robust stability. NY,Oxford: Oxford University Press, 1997. Symbolic and algebraic computation in robust stability analysis//ACM SIGSAM Bulletin 2000. Vol.34 P. 21-24 Search for the parameter values that make automatic control system stable, Autom. and Telem., 9, 1948, No.3, 190-203 (in Russian) Stability of the linearized systems, Leningrad: LKVVIA (in Russian). 1949. Parametric Robust Stability by Quantifier Elimination// Technical report no. 97-37 in RISC Report Series, University of Linz, Hagenberg, Austria. 1997. PhD Thesis Robust Stability and Control . Moscow, Nauka, 2002 (in Russian). A new decision method for elementary algebra.//Ann. Math. Vol.60(1954), P. 365–374. Algebraic Analysis on Asymptotic Stability of Continuous Dynamical Systems//Proceedings of the 36th international symposium on Symbolic and algebraic computation ISSAC-2011, San Jose, CA, USA - June 07 - 11, 2011 P. 313-320 A Decision Method for Elementary Algebra and Geometry, 2nd edn, Berkeley: University of California Press. A formula for computation of the real stability radius//Automatica 1995. Vol.31 P. 879-890. Mutual position of hypersurfaces in the projective space//Amer. Math. Soc. Transl. (2) 1998. Vol.186 P. 161-176; Sur la theorie generale des regulateurs//Comptes Rendues de l’Academie des Sciences, 1876. Vol.83 P. 318-321. Lower bounds for approximation by nonlinear manifolds//Transactions of the AMS. 1968. Vol.133 P. 167-178
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'The change detection problem is to determine if the Markov network structures of two Markov random fields differ from one another given two sets of samples drawn from the respective underlying distributions. We study the trade-off between the sample sizes and the reliability of change detection, measured as a minimax risk, for the important cases of the Ising models and the Gaussian Markov random fields restricted to the models which have network structures with $p$ nodes and degree at most $d$, and obtain information-theoretic lower bounds for reliable change detection over these models. We show that for the Ising model, $\Omega\left(\frac{d^2}{(\log d)^2}\log p\right)$ samples are required from each dataset to detect even the sparsest possible changes, and that for the Gaussian, $\Omega\left( \gamma^{-2} \log(p)\right)$ samples are required from each dataset to detect change, where $\gamma$ is the smallest ratio of off-diagonal to diagonal terms in the precision matrices of the distributions. These bounds are compared to the corresponding results in structure learning, and closely match them under mild conditions on the model parameters. Thus, our change detection bounds inherit partial tightness from the structure learning schemes in previous literature, demonstrating that in certain parameter regimes, the naive structure learning based approach to change detection is minimax optimal up to constant factors.'
author:
- |
[Aditya Gangrade, Bobak Nazer, and Venkatesh Saligrama]{}\
[Boston University]{}\
{`gangrade,bobak,srv`}`@bu.edu` [^1]
bibliography:
- 'main.bib'
title: 'Lower Bounds for Two-Sample Structural Change Detection in Ising and Gaussian Models'
---
Introduction
============
Markov random fields (MRFs) are a popular way to model the dependence between a set of random variables. Consider a class of MRFs, $\mathcal{M}$, on $p$ random variables, consisting of probability densities parametrised by a vector $\theta \in \mathbb{R}^{\binom{p}{2}}.$ Recall that such MRFs have distributions such that, for any $i,$ $X_i$ is conditionally independent of $X_j$ for $j$ such that $\theta_{ij} = 0$ given the random variables $\{X_k: k \textrm{ s.t. } \theta_{ik} \neq 0\}.$ Thus, with every $\theta,$ one can associate an undirected graph capturing the dependencies, called the Markov network structure of the distribution, $G(\theta) = ([1:p], E(\theta))$, where $E = \{(i,j) : \theta_{ij} \neq 0 \}.$ The Ising model and the Gaussian MRF are instances of MRFs for binary and real-valued random variables, respectively, and are employed in a variety of applications.\
An important problem in statistics is the inverse problem of determining the dependencies between a set of random variables given a set of samples drawn from their joint distributions. In the context of MRFs, let $\mathcal{M}$ be a class of MRFs as before, $P\in \mathcal{M}$ be parametrised by $\theta,$ and the random vector $X = (X_i)_{i \in [1\colon p]}$ be distributed according to $P$. The inverse problem can be stated in two closely related ways - given $n$ i.i.d. samples $X^n \sim P^{\otimes n}$, one may wish to either learn $\theta$, which is the *model selection* problem, or to learn $G(\theta)$, which is the *structure learning* problem. These problems are typically studied in the “*high-dimensional setting,*” where $n \ll p$, that is, the number of samples is much smaller than the number of variables involved. A large body of work has focussed on constructing structure learning schemes for the Gaussian and Ising models, including algorithms and regularised estimators, with a particular focus on the technical conditions and the number of samples required for consistent learning (see [@anandkumar2012high; @meinshausen2006; @ravikumar2010high] and references therein for the Gaussian MRF, and [@montanari2009graphical; @Bresler:2015:ELI:2746539.2746631; @ravikumar2010high; @anandkumar2012high_ising], and references therein for the Ising model). A relatively smaller set of papers has studied the hardness of structure learning in terms of simple graph properties, providing necessary and sufficient information-theoretic conditions on the number of samples required for reliable structure learning (see [@WanWaiRam] and [@SanWai; @TandomShanmugamRavikumarDimakis] for the Gaussian and Ising MRFs, respectively).\
Recently, the allied problem of change detection in MRFs has received attention. Let $P_1, P_2$ be some $\theta_1, \theta_2$ parametrised members of $\mathcal{M},$ respectively. The basic question is whether, given $n_1$ samples $X^{n_1} \sim P_1^{\otimes n_1}$, and $n_2$ samples [$\widetilde{X}^{n_2} \sim P_2^{\otimes n_2},$]{} one can estimate $\mathds{1}\{G(\theta_1) = G(\theta_2)\}$ (*structural change detection*) or $\mathds{1}\{\theta_1 = \theta_2\}$ (*general change detection*) well. In the following, we concentrate on structural change detection, and largely suppress the adjective structural, since all results hold for the general problem as well (and also since including it makes for clunky prose). Structural change detection has both received practical interest, for instance in the study of controlled experiments [@zhang2012learning]; in genetics [@zhao2014direct; @xia2015testing]; and in neuroscientific contexts [@belilovsky2016testing], and theoretical interest focused on change detection algorithms along with the study of their sample complexities and consistency conditions [@zhao2014direct; @FazBan; @liu2017; @liu2017learning]. [A related line of work is pursued in [@daskalakis2016testing], which studies, amongst other things, the sample complexity of goodness of fit testing for Ising models under statistical measures of difference, i.e., given $P$ and $X^n \sim Q^{\otimes n},$ where $P,Q$ are Ising models, it studies the problem of estimating $\mathds{1}\{D(P\|Q) + D(Q\|P) \ge \varepsilon\},$ where $D$ is the Kullback-Leibler (KL) divergence. We note that our proof technique also relies on analysing a goodness of fit test, but against structural distance measures. ]{}\
In this paper, we take an information-theoretic approach to the change detection problem for the Ising and the Gaussian MRFs. Following the structure learning literature, we study the problem for the simple case of distributions that have Markov network structures with maximum degree bounded by $d$, denoted $\mathcal{G}_{p,d}$, and study the same against a minimax risk of misdeclaring a change, or the lack thereof, which is denoted $R_{\textrm{cd}}.$ We allow Ising models where the non-zero $\theta_{ij}$ are lower bounded by some $\alpha$ and upper bounded by some $\beta$, and Gaussian MRFs where the ratios of the non-zero off-diagonal entries to the corresponding diagonal entries are lower bounded by some $\gamma$. Our results provide lower bounds on the sample complexity in terms of $p,d,$ and $(\alpha, \beta)$ or $\gamma$, for any detection method that makes $R_{\textrm{cd}}$ small, and use simple ensembles of possible changes to do so. Our proof technique is symmetric in the two datasets, and hence the results are lower bounds on $\min(n_1,n_2)$. Interestingly, under mild conditions on the parameters, these bounds improve upon all the known lower bounds on the sample complexity of structure learning, and are at most $\mathrm{poly}(d)$ separated from the sample complexity of the maximum likelihood structure learner. This suggests[^2] that, at least for the Ising and Gaussian MRFs on $\mathcal{G}_{p,d},$ change detection is as hard, in terms of its data cost, as structure learning.\
Comparison to Prior Work
------------------------
Note that one can naively perform change detection by estimating the network structures of the distributions underlying the two sets of samples and comparing the same. This method is generally considered profligate, especially in the pratically important case where the underlying models might be dense, but the differences between them are small[^3]. One possible justification for this comes from the compressed sensing literature, which has demonstrated that in certain model classes learning sparse changes between two models can be significantly more data efficient than learning either of the models. Thus, a widely believed “folk theorem” asserts the existence of schemes for change detection in the Ising and Gaussian MRFs that can handle wide ranges of model classes such that their sample requirements scale only with sparsity of the changes to be detected, rather than the complexity of the underlying models. The previous work mentioned, [@zhao2014direct; @FazBan; @liu2017; @liu2017learning], all develop algorithms that estimate certain functionals of the ratio of the probability densities of the underlying models via a regularised optimisation scheme, and detect change on the basis of this estimate with sample complexities that scale as $O(s^k\log p)$, where $s := \|\theta_1- \theta_2\|_0$. However, these results only hold under strong technical conditions, [e.g.]{} structural assumptions on the Fisher information matrices of the models considered, and as previously noted in [@montanari2009graphical; @anandkumar2012high], such conditions do not easily provide a clear description of which classes of graphs and models satisfy them in direct terms, that is, do not provide a non-trivial class $\Theta$ such that these results hold for every $\theta \in \Theta$. [Further, as [@Hein2016] points out for the case of Gaussian structure learning, such conditions can fail to hold in subtle ways for application relevant data.]{}\
The nature of our results is at odds with the conventional wisdom (e.g., the claims in [@FazBan; @liu2017learning]) that sparse changes can be easy to detect irrespective of the ranges underlying parameters $\theta_1,\theta_2$ and the graph classes they live in. We believe that this disconnect is because of the strictness of the conditions required for these results to hold, and the regimes they implicitly push the underlying models into. In particular, as previously noted, our lower bounds on sample complexity closely match all known lower bounds on structure learning in $\mathcal{G}_{p,d}$ under mild conditions on the parameters. Crucially, the ensembles we construct to demonstrate these bounds have the sparsest possible changes - $\|\theta_1 - \theta_2\|_0 = 1.$ We note that this dependence on the parameters and $d$ is nontrivial - for instance, for the Ising model, if $\beta >1$, then the sample complexity lower bound scales as $\Omega\left( e^{2\beta d} d^{-1} \log p\right)$, which is exponential in $d$, and even if one is allowed to cherry pick $(\alpha, \beta)$ to minimise the lower bound, one needs at least $\Omega\left( \frac{d^2}{(\log d)^2} \log p \right)$ samples to detect change, irrespective of any incoherence or dependency conditions that may be imposed. For the Gaussian, our bound matches the scaling of the corresponding the structure learning bound. Since this is known to be tight up to constant factors in certain regimes, this shows that the naive scheme for change detection is minimax optimal in at least some contexts.\
Lastly, our lower bounds on Ising model change detection sample complexity for $\mathcal{G}_{p,d}$ structured models are actually stronger than those in [@SanWai] in certain parameter regimes. Since change detection is reducible to structure learning, these are also new bounds for the latter. We further note that our proof technique, which is currently un-optimised, relies upon lower bounding the [$\chi^2$-distance of certain distributions,]{} and thus differs from the Fano bounds of the previous literature. Since the technique is largely elementary and provides improvements on the previous results, it may be of independent interest and merit refinement.\
*Organisation:* §\[sec:def\] defines the models considered, and precisely formulates the problem considered, while §\[sec:results\] states the results and compares them with previous work. §\[sec:proof\] begins by laying out the common technical development involved in the proof, and follows this with the actual proofs.\
*Notation:* For a natural $n$, we use $[1:n]$ as shorthand for the set $\{1, 2, \dots, n\}$. $X,\widetilde{X}$ denote random vectors, usually of dimension $d$ or $p$, with $\mathcal{X}$-valued entries. For a natural $i,$ $X_i$ is the $i^{\mathrm{th}}$ component of $X$. Similarly, for a set of naturals, $S,$ $X_S$ is the vector $(X_i)_{\{i \in S\}}$. We use $X^n$ to denote an $n$-length sequence of i.i.d. random vectors, frequently referred to as a ‘dataset’. For a distribution $P$, and given $X \sim P$, $P^{\otimes n}$ is the distribution of $X^n$. Further, given $X^n,$ the $t^{\mathrm{th}}$ sample in that dataset is denoted $X^{(t)}$. Lastly, we use $Z$ and diacritical modifications of the same for partition functions whenever they are relevant. The two element sets $\{i,j\}$ are interchangeably denoted $(i,j)$ when referring to edges in an undirected graph, and as just $ij$ when they appear in a subscript. Vectors in $\mathbb{R}^{\binom{n}{2}}$ are indexed by cardinality-two subsets of $[1:n]$. For instance, a vector $\theta \in \mathbb{R}^{\binom{3}{2}}$ is represented as $(\theta_{12}, \theta_{13}, \theta_{23})$. The identity matrix of size $p$ is denoted as $I_p$.\
For functions $f,g,$ $f = O(g)$ if there exists a positive constant such that $\lim_{n \to \infty} f(n)/g(n) \le C,$ and $f = o(g)$ if $\lim f/g \to 0$. Similarly, $f = \Omega(g)$ if $g = O(f),$ and $f = \omega(g)$ if $g = o(f)$.
Definitions and Problem Statement {#sec:def}
=================================
We begin the paper with a bit of background on Markov random fields (MRFs) and a precise problem statement, which also serve to establish notation followed in the rest of the paper. Note that while we define the problem over a general MRF, our results are restricted to the Ising and the Gaussian MRFs. We define the problem thus since we develop a single technique for establishing lower bounds on the sample complexity and apply it parallely to both the models considered.
Markov Random Fields
--------------------
Recall that an undirected finite simple graph $G= (V,E)$ consists of a finite vertex set $V$, here identified with $[1:p]$, and edge set $E$ which is a set of subsets of $V$ of cardinality $2$. For $u \in V$, we let $\partial u := \{v: \{u,v\} \in E\}$ be the set of neighbours of $u$, and the cardinality of $\partial u$ is said to be the degree of $u$. Lastly, for naturals $p,d \le p-1,$ we let $\mathcal{G}_{p,d}$ be the set of graphs on $p$ vertices such that every node’s degree is no bigger than $d$.\
For a set $\mathcal{X},$ an $\mathcal{X}$-valued *Markov random field* on a graph $G$ is a random vector $X = (X_v)_{v \in V} \in \mathcal{X}^{|V|}$ such that for every $v$, $X_v$ is conditionally independent of $X_{V \setminus (\{v\} \cup\partial v)}$ given $X_{\partial v}$. The graph $G$ is said to be the Markov network structure of the Markov random field. We define a class of MRFs as a set of probability distributions, each of which is an MRF. For a class of MRFs $\mathcal{M},$ and a family of graphs $\mathcal{G}$, $\mathcal{M}$ is said to be a class of MRFs on $\mathcal{G}$ if every MRF in $\mathcal{M}$ has a Markov network structure contained in $\mathcal{G}$.\
Given a graph $G$ and a vector $\theta \in \reals^{\binom{|V|}{2}}$ such that $\{i,j\} \not\in E \Rightarrow \theta_{\{i,j\}} = 0$, a $0$-external field *Ising model* on $G$ with parameter $\theta$ is a $\{\pm 1\}$-valued Markov random field with the distribution $$\mathbb{P}_{(G,\theta)}\left(X = x\right) = \frac{1}{Z(\theta)} {\operatorname{exp}\left( \sum_{\{u,v\}\in E} \theta_{uv} x_ux_v\right)}, \notag$$ where $Z(\theta)$ is a normalising constant commonly known as the partition function. We let $\mathcal{I}_{p,d}(\alpha, \beta)$ be the set of Ising models on graphs in $\mathcal{G}_{p,d}$ such that for every $u,v,$ either $\theta_{uv} = 0$, or $\alpha \le |\theta_{uv}| \le \beta$ holds. Note that every $\theta$ determines a network structure, which we refer to as $G(\theta)$.\
Similarly, given a graph $G,$ a symmetric, positive-definite matrix $A$ such that $A_{i,j} \neq 0$ only if $(i,j) \in E$, the $0$-mean *Gaussian Markov Random Field* on $G$ with parameter $A$ is the $\reals$-valued Markov random field with the distribution $$\P\left( X \in S\right) = \sqrt{\frac{\det A}{(2\pi)^{|V|}}} \int_{x \in S} {\operatorname{exp}\left(-\frac{x^{{\top}} A x }{2}\right)} \,{\mathrm{d}}x. \notag$$ The matrix $A$ is known as the precision matrix of $X$. Note that the Markov network structure of the distribution is determined by the non-zero entries of $A$. For $\gamma >0,$ let $\mathcal{N}_{p,d}(\gamma)$ be the set of Gaussian MRFs on graphs in $\mathcal{G}_{p,d}$ with $0$ mean and precision matrix $A$ such that all diagonal entries are non-zero, and for every $(i,j) \in V^2,$ $$\min_{A_{ij} \neq 0} \left|\frac{A_{ij}^2}{A_{ii}A_{jj}}\right| \ge \gamma^2 \notag$$
We note here that the value of $\gamma$ actually modulates the graphical structures allowed within $\mathcal{G}_{p,d}$. For instance, if we allow graphs such that even a single node may have $d$ neighbours in the Markov network structure of a distribution as above, then the condition of positivity of the precision matrices enforces $\gamma < 1/\sqrt{d}$, and if we allow the entirety of $\mathcal{G}_{p,d}$ unrestrictedly, the condition $\gamma < 1/d$ is required.\
Lastly, for the above classes of MRFs, we frequently describe distributions in terms of their Markov network structures. In particular, we say that a distribution has the edge $(i,j)$ with weight $w$ if $\theta_{ij} = w$ for the Ising model, and if $A_{ij} = A_{ji} = w$ for the Gaussian MRF. If all the edges of a distribution have the same weight, we say that the distribution has uniform edge weights.
The Change Detection Problem
----------------------------
Let $\mathcal{M}$ be a class of $\mathcal{X}$-valued Markov random fields with parameters $\theta \in \Theta$, and let $P_1, P_2 \in \mathcal{M},$ have parameters $\theta_1$ and $\theta_2$ respectively. Note that the distributions/parameters are unknown to us, and are potentially equal. For $P\in\mathcal{M}$, let $G(P)$ be its Markov network structure. We consider the structural change detection problem, which may be informally stated as follows: Given $n_1$ samples drawn according to $P_1$ independently and identically, and $n_2$ samples identically drawn according to $P_2$ independently of each other and of the other set of samples, can one determine if $G(P_1) = G(P_2)$ or not with high probability?\
Formally, for $\mathcal{X}$-valued Markov random fields $P_1, P_2 \in \mathcal{M},$ let $X^{n_1} \sim P_1^{\otimes n_1},$ and $\widetilde{X}^{n_2} \sim P_2^{\otimes n_2}$ be finite sets of samples, also referred to as datasets, drawn independently and identically from $P_1$ and $P_2$, respectively. An $(n_1,n_2)$-change detector for $\mathcal{M}$ is a map $\phi: \mathcal{X}^{n_1} \times \mathcal{X}^{n_2} \to \{0,1\}.$ Let $\mathbf{\Phi}_{n_1,n_2}$ be the set of all $(n_1,n_2)$-change detectors. Let the risk of a detector $\phi$, be $$\begin{aligned}
R(\phi;n_1,n_2,\mathcal{M}) := \sup_{ P_1, P_2 \in \mathcal{M}} \,\,\,\,&\P\left\{\phi(X_1^{n_1}, X_2^{n_2}) = 1 \mid G(P_1) = G(P_2)\right\} \\ +\, &\P\left\{\phi(X_1^{n_1}, X_2^{n_2}) = 0 \mid G(P_1) \neq G(P_2)\right\}.\end{aligned}$$ and the minimax change detection risk with $(n_1,n_2)$ samples over the class $\mathcal{M}$ be $$R_{\mathrm{cd}}(n_1,n_2;\mathcal{M}) := \inf_{\phi \in \mathbf{\Phi}_{n_1,n_2}} R(\phi;n_1,n_2,\mathcal{M}). \notag$$
Note that the above risk expressions are just the adaptation of the standard binary hypothesis testing risks to our situation.\
We say that an $(n_1,n_2)$-change detector is $\delta$-reliable over the class $\mathcal{M}$ if $R_{\textrm{cd}}\left(\phi ; n_1,n_2,\mathcal{M}\right) < \delta,$ and say that the change detection problem can be solved over $\mathcal{M}$ $\delta$-reliably with $(n_1,n_2)$ samples if there exists an $(n_1,n_2)$-change detector over $\mathcal{M}$ that is $\delta$-reliable or, equivalently, if $R_{\textrm{cd}}\left(n_1,n_2;\mathcal{M}\right) < \delta.$ The parameter $\delta$ is occasionally referred to as the reliability level.\
The main aim of this paper is to study the trade-off between the reliability level $\delta$ of change detection over a given class of MRFs and the sample size $(n_1,n_2).$ In particular, we provide necessary conditions on $(n_1,n_2)$ for $\delta$-reliable change detection with $(n_1,n_2)$ samples over the classes $\mathcal{I}_{p,d}(\alpha,\beta)$ and $\mathcal{N}_{p,d}(\gamma)$ in terms of their parameters.
Theorem Statements and Nature of Results {#sec:results}
========================================
As previously noted, our results are necessary conditions on the number of samples required for $R_{\textrm{cd}}$ to be smaller than a given $\delta$. The results are stated in separate subsections for the Ising and the Gaussian MRFs, respectively. A discussion comparing the results to parallel results in structure learning follows the theorem statements, while proofs are relegated to later sections. C.f. §\[sec:def\] for precise definitions of the models and graph class considered.
Ising Model
-----------
[myth]{}\[thm:ising\]
\[thm:ising\_easy\] Let $\alpha >0.$ For every $\delta \in [0,1],$ a necessary condition for $\delta$-reliable change detection with $(n_1,n_2)$ samples over $\mathcal{I}_{p,d}(\alpha,\beta)$ is $$\min(n_1,n_2) > \frac{\log\left(1 + 4(1-\delta)^2 \binom{p}{2}\right)}{\log (1 + \tanh^2 \alpha)} .$$
\[thm:ising\_1\] Let $d \ge 4$ and $ \beta (d-3)\ge \ln d.$ For every $\delta \in [0,1],$ a necessary condition for $\delta$-reliable change detection with $(n_1,n_2)$ samples over $\mathcal{I}_{p,d}(\alpha,\beta)$ is $$\min(n_1, n_2) > e^{2\beta d}\frac{\ln\left(1 + 4(1-\delta)^2 \left\lfloor p/(d+1)\right\rfloor\right)}{8(e^{4\beta} + d)}.$$
*Proof sketch:* While the exact proof is relegated to §\[sec:proof\], we loosely detail the strategy here. Let $P \in \mathcal{M}$, and let $\mathcal{Q} \subset \mathcal{M}$ be a set of distributions such that $G(P) \not\in G(\mathcal{Q}).$ Suppose we are given the information that the larger set of samples is drawn according to $P$, while the second set is either drawn from some $Q \in \mathcal{Q},$ or drawn from $P$. We have thus reduced the change detection problem to running a hypothesis test on the smaller dataset, with the simple null hypothesis that the same is drawn from $P$, and the composite alternate that it is drawn according to some $Q \in \mathcal{Q}$. Suppose further that we are supplied with a prior, $\pi_Q$ for the selection on $\mathcal{Q}$. Clearly, the average risk of the hypothesis testing problem under the prior $\pi_Q$ would be a lower bound for the minimax change detection risk. However, since the uniformly most powerful tests for such problems are known by the classical results of Neyman & Pearson, we can compute lower bounds on these average risks. In particular, we do this by applying a variation of Le Cam’s method, as outlined in [@arias-castro2012], and with a uniform prior on $\mathcal{Q}.$\
We call the pair $(P,\mathcal{Q})$ a change detection ensemble. The ensembles used to derive the above bounds are as follows.
- Theorem \[thm:ising\_easy\]: $P$ is the Ising model with no edges on $p$ nodes, and $\mathcal{Q}$ is the collection of Ising models with exactly one edge with edge weight $\alpha.$
- Theorem \[thm:ising\_1\]: $P$ is the Ising model with uniform edge weight $\beta$ on $\lfloor p/d+1 \rfloor$ separate cliques of size $d+1$ each, while $\mathcal{Q}$ is the collection of Ising models on the same graph as $P$ but with one known edge from exactly one of the cliques deleted, again with uniform weight $\beta.$
We note that the above proof technique directly allows us to state our results as structure learning bounds in the context of the recovery criterion defined in [@SanWai] as well, as will be discussed in §\[sec:proof\].\
*Remark:* The ensemble used for the proof of Theorem \[thm:ising\_1\] most likely does not satisfy the conditions required in work such as [@FazBan] or [@liu2017]. This is because these papers all require an ‘incoherence condition’, and, as pointed out in [@montanari2009graphical], these conditions essentially hold only if $\beta d < k$ for some $k<1$, while we need $\beta d = \Omega(\log d)$ for the results to follow.
### Remarks, and comparison with structure learning bounds
In the high-dimensional setting, one considers the behaviour of these bounds for large $p$. The above bounds above can be interpreted in three different contexts depending on which of the model parameters are allowed to change. In the following we set $\delta = 1/2,$ in order to discuss conditions necessary to beat a random coin, and $c$ is some arbitrary quantity that depends only on the non-changing parameters.
1. If the parameters $d,\alpha, \beta$ are given constants, then change detection requires at least $c\log p$ samples from each dataset to detect changes.
2. If we hold $\alpha, \beta$ as constants, and allow $d$ to grow with $p$, then for sufficiently large $d$, one needs $ce^{2\beta d}\frac{\log p/d}{d}$ samples from each dataset to detect changes.
3. If we allow $\beta$ and $\alpha$ as well as $d$ to change with $p$, then unless $\beta$ decays with $d,$ we are forced into the exponential growth in $d$ regime. Suppose we impose the requirement that the bounds grow at most polynomially in $d$. This can be done essentially by limiting $\beta \le k \log d/d$ for any constant $k$, which limits the second bound, although since $\alpha \le \beta$ must hold, this affects the first bound as well. Optimising for the $k$ which gives the lowest net growth, we get that in *any scenario* of parameter growth, we need at least $ \frac{(d-1)^2}{4 \log^2 d} \log p$ samples from each data set to detect changes.
We compare the bounds of Theorem \[thm:ising\] with two results due to Santhanam & Wainwright. Note that the statements have been modified to fit our notation.\
First we consider the necessary condition:
Consider $\mathcal{I}_{p.d}(\alpha, \beta)$ for $\beta d > 1.$ If the structure learning probability of error is smaller than $\delta - \frac{1}{\log {p}},$ then following condition on the number of samples, $n$, is necessary$$n> (1-\delta)\max \left\{\frac{\log p}{2 \alpha \tanh\alpha}, \frac{e^{\beta d} \log (pd/4 - 1)}{4\beta d e^\beta}, \frac{d}{8} \log \frac{p}{8d} \right\}.$$
Theorem \[thm:ising\_easy\] is the direct analogue of the first bound, and Theorem \[thm:ising\_1\] is the direct analogue of the second bound. The third bound is only active when all parameters are held constant, and even then is inactive for $d > 90$, and thus left largely unconsidered by us. However, the other two terms are weaker than Theorem \[thm:ising\] in the regime in which the latter hold. In particular, if $d = o(p),$ and $\beta d> \log d$, Theorem \[thm:ising\_1\] has an advantage of essentially $c e^{\beta d} \frac{\beta d e^{\beta}}{e^{4\beta} + d},$ which is unbounded in $d$ for $\beta > 2 \log d /d.$ [Note that while the $\beta d > \log d$ requirement for our second bound to hold may seem more stringent than the $\beta d > 1$ condition in [@SanWai], in the regime $\beta d < \log d$, the $\alpha^{-2} \log p$ bound dominates the exponential bound in both cases, and thus this distinction is rendered moot.]{} Lastly, observe that if we force $\alpha$ and $\beta$ to decay in a manner that allows at most polynomial growth with $d$, we see that structure learning requires $\Omega\left(\frac{d^2}{(\log d)^2} \log p\right)$ samples as in our case.\
In the same paper, Santhanam & Wainwright also show ([@SanWai Theorem 3a)]) that if the edge weights are given, it is possible to learn the structure with $n$ samples for $$n \gtrsim \frac{3(3e^{2\beta d} + 1)}{\sinh^2 (\alpha/2)} d\left(3 \log p + \log 2d + \log \frac{1}{\delta}\right). \notag$$ Ignoring the $\delta$ terms, the result above is separated from our lower bound by a factor of $c\frac{d^2 + e^{4\beta} d}{\sinh^2{\alpha/2}}.$ [Now if $1/\alpha$ and $e^{4\beta}$ are at most polynomial in $d,$ then this factor is at most polynomial in $d$, which is neglegible compared to the exponential in $d$ scaling forced under $\beta d = \omega( \log d)$. Since our bounds are computed with a change ensemble where we are aware of the edge weights, the closeness of these bounds implies that our technique cannot yield significantly stronger lower bounds in this regime. Lastly, recall that our change detection bounds can also be stated as structure learning bounds. Thus, our results close the exponential gap in the structure learning lower bounds of [@SanWai], which has persisted through all subsequent work on structure learning.]{}
Gaussian MRFs
-------------
\[thm:Gaussian\] Let $\gamma \le 0.39$. For every $\delta \in [0,1],$ a necessary condition for $\delta$-reliable change detection with $(n_1,n_2)$ samples over $\mathcal{N}_{p,d}(\gamma)$ is $$\min(n_1,n_2) > \frac{1}{2 \gamma^2} \log \left(1 + (1-\delta)^2p\right). \notag$$
*Remark:* As mentioned before, $\gamma$ controls the richness of the Markov network structures within $\mathcal{G}_{p,d}$ that are allowed, essentially since the two together determine the positivity of certain precision matrices in the class. In particular, if we allow even a single node to have $d$ neighbours, as we rightly should when considering models in $\mathcal{G}_{p,d}$, then $\gamma < 1/\sqrt{d}$ is forced, and if $d$ regular graphs are allowed, then $\gamma < 1/d$ is imposed. In light of this, the condition on $\gamma$ in the theorem statement is fairly benign. For instance, enforcing $\gamma < 1/\sqrt{d}$ for $d \ge 7$ already gives us $\gamma < 0.38.$\
As in the Ising case, the bound is proved by considering an explicitly stated restricted class of changes, and bounding the risk for them. Curiously, while we obtain the above bound by considering a simple ensemble of changes of the form independent versus one-edge, essentially the same result can be obtained up to constant factors[^4] in more richly connected classes of ensembles - for instance, by using repetitions as in §\[sec:K\_d\_extend\] of the star graph versus the star graph with one edge moved, or the complete graph versus the same with a known edge missing. This suggests that there might be some uniformity to the hardness of structure learning/change detection in Gaussian MRFs.\
Comparing the above bound to the corresponding structure learning bound, we note the following result of Wang, Wainwright, and Ramchandran, stated in our notation.
Consider the class $\mathcal{N}_{p,d}(\gamma)$ with $\gamma \in [0,1/2]$. A necessary condition for asymptotically[^5] reliable structure learning over this class is $$n > \max \left\{ \frac{\log\binom{p-d}{2} - 1}{4\gamma^2} , \frac{2\log \binom{p}{d} - 1}{\log \left(1 + \frac{d\gamma}{1- \gamma} \right) - \frac{d\gamma}{1 + (d-1) \gamma}}\right\} .$$
Note that the first term in the expression above dominates the second when $\gamma = o(1/\sqrt{d}),$ which, by the previous argument, is the range of $\gamma$ in which at least one node can be connected to $d$ other nodes in the Markov network structure of the graph. Thus, our lower bound matches the structure learning lower bound in the parameter region relevant for $\mathcal{G}_{p,d}$. [We note that this bound is near tight - for instance [@anandkumar2012high] achieves, under technical conditions, structure learning for Gaussian MRFs with sample complexity $O(\gamma^{-2}\log p/\delta)$.]{}
Proofs {#sec:proof}
======
We begin by detailing the general proof technique that we use, followed by the proofs of Theorems \[thm:ising\_easy\] and \[thm:Gaussian\], which are of the same flavour, and are rather simple. The proof of Theorem \[thm:ising\_1\] is relatively more involved and follows these sections.
Lower Bounding Technique {#sec:low_b_tech}
------------------------
We first note that any lower bound on sample complexities to achieve a given risk level must hold for both $n_1$ and $n_2$, since merely switching the labels of the two sets of samples should not affect anything. We hence set $n = \min(n_1, n_2)$, and derive lower bounds on $n$ by considering simpler hypothesis testing problems. In the following, $X$ denotes a random sample from the dataset with the smaller number of samples.\
Recall the proof strategy described after the statement of Theorem \[thm:ising\]. Continuing in the same vein, we consider the average risk for the hypothesis testing problem $$\begin{aligned}
H_0: \,\,\,&X^n \sim P^{\otimes n} \\
H_1: \,\,\,&X^n \sim Q^{\otimes n} \textrm{ for some $Q \in \mathcal{Q}$}
\end{aligned}$$ under the uniform prior on $\mathcal{Q}$. Recall that the average risk under this prior must be smaller than $R_{\textrm{cd}}$.\
By the results of Neyman & Pearson [@lehmann2006testing Ch. 3], we know that the most powerful tests for the above hypothesis test are of the form $$L_n \overset{H_1}{\underset{H_0}{\gtrless}} \tau, \notag$$ where $\tau$ is a positive real number, and $$L_n(X^n) := \frac{1}{|\mathcal{Q}|}\sum_{Q\in\mathcal{Q}} \prod_{t=1}^n \frac{{\mathrm{d}}Q}{{\mathrm{d}}P}(X^{(t)}) \notag$$ is the likelihood ratio of the distribution $P$ versus a distribution uniformly at random from set $\mathcal{Q}$. We refer to $L_n$ as the likelihood ratio of $P$ versus $\mathcal{Q}$.\
Note that for discrete distributions such as the Ising model, $\frac{{\mathrm{d}}Q}{{\mathrm{d}}P} = Q/P,$ and for distributions which admit a density with respect to a Euclidean space, such as the Gaussian, we have $\frac{{\mathrm{d}}Q}{{\mathrm{d}}P} = f_Q/f_P$, where $f_P$ and $f_Q$ denote the respective densities.\
Our main technical tool is captured by the following lemma, which allows us to compute lower bounds on $n$ required for $R_{\textrm{cd}}$ to be small by computing upper bounds on the variance of $L_n$ for well-chosen ${P}$ and $\mathcal{Q}$.
\[lem:lowb\_main\] Let $\mathcal{M}$ be a class of MRFs. For every $\delta \in [0,1],$ $P \in \mathcal{M}, \mathcal{Q} \subset \mathcal{M}$ such that $P \not\in \mathcal{Q}$, and $n = \min(n_1,n_2)$, if $L_n$ is the $n$-sample likelihood ratio of $P$ versus $\mathcal{Q},$ then $$R_{\textrm{cd}}(n_1,n_2; \mathcal{M}) \le \delta \implies \E_{P^{\otimes n}}\left[L_n^2\right] \ge 1 + 4(1-\delta)^2. \notag$$
For notational brevity, we let $$\widetilde{Q}^{n} := \frac{1}{|\mathcal{Q}|} \sum_{Q \in \mathcal{Q}} Q^{\otimes n}. \notag$$ Let $R^{\mathrm{avg}}_{\mathrm{hyp}}$ be the optimal average risk for the above test with $n$ samples when $Q$ is picked uniformly at random from $\mathcal{Q}$. By the previous discussion, $$\begin{aligned}
R_{\mathrm{cd}}(n_1,n_2;\mathcal{M}) \ge R^{\mathrm{avg}}_{\mathrm{hyp}} &= \inf_{\tau \ge 0} \P\left(L_n \ge \tau\mid H_0\right) + \P\left(L_n < \tau\mid H_1\right)\\
&= 1 - \sup_{\tau \ge 0} \left(\widetilde{Q}^{n}\left(L_n \ge \tau \right) - P^{\otimes n}\left(L_n \ge \tau\right) \right) \\
&\ge 1 - d_{\text{TV}}(P^{\otimes n},\widetilde{Q}^{n}) \\
&= 1 - \frac{1}{2} \int_{x^n \in \mathcal{X}^n} \left| \frac{{\mathrm{d}}\widetilde{Q}^{n}}{{\mathrm{d}}P^{\otimes n}}(x^n) -1\right| \,{\mathrm{d}}P^{\otimes n}(x^n)\\
&= 1 - \frac{1}{2}\E_{P^{\otimes n}}\left[|L_n - 1|\right] \\
&\ge 1 - \frac{1}{2}\sqrt{ \E_{P^{\otimes n}}\left[L_n^2\right] - 1}.
\end{aligned}$$ Since $R_{\mathrm{cd}} \le \delta,$ the proof concludes with a simple manipulation of the above inequality.
[*Remark:* The above proof technique can also be applied to obtain structure learning bounds. In particular, suppose we have access to a structure learning algorithm that uniformly over all distributions in $\mathcal{M}$ identifies the correct structure with probability greater than $1 - \delta'$. Then, for $(P, \mathcal{Q})$ such that $G(P) \neq G(Q)$ for all $Q \in \mathcal{Q}$, and under the distribution $\P(H_0) = \P(H_1) = 1/2,$ the above hypothesis test can be solved with probability of error smaller than $\delta'$ by learning the structure of the distributions. However, this must exceed $R_{\textrm{hyp}}^{\textrm{avg}}/2$. Thus, all our change detection bounds, which rely on bounding $R_{\textrm{hyp}}^{\textrm{avg}}$, can be converted to structure learning bounds by doubling the reliability level.]{}
Proof of Theorem \[thm:ising\_easy\]
------------------------------------
Let $P$ be the Ising model with no edges on $p$ nodes, and let $Q_{ij}$ be the Ising model with weight $\lambda \in [\alpha, \beta]$ on the edge $(i,j)$ and no other edges. We let the class $\mathcal{Q} := \{ Q_{ij} : 1 \le i < j \le p\}$, and consider the change detection ensemble $(P,\mathcal{Q})$. Lastly, we set $\eta = \frac{e^{-\lambda}}{e^{\lambda}+ e^{-\lambda}}.$ Note that the n-sample likelihood ratio is $$L_n(X^n) = \frac{1}{\binom{p}{2}} \sum_{i < j} \prod_{t= 1}^n 2\left( \eta + (1-2\eta) \mathds{1}\big\{X_i^{(t)} = X_j^{(t)}\big\} \right) . \notag$$
In order to apply Lemma \[lem:lowb\_main\], we need to compute the quantity $\E_{P^{\otimes n}}\left[L_n^2\right].$ Let $B_{ij}^t := \mathds{1}\{X_i^{(t)} = X_j^{(t)}\}$. We note that for $i<j$, $u<v$, since $P$ is just $p$ independent Bernoulli distributions together, we have $$\begin{aligned}
\E_P\left[B_{ij}^t\right] &=\,\,\, \frac{1}{4} \quad \textrm{ for every $(i,j)$, and} \\
\E_P\left[B_{ij}^t B_{uv}^t\right] &= \begin{cases} \frac{1}{2} & \textrm{if $(i,j) = (u,v) $} \\ \frac{1}{4} & \textrm{otherwise.}\end{cases}\end{aligned}$$
Now, $$\begin{aligned}
\label{eqn:ising_easy_proof}
\E_{P^{\otimes n}}\left[L_n^2\right] &= \frac{1}{\binom{p}{2}^2} \E_{P^{\otimes n}} \left[ \left( \sum_{i<j}\prod_{t= 1}^n 2\left( \eta + (1-2\eta) B_{ij}^t\right)\right)^2 \right] \notag\\
&=\frac{1}{\binom{p}{2}^2} \sum_{i<j} \sum_{u<v} \E_{P^{\otimes n}} \bigg[\bigg(\prod_{t= 1}^n 4\Big( \eta^2 + \eta(1-\eta)\left( B_{ij}^t + B_{uv}^t\right) + (1-2\eta)^2 B_{ij}^tB_{uv}^t\Big)\bigg) \bigg] \notag \\
&\overset{\mathrm{(i)}}= \frac{1}{\binom{p}{2}^2} \sum_{i<j} \sum_{u<v} \Big( 4\eta^2 + 4\eta(1-\eta)\mathbb{E}_P\left[ B_{ij}^1 + B_{uv}^1\right]+ 4(1-2\eta)^2 \mathbb{E}_P\left[B_{ij}^1B_{uv}^1\right] \Big)^n\notag \\
&\overset{\mathrm{(ii)}}= \frac{1}{\binom{p}{2}^2} \sum_{i<j} \sum_{u<v} \left( 1 + (1-2\eta)^2 \mathds{1}\{(i,j) = (u,v)\} \right)^n \notag \\
&= 1 + \frac{(1 + (1-2\eta)^2 )^n - 1}{\binom{p}{2}}\end{aligned}$$ where the equality $\mathrm{(i)}$ is since each of the $t$ samples are independent and have the distribution $P$, and $\mathrm{(ii)}$ comes from feeding in the moments computed before. Plugging equation $(\ref{eqn:ising_easy_proof})$ into the condition imposed by Lemma \[lem:lowb\_main\] we get that if the risk is smaller than $\delta,$ then we must have $$n \ge \frac{1 + 4(1-\delta)^2 \binom{p}{2}}{\log (1 + (1-2\eta)^2)}. \notag$$ We conclude by noting that $1- 2\eta = \tanh (\lambda),$ and that we may set $\lambda = \pm \alpha$. $\Box$
Proof of Theorem \[thm:Gaussian\]
---------------------------------
This result essentially follows the same arguments as in the proof of Theorem \[thm:ising\_easy\]. We identify the distributions of the Gaussian MRFs with their precision matrices. Let $P$ be the Gaussian MRF on $p$ nodes with no edges and unit variance, $Q_{ij}$ be the unit variance Gaussian MRF on $p$ nodes with the single edge $(i,j)$ with edge weight $\lambda$ s.t. $|\lambda| \ge \gamma$, and $\mathcal{Q} := \{Q_{ij}: 1\le i < j \le p\}.$ For convenience, we let $\Delta_{ij}$ be the matrix with the $(i,j)$ and $(j,i)$ entries equal to $\lambda$, and all other entries $0$. Note that the precision matrices are $P = I_p$ and $Q_{ij} = I_p + \Delta_{ij}$.
Again, consider the likelihood ratio. We have $$\begin{aligned}
L_n(X^n) = \frac{1}{\binom{p}{2}} \sum_{i<j} \prod_{t = 1}^n \frac{f_Q}{f_P}(X^{(t)}) = \frac{\left(\det(I + \Delta_{12}\right)^{n/2}}{\binom{p}{2}} \sum_{i<j} \prod_{t = 1}^n e^{ -\frac{\left(X^{(t)}\right)^{\top}\Delta_{ij} X^{(t)}}{2}}. \notag \end{aligned}$$
We first require a few computations. For convenience, let $C(\lambda) = \begin{pmatrix} 1 & \lambda \\ \lambda &1 \end{pmatrix}$ and let $i<j$ and $u<v$ be naturals in $[1:p]$
- $\det(I_p + \Delta_{ij}) = 1 - \lambda^2$ - simply observe that the matrix is similar to the block diagonal matrix ${\mathop\mathrm{diag}}(C(\lambda) , I_{p-2}),$ and thus the determinant is the product of $\det(C(\lambda)) = 1 - \lambda^2$ and $\det(I_{p-2}) = 1$.
- $\det(I_p + 2\Delta_{ij}) = 1 - 4\lambda^2,$ since this is similar to ${\mathop\mathrm{diag}}(C(2\lambda), I_{p-2}).$
- If $|\{i,j\} \cap \{u,v\}| = 0$, then $\det(I_p + \Delta_{ij} + \Delta_{uv}) = (1- \lambda^2)^2,$ since the two $\Delta$s each contribute separate blocks of $C(\lambda)$.
- If $|\{i,j\} \cap \{u,v\}| = 1$, then $\det(I_p + \Delta_{ij} + \Delta_{uv}) = 1 - 2\lambda^2$, since this matrix is similar to ${\mathop\mathrm{diag}}(D, I_{p-3})$ where $D = \begin{pmatrix} 1& \lambda & \lambda \\ \lambda & 1 & 0 \\ \lambda & 0 & 1 \end{pmatrix}.$
We are now in a position to bound $\mathbb{E}[L_n^2]$. The first few steps are parallel to the Ising model case, and are omitted. $$\begin{aligned}
\E_{P^{\otimes n}}\left[L_n^2\right] &= \frac{\left(\det(I + \Delta_{12})\right)^{n}}{\binom{p}{2}^2} \sum_{i<j}\sum_{u<v} \left(\mathbb{E}_P\left[ {\operatorname{exp}\left( - \frac{X^{\top}\left(\Delta_{ij} + \Delta_{uv}\right) X }{2} \right)}\right] \right)^n \notag\\
&\overset{\mathrm{(i)}}= \frac{\left(\det(I + \Delta_{12})\right)^{n}}{\binom{p}{2}^2} \sum_{i<j}\sum_{u<v} \left( \det(I_p + \Delta_{ij} + \Delta_{uv} )\right)^{-n/2} \notag \\
&= \frac{1}{\binom{p}{2}^2} \Bigg(\binom{p}{2} \binom{p-2}{2} + 2\binom{p}{2}\binom{p-2}{1} \left(\frac{1 - \lambda^2}{\sqrt{1 - 2\lambda^2}}\right)^n + \binom{p}{2} \left( \frac{1- \lambda^2}{\sqrt{1-4\lambda^2}}\right)^n \Bigg) \notag\end{aligned}$$ where $\mathrm{(i)}$ is due to the Gaussian integral, and the final equality is by simple counting. Note first that $1 - 2\lambda^2 > 1 - 4\lambda^2$. Since we are looking for an upper bound, we will set $a= \frac{1 - \lambda^2}{\sqrt{1 - 4\lambda^2}}$. We thus have $$\begin{aligned}
\label{eqn:gaussian_L}
\E_{P^{\otimes n}}\left[L_n^2\right] \le& \frac{1}{\binom{p}{2}} \left( (2p-3) a^n + \binom{p-2}{2}\right).\end{aligned}$$
Plugging (\[eqn:gaussian\_L\]) into the condition from Lemma \[lem:lowb\_main\], if the risk is smaller than $\delta$, it must be true that $$a^n \ge 1 + 4(1-\delta)^2 \frac{p(p-1)}{2(2p-3)} > 1 + (1-\delta)^2 p. \notag$$
Taking logarithms, we directly have that for every $|\lambda| \le \gamma,$ $$R_{\textrm{cd}}(n_1,n_2;\mathcal{N}_{p,d}(\gamma)) \le \delta \implies n > \frac{\log 1 + (1-\delta)^2 p}{\log \frac{1- \lambda^2}{\sqrt{1-4\lambda^2}}}. \notag$$ The stated result follows on noting that $\ln (1 - \lambda^2) - \frac{1}{2} \ln (1- 4 \lambda^2) \le 2\lambda^2$ for $|\lambda| \le 0.39$[^6], and since if $\gamma \le 0.39, $ we may set $\lambda = \pm \gamma$ above. Note that the quadratic denominator in $\gamma$ is retained as long as $\gamma$ is bounded away from $1/2$, although with a graceful weakening of the constants invovlved. $\Box$
Proof of Theorem \[thm:ising\_1\]
---------------------------------
To begin with, we will prove a likelihood ratio upper bound for a specific change detection ensemble on the graph class $\mathcal{G}_{d+1,d}$. We will next show a technique that allows us to obtain a closely related bound on the graph class $\mathcal{G}_{p,d}$, and follow this by a small section concluding the proof.
### A likelihood ratio bound
Let $P$ be the Ising model with uniform edge weight $\lambda$ on the graph $K_{d+1}$, the complete graph on $d+1$ nodes. Note that each node in $K_{d+1}$ has degree $d$. Similarly, let $Q$ be the Ising model with uniform edge weight $\lambda$ on the graph $K_d \setminus \{(1,2)\},$ i.e., the same graph as $P$ but with one edge deleted. We let $Z$ be the partition function for $P$, and $Z'$ be the partition function for $Q$.
Arithmetical manipulations show that for $K_{d+1},$ $$\sum_{u < v} \lambda x_ux_v = \frac{\lambda}{2}\left(\left( \sum x_u\right)^2 - (d+1)\right). \notag$$ We thus have $$P(X = x) = \frac{1}{Z} {\operatorname{exp}\left( \frac{\lambda}{2} \left( \left(\sum_{i = 1}^{d+1} x_i \right)^2 - (d+1) \right)\right)} , \notag$$ and that $$Q(X = x) = \frac{Z}{Z'} P(X= x) e^{-\lambda x_1 x_2}. \notag$$
We first compute a bound on the $n$ sample likelihood ratio for $P$ versus $Q$, denoted $L_n$. Observe that $$\begin{aligned}
L_n(X^n) &= \left(\frac{Z}{Z'}\right)^n \prod_{t = 1}^n e^{-\lambda X_1^{(t)}X_2^{(t)}}, \notag\\
\E_{P^{\otimes n}}\left[ L_n^2 \right] &= \left(\frac{Z}{Z'}\right)^{2n} \left(\E_P\left[ e^{-2\lambda X_1^{(1)}X_2^{(1)}}\right]\right)^n, \notag\end{aligned}$$ and we thus need a bound on $\mathbb{E}_P\left[ e^{-2X_1X_2}\right]$. This is the subject of the following lemma, the proof of which is relegated to the appendix.
\[lem:ising\_likelihood\] If $d \ge 4$ and $\lambda(d-3) \ge \log d,$ then $$\left(\frac{Z}{Z'}\right)^2 \E_P\left[ e^{-2X_1X_2}\right] \le 1 + 8\left(e^{4\lambda} + d\right)e^{-2\lambda d} \ .$$
### Lifting results on $d+1$ nodes to $p$ nodes {#sec:K_d_extend}
Let $P$ be a MRF on $d+1$ variables such that $G(P) = K_{d+1}$, and $Q$ be the MRF obtained by eliminating the dependence in $P$ between $X_1$ and $X_2$, i.e, clipping one edge from $G(P)$ with no change in the rest of the graph. Suppose that we have an estimate for the likelihood ratio variance of the form $$\E_{P^{\otimes n}}\left[L_n^2 \right] \le h(n,d) \notag$$ where $h$ is some positive function.\
Let $r := \left\lfloor \frac{p}{d+1} \right\rfloor, $ and consider the graph $G$ on $p$ nodes consisting of $p - r (d+1)$ singletons, and $r$ non-trivial completely connected components, each of which has exactly $d +1$ nodes. Observe that $G = \left(\bigoplus_{i = 1}^r K_d^i\right) \oplus N$, where for $i\in[1:r],$ $K_{d+1}^i$ is the complete graph on the nodes labelled $[1 + (i-1)(d+1): i(d+1)],$ and $N$ is the trivial graph on nodes labelled $[p - r(d+1) + 1, p].$\
Let $\overline{P}$ be the Markov random field that associates a copy of $P$ with each non-trivial connected component of $G$. In effect, we are sitting with $r$ independent copies of $P$. Further, for $\mu \in [1:r],$ let $\overline{Q}_\mu$ be the MRF that associates a copy of $P$ to each non-trivial component of $G$ except the $\mu^{\textrm{th}},$ with which it associates a copy of $Q$, and let $\overline{\mathcal{Q}}$ be the set of these $\overline{Q}_\mu$s. Suppose $X$ is the random vector associated with the whole graph. We refer to the $d+1$-dimensional random vector that consists of the $X_i$s corresponding to $i$s in the $\mu$th component of $G$ as $X(\mu)$. Finally, we set $\overline{L}_n$ to the likelihood ratio of $\overline{P}$ versus $\overline{\mathcal{Q}}$. Note that $$\overline{L}_n = \frac{1}{r}\sum_{\mu = 1}^r L_n(X^n(\mu)), \notag$$ and thus $$\begin{aligned}
\E_{\overline{P}^{\otimes n}}\left[L_n^2\right] &= \frac{1}{\rho^2} \sum_{\mu = 1}^\rho \sum_{\nu = 1}^\rho \E_{\overline{P}^{\otimes n}}\left[L_n(X^n(\mu)) L_n(X^n(\nu))\right] \\
&= \frac{1}{\rho^2} \sum_{\mu = 1}^\rho \sum_{\nu = 1}^\rho \left(\mathbb{E}_{P^{\otimes n}}\left[X^n(\mu)\right]\mathbb{E}_{P^{\otimes n}}\left[X^n(\nu)\right]\right)\mathbbm{1}_{\{\nu \neq \mu\}} \\ &\qquad \qquad \qquad \qquad \quad + \left( \mathbb{E}_{P^\otimes n}\left[\left( X^n(\mu) \right)^2\right]\right) \mathbbm{1}_{\{\nu = \mu\}}.\end{aligned}$$ where the final equality is because non-identical components are independent, and identically distributed according to $P$ in the null hypothesis. Now observe that $\mathbb{E}_{P^{\otimes n}} \left[ L_n(X(\mu))\right] = 1$. Consequently, $$\begin{aligned}
\E_{\overline{P}^{\otimes n}}\left[L_n^2\right] &= 1 - \frac{1}{r} + \frac{1}{r} \mathbb{E}_{P^{\otimes n}} \left[ L_n^2 \right] \notag \\
&\le 1 + \frac{h(n,d) - 1}{\lfloor p / (d+1) \rfloor} ,\notag\end{aligned}$$ which, by Lemma \[lem:lowb\_main\], leads to the change detection lower bound $$\label{eqn:trick}
R_{\mathrm{cd}}(n_1,n_2;\mathcal{M}') \le \delta \Rightarrow h(n,d) \ge 1 + 4(1-\delta)^2 \left\lfloor \frac{p}{d+1} \right\rfloor,$$ where $\mathcal{M}'$ is a class into which $\overline{P}$ and $\overline{Q}_\mu$ fit.\
### Finishing Up
We preserve the notation from the previous two subsections. Let $V:= 1 + 8\left(e^{4\lambda} + d\right)e^{-2\lambda d}$. Recall that by Lemma \[lem:ising\_likelihood\], $$\mathbb{E}_{P^{\otimes n}} \left[ L_n^2\right] \le V^n. \notag$$ We form an ensemble on $p$ nodes as in §\[sec:K\_d\_extend\], and use equation $(\ref{eqn:trick})$ with $h(n,d) = V^n$ and with $\mathcal{M}' = \mathcal{I}_{p,d}(\alpha, \beta)$, to get that for every $\lambda \in [\alpha, \beta] \cap [\frac{\log d}{d-3} , \infty),$ $$\begin{aligned}
\delta &\ge R_{\textrm{cd}}(n_1,n_2;\mathcal{I}_{p,d}(\alpha, \beta)) \\ \implies V^n &\ge 1 + 4(1-\delta)^2 \left\lfloor \frac{p}{d+1} \right\rfloor \\
\iff n &\ge \frac{\log 1 + 4(1-\delta)^2 \left\lfloor \frac{p}{d+1}\right\rfloor}{\log 1 + 8(e^{4\lambda} + d^2) e^{-2\lambda d}}.\end{aligned}$$
The proof is completed by recognising that for positive $x,$ $(\log 1 + x)^{-1} \ge 1/x,$ and noting that if $\beta(d-3) \ge \log d,$ then we may set $\lambda = \beta$ above. $\Box$
Discussion
==========
Our results, both independently and when coupled with the existence of algorithms with sample complexities depending only on the sparsity of changes and not on the maximum degrees of the underlying network structures, raise a few points of interest to both theorists and practitioners. We lay these out below as avenues for future work to explore.
1. As [@liu2017learning; @liu2017] state, it is believed that the change detection schemes in [@FazBan; @liu2017; @zhang2012learning; @zhao2014direct] are agnostic to the underlying models (modulo technical conditions). However, given our results, the technical conditions required for these results to hold cannot include the entirelty of $\mathcal{G}_{p,d}$, since any reliable change detector requires $\Omega\left(\frac{d^2}{(\log d)^2} \log p\right)$. It thus becomes important from the perspective of design and utilisation of change detection algorithms to precisely characterise the technical conditions made in these papers, and which graph classes they in-/preclude.
2. Following the initial work by Santhanam & Wainwright [@SanWai], lower bounds on the structure learning sample complexities for Ising models in a number of graph classes were determined by Tandon et. al. in [@TandomShanmugamRavikumarDimakis] via clever refinements and extensions of the original methods. Most of these happen to be exponential in the graph class parameters allowed, and grow at least super-linearly with the same when the model parameters are cherry picked to minimise the lower bounds. On the other hand, it has long been known that the structure learning of Ising models on trees is relatively easy. One then is led to wonder if a similar separation of ‘easiness’ driven by the graphical class of the models extends to the change detection problem, and if so, whether it differs from the parallel separation in structure learning. Concretely, we ask the questions -
1. For which graph classes is change detection as hard as structure learning? Again, due to the existence of schemes such as those in [@liu2017learning; @FazBan], change detection cannot be as data hungry as structure learning for every graph class. It would be interesting to identify the structural features that make the problem easy.
2. For which graph classes is change detection easy, in the sense of sample complexities depending only on sparsity of changes? Is this class larger than the ones determined by the previous work? Further, do larger changes always increase sample costs for detection? Given the practical needs for change detection algorithms, these questions are of significant importance.
Acknowledgement {#acknowledgement .unnumbered}
===============
The authors would like to express their gratitude to Or Ordentlich for many valuable discussions.
Proof of Lemma \[lem:ising\_likelihood\] {#appx:Ising_bound}
========================================
The proof essentially relies on a simple counting and bounding trick. We first demonstrate this trick to obtain an upper bound on $Z$. This is not directly used in the proof, but is the simplest way to demonstrate the result. This is followed by a quick bound on $Z/Z'$, and then a bound on $\E_P\left[ e^{-2X_1X_2}\right].$
1. Recall that $Z$ is the partition function for a complete $d+1$-clique with uniform edge weight $\lambda$. We assume, for simplicity, that $d$ is even, although the bound below also holds for odd $d$. $$\begin{aligned}
Z &= \sum_{\{\pm 1\}^{d+1}} e^{\frac{\lambda}{2} \left( \left(\sum x_i\right)^2 - (d+1) \right)} \\
&= \sum_{j = 0}^{d+1} \binom{d+1}{j} e^{ \frac{\lambda}{2} \left( (d+1 - 2j)^2 - (d+1)\right)} \\
&= 2 \sum_{j = 0}^{d/2} \binom{d+1}{j} e^{ \frac{\lambda}{2} \left( (d+1 - 2j)^2 - (d+1)\right) }.
\end{aligned}$$
Here, $j$ is the number of nodes that are $-1$. Let $T_j$ be the $j^{th}$ term in the second eqution above. Observe that as $j$ increases, the $\binom{d+1}{j}$ factors in $T_j$ increase polynomially, but the $e^{\frac{\lambda}{2}(d+1-2j)^2}$ factors decrease exponentially. Thus, for large $\lambda, d$, we expect that an exponentially large fraction of the sum is contributed by the $0^{th}$ term.\
[In order to show this, let, for $y \in [0,d/2],$ $$\tau(y) := \log (d+1 - y) - \log(1+y) - 2\lambda (d- 2y).$$ Note that at integer values of $y$, $ \tau(y) = \log \left({T_{y+1}}/{T_y}\right) ,$ and that $\tau(y) < 0$ means that $T_{y+1} < T_y$. By simple computation, it is easy to see that the derivative $\tau'(j)$ is a concave function with the maximum over the considered domain at $d/2.$ In particular, this maxima is positive for $\lambda d \ge 1$. If we now enforce $\tau(0) <0$ by setting appropriate bounds on $\lambda, d$, we find that $\tau$ either decreases and then increases, or just monotonically increases, and in either case it can have at most one root over $[0, d/2]$. However, $\tau(d/2) = 0,$ which means that for $j \in [0:d/2],$ the successive terms are all non-increasing as long as $T_0 > T_1.$ For reasons of convenience, we also demand $T_1 \ge d T_2.$ All the conditions required above are satisfied when $\lambda(d-2) \ge \log d+1 \ge 1$.]{}\
With the above in hand, we have, for $\lambda(d-2) \ge \log d+1$, $$2T_0 \le Z \le 2(T_0 + T_1 + \frac{d}{2}T_2) \le 2(T_0 + \frac{3}{2}T_1), \notag$$ and we hence have $$\label{eqn:Z}
1 \le \frac{Z}{2{\operatorname{exp}\left(\frac{\lambda}{2} ((d+1)^2 - (d+1))\right)}} \le 1 + 3d e^{-2\lambda d} .$$
2. Recall that $Z'$ is the partition function for the Ising distribution on $K_{d+1} \setminus (1,2)$ with uniform edge weight $\lambda$. Let $\sigma := \sum_{i = 3}^{d+1} x_i$. We have $$\begin{aligned}
Z' &= \sum_{\{\pm 1\}^{d+1}} e^{\frac{\lambda}{2}\left( (\sigma + x_1 + x_2)^2 -2x_1x_2 - (d+1) \right)} \notag\\
&\ge e^{-\lambda} \sum_{\{\pm 1\}^{d+1}} e^{\frac{\lambda}{2}\left( (\sigma + x_1 + x_2)^2 - (d+1) \right)} \notag\\
&= e^{-\lambda} Z \label{ineq:Z/Z'}
\end{aligned}$$
where the inequality holds since $\lambda \ge 0$ and $x_1x_2 = \pm 1$.
3. We now upper bound $V:= \left(\frac{Z}{Z'}\right)^2\E_P\left[ e^{-2X_1X_2}\right].$
For convenience, we assume below that $d$ is even, although the upper bound remain valid for odd $d$. Using the technique used in the sum in part a) above, $$\begin{aligned}
V &= \left(\frac{Z}{Z'}\right)^2\frac{2}{Z}\sum_{j=0}^{(d-2)/2} \binom{d-1}{j} \left(1 + 2e^{-2\lambda(d-2j -2) } + e^{-4\lambda(d-2j -1)}\right) e^{\frac{\lambda}{2}\left( (d+1-2j)^2 - (d+1) -4\right)} \notag \\
&\overset{\mathrm{(i)}}\le e^{2\lambda}\frac{2e^{-2\lambda}}{Z}\sum_{j=0}^{(d-2)/2} \binom{d-1}{j} e^{\frac{\lambda}{2}\left((d+1-2j)^2-(d+1)\right)} \left(1+e^{-2\lambda(d-2j-2)}\right)^2 \notag \\
&\overset{\mathrm{(ii)}}\le \frac{2}{Z}e^{\frac{\lambda}{2}((d+1)^2-(d+1))} \left(1+e^{-2\lambda(d-2)}\right)^2 \times \left(1 + 6(d-1) e^{-2\lambda d} \right) \notag \\
&\le \left(1+e^{-2\lambda(d-2)}\right)^2 \left(1 + 6d e^{-2\lambda d}\right) \notag \\
&\overset{\mathrm{(iii)}}{\le} 1 + 8(e^{4\lambda} + d)e^{-2\lambda d},\notag
\end{aligned}$$ where the inequality
1. invokes $(\ref{ineq:Z/Z'})$, and uses $\lambda >0.$
2. follows the strategy used to upper bound for $Z$ as in (\[eqn:Z\]), and holds if the $T_j$ are decreasing and $T_1 \ge \frac{d}{2} T_2,$ which is true if $ \lambda(d-3) \ge \log d$.
3. holds if $d\ge 3,$ and $\lambda(d-2) \ge 1$. $\Box$
[^1]: A. Gangrade was supported by DHS Contract: HSHQDC-15-C-B0003 and NSF grant CCF-1618800. B. Nazer was supported by NSF grant CCF-1618800. V. Saligrama was supported by NSF grant CCF-1320547.
[^2]: but does not prove
[^3]: All the applied work we have mentioned before falls into this category.
[^4]: and subject to the $\gamma$ conditions that allow these ensembles to exists,
[^5]: as $p$ grows large
[^6]: This inequality can be shown for $|\lambda| \le \frac{\sqrt{3 + \sqrt{7}}}{2} \approx 0.297$ by a standard positivity of derivative argument, and can be readily shown for $\lambda \le 0.39$ by any numerical equation solver.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'In this note, we go over the recent soft photon model and Faddeev-Jackiw quantization of the massless quantum electrodynamics in the eikonal limit to some extent. Throughout our readdressing, we observe that the gauge potentials in both approaches become pure gauges and the associated eikonal Faddeev-Jackiw quantum bracket matches with the soft quantum bracket. These observations and the fact that the gauge fields in two cases localize in two-dimensional plane (even if it is spatial in soft photon case and $1+1$-dimensional Minkowski in the eikonal case) imply that there might be an interesting relation between these two distinct perspectives.'
author:
- Suat Dengiz
title: 'Note on soft photons and Faddeev-Jackiw symplectic reduction of quantum electrodynamics in the eikonal limit'
---
For the recent couple of years, an elegant and affirmatively revolutionizing idea has been brought out by Strominger and his collaborators [@Strominger1; @Strominger2; @Strominger3; @Strominger4; @Strominger5; @Strominger6; @Strominger7; @Strominger8; @Strominger9; @Strominger10; @Strominger11; @Strominger12]. The idea is fundamentally based on the (extended) BMS symmetry which, together with the $SL(2, \mathbb{C})$ (or Poincare group), comprises an unignorable infinite family of diffeomorphism dubbed as supertranslations that *nontrivially* act on the physical data at the future and past null infinities leaving the asymptotic form of spacetime intact [@Bondi; @Barnich1; @Barnich2; @Barnich3]. Unlike the common consensus, taking all these nontrivial diffeomorphism compels that the vacuum of Einstein’s gravity is not unique; rather it is tremendously degenerated which are correlated to each others by virtue of the supertranslations. Here, the vacua are distinguished from each other by the Goldstone bosons of the broken symmetry. By assuming the antipodal sewing of past and future null structure around the spatial infinity, the charge conservation relations match the incoming radiation intersecting the conformal sphere at a particular angle with the outgoing one passing through the opposite angle. The quantum mechanical matrix components of those conservation laws provide infinite amount of Ward identities for the S-matrix which is shown to be the well-known Weinberg’s soft graviton theory [@Weinberg; @Strominger1; @Strominger2; @Strominger3; @Strominger4; @Strominger5; @Strominger6; @Strominger7; @Strominger8; @Strominger9; @Strominger10; @Strominger11; @Strominger12]. Thus, the corresponding Goldstone modes are nothing but the soft gravitons. Here, the null infinities of the vastly degenerated vacua are distinguished from each others via the creation/annihilation of soft gravitons possessing distinct angular momenta albeit zero energies. Having such an accurate relation between the (extended) BMS group and the soft model together with the gravitational memory which we do not discuss here reveals some strictly covered vital IR properties of quantum gravity. For example, as is mentioned above, unlike the generic presumption, the quantum-gravitational-vacuum is excessively *degenerated* rather than being unique. Despite the lack of some crucial points such as derivation of the Bekenstein-Hawking entropy in that aspect yet, by bearing in mind the appealing theoretical successes of the approach in various foremost theories hitherto, it seems the correct platform that one should study is indeed *not the spacelike but the null boundary* of the asymptotically flat spacetimes to enlighten the long-lived unsolved problems in theoretical physics. Particularly, this soft approach has great potential to provide viable explanations to the so-called Hawking’s information paradox that has been an unavoidable obstacle against a deterministic universe [@HawkingInf1; @HawkingInf2; @HawkingInf3]. Before going further, since it occupies a crucial place among the ultimate purposes of this recent soft approach, let us briefly recapitulate the point where the paradox comes about[^1]: as is well-known, at a specific time in the life time of stars, they throw away a substantial amount of ingredients which are necessary to trigger nuclear reactions and countervail their self-gravitational attractions. Due to the devoid of sufficient amount of balancing stuffs, the stars begin to collapse into themselves which can end up with either white dwarfs, neutron star or ultimately black holes. In the case of black hole, as is seen in Figure \[Figure1\], star creates an event horizon at a definite null surface and ultimately vanishes via the Hawking evaporation yielding Minkowski-like spacetimes both at the upper and lower regions.
![Gravitational collapse and black hole evaporation.[]{data-label="Figure1"}](figure1.png){width="50.00000%"}
Here, the evaporation seems to be natural consequence of the fact that black hole possesses the temperature $T=\frac{1}{8\pi {\cal M}_{BH}} $ and thus emits radiations *a la* Stefan-Boltzmann law[^2] $$\mbox{ER}= (\mbox{Surface Area}) \times T^4 \sim {\cal M}^2_{BH} \times T^4 \sim {\cal M}^{-2}_{BH}.$$ Therefore, black holes are actually mortal objects with the life durations around $t \sim {\cal M}^3_{BH }$. In the case of black holes with mass around solar mass, this duration is unimaginably large and thus ought to have not been such a big headache per sei. But, the problem inevitably turns up as one tries to interpret the process quantum mechanically. Strictly speaking, this course disrupts the quantum mechanical information transfer. To be more precise, let us take a look at Figure \[Figure1\] and consider some initial (incoming) and final (outgoing) states with density matrices $\varrho_i$ and $\varrho_f$, respectively. Due to the unitary time evolution condition, it is quantum mechanically required from those states to be unitarily related via ${\cal S}$-matrix as follows $$\varrho_f={\cal S}^+ \varrho_i \,{\cal S}.
\label{unitarity}$$ Yet, by analyzing the process more carefully, one will come to the conclusion that the mechanism actually does not satisfy the unitary time evolution criteria. To see this, let us first look at the three distinct hypersurfaces $\Sigma_1, \Sigma_2$ and $\Sigma_3$ in different regions of the Penrose diagram in Figure \[Figure1\]. Later on, let us try to find what observers at each of these surfaces will measure throughout the gravitational collapse ending up with the black hole that ultimately vanishes *a la* Hawking evaporation. Firstly, the observer at $\Sigma_1$ will notice that an extensive amount of information collapses and eventually goes into the black holes. As to the second observer, due to the No-Hair theorem which states that-*up to diffeomorphism*-the mass ${\cal M}$, charge $Q$ and angular momentum $J$ are the only quantities that entirely identify the physics of *stationary* black holes [@NoHair1; @NoHair2; @NoHair3; @NoHair4; @NoHair5], he/she will have just a few information about the black hole. So, due to the lack of being within the entire domain of dependence, this observer would not been able to know what information went into the black hole by solely looking at the $\Sigma_1$. Hence, the observer at $\Sigma_2$ will come across a severe inconsistency since almost infinite amount of states have gotten into black hole but he/she will be left with a couple of degrees of freedom (DOF) to label it. Finally, the third one at $\Sigma_3$ will observe only the corresponding thermal Hawking radiation. Having such a quantum mechanically disconnected information has brought out a weird result of the loss of information (namely, information paradox) which is not a tolerable flaw in the aspect of a deterministically working universe. To get out of this impasse, several alternative ideas have been introduced. As a first one, one could posit that quantum theory fails to be a well-behaved theory at some specific regions of spacetime and thus the Noether theorem is not valid with the existence black hole. That approach physically seems to be not a feasible one. On the other side, the idea of holography in string theory has undoubtedly brought out a leading alternative paradigm of how to handle the long-lived Hawking’s information paradox. As is well-known, the idea states that the black hole actually stocks almost all of its data to the Planckian grids in the holographic shell seating on the horizon [@Maldacena]. In this regard, it has been demonstrated that some particular sort of black holes interestingly supply adequately large rooms to reserve the information required by the Bekenstein-Hawking entropy [@StromingerVafa]. Thus, the holographic framework might be the correct way of how the universe acts. As an another but in some sense extremely radical view, one could expect that the Bekenstein-Hawking entropy formula fails to be true right after a critical size of black hole. In this case, black hole continuously calms down via Hawking radiation until it shrinks to a critical size at which it entirely relaxes as is demonstrated in Figure \[Figure2\].
![Penrose diagram for the case if Bekenstein-Hawking formula was partially valid.[]{data-label="Figure2"}](figure2.png){width="40.00000%"}
As a result, the black holes will not disappear but rather back up residual DOFs. However, since this approach allows even infinitely tiny black hole to inhabit infinite number of quantum mechanical bits and hence entropy, the idea cannot be accepted in a deterministic point of view. As a last but not least perception, one may examine the case wherein the Hawking radiation is not thermal; but instead it is nearly thermal, and thus search for any additional unrevealed DOFs. But, the No-Hair theorem does not allow this suggestion. Recently, Hawking, Perry and Strominger have implicitly demonstrated that, unlike the general perception, black holes actually possess countless number of soft graviton and photon hairs encoded in the holographic two-sphere seating on the future boundary of horizons [@Strominger1; @Strominger2]. The approach has a great potential to resolve the long-lived information paradox because these soft hairs can supply a sufficient amount of rooms wherein the information will be reserved. This opinion automatically comes to life as one starts to suspect about whether there could be any loop-hole in the No-Hair theorem or not. As it has been shown by Hawking, Perry and Strominger, there is indeed a serious and unrecognized shortcoming in the No-Hair theorem. That is, they have demonstrated that the disregarded diffeomorhisms in the original theorem *non-trivially* act on physical systems and thus cannot be moded out. Therefore, as one takes all these symmetries into account, one will get a vast amount of new DOFs (namely, soft particles) on black holes. The thing they have recognized is that as one analyzes the conformal compactification of gravitational shrinks carefully, one will see that black holes are actually *not* stationary but, rather, strongly *time-dependent* objects as is illustrated in Figure (\[Figure4\]). In other words, saying black holes are stationary right after the entire shrink does not say the black holes will be stationary forever. At the early time of the collapse, they might be taken as nearly stationary but the right point will be that it will keep evolving in time throughout their lifetimes.
![Time-dependent evolution of black-holes. To describe the fundamentals of black holes, since they are static or stationary throughout the self shrinks, one generally prefer to draw a spacelike hypersurface $\Sigma$ and determine what geometries are being observed at the $i^0$. The symmetry group there is Poincare symmetries. Thereby, the geometries of black holes observed at $i^0$ can be translated, boosted or rotated by interpreting the transformed states as them as completely dynamical process in such a way that the spacetime remains intact. But, actually this is not exactly the case. Observation of the process at $i^0$ in fact dismisses the dynamical evolution of the geometries. Hence, taking the dynamical evolution into account will naturally require to observe the process in the bulk of the spacetime at $\mathscr{I}^+$ that possess a larger group called (extended) BMS group composed of the product between Supertranslations and Poincare group [@Bondi; @Barnich1; @Barnich2; @Barnich3].[]{data-label="Figure4"}](figure3.png){width="45.00000%"}
This manifestly suggests that one needs to work in the null infinities rather than the spacelike ones in order to figure out the correct mechanisms behind the universe at least in the infrared limit of quantum gravity. Hence, one has to take the symmetries of null infinities into account in order to cope with the dynamical evolution of black holes. This leads us to the readdress the so-called BMS transformations which nontrivially act on the Hilbert spaces of the null-boundaries $\mathscr{I}^\pm$ of asymptotically flat spacetimes. In other words, the dynamics of time-dependent black holes can be studied via (extended) BMS symmetries along $\mathscr{I}^\pm$ [@Bondi; @Barnich1; @Barnich2; @Barnich3]. Therefore, let us take a short look at the transformations: the (extended) BMS is an infinite dimensional symmetry group belonging to the $4$-dimensional asymptotically Minkowski spacetimes. Together with the ordinary $SL(2, \mathbb{C})$ Lorentz boosts (that is, the global conformal group of the sphere at infinity), it also comprises an infinite family of allowed diffeomorphism (i.e., supertranslations) which *nontrivially* translate the information at the null boundaries $\mathscr{I}^\pm$ without altering the asymptotic form of the spacetime. To inquire into the basics of BMS transformations, as is seen in Figure (\[confcompmink\]), one shall set up the retarded (advanced) time coordinate $u$ ($v$) on $\mathscr{I}^+$ ( $\mathscr{I}^-$), a radial coordinate $r$, complex coordinates $z$ and $\bar{z}$ on the two-spheres at constant $u(v)$ and r.
![Conformal compactification of Minkowski and Antipodal Mapping.[]{data-label="confcompmink"}](figure4.png){width="50.00000%"}
Then, the asymptotic Minkowski metric can be recast into the Bondi metric as follows $$\begin{aligned}
ds^2&=-du^2-2dudr+2 r^2 \gamma_{z\bar{z}} dz d \bar{z}\\
&+\frac{2 m^+_B (z, \bar{z})}{r} du^2+r C_{zz} dz^2+r C_{\bar{z}\bar{z}} d\bar{z}^2\\
&+D^z C_{zz} du dz+D^{\bar{z}} C_{\bar{z}\bar{z}} du d\bar{z}+ \mbox{subsubleading terms}+\cdots,
\label{bondimetric}
\end{aligned}$$ in the upper-half coordinates of spacetime. Here, $\gamma_{z\bar{z}}=\frac{2}{(1+z\bar{z})^2}$ is the warped metric describing the conformal $S^2$, $m^+_B$ is the Bondi-mass aspect associated to the masses of the asymptotically Minkowski space at different retarded coordinates $\mathscr{I}^+ (u_1, u_2, \cdots)$ throughout the self gravitational shrink, $C_{zz}$ is the angular momentum aspect of flat spacetime, and $D_z$ is the covariant derivative assigned to $S^2$ at infinity. Here, the associated isometries of the Bondi metric are the elements of renowned (extended) BMS group that semi-directly reconciles the Poincare group and an infinite dimensional nontrivial family of diffeomorphisms dubbed as supertranslations. The most intriguing property of this group is undoubtedly the existence of the supertranslations $$u \rightarrow u- f(z, \bar{z}),$$ which, together with the following transformations $$r \rightarrow r-D^2_z f, \quad z \rightarrow z \rightarrow z+\frac{1}{z} D^z f,$$ leaves the leading piece (that is, the Minkowski part) of the Bondi metric in Eq.(\[bondimetric\]) intact but non-trivially evolves the subleading portion by virtue of arbitrary real $f(z, \bar{z})$. Here, the relevant vector field that creates these allowed infinite number of diffeomorphisms are $$\xi=f\partial_u+D^2_z f\, \partial_r-\frac{1}{r} (D^z f\, \partial_z + D^{\bar{z}} f \, \partial_{\bar{z}}),$$ which, as mentioned above, non-trivially changes the subleading Bondi mass and angular momentum inputs as follows $${\cal L}_\xi m^+_B= f \partial_u m^+_B, \qquad {\cal L}_\xi C_{zz}=f N_{zz}-2 D^2_z f.$$
The information encoded in the Bondi metric are linked to each others via the constraint equations which are the Einstein’s field equation $G_{\mu\nu}=T^{matter}_{\mu\nu}$. Accordingly, as one computes the $"uu"$ portion of Einstein equation, one will get $$\begin{aligned}
\partial_u m^+_B&=\frac{1}{4} (D^2_z N^{zz}+ D^2_{\bar{z}} N^{\bar{z}\bar{z}} )\\
&\hskip .2 cm -\frac{1}{4} N_{zz}N^{zz}-\frac{1}{2}\left. T^{matter}_{uu} \right\vert_{r \rightarrow \infty},
\end{aligned}$$ that gives the dynamical evolution of $m^+_B$ along $\mathscr{I}^+$. Here, $N_{zz}= \partial_u C_{zz}$ is the so-called Bondi news yielding the energy flux intersecting $\mathscr{I}^+$. Observe that the term $ \left.T^{matter}_{uu} \right\vert_{r \rightarrow \infty} $ is the ejected matter-energy from the bulk to $\mathscr{I}^+$ during the dynamical process. Additionally, the $"uz"$ component of Einstein equation will give $$\begin{aligned}
\partial_u N_z &=-\frac{1}{4} \Big[D_z D^2_{\bar{z}} C^{\bar{z}\bar{z}}-D^3_z C^{zz} \Big]-\left. T^{matter}_{uz} \right\vert_{r \rightarrow \infty}\\
&\hskip 0.4 cm+\partial_z m^+_B+\frac{1}{16} D_z \partial_u \Big[C_{zz} C^{zz} \Big] \\
&\hskip 0.4 cm-\frac{1}{4} N^{zz} D_z C_{zz}-\frac{1}{4} N_{zz} D_z C^{zz}\\
&\hskip 0.4 cm-\frac{1}{4} D_z \Big[C^{zz}N_{zz}-N^{zz}C_{zz} \Big].
\end{aligned}$$ Note that as long as one knows Bondi news, one can easily compute all the information at all the retarded times during entire dynamical process taken place [@Strominger1; @Strominger2; @Strominger3; @Strominger4; @Strominger5; @Strominger6; @Strominger7; @Strominger8; @Strominger9; @Strominger10; @Strominger11; @Strominger12; @Strominger13].
Hidden Soft Modes in the Massless Quantum Electrodynamics
=========================================================
Weinberg’s Soft Photon Theorem
------------------------------
As is shown in the Figure (\[WeinbergTruncation\]), by summing all the possible contributions coming from the vertices at which a photon ties the leg of outgoing particle via the truncation of the related propagators in the LSZ reduction formula, Weinberg evaluated the probability of getting a soft photon as $n$ incoming particles possessing charges $q^{(i)}_k$ and momenta $p^{(i)}_k $ interact and then scatter into $m$ outgoing particles possessing charges $q^{(o)}_k$ and momenta $p^{(o)}_k$ [@Weinberg]. Here, he showed that the amplitude of obtaining a soft photon with momentum $p_\lambda$ and polarization $\epsilon^+$ throughout this course is $$\begin{aligned}
&\lim_{\eta\rightarrow 0^+} (\eta M_+ [p_\gamma; \{p^{(i)}_k \}, \{p^{(o)}_k \}] )\\
&=e \lim_{\eta \rightarrow 0^+} \bigg (\sum^{m}_{k=1} \frac{\eta q^{(o)}_k p^{(o)}_k \cdot \epsilon^+(p_\gamma)}{p^{(o)}_k \cdot p_\gamma}\\
& \hskip 2 cm- \sum^{n}_{k=1} \frac{\eta q^{(i)}_k p^{(i)}_k \cdot \epsilon^+(p_\gamma)}{p^{(i)}_k \cdot p_\gamma} \bigg) \times M[\{p^{(i)}_k \}, \{p^{(o)}_k \}],
\label{weinbergsoft}
\end{aligned}$$ where $ M[\{p^{(i)}_k \}, \{p^{(o)}_k \}] $ is the scattering amplitude of incoming and outgoing particles, while $M_+ [p_\gamma; \{p^{(i)}_k \}, \{p^{(o)}_k \}] $ is the amplitude containing a soft photon. Strominger and his collaborators have explicitly shown that, with antipodal mapping of the past and future null infinities around $i^0$ which we review below, the emergent Ward identities associated to the quantum S-matrix turn out to be the celebrated Weinberg’s soft theory \[weinbergsoft\] [@Weinberg; @Strominger1; @Strominger2; @Strominger3; @Strominger4; @Strominger5; @Strominger6; @Strominger7; @Strominger8; @Strominger9; @Strominger10; @Strominger11; @Strominger12].
![The propagation of a soft photon through a scattering process.[]{data-label="WeinbergTruncation"}](figure5.png){width="42.00000%"}
Lienard-Wiechert potential near $i^0$ and Antipodal Mapping
-----------------------------------------------------------
As was mentioned above, there are actually two types of BMS groups which dynamically evolve the information on $\mathscr{I}^+$ and $\mathscr{I}^-$. These groups are totally distinct groups because of the singular point $i^0$ in the Einstein cylinder at which one clearly gets different results as one boosts toward and away from it. The discontinuity turns out to be a severe trouble particularly as one tries to link the initial states to the final ones via a convenient (classical or quantum mechanical) $S$-matrix. But, Strominger *et. al.* have recently shown that this obstacle can be overcome with a physically legitimate antipodal mapping which links the angle that incoming ray crosses two-sphere to the opposite angle of the outgoing ray [@Strominger1; @Strominger2; @Strominger3; @Strominger4; @Strominger5; @Strominger6; @Strominger7; @Strominger8; @Strominger9; @Strominger10; @Strominger11; @Strominger12; @Strominger13]. This can be seen by looking at the behavior of the Lienard-Wiechert potential around $i^0$ as follows: as is well-known, in the case of the Maxwell theory coupled to a massless charged scalar field, one has $$\nabla^\nu {\cal F}_{\nu \mu}= e^2 {\cal J}_\mu \quad \mbox{where} \quad \nabla^\mu {\cal J}_\mu=0.
\label{maxwelfeq}$$ Assuming a current associated to $m$ point-like sources whose wordline are described by proper time $\tau$ $${\cal J}_a(x)= \sum_{l=1}^m {\cal Q}_l \int d \tau\, \Theta(-\tau) \, \delta^4 [x^b-{\cal U}^{bl} \tau ]\, {\cal U}_{al},$$ where ${\cal U}^b_l=\gamma_l [1, \bm{\beta}_l ]$, $\gamma_l=[1-\bm{\beta}^2 ]^{-1/2}$, one gets the radial solution as follows $${\cal F}_{radial} (\vec{x}, t) = \frac{e^2}{4 \pi} \sum_{l=1}^m \frac{{\cal Q}_l \gamma_l \, [r- t \hat{\vec{x}} \cdot \bm{\beta}_l ]}{\Big( \gamma^2_l [t- r \hat{\vec{x}} \cdot \bm{\beta}_l]^2-[t^2-r^2]\Big)^{\frac{3}{2}}}.
\label{wiecherlineard}$$ Herein, $r=\sqrt{\vec{x}\cdot \vec{x}}$ and $\vec{x}=r\hat{\vec{x}}$. As Strominger *et. al.* showed, the above-mentioned discontinuity becomes quite manifest as one studies the solution about the spatial infinity $i^0$ [@Strominger7; @Strominger8; @Strominger9; @Strominger10; @Strominger13]. To see this, one needs to analyze the various limits of Eq.(\[wiecherlineard\]) in the conformal compactification of Minkowski spacetime. For this purpose, let us now approximate the Lienard-Wiechert potential toward $i^0$ through two distinct paths A and B as is illustrated in Figure (\[confcompmink\]): firstly, along the first path A, the potential in the retarded coordinates $(u=t-r, r, z, \bar{z})$ becomes $${\cal F}_{ru} = \frac{e^2}{4 \pi} \sum_{l=1}^m \frac{{\cal Q}_l \gamma_l \, [1-(1+\frac{u}{r}) \hat{\vec{x}} \cdot \bm{\beta}_l ]}{r^2\Big( \gamma^2_l [1+\frac{u}{r}- \hat{\vec{x}} \cdot \bm{\beta}_l]^2-[1+\frac{u}{r}]^2+1\Big)^{\frac{3}{2}}},
\label{wiecherlineard1}$$ which, for the limits $r \rightarrow \infty$ and $u=\mbox{constant}$, yields $${\cal F}_{ru} = \sum_{l=1}^m \frac{e^2 {\cal Q}_l}{4\pi \gamma^2_l r^2 [1- \hat{\vec{x}} \cdot \bm{\beta}_l]^2 },
\label{wiecherlineard2}$$ where the term $\hat{\vec{x}} \cdot \bm{\beta}_l$ corresponds to the electrical dipole moment. Observe that since $r$ in Eq.(\[wiecherlineard2\]) is *not* constant, the electric field here is *angle-dependent*. Similarly, through the path B, one will get the potential in limits $r \rightarrow \infty$ and $v=\mbox{constant}$ in the advanced coordinates $(v=t+r, r, z, \bar{z})$ as follows $${\cal F}_{rv} = \sum_{l=1}^m \frac{e^2 {\cal Q}_l}{4\pi \gamma^2_l r^2 [1+\hat{\vec{x}} \cdot \bm{\beta}_l]^2 }.
\label{wiecherlineard3}$$ Apparently, Eq.(\[wiecherlineard2\]) and Eq.(\[wiecherlineard3\]) are not equal to each others and thus the potential through paths A and B do not commute in the vicinity of $i^0$. That is, it is a singular point at least in the Poincare transformations aspect. This turns out to be a huge problem as one specifically tries to constitute a viable $S$-matrix connecting the incoming and outgoing modes. A possible way out of this problem is given with the following antipodal relationship that ties $\mbox{BMS}^-$ and $\mbox{BMS}^+$ belonging to $\mathscr{I}^-$ and $\mathscr{I}^+$, respectively $$\left. r^2 {\cal F}_{ru} (\hat{\vec{x}}) \right\vert_{\mathscr{I}^+_-}=\left. r^2 {\cal F}_{rv} (-\hat{\vec{x}}) \right\vert_{\mathscr{I}^-_+} \qquad \mbox{as}\qquad r \rightarrow \infty,$$ which has been shown to be a definite symmetry of S-matrix that leads to a countless amount of conserved charges matching to Weinberg’s soft photons [@Strominger1; @Strominger2; @Strominger7; @Strominger8; @Strominger9; @Strominger10; @Strominger13].
Field decays around $\mathscr{I}^\pm$, large gauge symmetries and the associated quantum brackets
-------------------------------------------------------------------------------------------------
Now that we have seen some peculiar features of the null boundaries of asymptotically Minkowski spacetimes, let us look at how the massless quantum electrodynamics acts in these infrared vicinities. (Here, we will totally follow [@Strominger7; @Strominger8; @Strominger9; @Strominger10; @Strominger13]). Recall that the Maxwell equation incorporated with a massless and charged scalar field in Eq.(\[maxwelfeq\]) has the ensuing gauge freedom $$\delta_\xi A_\mu=\partial_\mu \xi \quad \mbox{with} \quad \xi \sim \xi+2 \pi,$$ which can be fixed by imposing the following conditions $$A_r=0 \quad \mbox{and} \quad \left. A_u \right\vert_{\mathscr{I}^+} =0.
\label{decayu}$$ Let us now try to figure out how the dynamical quantities decay as one approximate them towards the future null infinity. To ends up with a definite energy flow through $\mathscr{I}^+$ varying according to $\int_{r \rightarrow \infty} F_u{^{z}} F_{uz} $ and satisfy Eq.(\[decayu\]) , one needs to respectively have $$\left. A_z \right\vert_{\mathscr{I}^+} \sim {\cal O}(1) \quad \mbox{and} \quad A_u \sim {\cal O}(\frac{1}{r}),$$ yielding $$\begin{aligned}
A_z({r, u, z, \bar{z}})&=\hat{A}_z({u, z, \bar{z}})+\sum_{n=1}^{\infty} \frac{\hat{A}_{z(n)}({u, z, \bar{z}})}{r^n}\\
A_u({r, u, z, \bar{z}})&=\frac{\hat{A}_u({u, z, \bar{z}})}{r}+\sum_{n=1}^{\infty} \frac{\hat{A}_{u(n)}({u, z, \bar{z}})}{r^{n+1}},
\label{decaygag}
\end{aligned}$$ in the neighborhood of $\mathscr{I}^+$. Here, with Eq.(\[decaygag\]), one will arrive at the following decay of the leading parts of the curvature $$F_{z\bar{z}}={\cal O}(1), \quad F_{ur}={\cal O}(\frac{1}{r^2}), \quad F_{uz}={\cal O}(1), \quad F_{rz}={\cal O}(\frac{1}{r^2}),$$ accompanied by the magnitudes corresponding to the curvatures on $\mathscr{I}^+$ $$\begin{aligned}
\hat{F}_{z\bar{z}}&=\partial_z \hat{A}_{\bar{z}}-\partial_{\bar{z}} \hat{A}_z, \quad \hat{F}_{uz}=\partial_u \hat{A}_z,\\
\hat{F}_{rz}&=-\hat{A}_z, \hskip 1.7 cm \hat{F}_{ur}=\hat{A}_u.
\end{aligned}
\label{strenghttensoronnull}$$ Using all these set-ups in the Maxwell field equations in Eq.(\[maxwelfeq\]) give rise to the following leading constraint $$\gamma_{z\bar{z}} \partial_u \hat{A}_u = \partial_u (\partial_z \hat{A}_{\bar{z}}+\partial_{\bar{z}}\hat{A}_z )+e^2 \gamma_{z\bar{z}} {\cal J}_u,
\label{maxwconstr}$$ where ${\cal J}_u({u, z, \bar{z}})= \left. r^2{\cal J}_u({u, r, z, \bar{z}}) \right\vert_{ \lim_{r\rightarrow \infty}}$.\
In fact, with the choices in Eq.(\[decayu\]), there still remain unadjusted redundant freedoms associated with the asymptotic angular transformations of the rounded sphere in the infinity. These leftover symmetries are dubbed as *large gauge transformations* and described as follows [@Strominger7; @Strominger8; @Strominger9; @Strominger10; @Strominger13] $$\delta_{\hat{\xi}} \hat{A}_z(u, z, \hat{z})=\partial_z \hat{\xi}(z, \bar{z}) \,\, \, \mbox{where} \,\, \, \hat{\xi}({z, \bar{z}}) \equiv \left. \xi ({r, z, \bar{z}}) \right\vert_{ \lim_{r\rightarrow \infty}},$$ which are induced by the charges $${\cal Q}^+_{\hat{\xi}} =\frac{1}{e^2}\int_{\mathscr{I}^+_-} d^2 z \gamma_{z\bar{z}}\, \hat{\xi}\, \hat{F}_{ru}.
\label{charge}$$ Note that, by virtue of the relevant leading component of the field-strength tensor in Eq.(\[strenghttensoronnull\]), Eq.(\[charge\]) can be extended throughout the entire $\mathscr{I}^+$ as $$\begin{aligned}
{\cal Q}^+_{\hat{\xi}} &=\frac{1}{e^2} \int_{\mathscr{I}^+} du d^2 z \, \hat{\xi}\, \gamma_{z\bar{z}}\, \partial_u \hat{F}_{ru}\\
&=\frac{1}{e^2} \int_{\mathscr{I}^+} du d^2 z \, \hat{\xi} \, \gamma_{z\bar{z}}\, \partial_u \hat{A}_u.
\end{aligned}$$ Observe that, with Eq.(\[maxwconstr\]), the charge operator will ultimately turn into $${\cal Q}^+_\xi =\frac{1}{e^2} \int_{\mathscr{I}^+} du d^2 z \, \hat{\xi}\, \Big(\partial_u(\partial_z \hat{A}_{\bar{z}}+\partial_{\bar{z}}\hat{A}_z )+e^2\gamma_{z\bar{z}}{\cal J}_u \Big),$$ which produces the large gauge transformations over the symplectic phase space of $\mathscr{I}^+$. Subsequently, setting $\hat{\xi}=1$ will yield the regular charge flow of the massless scalar through the conformal sphere $${\cal Q}^+_1 = \int_{\mathscr{I}^+} du d^2 z \, \gamma_{z\bar{z}}{\cal J}_u.$$ Alternatively, picking up the parameter as $\hat{\xi}(z, \bar{z})=\delta^2(z-w)$ will intriguingly give the *angle-relying* charge $${\cal Q}^+_{w\bar{w}} =\frac{1}{e^2} \int_{-\infty}^{\infty} du \Big(\partial_u(\partial_w \hat{A}_{\bar{w}}+\partial_{\bar{w}}\hat{A}_w )+e^2\gamma_{w\bar{w}}{\cal J}_u \Big),$$ which describes the whole energy emitted through the angles $(w, \bar{w})$ of the rounded sphere inhabiting in the infinity. Here, the total-derivative piece is the generator that creates the so-called zero-energy *soft photons* in the certain angles on the conformal sphere [@Strominger7; @Strominger8; @Strominger9; @Strominger10; @Strominger13].
As for the canonical quantization of the model, as one tries to take the mere forms of the existing nontrivial brackets in the literature in [@Ashtekar; @Frolov] as the basic brackets here, one will get $$\Big[\hat{F}_{uz}(u, z, \bar{z}), \hat{F}_{u^{'}\bar{w}} (u^{'}, w, \bar{w}) \Big]= \frac{i e^2}{2} \partial_u \delta(u-u^{'}) \delta^2(z-w),$$ which yields $$\Big[\hat{A}_z(u, z, \bar{z}), \hat{A}_{\bar{w}}(u^{'}, w, \bar{w}) \Big]=-\frac{i e^2}{4} \mbox{sgn}(u-u^{'}) \delta^2(z-w).
\label{commgag}$$ But, as is shown in [@Strominger7], with the above brackets, the charge operator cannot coherently and completely create the transformations due to the emerging coefficient $1/2$: $$\Big[{\cal Q}^+_{\hat{\xi}}, \hat{A}_z(u, z, \bar{z}) \Big]= \frac{i}{2}\partial_z {\hat{\xi}}(z, \bar{z})\neq i \delta_{\hat{\xi}} \hat{A}_z(u, z, \bar{z}),$$ This obstacle can be resolved by imposing the following physically viable extra boundary and decay restrictions: firstly, in order for the $\hat{A}_z$ component not to pick certain directions on the *spatially localized* sphere settling in the future and past boundaries of future null infinity $\mathscr{I}^+_\pm$, one needs to also assume the following restrain on the related portion of the curvature $$\left. \hat{F}_{z\bar{z}} \right\vert_{\mathscr{I}^+_\pm}=0,
\label{constriantcanquan}$$ which actually does not remain intact in Eq.(\[commgag\]). Accordingly, imposing the ensuing smooth relations $$\begin{aligned}
&\Big[\hat{A}^\pm_z(z, \bar{z}), \hat{A}_{\bar{w}}(u^{'}, w, \bar{w}) \Big]\\
&\qquad \qquad= \lim_{u \rightarrow \pm \infty}\Big[\hat{A}_z(u, z, \bar{z}), \hat{A}_{\bar{w}}(u^{'}, w, \bar{w}) \Big], \\
&\Big[\hat{A}^+_z(z, \bar{z})-\hat{A}^-_z(z, \bar{z}), \hat{A}^\pm_{\bar{w}}(w, \bar{w}) \Big]\\
&\qquad \qquad= \lim_{u^{'} \rightarrow \pm \infty} \Big[\hat{A}^+_z(z, \bar{z})-\hat{A}^-_z(z, \bar{z}), \hat{A}_{\bar{w}}(u^{'}, w, \bar{w}) \Big],
\label{extradecays}
\end{aligned}$$ eventually converts Eq.(\[constriantcanquan\]) into a theoretically and physically legitimate constraint on the phase space. Here, $\hat{A}^\pm_z(z, \bar{z})$ is the value of the $\hat{A}_z(u, z, \bar{z})$ at $\mathscr{I}^+_\pm$. Note that this extra condition imposes the gauge field to be a *pure-gauge* on the *spatially localized* two-sphere living in $\mathscr{I}^+_\pm$ as follows $$\hat{A}^\pm_z(z, \bar{z})=e^2\partial_z \hat{\Omega}_\pm(z, \bar{z}),
\label{puregauge}$$ with which one gets $$\begin{aligned}
\Big[\hat{\Omega}_\pm (z, \bar{z}), \hat{A}_w(u^{'}, w, \bar{w}) \Big]&=\mp \frac{i}{8 \pi} \frac{1}{z-w},\\
\Big[\hat{\Omega}_+(z, \bar{z}), \hat{\Omega}_-(w, \bar{w})\Big]&=\frac{i}{4 \pi e^2} \ln \lvert z-w \rvert^2.
\label{newcomu}
\end{aligned}$$ By recasting ${\cal Q}^+_\xi$ as $${\cal Q}^+_\xi=2 \int_{S^2}d^2z \, \xi \partial_z \partial_{\bar{z}} \Big(\hat{\Omega}_+-\hat{\Omega}_- \Big)+\int_{\mathscr{I}^+} du d^2z \, \gamma_{z\bar{z}} \xi {\cal J}_u
\label{newcharge}$$ one will finally achieve to produce the large gauge transformations without any problem as follows [@Strominger1; @Strominger2; @Strominger7; @Strominger8; @Strominger9; @Strominger10; @Strominger13] $$\Big[{\cal Q}^+_\xi, \, \hat{A}_z(u, z, \bar{z}) \Big]=i \partial_z \xi(z, \bar{z}).$$
Faddeev-Jackiw Brackets for massless QED in the Eikonal Limit
=============================================================
In this part, we will see that the basic quantum commutators for the soft-photons in Eq.(\[newcomu\]) can alternatively be obtained from the Faddeev-Jackiw quantization of the massless charged particle in the eikonal limit studied in [@JackiwKabat]. Here, we will also see that the gauge fields associated with the massless charged particle also localize on a two-dimensional $(t, z)$ plane and interestingly turn into *pure-gauges* in the eikonal limit in a similar manner to the soft-photons inhabiting over the spatially localized two-sphere at the $\mathscr{I}^+_\pm$ in Eq.(\[puregauge\]).
To see this explicitly, let us now reconsider the relevant computations in [@JackiwKabat] wherein the fields associated with the massless charged particle are obtained from the ones generated by the massive particle[^3]: recall that the fields in the frame of a massive charged point-like particle traveling in the $z$-direction and thereby admitting $\bm{\beta}=\frac{v}{c} \hat{\vec{z}}$ are described as follows $$\vec{E}^{'}=\frac{e\vec{r}^{'}}{r^{'3}}, \hskip 1 cm \vec{B}^{'}=0.
\label{rfram1}$$ Boosting Eq.(\[rfram1\]) yields $$\vec{E}=\gamma \vec{E}^{'}-\frac{\gamma^2}{\gamma+1} \bm{\beta} (\bm{\beta} \bm{\cdotp} \vec{E}^{'}), \hskip 1 cm \vec{B}=\bm{\beta} \times \vec{E}^{'},
\label{transeb}$$ which, with the $ \vec{r}^{'}=\vec{r}_\perp+\gamma (z-vt) \hat{\vec{z}} $ along $z$-direction, turns into $$\vec{E}=\frac{\gamma e\Big(\vec{r}_\perp+(z-vt) \hat{\vec{z}}\Big)}{\Big(r^2_\perp+\gamma^2 (z-vt)^2\Big)^{\frac{3}{2}}}.$$ As $v \rightarrow c$, $\beta \rightarrow 1$ (or $\gamma \rightarrow \infty $), the longitudinal portion goes away and thus all the fields concentrate throughout the transverse plane. Consequently, by making use of the relation $$\lim_{\Sigma\to\infty} \frac{\Sigma}{(1+\Sigma^2 \Delta^2)}=2 \delta(-\Delta),$$ one will eventually get the fields corresponding to the massless point-like charged particle as follows $$\vec{E}=\frac{2e\vec{r}_\perp}{r^2_\perp} \delta(ct-z) \quad \mbox{and} \quad \vec{B}=-\frac{2e\hat{\vec{v}} \times \vec{r}_\perp}{r^2_\perp} \delta(ct-z).
\label{massfed}$$ Furthermore, by benefiting from the Gauss’ identities $$\bm{\nabla}_\perp \bm{\cdotp} \Big(\frac{\vec{r}_\perp}{r^2_\perp} \Big)=2\pi \delta^{(2)}(\vec{r}_\perp) \quad \mbox{and} \quad (\hat{\vec{z}} \bm{\cdotp} \bm{\nabla}) \Big(\frac{\vec{r}_\perp}{r^2_\perp} \Big)=\partial_z\Big(\frac{\vec{r}_\perp}{r^2_\perp} \Big)=0,$$ one can easily show that for $${\cal J}^\mu=e c n^\mu \delta^{(2)}(\vec{r}_\perp)\delta(ct-z),$$ where $n^\mu=(1, \hat{\vec{n}})$, Eq.(\[massfed\]) obeys the Maxwell equation $$\partial_\nu F^{\nu \mu}= \frac{4 \pi}{c} {\cal J}^\mu.$$ It is straightforward to demonstrate that either of the following choices of the gauge fields generate Eq(\[massfed\]) $$\begin{aligned}
1^{st}: \hskip 0.3 cm A^0_1=A^z_1&=0, \qquad \vec{A}_\perp^1=-2e \theta(ct-z) \bm{\nabla}_\perp \ln(\mu r_\perp), \\
2^{nd}: \hskip 0.3 cm A^0_2=A^z_2&=-2e \delta(ct-z) \ln(\mu r_\perp), \quad \vec{A}_\perp^2 =0,
\label{gauge1}
\end{aligned}$$ which are linked to each others as follows $$A^\mu_1=A^\mu_2+\partial^\mu \Omega,$$ wherein the gauge parameter reads $$\Omega=2e\theta(ct-z)\ln(\mu r_\perp).$$ As is clear, for the first choice in Eq.(\[gauge1\]), the wave function does not see the potential when $ct<z$ whereas, for $ct>z$, it interacts with the potential and so $$\psi_{ct>z}=e^{[-i\frac{e e^{'}}{\hbar c} \ln(\mu^2 r^2_\perp)]}\psi_i.$$ Assuming $\psi_i$ to be a plane wave and making the appropriate chance of variables in [@tHooft1; @tHooft2], one ultimately gets $$F(s, t)= \frac{\Gamma(1+i\zeta)}{4 \pi i \mu^2 \Gamma(-i \zeta)} \Big (\frac{4 \mu^2}{-t} \Big)^{1+i\zeta},
\label{jackiweikonal}$$ that can be shown to be eikonal by slightly recasting the common formula as follows $$F_{eikonal}(s, t)= i \int \frac{d^2 b}{(2\pi)^2} e^{i \vec{q}_\perp \bm{\cdotp} \vec{b}} \Big (1-e^{-iee^{'}\int \frac{d^2 k_\perp}{(2\pi)^2} \frac{e^{i \vec{k} \bm{\cdotp}\vec{b}}}{k^2_\perp+\mu^2} } \Big).$$ Here, $s$ and $t$ are the Mandelstam quantities and $\mu$ is the IR regulator of the propagator [@JackiwKabat; @Eikonal].
As for the second choice, let us notice that, for the massless case, the current can be decomposed into the light-cone portion plus the transverse part as follows $${\cal J}_+(\vec{x})={\cal J}_+(x^+, \vec{r}_\perp), \,\,\, {\cal J}_-(\vec{x})={\cal J}_-(x^-, \vec{r}_\perp), \,\,\, {\cal J}^i(\vec{x})=0,$$ for which one could assume $$\vec{A}^\perp=0, \quad A_\pm=\partial_\pm \Omega.$$ Working in the Landau gauge $\partial_\mu A^\mu=0$ shows that the scalar potential is harmonic which allows one to decompose it into left and right modes as $$\Omega(\vec{x})=\Omega^+(x^+, \vec{r}_\perp)+\Omega^-(x^-, \vec{r}_\perp).
\label{eikone}$$ Accordingly, one can also recast the current in terms of wave numbers as follows $${\cal J}^\alpha=\epsilon^{\alpha \beta} \partial_\beta k \hskip 1 cm \alpha, \beta= +, -.
\label{currentk}$$ Here, one has $$k(\vec{x})=k^+(x^+, \vec{r}_\perp)-k^-(x^-, \vec{r}_\perp).$$ Notice that $\partial_\alpha {\cal J}^\alpha=0$ in this convention.
By referring [@JackiwKabat] for the details, let us recall that, in these settings, the Lagrangian reads $$\begin{aligned}
{\cal L}&=-\frac{1}{2} \partial_- \Omega^- \nabla^2 \partial_+ \Omega^+-\frac{1}{2} \partial_+ \Omega^+ \nabla^2 \partial_- \Omega^-\\
&-\partial_+ k^+ \partial_- \Omega^--\partial_- k^- \partial_+ \Omega^+\\
&=-\partial_- \Big( \frac{1}{2}\Omega^- \nabla^2 \partial_+ \Omega^++\partial_+k^+ \Omega_- \Big)\\
&-\partial_+ \Big(\frac{1}{2} \Omega^+ \nabla^2 \partial_- \Omega^-+\partial_- k^- \Omega^+ \Big),
\end{aligned}$$ which will lead to the action $$\begin{aligned}
S(\Omega, k)&= \oint d\tau \int d^2 r_\perp \Big(\frac{1}{2} \Omega^- \nabla^2 \dot{\Omega}^+-\frac{1}{2} \Omega^+ \nabla^2 \dot{\Omega}^- \\
&\hskip 2.7cm+\dot{k}^+ \Omega^--\dot{k}^- \Omega^+ \Big),
\label{eikoac}
\end{aligned}$$ that involves the transverse part and a localized $1+1$-dimensional surface $x^\alpha$ described by $\tau$ shown in Figure \[fjfig\]. Observe that the variations with respect to the $\Omega^-$ and $\Omega^+$ respectively give $$\nabla^2 \dot{\Omega}^+=-\dot{k}^+, \quad \nabla^2 \dot{\Omega}^-=-\dot{k}^-,$$ whose solutions are $$\begin{aligned}
\Omega^+(x^+, \vec{r}_\perp)&=-(\nabla^2)^{-1} k^+(x^+, \vec{r}_\perp)\\
\Omega^-(x^-, \vec{r}_\perp)&=-(\nabla^2)^{-1} k^-(x^-, \vec{r}_\perp).
\end{aligned}$$ By noting that Eq.(\[eikoac\]) is already in the desired symplectic form, one will get the fundamental Faddeev-Jackiw bracket [@FaddeevJackiw; @JackiwKabat] $$\Big[\Omega^+(x^+(\tau), \vec{r}_\perp),\, \Omega^-(x^-(\tau), \vec{r}_\perp) \Big]=\frac{i}{2 \pi} \ln{\left|\vec{r}_\perp-\vec{r}^{'}_\perp\right|},$$ which is apparently same as the one obtained in the soft approach via extra boundary conditions in Eq.(\[newcomu\]).
![Localization of the fields in the (1+1)-dimensional surface.[]{data-label="fjfig"}](figure6.png){width="30.00000%"}
Conclusions
===========
In this note, by reviewing some forthcoming topics in the literature, we try to understand whether the basic quantum commutator of the soft photon model [@Strominger7; @Strominger8; @Strominger9; @Strominger10] can be alternatively obtained from the Faddeev-Jackiw symplectic Hamiltonian reduction method [@FaddeevJackiw; @JackiwKabat] for the quantum electrodynamics in the eikonal limit or not. If this relation could be established entirely, one could then study the soft theories in the Faddeev-Jackiw framework which would more likely provide a more economical way without dealing with constraints and assuming extra decay conditions. Throughout the analysis, we observe that there are indeed some intriguing similarities between either perspectives such as having the same fundamental brackets or being pure gauge. But, for a concrete understanding of these observations and hence the existence or non-existence of such a relation, one undoubtedly needs to elaborate the topics in all aspects.
\[ackno\] Acknowledgments
=========================
We would like to thank Roman Jackiw and Daniel Kabat for the discussions during our postdoc at MIT/LNS. We would also like to thank Ercan Kilicarslan for critical reading of the note and Department of Physics at Eastern Mediterranean University (N. Cyprus) for their warm hospitality.
[0]{}
S. W. Hawking, M. J. Perry and A. Strominger,“Superrotation Charge and Supertranslation Hair on Black Holes,” JHEP [**1705**]{}, 161 (2017).
S. W. Hawking, M. J. Perry and A. Strominger, “Soft Hair on Black Holes,” Phys. Rev. Lett. [**116**]{}, no. 23, 231301 (2016).
T. He, V. Lysov, P. Mitra and A. Strominger, “BMS supertranslations and Weinberg’s soft graviton theorem,” JHEP [**1505**]{}, 151 (2015).
A. Strominger, “On BMS Invariance of Gravitational Scattering,” JHEP [**1407**]{}, 152 (2014).
D. Kapec, V. Lysov, S. Pasterski and A. Strominger, “Semiclassical Virasoro symmetry of the quantum gravity $ \mathcal{S}$-matrix,” JHEP [**1408**]{}, 058 (2014).
A. Strominger and A. Zhiboedov,“Gravitational Memory, BMS Supertranslations and Soft Theorems,” JHEP [**1601**]{}, 086 (2016).
T. He, P. Mitra, A. P. Porfyriadis and A. Strominger, “New Symmetries of Massless QED,” JHEP [**1410**]{}, 112 (2014).
V. Lysov, S. Pasterski and A. Strominger, “Low’s Subleading Soft Theorem as a Symmetry of QED,” Phys. Rev. Lett. [**113**]{}, no. 11, 111601 (2014).
D. Kapec, M. Pate and A. Strominger, “New Symmetries of QED,” Adv. Theor. Math. Phys. [**21**]{}, 1769 (2017).
D. Kapec, M. Perry, A. M. Raclariu and A. Strominger, “Infrared Divergences in QED, Revisited,” Phys. Rev. D [**96**]{}, no. 8, 085002 (2017).
A. Strominger,“Asymptotic Symmetries of Yang-Mills Theory,” JHEP [**1407**]{}, 151 (2014).
T. He, P. Mitra and A. Strominger, “2D Kac-Moody Symmetry of 4D Yang-Mills Theory,” JHEP [**1610**]{}, 137 (2016).
H. Bondi, M. G. J. van der Burg and A. W. K. Metzner,“Gravitational waves in general relativity. 7. Waves from axisymmetric isolated systems,” Proc. Roy. Soc. Lond. A [**269**]{}, 21 (1962); R. K. Sachs,“Gravitational waves in general relativity. 8. Waves in asymptotically flat space-times,” Proc. Roy. Soc. Lond. A [**270**]{}, 103 (1962).
G. Barnich and C. Troessaert, “Supertranslations call for superrotations,” PoS, 010 (2010). \[Ann. U. Craiova Phys. [**21**]{}, S11 (2011)\].
G. Barnich and C. Troessaert, “Symmetries of asymptotically flat 4 dimensional spacetimes at null infinity revisited,” Phys. Rev. Lett. [**105**]{}, 111103 (2010).
G. Barnich and C. Troessaert, “BMS charge algebra,” JHEP [**1112**]{}, 105 (2011).
S. Weinberg, “Infrared photons and gravitons,” Phys. Rev. [**140**]{}, B516 (1965).
S. W. Hawking, “Breakdown of Predictability in Gravitational Collapse,” Phys. Rev. D [**14**]{}, 2460 (1976).
S. W. Hawking, “Black hole explosions,” Nature [**248**]{}, 30 (1974).
S. W. Hawking, “Particle Creation by Black Holes,” Commun. Math. Phys. [**43**]{}, 199 (1975). Erratum: \[Commun. Math. Phys. [**46**]{}, 206 (1976)\].
S. Hawking, and G. Ellis, “The Large Scale Structure of Space-Time,” (Cambridge Monographs on Mathematical Physics). Cambridge: Cambridge University Press (1973).
Malcolm Perry ,“The collection of lecture videos on black holes (DAMTP, 2006),”Centre for Applied Research in Educational Technologies, University of Cambridge. (https://www.sms.csx.cam.ac.uk/collection/661512).
A. Strominger. 2016 ,“Infrared structure of gravity and gauge theories, Physics 211r lecture videos online at Harvard University,” U.S.A Spring. (https://www.youtube.com/playlist?list=PLwLjkVy3evOazQ3FoRH-Sz8Eoxx2oriXL)
A. Strominger, “Lectures on the Infrared Structure of Gravity and Gauge Theory,” arXiv:1703.05448 \[hep-th\].
W. Israel, “Event horizons in static vacuum space-times,” Phys. Rev. [**164**]{}, 1776 (1967).
B. Carter, “Axisymmetric Black Hole Has Only Two Degrees of Freedom,” Phys. Rev. Lett. [**26**]{}, 331 (1971).
S. W. Hawking, “Black holes in general relativity,” Commun. Math. Phys. [**25**]{}, 152 (1972).
J. M. Bardeen, B. Carter and S. W. Hawking, “The Four laws of black hole mechanics,” Commun. Math. Phys. [**31**]{}, 161 (1973).
D. C. Robinson, “Uniqueness of the Kerr black hole,” Phys. Rev. Lett. [**34**]{}, 905 (1975).
J. M. Maldacena,“The Large N limit of superconformal field theories and supergravity,” Int. J. Theor. Phys. [**38**]{}, 1113 (1999) \[Adv. Theor. Math. Phys. [**2**]{}, 231 (1998)\].
A. Strominger and C. Vafa, “Microscopic origin of the Bekenstein-Hawking entropy,” Phys. Lett. B [**379**]{}, 99 (1996).
A. Ashtekar, “Asymptotic Quantization: Based on $1984$ Naples Lectures,” Naples, Italy: Bibliopolis (1987).
V. P. Frolov, “Null Surface Quantization and Quantum Field Theory in Asymptotically Flat Space-Time,” Fortsch. Phys. [**26**]{}, 455 (1978).
R. Jackiw, D. N. Kabat and M. Ortiz, “Electromagnetic fields of a massless particle and the eikonal,” Phys. Lett. B [**277**]{}, 148 (1992).
J. Jackson, “Classical electrodynamics,” Wiley, New York, NY, 3rd ed. edition, (1999).
G. ’t Hooft, “Graviton Dominance in Ultrahigh-Energy Scattering,” Phys. Lett. B [**198**]{}, 61 (1987).
T. Dray and G. ’t Hooft, “The Gravitational Shock Wave of a Massless Particle,” Nucl. Phys. B [**253**]{}, 173 (1985).
H. Fried, “Functional Methods and Models in Quantum Field Theory,” MIT Press, Cambridge, MA, 1972.
L. D. Faddeev and R. Jackiw, “Hamiltonian Reduction of Unconstrained and Constrained Systems,” Phys. Rev. Lett. [**60**]{}, 1692 (1988).
[^1]: Here, we will briefly go over the literature. For this part, although one can enumerate abundant amount of relevant works, we particularly suggest [@HawkingBook], Malcolm Perry’s lectures on black holes [@PerryLecturesDAMPT] and Strominger’s lectures [@StromingerYoutube; @Strominger13].
[^2]: Here, observe that they are in the natural unit system.
[^3]: In fact, some portion of the derivations that we will review in this part is also given as an exercise question in [@JacksonBook].
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We investigate the $p(\gamma,K^+)\Lambda$ and $p(\gamma, K^+)\Sigma^0$ reactions within a coupled-channels effective-Lagrangian method which is based on the K-matrix approach. The two-body final channels included are $\pi N$, $\eta N$, $\phi N$, $\rho N$, $\gamma N$, $K \Lambda$, and $K \Sigma$. Non-resonant meson-baryon interactions are included in the model via nucleon intermediate states in the $s$- and $u$-channels and meson exchanges in the $t$-channel amplitude and the $u$-channel resonances. The nucleon resonances $S_{11}$(1535), $S_{11}$(1650), $S_{31}$(1620), $P_{11}$(1440), $P_{11}$(1710), P$_{13}$(1720), $P_{33}$(1232), $P_{33}$(1600), D$_{13}$(1520), $D_{13}$(1700), and $D_{33}$(1700) are included explicitly in the calculations. With a single parameter set which was derived earlier from our analysis of the $\eta$ meson photoproduction, the model describes well all the available cross section and polarization data of the SAPHIR collaboration for the two investigated channels. The description of the data of the CLAS collaboration, however, is not of the same quality. In contrast to some previous studies, we do not find any compelling need for including a $D_{13}$ state with mass of around 2.0 GeV in order to reproduce the data for the $p(\gamma,K^+)\Lambda$ reaction at photon energies corresponding to the invariant mass around 1.9 GeV.'
author:
- 'R. Shyam$^{1,3}$, O. Scholten$^2$, and H. Lenske$^1$'
title: 'The associated photoproduction of $K^+$ meson off proton within a coupled-channels [$\mathsf K$]{}-matrix approach'
---
Introduction
============
Recently, several laboratories have devoted considerable amount of effort to the investigation of the photoproduction of strangeness on the nucleon. Experiments performed at JLab-CLAS [@bra06; @mcn04], ELSA-SAPHIR [@gla04; @law05] and SPring8/LEPS [@zeg03; @sum06] have produced high quality data on the associate strangeness production reactions $p(\gamma,K^+)\Lambda$ and $p(\gamma, K^+)\Sigma^0$ covering the photon energy regime from threshold to upto 3.0 GeV. Furthermore, data on single and double polarization [@bra07] observables have also become available. The motivation behind these studies has been the fact that a considerable part of the excitation spectrum of the nucleon can, in principle, participate in the production process of associated strangeness ($K^+\Lambda$ and ${\ensuremath{\mathsf K}}^+\Sigma^0$) through these reactions because even at the threshold they involve invariant masses that exceed those of several baryonic resonances. It is hoped that these studies could help in investigating the so-called missing baryonic resonances that are predicted by the quark models (see, e.g., Ref. [@cap00]) but not observed in non-strange meson photoproduction. Some of these resonances may couple strongly to the $K\Lambda$ and $K\Sigma$ channels.
The determination of the properties of the nucleon resonances (e.g., their masses, widths, and coupling constants to various decay channels) is an important issue in hadron physics. This will help in testing the predictions of lattice quantum chromodynamics (LQCD) which is the only theory that tries to calculate these properties from first principles. Even though, the requirement of computational power is enormous for their numerical realization, such calculations have started to provide results for the properties of nucleon ground as well as excited states. [@and09; @bas07; @bur06; @mat05; @lei05]. Furthermore, reliable nucleon resonance data are important for testing the “quantum chromodynamics (QCD) based” quark models of the nucleon (see, [*e.g.*]{}, [@cap00; @lor01]) and also the dynamical coupled-channels models of baryonic resonances [@lut05].
One of the major challenges of this field is the extraction of reliable information about nucleon resonance properties from the photoproduction data. In experiments, these resonances are excited as intermediate states before decaying into the final meson and baryon channels. A good description of intermediate-energy scattering is still not amenable to the LQCD calculations. Therefore, at this stage the effective methods are usually employed to describe the dynamics of the meson production reactions. Baryon resonance states are included explicitely in these approaches and their properties are obtained by comparing the predictions of the theory with the experimental data [@ben95; @arn00; @tia99; @dre99; @shy07; @jan01].
In order to determine resonance properties reliably from the experimental measurements one requires a model that can analyze the different reactions over the entire energy range using a single Lagrangian density that generates all non-resonance contributions from Born, $u$- and $t$-channel contributions without introducing new parameters. At the same time, the Lagrangian should also satisfy the symmetries of the fundamental theory (i.e. QCD) while retaining only mesons and baryons as effective degrees of freedom.
The coupled-channels method within the [$\mathsf K$]{}-matrix approach [@feu98; @kor98; @pen02; @uso05; @shk05; @uso06; @shy08] provides a way to analyze simultaneously all reaction data for a multitude of observables in different reaction channels while respecting the constraints described above. This method is attractive because it is based on an effective-Lagrangian framework that is gauge invariant and is consistent with chiral symmetry. It also provides a convenient way of imposing the unitarity constraint. This results from the Bethe-Saltpeter equation in the approximation where particles forming the loop are taken on the mass shell i.e. only the discontinuity part of the loop integral is retained. The $S$ matrix in this approach is unitary provided the [$\mathsf K$]{}-matrix is taken to be real and Hermitian.
Alternatively, the dynamical coupled channels models within the Hamiltonian formalism have also been used to describe the meson-production reactions [@dur08; @jul08; @jul07; @sag07; @jul05; @jul06; @chi01]. Isobaric models such as Kaon-Maid [@mar01a] and Saclay-Lyon [@dav96] have been utilized in Ref. [@mar01b; @mar06] to describe the $p(\gamma,K^+)\Lambda$ reaction. This reaction has also been studied within a variety of tree-level isobar models [@ade85; @ade90; @wil91; @han01] and in the quark models [@li95; @lu95; @he08]. In Ref. [@bor07] a gauge-invariant chiral unitary framework is used to describe the photoproduction data of both SAPHIR and CLAS collaborations.
The main objective of this paper is to study photoproduction reactions $p(\gamma,K^+)\Lambda$ and $p(\gamma, K^+)\Sigma^0$ for photon energies ranging from threshold to about 3 GeV in a coupled-channels formalism of Refs. [@uso05; @uso06; @shy08] which is based on the [$\mathsf K$]{}-matrix approach. This is an effective Lagrangian model which is gauge invariant and obeys the low-energy theorem. We aim at describing simultaneously the data on total and differential cross sections as well as on polarization observables for both the reactions within the same framework with a single parameter set. The $\Lambda$ and $\Sigma^0$ hyperons have isospins of 0 and 1, respectively. Thus, the intermediate states of $K^+\Lambda$ have the isospin $\frac{1}{2}$ only ($N^*$) while those of $K^+\Sigma^0$ can have both $\frac{1}{2}$ and $\frac{3}{2}$ isospins ($N^* $ and $\Delta$). Therefore, a combined description of all the available data for both these channels is indeed quite interesting and is a challenge to any theoretical model.
We would like to add that a subset of SAPHIR and CLAS data for the $p(\gamma,K^+)\Lambda$ reaction has been investigated previously in Ref. [@shk05] within the Giessen model which is also a coupled-channels effective-Lagrangian [$\mathsf K$]{}-matrix approach. Despite some differences in details (see the discussions in Ref. [@shy08]), our method is similar to that of the Giessen model. The calculations presented in Ref. [@shk05] have reproduced the $p(\gamma,K^+)\Lambda$ data well with a slight preference to the SAPHIR data. In the present work we have attempted to describe the $p(\gamma,K^+)\Sigma^0$ reaction also along with the $p(\gamma,K^+)\Lambda$ one within the same framework using the same parameter set which already makes our analysis self-contained. Nevertheless, by comparing our results with those of the Giessen model we expect to gain further insight in the mechanism of the strangeness photoproduction off proton.
Our paper is organized in the following way. An overview of our model is given in section II. This consists of a short discussion of the [$\mathsf K$]{}-matrix formalism, the model space and the channels included, the Lagrangians, and the form factors. Our results and their discussions are presented in section III. Summary and conclusions of our work are presented in section IV. Finally form of Lagrangians at various vertices are given in appendix A.
Description of the Model
========================
This work is based on an effective-Lagrangian model. The kernel in the [$\mathsf K$]{}-matrix approach is built by using the effective Lagrangian as given in Appendix A. We have taken into account contributions from (i) the nucleon Born term, (ii) $t$-channel exchanges of mesons, (iii) nucleon and resonance terms in the $u$-channel, and (iv) baryonic resonance in the $s$-channel (see Fig. 1). The sum of amplitudes (i), (ii) and (iii) is termed as the background contribution. As is discussed below, this approach allows to account for coupled-channels effects while preserving many symmetries of a full field-theoretical method.
-0.1in
\[Fig1\]
[$\mathsf K$]{}-matrix model {#sec:Kmatr}
----------------------------
The coupled-channels (or re-scattering) effects are included in our model via the [$\mathsf K$]{}-matrix formalism. In this section we present a short overview of this approach; a more detailed description can be found in Refs. [@kor98; @uso05; @sch02; @new82].
In the [$\mathsf K$]{}-matrix formalism the scattering matrix is written as $${\label{eq:T-matr}}
{\ensuremath{\mathsf T}}= \frac{{\ensuremath{\mathsf K}}}{1-i{\ensuremath{\mathsf K}}} \,.$$
It is easy to check that the resulting scattering amplitude $S=1+2i{\ensuremath{\mathsf T}}$ is unitary provided that [$\mathsf K$]{} is Hermitian. The construction in [Eq. (\[eq:T-matr\])]{} can be regarded as the re-summation of an infinite series of loop diagrams by making a series expansion, $${\ensuremath{\mathsf T}}= {\ensuremath{\mathsf K}}+ i{\ensuremath{\mathsf K}}{\ensuremath{\mathsf K}}+ i^2{\ensuremath{\mathsf K}}{\ensuremath{\mathsf K}}{\ensuremath{\mathsf K}}+ \cdots \,.$$ The product of two [$\mathsf K$]{}-matrices can be rewritten as a sum of different one-loop contributions (three- and four-point vertex and self-energy corrections) depending on Feynman diagrams that are included in the kernel [$\mathsf K$]{}. However, not the entire spectrum of loop corrections present in a true field-theoretical approach, is generated in this way and the missing ones should be accounted for in the kernel. In constructing the kernel, care should be taken to avoid double counting. For this reason we include in the kernel tree-level diagrams only \[Figs. 1(a)-1(c)\], modified with form-factors and contact terms \[Fig. 1(d)\]. The contact terms (or four-point vertices) ensure gauge invariance of the model and express model-dependence in working with form factors (see [Section \[sec:ff\]]{}). Inclusion of both $s$- and $u$-channel diagrams \[Figs 1(a) and 1(b), respectively\] in the kernel insures the compliance with crossing symmetry.
[C|x3|x3|d|d|d|d|d|d]{} $L_{IJ}$ &&& & &&& &\
[$S_{11}$]{}(1535) &1.525 &0.0 &0.6 &-0.60& &0.1 &0.0 & 2.2\
[$S_{11}$]{}(1650) &1.690 &0.030&1.0 &-0.45& &-0.1 &0.0 & -0.8\
[$S_{31}$]{}(1620) &1.630 &0.100&3.7 &-0.12& & &-0.8&\
[$P_{11}$]{}(1440) &1.520 &0.200&5.5 & 0.65& &0.0 &-2.0& 0.0\
[$P_{11}$]{}(1710) &1.850 &0.300&3.0 & 0.25& &0.0 &-3.0& 2.0\
[$P_{13}$]{}(1720) &1.750 &0.300&0.12 &-0.75&0.25 &-0.05 &0.0 & 0.12\
[$P_{33}$]{}(1230) &1.230 &0.0 &1.7 &-2.2 &-2.7 & &0.0 &\
[$P_{33}$]{}(1600) &1.855 &0.150&0.0 &-0.4 &-0.6 & &0.55&\
[$D_{13}$]{}(1520) &1.515 &0.050&1.2 & 2.6 & 2.5 &2.0 &0.0 &1.2\
[$D_{13}$]{}(1700) &1.700 &0.090&0.0 &-0.5 & 0.0 &0.0 &0.3 &-0.04\
[$D_{33}$]{}(1700) &1.670 &0.250&0.8 & 1.5 & 0.6 & &-3.0&\
To be more specific, the loop corrections generated in the [$\mathsf K$]{}-matrix procedure include only those diagrams which correspond to two on-mass-shell particles in the loop [@kon00; @kon02]. This is the minimal set of diagrams one has to include to ensure two-particle unitarity. Thus, all diagrams that are not two particle reducible (e.g. $\gamma N \to 2\pi N$ channels), are not included. In addition, only the convergent pole contributions i.e. the imaginary parts of the loop correction, are generated. The omitted real parts are important to guarantee analyticity of the amplitude and may have complicated cusp-like structures at energies where other reaction channels open. In principle, these can be included as form factors as is done in the dressed [$\mathsf K$]{}-matrix procedure [@kon00; @kor03]. We have chosen to work with purely phenomenological form factors for the reason of simplicity. An alternative procedure to account for the real-loop corrections is offered in Refs. [@sat96; @chi04; @lut02].
The strength of the [$\mathsf K$]{}-matrix procedure is that in spite of its simplicity, several symmetries are obeyed by it [@sch02]. As was already noted, the resulting amplitude is unitary provided that [$\mathsf K$]{} is Hermitian, and it obeys gauge invariance provided the kernel is gauge-invariant. In addition, the scattering amplitude complies with crossing symmetry when the kernel is crossing symmetric. This property is crucial for a proper behavior of the scattering amplitude in the low-energy limit [@kon02; @kon04]. Coupled-channels effects are automatically accounted for by this approach for the channels explicitly included into the [$\mathsf K$]{}-matrix as the final states.
As a result of this channel coupling, the resonances generate widths which are compatible with their decays to channels included in the model space. For some resonances, such as the $\Delta$ and the $S_{11}$(1535), this corresponds to their total width. Other resonances, particularly the high lying ones, may have important decay branches to states that are not included in the model basis. To account for this in our calculations, we have added an explicit dissipative part to the corresponding propagators. The magnitudes of these widths are equivalent to decay widths of the resonances to states outside of our model space.
The resonances which are taken into account in building the kernel are summarized in [Table \[tbl:resonances\]]{}. In the current work we limit ourselves to the spin-$\frac12$ and spin-$\frac32$ resonances as in this energy regime higher spin resonances are known [@shk05] to give only a minor contribution to the $K^+\Lambda$ channel. Spin-$\frac32$ resonances are included with so-called gauge-invariant vertices which have the property that the coupling to spin-$\frac12$ pieces in the Rarita-Schwinger propagator vanish [@pas00; @kon00; @pas01]. We have chosen this prescription since it reduces the number of parameters as we do not have to deal with the off-shell couplings. The effects of these off-shell couplings can be absorbed in contact terms [@pas01] which we prefer, certainly within the context of the present work.
The masses of the resonances given in [Table \[tbl:resonances\]]{} are bare masses and they thus may deviate from the values given by the Particle Data Group [@PDG]. Higher-order effects in the [$\mathsf K$]{}-matrix formalism do give rise to a (small) shift of the pole-position with respect to the bare masses. The masses of very broad resonances, in particular the $P_{11}$, are not well determined - values lying in a broad range (typically a spread of the order of a quarter of the width) give comparable results. The width quoted in [Table \[tbl:resonances\]]{} corresponds to the partial width for decay to states outside our model space. The parameters as quoted in [Table \[tbl:resonances\]]{} are mostly unchanged as compared to those presented in previous calculations within this model [@uso05; @uso06; @shy08]. The $t$-channel contributions which are included in the kernel, are summarized in [Table \[tbl:particles\]]{}.
[C|d|CC|c]{} && S$^\pi$ & I & t-ch contributions\
$\pi$ & 0.135 & 0$^-$ & 1 & (N N), ( N N)\
K & 0.494 & 0$^-$ &[$\frac12$]{}& (N K ), (N K)\
$\phi$ & 1.019 & 1$^-$ & 0 &\
$\eta$ & 0.547 & 0$^-$ & 0 & (N N)\
$\rho$ & 0.770 & 1$^-$ & 1 & (N N), (N N), (KK),\
& & & & (KK), (NK),\
& & & & (NN), (NN)\
$\omega$ & 0.781 & 1$^-$ & 0 & (NN), (NN)\
$\sigma$ & 0.760 & 0$^+$ & 0 & (NN), (NN)\
K$^*$ & 0.892 & 1$^-$ &[$\frac12$]{}& (NK), (NK),\
& & & & (KN), (KN),\
& & & & (NK)\
Model space, channels included
------------------------------
To keep the model manageable and relatively simple, we consider only stable particles or narrow resonances in two-body final states. The $\Lambda K$, $\Sigma K$, $N\phi$, $N\eta$ and $N\gamma$ are the final states of primary interest, and the $N\pi$ final state is included for its strong coupling to most of the resonances. Three-body final states, such as $2\pi N$, are not included explicitly as stated above. Their influence on widths of resonances is taken into account by assigning an additional (energy dependent) width to them [@kor98]. To investigate the effects of coupling to more complicated states, we have also included the $N\rho$ final state. As was shown in Ref. [@uso05], inclusion of the $\rho$ channel has a strong influence on the pion sector but only a relatively minor effect on $\Lambda$ and $\Sigma$ photoproduction.
The components of the kernel which couple the different non-electromagnetic channels are taken as the sum of tree-level diagrams, similar to what is used for the photon channels. For these other channels no additional parameters were introduced and they thus need no further discussion.
Form-factors & gauge restoration [\[sec:ff\]]{}
-----------------------------------------------
Without the introduction of form factors, calculations with Born terms strongly overestimates the cross section at higher energies. Although inclusion of coupled-channels effects reduces the cross section at high energies yet disagreement with the experimental data still persists. Therefore, the Born contribution will have to be quenched with form factors. There are two physical motivations for introducing form factors (or vertex functions). First of all, at high photon energies one may expect to become sensitive to the short-range quark structure of the nucleon. Because this physics is not included explicitly in our model, we can only account for it through the introduction of phenomenological vertex functions. The second reason has to do with the intermediate-range effects because of meson-loop corrections which are not generated through the [$\mathsf K$]{}-matrix formalism. Examples of these are given in Refs. [@kon00; @kor03].
In our approach as well as that of Ref. [@pen02], the form-factors are not known *a priori* and thus they introduce certain arbitrariness in the model. In the current paper we limit ourselves to dipole form-factors in $s$-, $u$-, and $t$-channels because of their simplicity, $${\label{eq:ff-dipole}}
F_m(s)=\frac{\lambda^2}{\lambda^2+(s-m^2)^2} \;,$$ where $m$ is the mass of the propagating particle and $\lambda$ is the cut-off parameter. For ease of notation we introduce the subtracted form factors $${\label{eq:ff-twiddle}}
{\ensuremath{\tilde f}}_m(s)=\frac{1-F_m(s)}{s-m^2} \,,$$ where $F_m(s)$ is normalized to unity on the mass-shell, $F_m(m^2)=1$, and ${\ensuremath{\tilde f}}_m(m^2)$ is finite.
However, only in the kaon sector we use a different functional form for the $u$-channel form-factors $${\label{eq:ff-u-channel}}
H_m(u)=\frac{u\lambda^2}{\big( \lambda^2+(u-m^2)^2 \big)m^2}.$$ The argumentation for this different choice is presented in the discussion of the $\Sigma$-photoproduction results in Ref. [@uso05]. Often a different functional form and cut-off values are introduced for the $t$-channel form factors. Although this can easily be motivated, it introduces additional model dependence and increases the number of free parameters. To limit the overall number of parameters we have taken the same cut-off value \[$\lambda=1.2\; \mbox{GeV}^2$, see [Eq. (\[eq:ff-dipole\])]{}\] for all form-factors except for the Born contributions in kaon channels where we used $\lambda=1.0\;\mbox{GeV}^2$.
Inclusion of form-factors will in general break electromagnetic gauge-invariance of the model. Therefore, a gauge-restoration procedure should be applied. In Ref. [@uso05], the implications of various gauge-restoration procedures was studied for the $\gamma p \to K \Sigma$ amplitude. It was observed that the gauge-invariance restoration procedure is model dependent which may give rise to strongly different Born contributions to the amplitude. Therefore, the choice of a procedure to be adopted is guided by its ability to describe the experimental data. It was found that the gauge-restoration procedure of Davidson and Workman [@dav01] provided the best description of the data on the $K \Sigma$ photoproduction. We have used this procedure in the present work also.
We note that fitting the pion-scattering and pion-photoproduction amplitudes fixes masses as well as pion- and photon-coupling constants for most of the resonances. This limits strongly the number of free parameters for the kaon-production channels.
Results and Discussions
=======================
Our aim in this paper is to use the data base on $p(\gamma,K^+)\Lambda$ and $p(\gamma, K^+)\Sigma^0$ reactions of both SAPHIR and CLAS collaborations to check various ingredients of our unitary coupled-channels field theoretic model of meson production in photon induced reactions on nucleons. In particular we are interested in checking to what extent a simultaneous fit to data for a multitude of observables for both the reactions can be obtained with a single set of input parameters. This is expected to provide a strong constraint on the model parameters thus reducing the model dependence to a minimum. It is also likely to highlight the role of channel couplings in various regions of photon energies because several calculations of the associated kaon production reactions have neglected these effects.
We emphasize however, that even the large experimental data base may not allow to fix the extracted parameters uniquely within the unitary coupled-channels effective-Lagrangian model [@pen02]. This is due to the fact that it is necessary to include empirical form factors in the model to regularize the amplitudes at higher energies. As mentioned before These form factors require a gauge-invariance restoration procedure which involves ambiguities. Nevertheless, confronting the model with a large data base in several reaction channels is expected to provide a means to overcome this problem.
The parameters in the model have been adjusted [@uso06] to reproduce the Virginia Tech partial wave amplitudes of Arndt et al. [@vir96]. In Ref. [@shy08] we have presented a comparison of our calculated $S$-, $P$-, and $D$-wave amplitudes for pion-nucleon scattering for isospins $I$ = 1/2 and 3/2 channels with those of the FA08 single-energy partial wave amplitudes of Ref. [@vir96]. The corresponding results for pion photoproduction and the Compton scattering are given in Ref. [@uso06]. There we noted that both real and imaginary parts of the pion-nucleon scattering amplitudes are described well although some differences start to show up at the upper limit of the energy range considered.
The data for $p(\gamma,K^+)\Lambda$ and $p(\gamma, K^+)\Sigma^0$ reactions consist of total and differential cross sections and hyperon polarizations measured at CB-ELSA (SAPHIR collaboration [@gla04]) and at JLab (CLAS collaboration [@mcn04; @bra06]) for photon energies ranging from respective thresholds to about 3 GeV. Moreover, beam asymmetry data are available for 9 photon energy beans between 1.5 GeV to 2.4 GeV from SPring8/LEPS [@zeg03; @sum06] group. These data, therefore, cover not only the entire resonance region but also the region where the background contributions are expected to be dominant.
--------------------- ---------------------
 
--------------------- ---------------------
In Fig. 2, we compare the results of our calculations for the total cross sections with the corresponding data of SAPHIR collaboration [@gla04] for $p(\gamma,K^+)\Lambda$ and $p(\gamma, K^+)\Sigma^0$ reactions. Photon energies ($E_\gamma$) range from threshold to about 2.6 GeV. The experimental cross sections for the $p(\gamma,K^+)\Lambda$ reaction show a steep rise as $E_\gamma$ increases from threshold to about 1.1 GeV. The latter corresponds to a $\gamma p$ channel total invariant mass ($W$) of $\approx$ 1.7 GeV which coincides with the masses of $S_{11}(1650)$, $P_{11}(1710)$ and $P_{13}(1720)$ resonances. The decrease in the cross sections just before the threshold for the $K^+\Sigma^0$ channel indicates a cusp due to the opening of this channel which has already been indicated in Refs. [@kai97; @uso05]. The data in this region are well reproduced by our calculations. There is also a second peak in the data at $E_\gamma \approx$ 1.5 GeV ( $W$ $\approx$ 1.9 GeV). Our calculations describe the data well also in this region.
In contrast to studies reported in Ref. [@mar06; @mar01b; @jan01; @sar05], we do not require an additional baryonic resonance $D_{13}$(1895) in order to explain the data in the second peak region. Looking at the cross sections for the $p(\gamma, K^+)\Sigma^0$ reaction (right panel) one notices that the second maximum in the $p(\gamma,K^+)\Lambda$ data is centered around the same value of $W$ where the only peak is observed in the $K^+\Sigma^0$ channel. Hence, peaks in the $p(\gamma,K^+)\Lambda$ total cross section are more likely to be the consequences of unitarity and multi-channel dynamics. This is further supported by the coupled channel analysis of this reaction as reported in Ref. [@shk05] where the second peak in the $K^+\Lambda$ total cross section data is explained as resulting from the interference between background and resonance contributions and not due to the presence of a $D_{13}$(1895) resonant state. The same conclusion was arrived also in the previous coupled-channels studies [@pen02].
On the other hand, the total cross sections of the $p(\gamma, K^+)\Sigma^0$ reaction rise smoothly from threshold to its peak at $E_\gamma \approx$ 1.45 GeV ($W \approx$ 1.9 GeV). The cross sections drop smoothly for $E_\gamma$ larger than this value. Our calculations are able to reproduce the data well in the entire region of photon energies with the exception of some far lower and some higher photon energies where the data are somewhat underestimated.
--------------------- ---------------------
 
--------------------- ---------------------
In Fig. 3, we show the contribution of various resonances and background terms to the total cross sections of $p(\gamma,K^+)\Lambda$ and $p(\gamma, K^+)\Sigma^0$ reactions as a function of $E_\gamma$. It is clear from this figure that while the background contributions dominate the $p(\gamma,K^+)\Lambda$ cross sections in the entire range of photon energies, they do so only for $E_\gamma >$ 1.5 GeV in case of the $p(\gamma, K^+)\Sigma^0$ reaction. It is interesting to note that contributions of $P_{13}(1720)$, $S_{11}(1535)$, $S_{11}(1650)$, $P_{13}(1720)$ and $D_{13}(1520)$ (depicted as $P13$, $S11$-1, $S11$-2, $D13$-1, respectively in the left panel of Fig. 3) resonances peak at about the same value of $E_\gamma$ ($\approx$ 1.1 GeV) in the $p(\gamma,K^+)\Lambda$ total cross section. In the region around this energy, the resonance contributions are comparable to those of the background terms and they combine together to constitute the structure of the first peak in the data. Furthermore, no one resonance individually dominates in this region which is in contrast to the results of Ref. [@shk05]. The contributions of the $P_{11}$(1710) resonance are very weak and are not included in our study. This is in agreement with the results Ref. [@shk05]. We have also not included a third $S_{11}$ resonance with mass and width around 1.780 GeV and 0.28 GeV, respectively which was considered in descriptions of $p(\gamma,\eta)p$ and $p(\gamma,K^+)\Lambda$ reactions in Refs. [@he08; @jul05].
--------------------- ---------------------
 
--------------------- ---------------------
The peak region in the $p(\gamma, K^+)\Sigma^0$ total cross section, on the other hand, is dominated by the contributions from the spin-$\frac{3}{2}$, isospin-$\frac{3}{2}$ $P_{33}$(1600) resonance. Apart from the background and to a lesser extent $D_{33}$(1700) terms other resonances are almost unimportant in this region. Furthermore, magnitudes of $D_{13}$(1700), $P_{11}$(1440) and $P_{11}$(1710) resonances (depicted as $D13$-2, $P11$-1 and $P11$-2, respectively in the right panel of Fig. 3) are comparatively small in the entire range of photon energies. However, for $E_\gamma$ very close to threshold the $S_{31}$(1620) resonance is most important.
For both the reactions shown in Fig. 3, we note that the total cross sections beyond 2 GeV are almost solely governed by the contributions of the background terms. In this region all resonance contributions are small and comparable to each other.
--------------------- ---------------------
 
--------------------- ---------------------
Differential cross sections (DCS) provide more valuable information about the reaction mechanism. They reflect the quantum number of the excited state (baryonic resonance) when the cross section is dominated by it. DCS include terms that weigh the interference terms of various components of the amplitude with the outgoing $K^+$ angles. Therefore, the structure of interference terms could highlight the contributions of different resonances in different angular regions. For $p(\gamma,K^+)\Lambda$ and $p(\gamma, K^+)\Sigma^0$ reactions DCS data of ELSA-SAPHIR group exist for 36 and 35 photon energy bins, respectively in the range of respective thresholds to about 2.60 GeV covering a wide range of $K^+$ center of mass (c.m.) angles [@gla04]. In the left panels of Figs. 4-6 we show comparisons of our calculations for DCS with the corresponding SAPHIR data for the $p(\gamma,K^+)\Lambda$ reaction for energy bins in the range of 0.9 GeV – 1.4 GeV, 1.4 GeV – 2.0 GeV, and 2.0 GeV – 2.6 GeV, respectively, while in the right panels the same are shown for the $p(\gamma, K^+)\Sigma^0$ reaction for photon energy bins in the range of 1.05 GeV – 1.55 GeV, 1.55 GeV – 2.15 GeV, and 2.15 GeV – 2.55 GeV, respectively.
--------------------- ---------------------
 
--------------------- ---------------------
It is seen that the differential cross sections are flat as a function of kaon angle near the respective thresholds which signifies the dominance of $S$-wave resonances near these energies. As $E_\gamma$ rises further the DCS develop a significant forward peaking which is consistent with the domination of the background terms or the interference between background and $s$-channel resonance contributions. At still higher energies, the data show a tendency of a slow rise at the extreme forward angles.
It is clear that our model describes general trends of the data well in the complete photon energy regime of the SAPHIR measurement. However, a few specific details of the data are missed for some energy bins. In the regions of 1.0–1.20 GeV and 1.45–1.8 GeV, the angular distributions at the extreme forward angles are not properly reproduced for the $p(\gamma,K^+)\Lambda$ reaction. However, those of the $p(\gamma, K^+)\Sigma^0$ reaction are well reproduced in these regions. At photon energies $>$ 2.0 GeV, the $p(\gamma,K^+)\Lambda$ cross sections show a tendency of peaking at extreme forward angles while the $p(\gamma, K^+)\Sigma^0$ data do not seem to do so. Within statistical errors our calculations are consistent with this trend of the data, although for the $K^+\Lambda$ channel the agreement is of a lesser quality as compared to that for the $K^+\Sigma^0$ one. In view of the fact that our background terms include both $K$ and $K^*$ exchange diagrams with the same couplings for both the reactions, we have obtained a reasonably good agreement with the data.
--------------------- ---------------------
 
--------------------- ---------------------
-0.1in
There are some discrepancies between the data of ELSA-SAPHIR and CLAS collaborations. The CLAS group has reported consistently larger cross sections at most kaon angles for $E_\gamma >$ 1.19 GeV. For certain forward angles and photon energy bins one notices a large difference in the SAPHIR and CLAS data for DCS. Therefore, a simultaneous description of the data of two collaborations has often been problematic for the theoretical models [@mar06; @jul06]. Our model is no exception to this. We show in Figs. 7 and 8, comparisons between our calculations and the CLAS data for differential cross sections for $p(\gamma,K^+)\Lambda$ and $p(\gamma, K^+)\Sigma^0$ reactions for 12 chosen photon energy bins. In these figures the SAPHIR data points are also given for the comparison purpose. Since, the CLAS data are given in energy and angular bins different from those of the SAPHIR one, we have chosen those bins which are nearly equal in two cases.
--------------------- ---------------------
 
--------------------- ---------------------
We note that CLAS and SAPHIR data are identical for $E_\gamma <$ 1.1 GeV for both reactions. However, at larger energies considerable difference are seen between the two data sets particularly at forward angles. The CLAS DCS data for the $K^+\Lambda$ channel are more forward peaked for $E_\gamma$ $\approx$ 1.2 - 1.4 GeV as compared to the SAPHIR data as well as our calculated cross sections. For the $K^+\Sigma^0$ case, the CLAS DCS are consistently larger than the SAPHIR ones for all angles except for the backward ones for $E_\gamma$ between 1.25 to 1.5 GeV. Our calculations are unable to reproduce this feature.
For $E_\gamma$ between 1.4 -1.7 the $K^+\Lambda$ channel CLAS data have the tendency of backward peaking as well along with the stronger forward peaking which is again in contrast to the SAPHIR data as well as our calculations. The backward peaking of the CLAS data disappears for photon energies between 2.0 - 2.6 GeV, however, stronger forward peaking still remains and our calculations are unable to reproduce it fully. For the $p(\gamma,K^+)\Sigma^0$ reaction the CLAS data show some backward peaking at photon energies between 2.2 - 2.4 GeV. At other energies the differences between the data of the two groups are less noticeable for this channel. Our calculations are in better agreement with the CLAS data for these cases.
--------------------- ---------------------
 
--------------------- ---------------------
Resonance structure in the $s$-channel should appear more clearly in the $W$ dependence of the differential cross sections at various $K^+$ angles. In Fig. 9 (Fig. 10) we show the DCS for $p(\gamma,K^+)\Lambda$ and $p(\gamma,K^+)\Sigma^0$ reactions as a function of $W$ for positive (negative) values of COS$(\theta_{K^+})$. The experimental data are from the SAPHIR collaboration. From these figures we note that overall shapes of the $W$ distributions are reproduced well by our calculations for all the angles for both the reactions.
Nevertheless, we also notice that for the case of the $K^+\Lambda$ production our calculations underestimate the data somewhat for $W$ between 1.7-2.0 GeV at forward angles \[for COS$(\theta_{K^+})$ = 0.75, 0.85, and 0.95\]. As bulk of the forward peaking is due to $t$-channel exchanges which are well understood, this leads to the suggestion that a contribution to the background is still missing. It this context supplementing the high energy parts of the $t-$channel exchanges by Regge-trajectory exchange [@gui03; @cor06; @cor07a; @cor07b; @van09] may be an interesting option. Work is in progress to include the Regge-trajectory exchange in our model.
---------------------- ----------------------
 
---------------------- ----------------------
-0.1in
For the backward angles (see Fig. 9), the data on the $p(\gamma,K^+)\Lambda$ reaction show two peaks at $W$ around 1.7 GeV and 1.9 GeV. Our calculations are able to reproduce the structure of both the peaks for all the angles except for the most backward one \[COS$(\theta_{K^+})$ = 0.85\] where the peak at $W \approx$ 1.9 GeV is underestimated. We would like to stress that unlike Ref. [@jul06], we find no need of including an additional $D_{13}$ resonance (with mass $\approx$ 1.9 GeV and width 0.316 GeV) to reproduce the data around $90^0$. The underestimation of the peak for $W$ around 1.9 GeV for COS$(\theta_{K^+})$ = 0.85 might be seen as an indication of the need to include such a resonance. However, with the inclusion of this resonance the cross section increases for other backward angles also around this value of $W$ [@mar01b] and the signs of the beam asymmetries come out to be opposite to what has been observed experimentally for this $W$ [@zeg03].
On the other hand, for the $K^+\Sigma^0$ channel, the $W$ dependence of the DCR (see Figs. 9 and 10) is reproduced very well by our calculations for all the angles \[corresponding to both negative and positive values of COS$(\theta_{K^+})$\] with the exception of one very forward $\theta_{K^+}$.
Polarization observables provide more sensitive tests of reaction models. The reason for this lies in the fact that these observables are generally very sensitive to the imaginary parts of the amplitudes which are governed by coupling to other channels via the optical theorem. Data on hyperon recoil polarization ($P_Y$, with $Y$ being $\Lambda$ or $\Sigma^0$) have been reported by both SAPHIR and CLAS collaborations [@gla04; @mcn04]. $P_Y$ is related to interferences of the imaginary parts of the resonant amplitudes with real parts of other amplitudes including those of the back ground terms. In Fig. 11 we compare the results of our calculations $P_Y$ with the corresponding data of the SAPHIR collaboration for $p(\gamma,K^+)\Lambda$ and $p(\gamma,K^+)\Sigma^0$ reactions. One notices that experimental $P_\Lambda$ tend to be positive at backward angles, nearly zero at angles around zero and negative at forward angles. In contrast, $P_{\Sigma^0}$ data show nearly opposite trend. The CLAS $P_Y$ data are used in the analyses of Refs. [@jul06; @mar01b].
---------------------- ----------------------
 
---------------------- ----------------------
Our calculations reproduce approximately the trends seen in the data. The opposite signs of the observed $P_Y$ of two channels are nearly reproduced. The agreement with data is relatively better for $E_\gamma$ below 1.6 GeV. The large positive back angle polarizations seen in the $P_\Lambda$ data are reproduced for all the photon energy bins. Similarly the large negative experimental $P_{\Sigma^0}$ at these angles are also nearly reproduced. At forward angles our model is relatively less successful in reproducing the $P_\Lambda$ data at $E_\gamma$ above 1.6 GeV. This again indicates perhaps some inadequacy of the background terms within our model.
Beam asymmetry ($\Sigma_B$) is the measure of the azimuthal anisotropy of a reaction yield relative to the linear polarization of the incoming photon. In Fig. 12 we compare the results of our calculations for this observable for both $p(\gamma,K^+)\Lambda$ and $p(\gamma,K^+)\Sigma^0$ reactions with the corresponding data taken from Ref. [@zeg03; @sum06] which are available for nine energy bins for photon energies between 1.5 GeV to 2.3 GeV. Available experimental $\Sigma_B$ are positive for both the reactions. Our calculations reproduce the data with varying degrees of success. For the $p(\gamma,K^+)\Lambda$ reaction the agreement with the data is relatively better for $E_\gamma <$ 2.0 GeV. For larger $E_\gamma$ the data are underestimated by our model. On the other hand, for the $p(\gamma,K^+)\Sigma^0$ case the data are described better for $E_\gamma >$ 1.9 GeV while at lower photon energies they are overestimate. In calculations reported in Ref. [@shk05] comparison with the data is shown for the $K^+\Lambda$ channel for one photon energy ($E_\gamma =$ 1.946 GeV) only. In the coupled-channels analysis of Ref. [@jul06], the agreement with the $\Sigma_B$ data for the $p(\gamma,K^+)\Lambda$ reactions is of the same quality as that achieved by us – there too the data are better reproduced for $E_\gamma <$ 2.0 GeV while they are underpredicted at energies larger than this. However, we emphasize that such data for both $K^+\Lambda$ and $K^+\Sigma^0$ channels have not been simultaneously analyzed with one parameter set in any other coupled-channels model.
---------------------- ----------------------
 
---------------------- ----------------------
-0.1in
Summary and Conclusions
=======================
In this paper we investigated the photoproduction reactions $p(\gamma,K^+)\Lambda$ and $p(\gamma,K^+)\Sigma^0$ within a coupled-channels effective-Lagrangian approach which is based on the [$\mathsf K$]{}-matrix method. Unitarity effects are correctly taken into account, since all important final channels (consisting of two-body systems $\pi N$, $\eta N$, $\phi N$, $\rho N$, $\gamma N$, $K \Lambda$, and $K \Sigma$) are included in the [$\mathsf K$]{}-matrix kernel. We build this kernel by using effective Lagrangians for the Born, $u$-channel, $t$-channel, and spin-$\frac{1}{2}$ and spin-$\frac{3}{2}$ resonance contributions. Thus, the background contributions are generated consistently and crossing symmetry is obeyed. The advantage of a full coupled-channel calculation is that it allows for the simultaneous calculation of observables for a large multitude of reactions with considerably fewer parameters than would be necessary if each reaction channel is fitted separately. More significantly, the implementation of unitarity ensures that the imaginary parts of the amplitudes are compatible with the cross sections for other channels.
Our model provides reasonable description of the experimental data of the SAPHIR group on total and differential cross sections as well as on hyperon polarizations for photon energies ranging from threshold to up to $\approx$ 3 GeV for both $p(\gamma,K^+)\Lambda$ and $p(\gamma,K^+)\Sigma^0$ reactions. The beam asymmetry data of the SPring8/LEPS group for the two reactions are also reasonably reproduced.
An important point about our study is that the same parameter set was used in calculating observables for both the reactions.
We made a detailed investigation of the contribution of various partial waves and showed that peaks seen in the total $p(\gamma,K^+)\Lambda$ cross sections for photon energies around 1.5 GeV (invariant mass of 1.9 GeV) are largely due to coupled-channels effects rather than the contribution of a $D$-wave resonance with mass around 1.9 GeV. A major part of cross sections of this reaction is generated via the background terms and resonances give prominent contributions only in the first peak region (for $E_\gamma$ around 1.1 GeV). On the other hand, the total $p(\gamma,K^+)\Sigma^0$ cross sections are dominated by the $P_{33}$ (1600) resonance in its peak region ($E_\gamma$ around 1.5 GeV).
The agreement between our calculations and the differential cross section data of the CLAS collaboration is not of the same quality as that seen in case of the SAPHIR data, particularly at extreme forward angles. At some backward angles the CLAS data show pronounced peaks for some photon energies but no peak is seen in the corresponding SAPHIR data or in our theoretical cross sections. Whether or not these peaks are an indication for an additional $P$ or $D$ wave resonance is a matter of debate and additional investigations. In this context, it is quite desirable to settle the issue of mutual inconsistency between the CLAS and SAPHIR data sets.
Our work shows that it is indeed possible to fit meson photoproduction data of many channels simultaneously with a single parameter set within a coupled-channels model. Further improvement of our model is however, necessary for a better description of the data at extreme forward angles in case of higher photon energies.
Acknowledgments
===============
This work has been supported by Sonderforschungsbereich/Transregio 16, Bonn–Giessen–Bochum of the German Research Foundation (DFG) and Dutch Organization for Scientific Research (NWO).
Effective Lagrangians
=====================
We list here the effective Lagrangians for various vertices. $p$, $k$, $p\prime$ and $-q$ represent four momenta of the initial nucleon, final meson, final nucleon and photon, respectively. We assume that meson momenta are directed into the vertex, so that energy momentum conservation reads as $p + k = p\prime - q $.
For the nucleon vertices the following couplings were used $${\label{eq:LagrangianN}}
\begin{aligned}
{\ensuremath{\mathcal L}}_{NN\pi} &= i g_{NN\pi}{\bar{\Psi}}_N {\frac{(\chi {{\mbox{\boldmath $\varphi$}}}_\pi + i {/\llap{$\partial$}} {{\mbox{\boldmath $\varphi$}}}_\pi /2m_N) \cdot {{\mbox{\boldmath $\tau$}}}}
{\chi+1} {\gamma_5}}\Psi_N \\
{\ensuremath{\mathcal L}}_{NN\eta} &= i g_{NN\eta}{\bar{\Psi}}_N {\frac{\chi \varphi_\eta + i {/\llap{$\partial$}} \varphi_\eta /2m_N}
{\chi+1} {\gamma_5}} \Psi_N \\
{\ensuremath{\mathcal L}}_{NN\sigma} &= -g_{NN\sigma}{\bar{\Psi}}_N \varphi_\sigma \Psi_N \\
{\ensuremath{\mathcal L}}_{NN\rho} &= -g_{NN\rho}{\bar{\Psi}}_N {\left(\gamma_\mu {{\mbox{\boldmath $\varphi$}}}_\rho^\mu
+\frac{\kappa_{\rho}}{2m_N} \sigma_{\mu\nu}\partial^\nu{{\mbox{\boldmath $\varphi$}}}_\rho^\mu
\right) \cdot {{\mbox{\boldmath $\tau$}}}} \Psi_N
\\
{\ensuremath{\mathcal L}}_{NN\omega} &= -g_{NN\omega}{\bar{\Psi}}_N {\left(\gamma_\mu \varphi_\omega^\mu
+\frac{\kappa_\omega}{2m_N} \sigma_{\mu\nu}\partial^\nu \varphi_\omega^\mu
\right)}\Psi_N \\
{\ensuremath{\mathcal L}}_{NN\phi} &= -g_{NN\phi}{\bar{\Psi}}_N {\left(\gamma_\mu \varphi_\phi^\mu
+\frac{\kappa_\phi}{2m_N} \sigma_{\mu\nu}\partial^\nu \varphi_\phi^\mu
\right)}\Psi_N \\
{\ensuremath{\mathcal L}}_{NN\gamma} &= -e {\bar{\Psi}}_N \left( \frac{1+\tau_0}{2}
\gamma_\mu A^\mu +\frac{\kappa_\tau}{2m_N}
\sigma_{\mu\nu} \partial^\nu A^\mu \right)\Psi_N \\
{\ensuremath{\mathcal L}}_{NN\gamma\varphi}&= -e\frac{g_{NN\phi}}{2m_N}{\bar{\Psi}_N} \gamma_5
\gamma_\mu[{{\mbox{\boldmath $\tau$}}}\times {{\mbox{\boldmath $\varphi$}}}]A^\mu \;.
\end{aligned}$$ The parameter $\chi$ controls the admixture of pseudoscalar and pseudovector components in the corresponding Lagrangian. Its value is taken to be 0.5. This value was obtained in our previous study of photoproduction of associated strangeness [@uso05] and has been held fixed in the study of all other reactions within our model. Nucleon spinors are depicted by $\Psi$ and meson fields by $\varphi$. The magnetic moments are represented by $\kappa$. ${\ensuremath{\mathcal L}}_{NN\gamma\varphi}$ generates the seagull or the contact term diagrams. We have followed the notations of Ref. [@bjo64].
The Lagrangians for the meson vertices are $${\label{eq:LagrangianM}}
\begin{aligned}
{\ensuremath{\mathcal L}}_{\rho\pi\pi} &= - g_{\rho\pi\pi}
{{{\mbox{\boldmath $\varphi$}}}_\rho}_\mu \cdot ({{\mbox{\boldmath $\varphi$}}}_\pi \times {\tensor\partial}^\mu
{{\mbox{\boldmath $\varphi$}}}_\pi) /2 \\
{\ensuremath{\mathcal L}}_{\gamma\pi\pi} &= e {\varepsilon}_{3ij} A_\mu
(\varphi_{\pi_i} {\tensor\partial}^\mu \varphi_{\pi j}) \\
{\ensuremath{\mathcal L}}_{\rho\gamma\pi} &= e \frac{g_{\rho\gamma\pi}}{m_\pi}{{{\mbox{\boldmath $\varphi$}}}_\pi\cdot {\left({\varepsilon}_{\mu\nu\rho\sigma}
(\partial^\rho A^\mu) (\partial^\sigma {{\mbox{\boldmath $\varphi$}}}_\rho\,^\nu) \right)}} \\
{\ensuremath{\mathcal L}}_{\omega\gamma\pi} &= e \frac{g_{\omega\gamma\pi}}{m_\pi}
\varphi_{\pi^0}{\left({\varepsilon}_{\mu\nu\rho\sigma}
(\partial^\rho A^\mu) (\partial^\sigma \omega^\nu) \right)} \\
{\ensuremath{\mathcal L}}_{\phi\gamma\pi} &= e \frac{g_{\phi\gamma\pi}}{m_\pi}
{{\mbox{\boldmath $\varphi$}}}_{\pi^0}
{\left({\varepsilon}_{\mu\nu\rho\sigma}
(\partial^\rho A^\mu) (\partial^\sigma \phi^\nu) \right)} \\
{\ensuremath{\mathcal L}}_{\phi\gamma\eta} &= e \frac{g_{\phi\gamma\eta}}{m_\pi} \varphi_\eta
{\left({\varepsilon}_{\mu\nu\rho\sigma}
(\partial^\rho A^\mu) (\partial^\sigma \varphi_\phi^\nu) \right)} \\
{\ensuremath{\mathcal L}}_{\rho\gamma\eta} &= e \frac{g_{\rho\gamma\eta}}{m_\pi} \varphi_\eta
{\left({\varepsilon}_{\mu\nu\rho\sigma}
(\partial^\rho A^\mu) (\partial^\sigma {\varphi_{\rho^0}}^\nu) \right)} \\
{\ensuremath{\mathcal L}}_{\rho\gamma\sigma} &= e \frac{g_{\rho\gamma\sigma}}{m_\rho}
(\partial^\mu \varphi_{\rho^\nu} \partial_\mu A_\nu
- \partial^\mu \varphi_{\rho^\nu} \partial_\nu A_\mu ) \\
{\ensuremath{\mathcal L}}_{\rho\rho\gamma} &= 2e \big(
A^\mu (\partial_\mu {\varphi_\rho}_\nu) \tau_0 {\varphi_\rho}^\nu
-(\partial^\nu A^\mu) {\varphi_\rho}_\nu \tau_0 {\varphi_\rho}_\mu \\
&\quad +(\partial^\nu A^\mu) {\varphi_\rho}_\mu \tau_0
{\varphi_\rho}_\nu \big)\\
{\ensuremath{\mathcal L}}_{\phi {\ensuremath{\mathsf K}}{\ensuremath{\mathsf K}}} &= -ig_{\phi {\ensuremath{\mathsf K}}{\ensuremath{\mathsf K}}}{\bar {{\mbox{\boldmath $\varphi$}}}_{\ensuremath{\mathsf K}}} {\tensor\partial}^\mu
{{\mbox{\boldmath $\varphi$}}}_{\ensuremath{\mathsf K}}) \phi_\mu\\
{\ensuremath{\mathcal L}}_{\eta{\ensuremath{\mathsf K}}^*{\ensuremath{\mathsf K}}}&= -i g_{\eta{\ensuremath{\mathsf K}}{\ensuremath{\mathsf K}}^*} {{\mbox{\boldmath $\varphi$}}}_{\ensuremath{\mathsf K}}{\tensor\partial}^\mu
\varphi_\eta {\bar {{{\mbox{\boldmath $\varphi$}}}_{{\ensuremath{\mathsf K}}^*}}}_\mu\\
{\ensuremath{\mathcal L}}_{\pi{\ensuremath{\mathsf K}}^*{\ensuremath{\mathsf K}}}& = -i g_{\pi{\ensuremath{\mathsf K}}{\ensuremath{\mathsf K}}^*} {\bar {{\mbox{\boldmath $\varphi$}}}_{\ensuremath{\mathsf K}}} {\tensor\partial}^\mu
{{\mbox{\boldmath $\varphi$}}}_\pi \cdot {{\mbox{\boldmath $\tau$}}}{{{\mbox{\boldmath $\varphi$}}}_{{\ensuremath{\mathsf K}}^*}}_\mu \\
{\ensuremath{\mathcal L}}_{\rho\pi\eta}&= -i g_{\rho\pi\eta} (\varphi_\eta {\tensor\partial}^\mu
{{\mbox{\boldmath $\varphi$}}}_\pi) {{{\mbox{\boldmath $\varphi$}}}_\rho}_\mu\\
{\ensuremath{\mathcal L}}_{{\ensuremath{\mathsf K}}^*{\ensuremath{\mathsf K}}^0\gamma}& = \frac{g_{{\ensuremath{\mathsf K}}^*{\ensuremath{\mathsf K}}\gamma}}{m_\pi}
{\bar {{\mbox{\boldmath $\varphi$}}}_{{\ensuremath{\mathsf K}}^0}}
{\left({\varepsilon}_{\mu\nu\rho\sigma}
(\partial^\rho A^\mu) (\partial^\sigma {{\mbox{\boldmath $\varphi$}}}_{{\ensuremath{\mathsf K}}^*}^\nu) \right)}\\
{\ensuremath{\mathcal L}}_{{\ensuremath{\mathsf K}}^*{\ensuremath{\mathsf K}}^\pm\gamma}& = \frac{g_{{\ensuremath{\mathsf K}}^*{\ensuremath{\mathsf K}}\gamma}}{m_\pi}
{\bar {{\mbox{\boldmath $\varphi$}}}_{{\ensuremath{\mathsf K}}^\pm}}
{\left({\varepsilon}_{\mu\nu\rho\sigma}
(\partial^\rho A^\mu) (\partial^\sigma {{\mbox{\boldmath $\varphi$}}}_{{\ensuremath{\mathsf K}}^*}^\nu) \right)} \;.
\end{aligned}$$ The coupling constants entering into Eqs. (A.1) and (A.2) together with baryon magnetic moments are listed in Table III. We have taken positive values for all primary coupling constants involving the nucleon. In particular, the sign of $g_{NK \Lambda}$ differs from its customary negative value [@cot04] However, we would like to stress that in a calculation like ours and also in many of those cited in Ref. [@cot04] this sign is undetermined. Changing the sign of all the coupling constants involving a single $\Lambda$-field leaves the calculated observables invariant since it corresponds to a sign redefinition of this field. The magnitudes of the couplings are within the broad range specified in [@cot04].
[C|C|C|C]{} g$_{NN\pi}$ & 13.47 & g$_{NN\eta}$ & 0.85\
g$_{NN\sigma}$ & 10.0 & g$_{NN\rho}$ & 4.2\
g$_{NN\omega}$ & 3.0 & g$_{NN\phi}$ & -0.0\
g$_{\Sigma\sigma\rho}$ & 33.0 & g$_{\Sigma\Lambda\rho}$&-27.0\
g$_{\rho\pi\pi}$ & 6.0 & g$_{\rho \pi \eta}$ & 0.0\
g$_{\rho\pi^0\gamma}$ & -0.12 & g$_{\rho\pi^\pm\gamma}$& -0.10\
g$_{\rho\eta\gamma}$ & -0.21 & g$_{\omega \eta\gamma}$& -0.12\
g$_{\omega\pi\gamma}$ & 0.32 & g$_{\rho\sigma\gamma}$ & 12.0\
g$_{N\Lambda {\ensuremath{\mathsf K}}}$ & 10.0 & g$_{N\Sigma K}$ & 14.5\
g$_{N\Lambda {\ensuremath{\mathsf K}}^*}$ & -3.3 & g$_{N\Sigma K^*}$ & 0.0\
g$_{\phi{\ensuremath{\mathsf K}}{\ensuremath{\mathsf K}}}$ & -4.5 & g$_{\rho {\ensuremath{\mathsf K}}{\ensuremath{\mathsf K}}}$ & -3.0\
g$_{\pi {\ensuremath{\mathsf K}}{\ensuremath{\mathsf K}}^*}$ & -3.26 & g$_{\eta {\ensuremath{\mathsf K}}{\ensuremath{\mathsf K}}^*}$ & -3.2\
g$_{{\ensuremath{\mathsf K}}^*{\ensuremath{\mathsf K}}^0\gamma}$ & 0.177 & g$_{{\ensuremath{\mathsf K}}^*{\ensuremath{\mathsf K}}^\pm\gamma}$ & -0.177\
$\kappa_p$ & 1.79 & $\kappa_n$ & -1.91\
$\kappa_\Lambda$ &-0.613 & $\kappa_{\Sigma^0}$ & 0.79\
$\kappa_{\Sigma^+}$ & 1.45 & $\kappa_{\Sigma^-}$ & -0.16\
$\kappa_{\Sigma^0 \rightarrow \Lambda \gamma}$ & -1.61 &\
For the $S_{11}$, $S_{31}$,$P_{11}$ and $P_{3,1}$ resonances the hadronic couplings are written as $$\begin{aligned}
{\ensuremath{\mathcal L}}_{\varphi NR_{1/2}} & = & -g_{\varphi NR} {\bar{\Psi}}_R [\chi{i\Gamma}
{\varphi}+(1-\chi)\frac{1}{M}\Gamma \gamma_\mu(\partial^\mu
\varphi)]
\Psi_N + {\rm H.c.},\end{aligned}$$ where $M \,=\,(m_R \,\pm\,m_N)$, with upper sign for even parity and lower sign for odd parity resonance. The operator $\Gamma$ is $\gamma_5$ and unity for even and odd parity resonances, respectively. For isovector mesons, $\varphi$ in Eq. (A.3) needs to be replaced by ${{\mbox{\boldmath $\tau$}}}\cdot {{\mbox{\boldmath $\varphi$}}}$ for isospin-$\frac{1}{2}$ resonances and by $\bf T \cdot {{\mbox{\boldmath $\varphi$}}}$ otherwise.
The corresponding electromagnetic couplings are $$\begin{aligned}
{\ensuremath{\mathcal L}}_{\gamma NR_{1/2}} & = & -eg_1 {\bar{\Psi}_R} \frac{\Gamma}{4m_N}
\sigma_{\mu\nu}\Psi_N F^{\mu\nu} + H.c.,\end{aligned}$$ where $\Psi_R$ is the resonance spinor and $F^{\mu\nu} = \partial^\mu A^\nu - \partial^\nu {\ensuremath{\mathsf A}}^\mu$. The operator $\Gamma$ is 1 for the positive parity resonance and $-i\gamma_5$ for the negative parity one.
For spin-$\frac{3}{2}$ resonances, we have used the gauge-invariant effective Lagrangians as discussed in Refs. [@kon00; @pas00; @pas01; @pas98; @luk06]. We write here the vertex functions used by us in computation involving these vertices. The resonance-nucleon-pion vertex function (e.g.) is given by $$\begin{aligned}
\Gamma_{R_{3/2} \to N\pi}^\alpha &=& {\frac{g_1}{m_\pi}}\,
\Big[ \gamma^\alpha (q \cdot p) -{p\!\!\!/} q^\alpha \Big]
[(1-\chi) + \chi {p\!\!\!/} /M_p],\end{aligned}$$ and the corresponding electromagnetic vertices are $$\begin{aligned}
\Gamma_{R_{3/2}\to N\gamma}^{\alpha \mu} &=& \Bigg{\{}
(g_2 + 2 g_1) \Big[ q^\alpha p^\mu -g^{\alpha\mu} p\cdot q \Big] +
\nonumber \\
&& g_1 \Big[ g^{\alpha\mu}{p\!\!\!/}{q\!\!\!/} - q^\alpha {p\!\!\!/}
\gamma^\mu + \gamma^\alpha(\gamma^\mu p\cdot q - p^\mu {q\!\!\!/}) \Big]
+ \nonumber \\
&& g_3 \Big[(-q^2 g^{\alpha\mu} + q^\mu q^\alpha){p\!\!\!/} +
(q^2 p^\mu - q^\mu p\cdot q) \gamma^\alpha \Big]
\Bigg{\}} \nonumber \\
&&\times \gamma_5 [(1-\chi) + \chi {p\!\!\!/} /M_p].\end{aligned}$$ Here $p$ is the four-momentum of the resonance and $q$ is that of the meson. Index $\alpha$ belongs to the spin-$\frac{3}{2}$ spinor and $\mu$ to photon. Interesting property of these vertices is that the product, $p \cdot \Gamma$ = 0, where $\Gamma$ defines the vertices on the left hand side of Eqs. (A.5) and (A.6). As a consequence, the spin-$\frac{1}{2}$ part of the corresponding propagator becomes redundant as its every term is proportional to either $p_\mu$ or $p_\nu$. Thus only spin-$\frac{3}{2}$ part of this propagator gives rise to non-vanishing matrix elements.
[99]{}
[bra06]{} R. Bradford [*et al.*]{}, Phys. Rev. C[**73**]{}, 035202 (2006), R. A. Schumacher, Private communication (2009).
[mcn04]{} J.W.C. McNabb [*et al.*]{}, Phys. Rev. C [**69**]{}, 042201 (2004).
[gla04]{} K.H. Glander [*et al.*]{}, Eur. Phys. J. A [**19**]{}, 251 (2004).
[law05]{} R. Lawall [*et al.*]{}, Eur. Phys. J. A [**24**]{}, 275 (2005).
[zeg03]{} R.G.T. Zengers [*et al.*]{}. Phys. Rev. Lett. [**91**]{}, 092001 (2003).
[sum06]{} M. Sumihama [*et al.*]{}, Phys. Rev. C [**73**]{}, 035214 (2006).
[bra07]{} R. Bradford [*et al.*]{}, Phys. Rev. C[**75**]{}, 035205 (2007)
[cap00]{} S. Capstick and W. Roberts, Prog. Part. Nucl. Phys. [**45**]{}, 241 (2000) and references therein.
[and09]{} Andre Walker-Loud, Huey-Wen Lin, D.G. Richards, R.G. Edwards, M. Engelhardt, G.T. Flemming, Ph. Hagler, B. Musch, M.F. Lin, Harvey B. Meyer, John W. Negele, A.V. Pochinsky, Massimiliano Procura, Sergey Syritsyn, C.J. Morningstar, K. Orginos, D.B. Renner, W. Schroers, Phys.Rev.D [**79**]{}, 054502 (2009).
[bas07]{} S. Basak, R. G. Edwards, G. T. Fleming, K. J. Juge, A. Lichtl, C. Morningstar, D. G. Richards, I. Sato, S. J. Wallace, Phys.Rev.D [**76**]{}, 074504 (2007).
[bur06]{} T. Burch, C. Gattringer, L. Ya. Glozman, C. Hagen, D. Hierl, C. B. Lang, and A. Schäfer, Phys. Rev. D. [**74**]{}, 014504 (2006).
[mat05]{} N. Mathur, Y. Chen, S.J. Dong, T. Draper, I. Horváth, F.X. Lee, K.F. Liu, and J.B. Zhang, Phys. Lett. [**B605**]{}, 137 (2005).
[lei05]{} D. B. Leinweber, W. Melnitchouk, D. G. Richards, A. G. Williams, and L. M. Zanotti, in [*Lattice Hadron Physics*]{}, edited by A. Kalloniatis, D. Leinweber, A. Williams (Springer, Berlin, 2005), P. 71; J. M. Zanotti, B. Lasscock, D. B. Leinweber and A. G. Williams, Phys. Rev. [**D71**]{}, 034510 (2005); R.D. Young, D. B. Leinweber, and A. W. Thomas, Phys. Rev. [**D71**]{}, 014001 (2005).
[lor01]{} U. Löring, K. Kretzschmar, B. C. Metsch, H. R. Petry, Eur. Phys. J A [**10**]{}, 395 (201).
[lut05]{} M. F. M. Lutz and E. E. Kolomeitsev, Nucl. Phys. [**A755**]{}, 29 (2005); J. Hofmann and M. F. M. Lutz, Nucl. Phys. [**A763**]{}, 90 (2005); J. Hofmann and M. F. M. Lutz, Nucl. Phys. [**A776**]{}, 17 (2006)
[ben95]{} M. Benmerrouche, N. C. Mukhopadhyay, J. F. Zhang, Phys. Rev. D [**51**]{}, 3237 (1995); N. C. Mukhopadhyay and N. Mathur, Phys. Lett. [**B444**]{}, 7 (1998); R. M. Davidson, N. Mathur, N. C. Mukhopadhyay, Phys. Rev. C [**62**]{}, 058201 (2000).
[arn00]{} R. Workman, R. A. Arndt, I. I. Strakovsky, Phys. Rev. C [**62**]{}, 048201 (2000); R. A. Arndt, W. J. Briscoe, I. I. Strakovsky, R. L. Workman, Phys. Rev. C [**66**]{}, 055213; Ya. I. Azimov, R. A. Arndt, I. I. Strakovsky, and R. L. Workman, Phys. Rev. C [**68**]{}, 045204 (2003); R. A. Arndt, W. J. Briscoe, I. I. Strakovsky, and R. L. Workman, Phys.Rev. C [**72**]{}, 058203 (2005).
[tia99]{} L. Tiator, D. Drechsel, G. Knöchlein, and C. Bennhold, Phys. Rev. C [**60**]{}, 035210 (1999).
[dre99]{} D. Drechsel, O. Hanstein, S. S. Kamalov, L. Tiator, Nucl. Phys. [**A 645**]{}, 145 (1999)
[shy07]{} R. Shyam, Phys. Rev. C [**75**]{}, 055201 (2007); R. Shyam, Phys. Rev. C [**60**]{}, 055213 (1999); R. Shyam, G. Penner and U. Mosel, Phys. Rev. C [**63**]{}, 022202(R) (2001); R. Shyam, Phys. Rev. C [**73**]{}, 035211 (2006).
[jan01]{} Stijn Janssen, Jan Ryckebusch, Dimitri Debruyne, and Tim Van Cauteren, Phys. Rev. C [**65**]{}, 015201 (2001).
[feu98]{} T. Feuster and U. Mosel, Phys. Rev.C [**58**]{}, 457 (1998); T. Feuster and U Mosel, Phys. Rev.C [**59**]{},460 (1999).
[kor98]{} A. Yu. Korchin, O. Scholten, and R. G. E. Timmermans, Phys. Lett. [**B438**]{}, 1 (1998).
[pen02]{} G. Penner and U. Mosel, Phys. Rev. C [**66**]{}, 055211 (2002); [**66**]{}, 055212 (2002); G. Penner, Ph.D. thesis (in English), Universität Giessen, 2002, available at the URL http://theorie.physik.uni-giessen.de
[uso05]{} A. Usov and O. Scholten, Phys. Rev. C[**72**]{}, 025205 (2005).
[shk05]{} V. Shklyar, H. Lenske, and U. Mosel, Phys. Rev. C [**72**]{}, 015210 (2005).
[uso06]{} A. Usov and O. Scholten, Phys. Rev. C[**74**]{}, 015205 (2006).
[shy08]{} R. Shyam and O. Scholten, Phys. Rev. C [**78**]{}, 065201 (2008).
[dur08]{} J. Durand, B. Julia-Diaz, T.-S. H. Lee, B. Saghai, T. Sato, Phys.Rev. C [**78**]{}, 025204 (2008).
[jul08]{} B. Julia-Diaz, T.-S. H. Lee, A. Matsuyama, T. Sato, L. C. Smith, Phys. Rev. C [**77**]{}, 045205 (2008).
[jul07]{} B. Julia-Diaz, T.-S. H. Lee, A. Matsuyama, T. Sato, Phys. Rev. C [**76**]{}, 065201 (2007).
[sag07]{} B. Saghai, J.-C. David, B. Julia-Diaz, and T.-S. H. Lee, Eur. Phys. J. A [**31**]{}, 512 (2007).
[jul05]{} B. Julia-Diaz, B. Saghai, F. Tabakin, W.-T. Chiang, T.-S. H. Lee, and Z. Li, Nucl. Phys. [**A755**]{}, 463c (2005). [jul06]{} B. Julia-Diaz, B. Saghai, T.-S. H. Lee, and F. Tabakin, Phys. Rev. C [**73**]{}, 055204 (2006).
[chi01]{} Wen-Tai Chiang, F. Tabakin, T.-S. H. Lee, and B. Saghai, Phys. Lett. [**B517**]{}, 101 (2001).
[mar01a]{} T. Mart, C. bennhold, H. Haberzettl, and L. Tiator, http://www.kph.uni-mainz.de/kaonmaid.html.
[dav96]{} J. C. David, C. Fayard, G.-H. Lamot, B. Saghai, Phys. Rev. C [**53**]{}, 2613 (1996).
[mar01b]{} T. Mart and C. Bennhold, Phys. Rev. C [**61**]{}, 012201(R) (1999).
[mar06]{} T. Mart and A. Sulaksono, Phys. Rev. C [**74**]{}, 055203 (2006); P. Bydzovsky and T. Mart, Phys. Rev. C [**76**]{}, 065202 (2007).
[ade85]{} R. A. Adelseck, C. Bennhold, and L. E. Wright, Phys. Rev. C [**32**]{}, 1681 (1985).
[ade90]{} R. A. Adelseck, and B. Saghai, Phys. Rev. C [**42**]{}, 108 (1990).
[wil91]{} R. A. Williams, C. R. Ji, and S. R. Cotanch, Phys. Rev. C [**43**]{}, 452 (1991).
[han01]{} B. S. Han, M. K. Cheoun, K. S. Kim, and I. T. Cheon, Nucl. Phys. [**A691**]{}, 713 (2001).
[li95]{} Z. Li, Phys. Rev. C [**52**]{}, 1648 (1995).
[lu95]{} D. Lu, R. H. Landau, and S. C. Phatak, Phys. Rev. C [**52**]{}, 1662 (1995).
[he08]{} J.He, B. Saghai, Z. Li, Phys. Rev. C [**78**]{}, 035204 (2008).
[bor07]{} B. Borasoy, P. C. Bruns, U.-G. Meissner and R. Nissler, Eur. Phys. J. A [**34**]{}, 161 (2007).
[sch02]{} O. Scholten, S. Kondratyuk, L. Van Daele, D. Van Neck, M. Waroquier and A.Yu. Korchin, Acta Phys. Pol.[**B33**]{}, 847 (2002).
[new82]{} R.G. Newton, *Scattering theory of Waves and Particles* (Springer, New York, 1982).
[kon00]{} S. Kondratyuk and O. Scholten, Nucl. Phys. [**A677**]{},396 (2000); S. Kondratyuk and O. Scholten, Phys. Rev. C [**62**]{}, 025203 (2000).
[kon02]{} S. Kondratyuk and O. Scholten, Phys. Rev. C [**65**]{}, 038201; [*ibid.*]{} Phys. Rev. C [**64**]{}, 024005 (2001)
[kor03]{} A.Yu. Korchin and O.Scholten, Phys. Rev. C [**68**]{}, 045206 (2003).
[sat96]{} T. Sato and T.-S. H. Lee, Phys. Rev. C [**54**]{}, 2660 (1996)
[chi04]{} W. T. Chiang, B. Saghai, F. Tabakin, T.-S. H. Lee, Phys. Rev. C [**69**]{}, 065208 (2004).
M.F.M. Lutz and E.E. Kolomeitsev, Nucl. Phys. [**A700**]{}, 193 (2002).
[kon04]{} S. Kondratyuk, K. Kubodera, F. Myhrer and O.Scholten, Nucl. Phys. [A736]{}, 339 (2004).
[pas00]{} V. Pascalutsa, Nucl. Phys. [**A680**]{}, 76 (2000).
[pas01]{} V. Pascalutsa, Phys. Lett. [**B503**]{}, 85 (2001).
[PDG]{} W. M. Yao [*et al.*]{} ( Particle Data Group), J. Phys. G: Nucl. Part. Phys. [**33**]{}, 1 (2006).
[dav01]{} R. M. Davidson and R. Workman, Phys. Rev. C [**63**]{}, 025210 (2001).
[vir96]{} Virginia Tech SAID data base, see the URL http://gwdac.phys.hwu.edu/, FA08 solution R. A.Arndt, I. I. Strakovsky, and R. L. Workman, Phys. Rev. C [**53**]{}, 430 (1996); R. A. Arndt, W. J. Briscoe, I. I. Strakovsky and R. L. Workman, Phys. Rev. C [**74**]{}, 045205 (2006).
[kai97]{} N. Kaiser, T. Waas and W. Weise, Nucl. Phys. [**A612**]{}, 297 (1997).
[sar05]{} A. V. Sarantsev, V. A. Nikonov, A. V. Anisovich, E. Klempt, and U. Thoma, Eur. Phys. J. A [**25**]{}, 441 (2005).
[gui03]{} M. Guidal, J. M. Laget, M. Vanderhaeghen, Nucl. Phys. [**A627**]{}, 645 (1997).
[cor06]{} T. Corthals, J. Ryckebusch, and T. Van Cauteren, Phys. Rev. C [**73**]{}, 045207 (2006).
[cor07a]{} T. Corthals, T. Van Cauteren, J. Ryckebusch, and D. G. Ireland, Phys. Rev. C [**75**]{}, 045204 (2007).
[cor07b]{} T. Corthals, T. Van Cauteren, P. Vancraeyveld, J. Ryckebusch, and D. G. Ireland, Phys. Lett. [**B 656**]{}, 186 (2007).
[van09]{} P. Vancraeyveld, L. De Cruz, J. Ryckebusch, and T. Van Cauteren, Phys. Lett. [**B681**]{}, 428 (2009).
[bjo64]{} J. D. Bjorken and S. D. Drell, [*Relativistic Quantum Mechanics*]{} (McGraw-Hill, New York, 1964).
[cot04]{} I.J. General and S.R. Cotanch, Phys. Rev. C [**69**]{}, 035202 (2004).
[pas98]{} V. Pascalutsa, Phys. Rev. D [**58**]{}, 096002 (1998).
[luk06]{} L. Jahnke and S. Leupold, Nucl. Phys. [**A778**]{}, 63 (2006).
| {
"pile_set_name": "ArXiv"
} |
---
author:
- Sascha Trippe
date: 'Received ; accepted '
title: ' AGN Broad Line Regions Scale with Bolometric Luminosity[^1] '
---
Introduction \[sec:intro\]
==========================
Supermassive black holes, with masses ranging from millions to billions of solar masses, are components of probably all galaxies [e.g., @fletcher2003; @ferrarese2005]. Accretion of gas onto such black holes gives rise to the phenomenon of active galactic nuclei (AGN) which in turn are able to influence the evolution of their host galaxies via “AGN feedback” [e.g., @fabian2012]. Accordingly, knowledge of the masses of supermassive black holes is of great astrophysical importance.
For nearby galaxies, black hole masses $M$ can be derived from the dynamics of stars [e.g., @bender2005; @oh2009], of gas [e.g., @walsh2013], or of interstellar masers [e.g., @herrnstein2005]. For distant AGN, black hole masses can be derived from spectroscopy of broad optical/ultraviolet emission lines. Such mass estimates make use of a virial relation of the form $$M = f\,\frac{v^2 R}{G}
\label{eq:virial}$$ where $v$ is the root-mean-squared line-of-sight gas velocity, $R$ is the (effective) radius of the broad line region, $G$ is Newton’s constant, and $f$ is a geometry factor of order unity. The velocity $v$ is derived from the widths of the emission lines. The radius $R$ can be derived via reverberation mapping [@peterson1993]. As radii derived from reverberation mapping are available only for a rather small number of AGN, the monochromatic luminosity of the AGN continuum is commonly used as a proxy for the radius. The virial mass estimator then takes the form $$M = \kappa\,v^2\,(\lLl)^{\alpha}
\label{eq:mass}$$ where $\Ll$ is the specific continuum luminosity at wavelength $\lambda$, $\kappa$ is a constant to be calibrated, and $\alpha\approx0.5$ [@kaspi2007; @bentz2009]. The monochromatic luminosities $\lLl$ are estimated from the continuum close to the emission line from which $v$ is derived. Estimators of this type, for various lines, have been used successfully for deriving black hole masses for large samples of AGN [@greene2005; @vestergaard2006; @kim2010; @park2013; @jun2015].
Currently used mass estimators are based on the expectation that the best proxy for the radius of the broad line region for a given line is $R\propto(\lLl)^{\alpha}$. A priori however, this is an ad-hoc assumption that requires justification, and alternative scalings might be possible. In the following, I thus explore the hypothesis that the sizes of broad line regions do not scale with $\lLl$ but rather with *bolometric* luminosity.
Color Effects \[sec:color\]
===========================
I begin with the assumption that the correct mass estimator based on an optical/ultraviolet emission line $i$ is $$M_i = \kappa_i\,v^2_i\,L^{\alpha}
\label{eq:mass0}$$ where $L$ is the *bolometric* luminosity of the optical/ultraviolet AGN continuum and $\alpha\approx0.5$ for all lines. At least in some cases, we can analyze the same target using two lines $a, b$, resulting in two mass estimates $M_a, M_b$. Provided that both mass estimators have been calibrated correctly, the results for $M_a$ and $M_b$ should be identical (within errors) for any given target, meaning $$\label{eq:mass0ratio}
\frac{M_a}{M_b} = \frac{\kappa_a}{\kappa_b} \left(\frac{v_a}{v_b}\right)^2 = 1 \ .$$ If we replace the bolometric luminosity $L$ by the monochromatic luminosity derived from a narrow band centered on a wavelength $\lambda$ located close to the emission line to be analyzed, $\lLl$, we find a modified mass estimator $$M'_i = \kappa_i\,v^2_i\,\xi_i\,L_i^{\alpha}
\label{eq:mass1}$$ with $L_i\equiv(\lLl)_i$ for convenience and $\xi_i$ being another constant. When analyzing the same target using two different lines $a, b$, we find $$\label{eq:mass1ratio}
\frac{M'_a}{M'_b} = \frac{\kappa_a}{\kappa_b} \left(\frac{v_a}{v_b}\right)^2 \frac{\xi_a}{\xi_b} \left(\frac{L_a}{L_b}\right)^{\alpha} .$$ Applying Equation (\[eq:mass0ratio\]), we can reduce Equation (\[eq:mass1ratio\]) to $$\label{eq:mass1ratio2}
\frac{M'_a}{M'_b} = \frac{\xi_a}{\xi_b} \left(\frac{L_a}{L_b}\right)^{\alpha} = \xi_{ab} \left(\frac{L_a}{L_b}\right)^{\alpha}$$ using here $\xi_{ab}\equiv(\xi_a/\xi_b)$ for convenience.
Assuming that (i) both mass estimators $M_{a,b}$ have been calibrated independently and using sufficiently representative AGN samples, and (ii) the ratio $M'_a/M'_b$ can be derived for a sufficiently large and representative sample of AGN, the *ensemble average* of the mass ratio, $\langle M'_a/M'_b\rangle$, will be unity (within errors) by construction: $$\label{eq:mass1meanratio}
\left\langle\frac{M'_a}{M'_b}\right\rangle = \left\langle \xi_{ab} \left(\frac{L_a}{L_b}\right)^{\alpha} \right\rangle = \xi_{ab} \left\langle \left(\frac{L_a}{L_b}\right)^{\alpha} \right\rangle = 1 \ .$$ This relation implies $\xi_{ab}=\langle(L_a/L_b)^{\alpha}\rangle^{-1}$, at least if the values $M'_a/M'_b$ are derived for a sufficiently large and representative sample. In case of small samples (as frequently the case), $\xi_{ab}$ needs to be estimated separately from the typical optical/ultraviolet color of the AGN continuum.
Comparison to Observations \[sec:obs\]
======================================
Arguably the most careful cross-analysis of different mass estimators is provided by @assef2011 who compare mass estimates based on the [[C]{}[iv]{}]{}$\lambda$1549 line to those based on the hydrogen Balmer lines $\lambda$6565 and $\lambda$4863 (with all wavelengths in Å). @assef2011 apply the mass estimators $$\begin{aligned}
\label{eq:assefmasses}
\frac{M'\Hbb}{\Msol} \hskip-5pt &=& \hskip-5pt 6.71\hskip-3pt\times\hskip-3pt 10^6 f \left(\frac{v\Hbb}{10^6\,{\rm m\,s^{-1}}}\right)^2 \left(\frac{\LH}{10^{37}\,{\rm W}}\right)^{0.52} \nonumber \\
\frac{M'\Haa}{\Msol} \hskip-5pt &=& \hskip-5pt 7.68\hskip-2pt\times\hskip-2pt 10^6 f \left(\frac{v\Haa}{10^6\,{\rm m\,s^{-1}}}\right)^{2.06} \hskip-2pt \left(\frac{\LH}{10^{37}\,{\rm W}}\right)^{0.52} \nonumber \\
\frac{M'\C}{\Msol} \hskip-5pt &=& \hskip-5pt 10^{\epsilon} \, \left(\frac{v\C}{10^6\,{\rm m\,s^{-1}}}\right)^2 \left(\frac{\LC}{10^{37}\,{\rm W}}\right)^{0.53} ,
$$ drawn from @greene2005, @vestergaard2006, and @bentz2009, to a sample of 12 lensed high-redshift quasars. Here $\LC$ and $\LH$ denote the continuum luminosities $\lLl$ derived at 1350Å and 5100Å, respectively; the values of the constants $f$ and $\epsilon$ depend on if $v$ is derived from the dispersion or the FWHM of a line. The relation between $M'\Haa$ and $v\Haa$ is actually $M'\Haa\propto v\Haa^{2.06\pm0.06}$, i.e., consistent with $M\propto v^2$. Likewise, the power law indices of $\LC$ and $\LH$ are consistent with 0.5 within errors.
From their analysis, @assef2011 find a systematic discrepancy between the masses derived from the Balmer lines on the one hand and from [[C]{}[iv]{}]{} on the other hand. This discrepancy follows the relation (their Equation (8)) $$\frac{M'\H}{M'\C} = 10^{-y} \left(\frac{\LC}{\LH}\right)^{-x}
\label{eq:assefratio}$$ where the subscript H denotes either or . Depending on which Balmer line and which velocity indicator (line dispersion or line FWHM) is used, best-fit values range from $0.51\pm0.14$ to $0.95\pm0.22$ for $x$ and $-0.11\pm0.06$ to $-0.27\pm0.08$ for $y$.
If indeed the broad line region radius scales with bolometric luminosity rather than with $\lLl$, i.e., if Equation (\[eq:mass0\]) gives the true underlying relation, then Equations (\[eq:mass1ratio2\]) and (\[eq:assefratio\]) are equivalent. This can be checked in a straightforward manner by identifying either or with “line $a$” and [[C]{}[iv]{}]{} with “line $b$”. Evidently, Equation (\[eq:mass1ratio2\]) predicts $\alpha=x$. Indeed, the values observed by @assef2011 for $x$ are in agreement with $\alpha\approx0.5$ within errors.
The *ensemble averaged* optical/ultraviolet continuum of quasars is known to follow the relation $\lLl\propto\lambda^{-0.56}$ in the wavelength range 1300–5500Å [@vandenberk2001]. Using $\alpha=0.5$, this implies[^2] $\xi_{ab} \approx [(5100\,{\textrm\AA}/1350\,{\textrm\AA})^{-0.56}]^{-0.5} \approx 1.45$. Comparison of Equations (\[eq:mass1ratio2\]) and (\[eq:assefratio\]) leads to the prediction $\log\xi_{ab}\approx0.16=-y$. Within errors, this is in agreement with the values observed by @assef2011.
Discussion and Conclusions \[sec:discuss\]
==========================================
From the analysis provided in Section \[sec:obs\], it is straightforward to see that the observed systematic discrepancy between mass estimates based on different emission lines is in agreement with AGN broad line regions scaling with bolometric rather than monochromatic luminosity. On the one hand, the fact that AGN broad line regions are shaped by both the optical/ultraviolet and the ionizing Lyman continuum was already noted by @peterson1993 and @baldwin1995. On the other hand, little effort has been made so far to distinguish scalings with $L$ from those with $\lLl$ observationally. It now seems that the data of @assef2011 provide the as yet most direct evidence for a scaling of the radii of AGN broad line regions with bolometric luminosity.
Comparisons of black hole masses derived from optical and ultraviolet lines usually focus on the line widths rather than the broad line region radii (cf., e.g., @ho2012 vs. @runnoe2013); this might explain why color effects have largely been overlooked so far. Even though, an independent hint might have been provided by the *Baldwin effect* [@baldwin1977], a characteristic anticorrelation between the equivalent widths of ultraviolet emission lines and continuum luminosity: whereas equivalent widths are usually discussed as function of $\lLl$, the actual underlying correlation might be the one between equivalent width and bolometric luminosity [cf., @xu2008 and references therein].
When expressed as functions of bolometric AGN luminosity (to ease comparisons of different lines), characteristic broad line region radii can be assigned to individual emission lines; for the case of vs. [[C]{}[iv]{}]{}, the radii of the former are about three times larger than the radii of the latter in general (cf. Chapter 7.1.8 of @netzer2013). Such a gas stratification is the natural consequence of the differences in ionization potential; lines requiring higher excitation or ionization energies have to be located closer to the central source of radiation. Indeed, studies of the radius–luminosity relationships of broad line regions assume (at least implicitly) that $\lLl$ is a sufficient proxy for the bolometric continuum luminosity [e.g., @kaspi2007; @bentz2009]. Such studies derive radius–luminosity relationships, as well as black hole mass–luminosity relationships, by analyzing radius or mass as function of $\lLl$ and applying global scaling factors that implicitly include an *ensemble averaged* bolometric correction. However, if broad line regions scale with bolometric AGN luminosity, bolometric corrections need to be applied *source by source*. As demonstrated by @assef2011, neglecting colors and bolometric corrections leads to differences up to a factor of four between black hole masses estimated from the Balmer lines and [[C]{}[iv]{}]{}, respectively. This also suggests that *all* emission-line based mass estimators are biased to varying degrees.
Assuming a scaling of mass estimates with bolometric luminosity, the cross-comparison of non-simultaneous virial mass estimates based on different lines is affected by the variability of the AGN luminosity. Quasars are known to vary in luminosity by factors up to about three on time scales of years [e.g., @schramm1993]; AGN variability in general obeys *red noise* statistics, meaning that stronger variations occur on longer time scales [@park2012; @park2014; @kim2013]. However, such variability cannot introduce systematic trends into samples of sources; rather, it increases the measurement error for individual sources. In general, variability effects are relatively moderate because only the square root of the luminosity enters the mass estimate. The individual mass estimates used by @assef2011 come with errors of around 0.3 dex; accordingly, variations in luminosity up to factors of about 5 would be within the measurement errors (because $\log(\sqrt{5})=0.35$).
The discussion provided in this paper, along with the one by @assef2011, illustrates once more the fact that any systematic scatter in a given relation, like the one between mass estimates based on / and [[C]{}[iv]{}]{}, is indicative of hidden parameters. Only a careful analysis of the impact of AGN continuum color eventually leads to the insight that AGN broad line regions scale with bolometric luminosity rather than $\lLl$. (Another recent example is provided by the relation between Bondi accretion rate and kinetic jet power in radio galaxies; @trippe2014.) In turn, blind application of “standard” relations despite the presence of obvious systematic scatter comes with the risk of producing results that are seriously biased.
I am grateful to [Marios Karouzos]{} (SNU Seoul) for pointing out the work by @assef2011 to me and for valuable discussion, and to an anonymous referee for an encouraging report. I acknowledge financial support from the Korean National Research Foundation (NRF) via Basic Research Grant 2012-R1A1A-2041387.
Assef, R. J., Denney, K. D., Kochanek, C. S., et al. 2011, Black Hole Mass Estimates Based on [[C]{}[iv]{}]{} Are Consistent with Those Based on the Balmer Lines, ApJ, 742, 93
Baldwin, J. A. 1977, Luminosity Indicators in the Spectra of Quasi-Stellar Objects, ApJ, 214, 679
Baldwin, J., Ferland, G., Korista, K., & Verner, D. 1995, Locally Optimally Emitting Clouds and the Origin of Quasar Emission Lines, ApJ, 455, L119
Bender, R., Kormendy, J., Bower, G., et al. 2005, HST STIS Spectroscopy of the Triple Nucleus of M 31: Two Nested Disks in Keplerian Rotation around a Supermassive Black Hole, ApJ, 631, 280
Bentz, M. C., Peterson, B. M., Netzer, H., et al. 2009, The Radius–Luminosity Relationship for Active Galactic Nuclei: The Effect of Host-Galaxy Starlight on Luminosity Measurements. II. The Full Sample of Reverberation-Mapped AGNs, ApJ, 697, 160
Fabian, A. C. 2012, Observational Evidence of Active Galactic Nuclei Feedback, ARAA, 50, 455
Ferrarese, L. & Ford, H. 2005, Supermassive Black Holes in Galactic Nuclei: Past, Present and Future Research, Space Sci. Rev., 116, 523
Fletcher, A. B. 2003, Massive Black Hole Evolution in Radio-Loud Active Galactic Nuclei, JKAS, 36, 177
Greene, J. E. & Ho, L. C. 2005, Estimating Black Hole Masses in Active Galaxies Using the Emission Line, ApJ, 630, 122
Herrnstein, J. R., Moran, J. M., Greenhill, L. J., & Trotter, A. S. 2005, The Geometry of and Mass Accretion Rate Through the Maser Accretion Disk in NGC 4258, ApJ, 629, 719
Ho, L. C., Goldini, P., Dong, X.-B., et al. 2012, Simultaneous Ultraviolet and Optical Emission-Line Profiles of Quasars: Implications for Black Hole Mass Determination, ApJ, 754, 11
Jun, H. D., Im, M., Lee, H. M., et al. 2015, Rest-Frame Optical Spectra and Black Hole Masses of $3<z<6$ Quasars, `arXiv:1504.00058`
Kaspi, S., Brandt, W. N., Netzer, H., et al. 2007, Reverberation Mapping of High-Luminosity Quasars: First Results, ApJ, 659, 997
Kim, D., Im, M., & Kim, M. 2010, New Estimators of Black Hole Mass in Active Galactic Nuclei with Hydrogen Paschen Lines, ApJ, 724, 386
Kim, J.-Y. & Trippe, S. 2013, How to Monitor AGN Intra-Day Variability at 230 GHz, JKAS, 46, 65
Netzer, H. 2013, The Physics and Evolution of Active Galactic Nuclei (Cambridge: Cambridge University Press)
Oh, S., Kim, S. S., & Figer, D. F. 2009, Mass Distribution in the Central Few Parsecs of Our Galaxy, JKAS, 42, 17
Park, J.-H. & Trippe, S. 2012, Multiple Emission States in Active Galactic Nuclei, JKAS, 45, 147
Park, J.-H. & Trippe, S. 2014, Radio Variability and Random Walk Noise Properties of Four Blazars, ApJ, 785, 76
Park, D., Woo, J.-H., Denney, K. D., & Shin, J. 2013, Calibrating [[C]{}[iv]{}]{}-Based Black Hole Mass Estimators, ApJ, 770, 87
Peterson, B. M. 1993, Reverberation Mapping of Active Galactic Nuclei, PASP, 105, 247
Runnoe, J. C., Brotherton, M. S., Shang, Z., and DiPompeo, M. A. 2013, Rehabilitating [[C]{}[iv]{}]{}-Based Black Hole Mass Estimates in Quasars, MNRAS, 434, 848
Schramm, K.-J., Borgeest, U., Camenzind, M., et al. 1993, Recent Activity in the Optical and Radio Lightcurves of the Blazar 3C 345: Indications for a ‘Lighthouse Effect’ due to Jet Rotation, A&A, 278, 391
Trippe, S. 2014, Does the Jet Production Efficiency of Radio Galaxies Control Their Optical AGN Types?, JKAS, 47, 159
Vanden Berk, D. E., Richards, G. T., Bauer, A., et al. 2001, Composite Quasar Spectra from the Sloan Digital Sky Survey, AJ, 122, 549
Vestergaard, M. & Peterson, B. M. 2006, Determining Central Black Hole Masses in Distant Active Galaxies and Quasars. II. Improved Optical and UV Scaling Relations, ApJ, 641, 689
Walsh, J. L., Barth, A. J., Ho, J. C., & Sarzi, M. 2013, The M87 Black Hole Mass from Gas-Dynamical Models of Space Telescope Imaging Spectrograph Observations, ApJ, 770, 86
Xu, Y., Bian, W.-H., Yuan, Q.-R., & Huang, K.-L. 2008, The Origin and Evolution of [[C]{}[iv]{}]{} Baldwin Effect in QSOs from the Sloan Digital Sky Survey, MNRAS, 389, 1703
[^1]: Rapid Communication
[^2]: I use here $\langle(L_a/L_b)^{\alpha}\rangle^{-1} \approx \langle L_a/L_b\rangle^{-\alpha}$. This is possible because the ratio $L_a/L_b$ is of order unity always and $0.56\alpha<1$.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We develop and discuss the properties of a new class of lattice-based avalanche models of solar flares. These models are readily amenable to a relatively unambiguous physical interpretation in terms of slow twisting of a coronal loop. They share similarities with other avalanche models, such as the classical stick–slip self-organized critical model of earthquakes, in that they are driven globally by a fully deterministic energy loading process. The model design leads to a systematic deficit of small scale avalanches. In some portions of model space, mid-size and large avalanching behavior is scale-free, being characterized by event size distributions that have the form of power-laws with index values, which, in some parameter regimes, compare favorably to those inferred from solar EUV and X-ray flare data. For models using conservative or near-conservative redistribution rules, a population of large, quasiperiodic avalanches can also appear. Although without direct counterparts in the observational global statistics of flare energy release, this latter behavior may be relevant to recurrent flaring in individual coronal loops. [[This class of models could provide a basis for the prediction of large solar flares]{}]{}.'
author:
- 'Antoine Strugarek ,Paul Charbonneau ,Richard Joseph , Dorian Pirot$^{1}$'
title: Deterministically Driven Avalanche Models of Solar Flares
---
Reconnection Avalanches and Parker’s Nanoflare Hypothesis
=========================================================
Identifying the physical mechanism responsible for coronal heating is still standing as one of the grand challenges of contemporary solar physics [see @Klimchuk:2006ka; @Aschwanden:2007kq; @DePontieu:2011fk for different theoretical and observational viewpoints]. Nearly a quarter of a century ago, @Parker:1983hf began exploring in earnest the possibility that coronal heating could be due to thermal energy released by small magnetic reconnection events occurring continuously throughout the magnetized corona [see @Parker:1983gr; @Parker:1988ia; @Priest:2002hx and references therein for related physical scenarii]. His physical picture for nanoflare production is now well-known. Photospheric fluid motions associated with turbulent convection randomly shuffle the magnetic footpoints of coronal loop leading to a braiding of fieldlines throughout the volume of the loop, with which is associated an inexorable buildup of current sheets, as individual fieldlines try to return to dynamical equilibrium under the topological constraints imposed by flux-freezing. As these localized current systems build up, plasma instabilities eventually set in and lead to energy release through magnetic reconnection. Parker dubbed these elementary energy release events “nanoflares”, went on to argue that they occur continuously throughout the corona, and conjectured that they are responsible for coronal heating.
Intense efforts have been expended to observationally verify or contradict Parker’s conjecture. This is no easy task, as it involves measuring the energy output of flares very close to the detection limit of the best current solar imagers in space. On the other hand, decades of flare observations have shown that the probability distribution function \[$f(E)$\] for flare energy \[$E$\] takes the form of a power law \[$f(E)\propto E^{-\alpha}$\] spanning some eight orders of magnitude in energy, with $\alpha$ estimated in the range $1.65$ – $1.95$ [*e.g.* @Aschwanden:2000gc; @Aschwanden:2002fl; @Aschwanden:2011hw and references therein]. This is too low for small flares to dominate energy input into the corona (requiring $\alpha>2$), unless the PDF steepens abruptly below the current detection limit. This is possible but increasingly unlikely as instruments such as the EUV imagers on the *Transition Region And Coronal Explorer* (TRACE) and *Solar Dynamics Observatory* (SDO) have pushed the detection limit right up to the boundary of the nanoflare regime. Another possibility is that physical models used to infer thermal energy release from EUV or soft X-ray fluxes are in error, due to erroneous assumptions regarding the intrinsic geometry of the flaring volume [@McIntosh:2001jt], or of the physical conditions therein, required to infer thermal energy release from observed radiative fluxes [see @McIntosh:2000kn; @Aschwanden:2002jy; @Aschwanden:2013kb and references therein].
Whatever recent observational determination of $f(E)$ may imply for coronal heating, Parker’s scenario certainly remains a fine model for solar flares in general. Although not originally emphasized by Parker, his nanoflare scenario contains all ingredients now understood to be required to lead to self-organized criticality [hereafter SOC; see, *e.g.* @Bak:1987ef; @Jensen:1998ww]: an open dissipative system – here a coronal loop – slowly driven – here by random footpoint motions – and subjected to a self-limiting local threshold instability – here magnetic reconnection. Despite the slow and gradual loading of energy, such a physical system will release energy in an intermittent manner, in the form of avalanches that can span up to the whole system.
@Lu:1991hp have proposed an avalanche-type model for solar flares that has become a reference model in the solar context, although numerous variations have now been developed over the intervening years. The present article describes one such variation that (we believe) is particularly interesting in that i) it is readily amenable to (relatively) unambiguous physical interpretation in terms of twisting of a coronal loop, and ii) it can yields qualitatively distinct flaring behavior as the degree of conservation is allowed to vary. Section \[sec:newmod\] briefly describes the original Lu and Hamilton model [more specifically, the version put forth by @Lu:1993cy hereafter the “LH model”], along with the various manners in which we have altered the forcing and redistribution rules. Section \[sec:results\] provides modelling results, comparing and contrasting the statistical properties of avalanches in the LH model and our variations thereof. We conclude in Section \[sec:conclusion\] with a discussion comparing and contrasting our models with other classes of globally and deterministically driven avalanches models developed in different physical contexts, outlining some possible model improvements, and by summarizing the implications of our modelling results for coronal heating (and more specifically with regards to the flaring behavior of small coronal loops) as well as solar flares predictions.
A Class of Avalanche Models With Deterministic Driving {#sec:newmod}
======================================================
The Lu and Hamilton Model \[ssec:LH\]
-------------------------------------
Following in part @Kadanoff:1989ki, @Lu:1991hp developed a SOC avalanche model for solar flares that by now has become a kind of “standard” [for a review, see @Charbonneau:2001fv]. The original Lu and Hamilton model is a cellular automaton defined over a 3D regular cartesian grid with nearest neighbour connectivity (top+down+right+left+front+back) over which a vector field \[${{\mathbfit A}}$\] is defined. Here we will consider instead a 2D version of the model, where a scalar nodal quantity \[$A_{i,j}^n$\] is defined over the lattice. @Robinson:1994cf has shown that for the type of driving used in the Lu and Hamilton model, the use of a vector nodal variable yields results identical (statistically) to the use of a scalar variable. Also, strictly speaking the term “cellular automaton” should be restricted to lattice models where the nodal variable is discrete (*i.e.* integers mapping to a finite number of possible nodal states); in cases where the nodal variable is continuous, one is dealing with a “coupled map lattice”. We retain “cellular automaton” here, because it has become common usage in this context. The superscript $n$ is a discrete time index, and the subscript pair \[$i,j$\] identifies a single node on the 2D lattice. Keeping $A=0$ on the lattice boundaries, the cellular automaton is driven by adding one small increment in $A$ per time step, at some randomly selected node that changes from one time step to the next. A deterministic stability criterion is defined in terms of the local curvature of the field at node $(i,j)$: $$\label{eq:stab1}
\Delta A_{i,j}^n\equiv A^n_{i,j}
-\frac{1}{4}\sum_kA_k^n \, ,$$ where the sum runs over the four nearest neighbours at nodes $(i,j\pm 1)$ and $(i\pm 1,j)$. If this quantity exceeds some pre-set threshold \[$Z_c$\] then an amount of nodal variable \[$Z$\] is redistributed to the same four nearest neighbours according to the following discrete, deterministic rules: $$\label{eq:redis1}
A_{i,j}^{n+1}= A_{i,j}^n-\frac{4}{5} Z~,$$ $$\label{eq:redis2}
A_{i\pm 1,j\pm 1}^{n+1}= A_{i\pm 1,j\pm 1}^n+\frac{1}{5} Z~,$$ with $Z=\rm{sign}(\Delta A)Z_c$. Following this redistribution, it is possible that one of the nearest neighbour nodes now exceeds the stability threshold. The redistribution process begins anew from this node, and so on in classical avalanching manner. Driving is suspended during avalanching, implicitly implying a separation of timescales between driving and avalanching dynamics, and all nodal values are updated synchronously during an avalanche to avoid introducing a directional bias in avalanche propagation.
It is readily shown that these redistribution rules, while conservative in $A$, lead to a decrease in $A^2$ summed over the five nodes involved by an amount $$\label{eq:erel1}
\Delta e_{i,j}^{n}=\frac{4}{5}\left(2{\left|\Delta A_{i,j}^n\right|\over Z_{c}}-1\right)Z_{c}^2 ~,$$ with the energy released being “assigned” to the unstable node $(i,j)$. If one identifies $A^2$ with a measure of magnetic energy (more on this shortly), the total energy liberated by all unstable node at a given iteration is then equated to the energy release per unit time in the flare: $$\label{eq:elatt}
(\Delta E)^n=\sum_{\rm unstable}\Delta e_{i,j}^{n}~.$$ A natural energy unit here, used in all that follows, is the quantity of energy \[$e_0$\] liberated by a single node exceeding the stability threshold by an infinitesimal amount; setting $|\Delta A_{i,j}|\approx Z_{c}$ in Equation (\[eq:erel1\]) immediately leads to $$\label{eq:eunit}
e_0=(4/5)Z_c^2~.$$ This very simple model yields a strikingly good representation of flare statistics, namely the observed power law form (and associated exponents) of the frequency distributions of flare peak energy release \[$P$\], duration \[$T$\], and total energy release [see @Lu:1991hp; @Charbonneau:2001fv; @Aschwanden:2002fl]. However, Equations (\[eq:stab1\]) – (\[eq:redis2\]) are admittedly very far removed from the MHD equations governing the process of magnetic reconnection, but there exist plausible bridges across this conceptual chasm.
Physical Interpretation\[ssec:phys\]
------------------------------------
We adopt here a variation [@Charbonneau:2013vt] of the interpretative picture originally put forth by LH. The 2D lattice will be considered as a slice across a coronal loop, *i.e.* in cylindrical coordinates $(\varpi,\phi,z)$ (see schematic in Figure \[fig:schematic\_global\_forcing\]A), a slice in the $\varpi\phi$-plane, with the loop axis in the $z$-direction, and the lattice variable as the $z$-component of the magnetic vector potential ${{\mathbfit A}}$ such that ${{\mathbfit B}}=\nabla\times {{\mathbfit A}}$. Such a magnetic field automatically satisfies the solenoidal constraint $\nabla\cdot {{\mathbfit B}}=0$. The vector potential component \[$A_z$\] then defines the magnetic field component in the $\varpi\phi$-plane, which can be directly related to the degree of twist in the loop, *i.e.* the ratio of transversal to longitudinal field components. As argued by LH, adding small random increments of $A$ at one lattice node then corresponds to adding a small amount of twist somewhere in the loop, which fits nicely the Parker scenario outlined above. Moreover, the stability criterion becomes a threshold condition on $\nabla^2(A_z \hat {{{\mathbfit z}}})$, *i.e.* the magnitude of the longitudinal component of the electric current, which is also satisfying from the point of view of MHD stability.
This [*ansatz*]{} suffers from one significant difficulty, however: there is not necessarily a one-to-one correspondence between the squared nodal variable [[(squared vector potential)]{}]{} and magnetic energy as conventionally defined \[$\propto \int B^2{\rm d}V$\]. It was nevertheless shown [see Figure 12.6 in @Charbonneau:2013vt] that the correspondance of $A^{2}$ to $B^{2}$ is empirically almost always satisfied[[, in the sense that the variations of $A^{2}$ almost always follow the variations of $B^{2}$ in our model]{}]{}.
Deterministic Global Driving\[ssec:detdrive\]
---------------------------------------------
The Parker picture of random shuffling of a loop’s magnetic footpoint by photospheric flows makes sense if the loop is of diameter comparable to (or preferably larger) than the granular length scale, and if no larger fluid motions co-exist near the base of the loop. The granular or larger scale flow otherwise displaces the footpoints in a spatially coherent manner far removed from random shuffling. One particularly interesting form of such global forcing is a twisting of the loop’s footpoints (illustrated in Figure \[fig:schematic\_global\_forcing\]), which then propagates upwards and accumulates along the length of the loop \[$A\rightarrow B$\]. Such a twisted magnetic loop can become unstable [e.g. to a kink-type instability, $B\rightarrow C$, see @Baty:1996we; @Bareford:2010ka and references therein]. Local tangential discontinuities arise and associated current sheets appear. Magnetic reconnections release energy locally and reduce the effective stresses \[$C\rightarrow D$\], possibly restoring the loop stability \[$D\rightarrow
E$\]. [[Note also that the final stable state \[$E$\] is not necessarily twist-free; it is generally composed of a particular pattern of stresses resulting from the past reconnection history in the loop.]{}]{}
![Cartoon of a physical interpretation of the deterministic forcing and its associated avalanches in the context of a solar coronal loop. See text for a detailed description.[]{data-label="fig:schematic_global_forcing"}](fig_tubes2){width="\linewidth"}
This form of global forcing and energy release mechanism has a direct equivalent in our 2D cellular automaton. Working in cylindrical coordinates $(\varpi,\phi,z)$, with the loop axis in the $z$-direction, the total magnetic field can be written as $$\label{eq:bdef}
{{\mathbfit B}}(\varpi,\phi,t)=\nabla\times (A_z(\varpi,\phi,t){\hat {{\mathbfit z}}})+B_z(\varpi){\hat {{\mathbfit z}}}~.$$ Identifying now the nodal variable $A$ with $A_z$ sampled on the lattice, consider the following deterministic, global driving mechanism: $$\label{eq:detdrive}
A_{i,j}^{n+1}= A_{i,j}^n \times (1+\varepsilon)~,
\qquad 0<\varepsilon\ll 1~,\qquad \forall\,(i,j)~,$$ where the parameter $\varepsilon$ ($\ll 1$) is a measure of the driving rate. With $A_z=0$ on the boundaries, the SOC state is peaked at the center of the lattice, and far enough from the latticed boundaries is characterized by near-axisymmetry about the lattice center. This implies that $\nabla\times (A_z{\hat{{{\mathbfit z}}}})$ is primarily in the $\phi$-direction. The above driving mechanism thus amounts to a geometrical increase of $B_\phi$ by a factor $(1+\varepsilon)^n$ after $n$ driving iterations, while $B_z$ remains unaffected. In other words, the degree of twist within the loop increases geometrically as well. The stability criterion based on the magnitude of the longitudinal component of the electric current then becomes a natural criterion on the intensity of the current sheet(s) developing in the loop.
In addition, $B_\phi$ is increasing in direct proportion to $A_z$ everywhere in the loop, which means that associating $A^2$ summed over the lattice with magnetic energy is entirely realistic in this context. All avalanche models discussed in the remainder of this article make use of this global, deterministic driving scheme. As in the LH model, driving is interrupted during avalanching, which amounts to assuming that the driving timescale is much longer than the avalanching timescale, a reasonable assumption in the solar coronal context [for a discussion, see @Lu:1995ce].
Similar global, deterministic driving schemes have been used extensively in the context of SOC models of seismic faults, based on a cellular automaton representation of the Burridge–Knopoff stick–slip model [see @Olami:1992do; @Lise:2001du and references therein]. The main difference from our model resides in their stability criterion that applies to the nodal variable rather than, in our case, to its curvature. A global, deterministic driver has also been used by @Liu:2006cp [@VallieresNollet:2010ed; @Liu:2011jr] in their cellular automaton model of magnetospheric substorms generation in the Earth’s central plasma sheet. [[To our knowledge, @Hamon:2002hs have been the only authors to explore deterministically driven avalanche models in the context of solar flares. They used a slightly modified version of the original model of @Olami:1992do, in which the threshold triggering the avalanching behavior applies is defined on the nodal variable rather than its curvature. In the present study, we use the latter criterion, which is naturally inherited from the original LH model.]{}]{} Evidently, the form of deterministic driving embodied in Equation (\[eq:detdrive\]) precludes using the initial condition $A_{i,j}^0=0$. In what follows we typically use a lattice having reached the SOC state via the original LH model, and then switch to deterministic driving. This will typically lead to a transient phase during which the lattice reajusts to the new driving scheme, and/or other altered model components, as discussed presently. This transient phase is usually accompanied by a gradual change in the total lattice energy, so that monitoring the latter is the best way to ascertain the recovery to a statistically stationary avalanching state.
Stochastic Redistribution Rules\[ssec:storedst\]
------------------------------------------------
Directly incorporating the deterministic driving rule given by Equation (\[eq:detdrive\]) within the Lu and Hamilton model yields a deterministic cellular automaton characterized by a periodic cycle of energy loading/unloading, all energy release events having the same magnitude. To recover SOC-like behavior, some stochastic element need be reintroduced somewhere in the model.
### Random Redistribution\[sssec:rr\]
Consider the following, plausible stochastic variation on the Lu and Hamilton redistribution rules: the same amount of nodal variable \[$Z$\] is always removed from any unstable site, following the @Lu:1993cy prescription, but it is distributed randomly, rather than equally, amongst its four nearest neighbours: $$\begin{aligned}
\label{eq:redis3}
A_{i,j}^{n+1}&=& A_{i,j}^n-\frac{4}{5}Z~, \\
\label{eq:redis4}
A_{i\pm 1,j\pm 1}^{n+1}&=& A_{i\pm 1,j\pm 1}^n+{4\over 5}\left({r_k\over R}\right)Z~,\end{aligned}$$ where the $r_k$ ($k=1,...,4)$ are random deviates uniformly distributed in $[0,1]$, and $R=\sum r_k$. This scheme will be referred to as [*random redistribution*]{}. Note that here that some particular random sequences and node configurations may lead to negative energy release from the redistribution process. We avoid such unphysical events by permuting the four random numbers over the neighbouring nodes and then sometimes generate four new random numbers until a positive energy is released by the avalanche.
### Random Extraction\[sssec:re\]
Another means of introducing stochasticity is to retain equal redistribution to nearest neighbours, but let the quantity $Z$ extracted from the unstable node be drawn from a bounded distribution of uniform random deviates $$\label{eq:redis5}
Z\in [|\Delta A_{i,j}^n|-Z_c, Z_c]~,$$ while retaining Equations (\[eq:redis1\]) – (\[eq:redis2\]) for the redistribution [*per se*]{}. The lower bound in Equation (\[eq:redis5\]) ensures that the redistribution will, at worst, restore marginal stability, while the upper bound corresponds to the setting of the Lu and Hamilton model. This scheme will be referred to as [*random extraction*]{}, and, defined in the above manner, involves no adjustable parameters once $Z_c$ has been specified.
### Nonconservative Redistribution\[sssec:nc\]
A third approach is to introduce [*non-conservative*]{} redistribution rules. In all rules discussed so far – including the original Lu and Hamilton rules – redistribution is conservative, in that whatever quantity of $A$ being extracted from an unstable node ends up in the nearest neighbours, whether in equal (Equations (\[eq:redis1\]) – (\[eq:redis2\])) or unequal (Equations (\[eq:redis3\]) – (\[eq:redis4\])) proportions. This conservation property is basically inspired by the sandpile analogy, where avalanches redistribute sand grains without creating or destroying any. With a nodal variable associated with a single component of the magnetic vector potential, there exist no physical requirement for conservative redistribution of $A$; the important physical requirement, namely $\nabla\cdot{{\mathbfit B}}=0$, is already taken care of under the current interpretive scheme, see Equation (\[eq:bdef\]).
A nonconservative version of the Lu and Hamilton redistribution rules can be defined as follows: $$\label{eq:redis8}
A_{i,j}^{n+1}= A_{i,j}^n-\frac{4}{5}\,Z~,$$ $$\label{eq:redis9}
A_{i\pm 1,j\pm 1}^{n+1}= A_{i\pm 1,j\pm 1}^n+{r_0\over 5} Z~,$$ where $r_0\in[D_{\rm nc},1]$ is again extracted from a uniform distribution of random deviates with a lower bound \[$D_{\rm nc}$ ($<1$)\], such that $1-D_{\rm nc}$ is the fraction of the redistributed quantity $Z$ that is lost rather than redistributed. This rule thus involves one free parameter, namely the conservation parameter \[$D_{\rm nc}\in [0,1[$\]. A nonconservative model of this type, using fully deterministic driving, redistribution, and stability criteria, has been developed by @Liu:2006cp [@VallieresNollet:2010ed] in the context of magnetospheric substorms, and has been shown to produce avalanches with SOC-like power-law distribution for event sizes.
Stochastic Threshold\[ssec:stothresh\]
--------------------------------------
Another interesting possibility is to introduce stochasticity at the level of the stability threshold. This can be achieved, for example, by replacing the deterministic threshold rule given by Equation (\[eq:stab1\]) by extracting a value $Z_c$ anew at each node of each temporal iteration from a sharply peaked normal distribution centered on some mean value ${\bar Z_c}$. An important parameter in such a scheme is the width at half-maximum \[$\sigma$\] of the Gaussian distribution; if $\sigma$ is very small (in the sense that $\sigma/{\bar Z_c}\ll 1$, one expects behavior similar to the conventional, fixed threshold stability rule, while a very wide Gaussian turns the stability threshold into a random triggering process driven by noise.
@Lu:1993cy experimented with similar random variations of the threshold rules in the context of their stochastically driven model, but report no significant variations of the results as compared to their deterministic threshold rule. As will become apparent further below, in the context of a deterministically driven model, however, interesting effects are produced in the regime $\sigma/{\bar Z_c}\ll 1$.\
To sum up, we are displacing the stochastic element in the Lu and Hamilton model from the forcing rule to either the redistribution rule or the threshold rule. In the original Lu and Hamilton model, avalanche dynamics is fully deterministic but the driving is stochastic. The model variations outline above move the stochastic element to the avalanche dynamics, in the context of a spatially global and fully deterministic driving mechanism that can be interpreted plausibly as global twisting of a coronal loop.
Results\[sec:results\]
======================
In principle, fully deterministic avalanche models, *e.g.* the LH model with the deterministic driving defined by Equation (\[eq:detdrive\]) and a smooth initial condition, lead to a regular cycle of energy loading/unloading, with avalanches equally spaced in time and all liberating the same amount of energy. The challenge in reintroducing stochasticity in the stability and/or redistribution rules is to break this deterministic loading/unloading cycle [in this context, see also @Wheatland:1998gx and references therein].
In the preceding section we have introduced various classes of redistribution rules based on a number of stochastic elements, as well as a stochastic threshold rule. A large number of distinct avalanche models can be constructed by picking and combining this or that model element. One could, for instance, build a fixed-threshold model using nonconservative redistribution rules (Section \[sssec:nc\]) with equal distribution to nearest neighbors; or a conservative rule with random distribution to nearest-neighbors (Section \[sssec:rr\]), in conjunction with a stochastic threshold rule (Section \[ssec:stothresh\]). In addition, each such model may involve adjustable parameters (*e.g.* the width \[$\sigma$\] of the Gaussian distribution from which the threshold values are extracted). Although we have explored a vast portion of the relevant model space, in what follows we focus on a relatively small subset of such models, chosen so as to illustrate the range of behaviours possible in deterministically driven avalanche models. In order to expedite this exploration of parameter space, most model runs were carried out on a small grid ($48^{2}$), but we have also recomputed a number of runs on larger grids (up to $384^{2}$) to ensure that the results reported upon below are robust with respect to lattice size.
Exploration of model space quickly reveals that an important behavioral discriminant is the conservative property (or lack thereof) of the redistribution rule. Consequently we first discuss results for a variety of conservative models (Section \[ssec:rescons\]), and turn subsequently to their non-conservative counterparts (Section \[ssec:resncons\]). We finally considers (Section \[ssec:fss\]) the origin of the break of finite size scaling that takes place in our models.
Conservative Redistribution\[ssec:rescons\]
-------------------------------------------
Table \[tab:cons\] lists the defining rules and parameter values of a set of representative conservative avalanche models, whose properties will be discussed in the remainder of this section. Here and in all that follows, unless noted otherwise all models are run on a $48^{2}$ lattice, use the driving scheme described by Equation (\[eq:detdrive\]), and a stability threshold $Z_c=1$ (or ${\bar Z_c}=1$ for models with a stochastic threshold). A “0” entry in the second column indicates a deterministic threshold rule. The entry “random” in the third and fourth columns indicates use of Equations (\[eq:redis5\]) and (\[eq:redis3\]) – (\[eq:redis4\]), respectively. The last line in the Table lists the “ingredients” of the LH model, for comparison.
$\sigma/Z_c$ Extraction Redistribution $\alpha_E$ $\alpha_P$ $\alpha_T$
---- -------------- ------------ ---------------- ------------------ ------------------ ------------------
C1 0.0 Random $4\times 1/5$ $1.43 \pm 0.044$ $1.49 \pm 0.243$ $1.94 \pm 0.136$
C2 0.01 $Z_c$ $4\times 1/5$ $1.47 \pm 0.008$ $1.84 \pm 0.019$ $1.92 \pm 0.023$
C3 0.0 $Z_c$ Random $1.29 \pm 0.026$ $1.42 \pm 0.045$ $1.38 \pm 0.022$
C4 0.01 Random $4\times 1/5$ $1.40 \pm 0.013$ $1.56 \pm 0.087$ $1.77 \pm 0.033$
C5 0.01 $Z_c$ Random $1.29 \pm 0.015$ $1.47 \pm 0.037$ $1.38 \pm 0.011$
C6 0.0 Random Random $1.22 \pm 0.009$ $1.12 \pm 0.037$ $1.47 \pm 0.008$
C7 0.01 Random Random $1.19 \pm 0.015$ $1.08 \pm 0.034$ $1.42 \pm 0.004$
LH 0.0 $Z_c$ $4\times 1/5$ $1.48 \pm 0.005$ $1.72 \pm 0.143$ $1.63 \pm 0.023$
: Conservative avalanche models, with a driving parameter $\varepsilon =10^{-6}$. Error bars were obtained with ten different random number sequences.[]{data-label="tab:cons"}
Figure \[fig:tscons\] shows a short segment of the time series for total lattice energy (a) and energy release (b) in model C3. This is a model using deterministic extraction, random redistribution, a fixed stability threshold $Z_c$, and a deterministic driving parameter $\varepsilon=10^{-6}$. The lattice energy time series exhibits a pattern of quasiperiodic energy loading and unloading. As can be seen in Figure \[fig:tscons\]b, the unloading takes place via a sequence of large avalanches of roughly similar amplitudes and durations taking place at more or less regular time intervals. Comparison between panels c and a reveals that a small but significant fraction ($\lesssim 1 \%$) of the total lattice energy is dissipated by any one of these large avalanches, in between which lattice energy grows almost steadily in response to driving. Careful scrutiny of Figure \[fig:tscons\]b also reveals a population of much smaller avalanches interspersed between these large ones. Figure \[fig:tscons\]d is a zoom with an expanded vertical scale, spanning an interval between two large avalanches and showing the smaller ones in better detail. Unlike the quasiperiodic large avalanches, these are seen to span a wide range of amplitudes, and show no hint of any (quasi)periodicities. Even the largest avalanches in this second population only release a minuscule fraction of the lattice energy, and are hardly visible during the energy loading rising phases in Figure \[fig:tscons\]a.
![Sample time series for avalanche model C3. (a) Temporal evolution of total lattice energy. (b) Energy released by avalanches. There are two populations of avalanches here: quasiperiodic large avalanches (c), in between which occurs a population of smaller avalanches spanning a wide range of scales in size (d). Energy is measured in the elementary unit $e_0$ given by Equation (\[eq:eunit\]).[]{data-label="fig:tscons"}](fig1_C3){width="\linewidth"}
All of this suggests the presence of (at least) two dynamically distinct avalanche populations. This impression is confirmed upon computing the probability distribution functions (hereafter PDF) for total energy release, *i.e.* the probability of finding an avalanche having released a total energy within the interval $[E,E+{\rm d}E]$. This PDF, as well as the corresponding PDF for peak energy release \[$P$\] and avalanche duration \[$T$\], are shown on Figure \[fig:pdfcons\] for a set of models. The bulk part of all three PDFs is well fitted by a power-law for all models, except at the high ends of the distributions. A marked excess of large, long duration events shows up as a bump – well fitted by a Gaussian –, and for cases with random redistribution/extraction a significant depletion for very low avalanches occurs (a detailed discussion on to this particular population is given at the end of this section). It is readily verified that the bump at the lower end of the PDF corresponds to the large avalanches, while the power-law portion of the PDF is made up of all avalanches taking place in between, as in Figure \[fig:tscons\]d.
![Statistical properties of avalanches in models C1, C3, C7, and LH, as labeled. The various panels show the probability distribution functions of (a) total energy released by avalanches \[$E$\], (b) peak energy release \[$P$\], and (c) duration \[$T$\]. A close-up of the high-$E$ end of the total energy PDF is also shown. The straight lines are guides to the eye, showing the slopes associated with the mid-sized avalanches of the distributions. All PDFs are constructed from $10^8$ iterations runs, initialized from prior runs of the same models having already reached a statistically stationary state. The vertical dashed lines label the transition from small to power-law avalanches (see discussion a the end of section \[sec:results\]).[]{data-label="fig:pdfcons"}](fig2){width="\linewidth"}
Evidently, in the case of model C3, the stochasticity introduced at the level of redistribution is insufficient to break the quasiperiodic pattern of energy loading and unloading, as manifested by the population of large avalanches. Indeed, this pattern proves surprisingly difficult to break. Figure \[fig:3ts\] shows $10^5$ iterations segments of energy release time series extracted from a sequence of three avalanche models, all operating at the same deterministic driving rate $\varepsilon=10^{-6}$, but including more and more stochastic components (specifically, models C3, C6, and C7 in Table \[tab:cons\]). Comparing the three panels reveals immediately that increasing stochasticity gradually breaks the quasiperiodicity of large avalanches, while blurring the energy contrast between the populations of large and small avalanches, mostly by reducing the size of the large loading/unloading avalanches. Corresponding PDFs for the integrated energy release \[$E$\], peak energy release \[$P$\] and avalanche duration \[$T$\] are also shown on Figure \[fig:pdfcons\]. Even in the case of the strongly stochastic model C7, these PDFs (solid green histograms in Figure \[fig:pdfcons\]) still show an excess of large events as compared to a pure, scale-free power-law distribution.
![Extracts of energy release time series in models C3, C6, and C7 (panels a through c). This represents a sequence of increasing stochasticity built into the avalanche models (cf. Table \[tab:cons\]). All models use the same deterministic driving rate, namely $\varepsilon=10^{-6}$ in Equation (\[eq:detdrive\]).[]{data-label="fig:3ts"}](fig3_C367_eps1e-06){width="\linewidth"}
Further insight into this dichotomy in avalanching behavior is obtained by computing a waiting time distribution (WTD) separately for each class of avalanches. The waiting time \[$\Delta T$\] associated with the $k{\rm th}$ avalanche is defined as the time elapsed since the end of the previous avalanche and the beginning of that under consideration. The distinction between “large” and “small” avalanches is made on the basis of the PDF for peak energy release, which typically shows most clearly the bump associated with the population of large avalanches; as the PDF is scanned from the low end of the size distribution towards the high end, a point is reached where the PDF shows a local minimum (*e.g.* at $P/e_0\approx 2\times 10^2$ for model C1 in Figure \[fig:pdfcons\]b). All avalanches located right of this point are considered to belong to the population of “large” avalanches, even though they likely contain a few of the largest avalanches belonging to the second population of “small” avalanches, but there is simply no way to reliably distinguish these.
The WTDs for large and small avalanches are plotted in Figure \[fig:wtdcons\]a – b, here for C1-type models with driving rates $\varepsilon=10^{-7}$, $10^{-6}$, and $10^{-5}$. Large avalanches have a well-defined mean wait time and their WTD is well fitted by a Gaussian, while the population of smaller avalanches has an exponential WTD, indicative of triggering by a stationary random process. Increasing the driving reduces the mean wait time for large avalanches in direct proportion, while steepening the exponential WTD for the population of small avalanches. This indicates that the large avalanches are the (perturbed) signature of the energy loading/unloading process that would otherwise characterize a fully deterministic model. For the population of small avalanches, increasing the driving rate merely steepens the WTD while maintaining its exponential form. This is the same behavior observed in the LH model when the mean amplitude of the spatially random increments remains constant in time and space [on the WTD in the LH model, see also @Norman:2001co; @Wheatland:2000if].
![Waiting-time distributions for the populations of (a) large and (b) small avalanches, for three C1-type models with deterministic driving rates of $\varepsilon=10^{-7}$, $10^{-6}$, and $10^{-5}$, as labeled. The WTDs for large avalanches are well fitted with a Gaussian (solid lines on panel a), while those characterizing small avalanches have an exponential form.[]{data-label="fig:wtdcons"}](fig4_C1_eps){width="\linewidth"}
This is confirmed upon constructing scatter plots between the energy of large avalanches and the wait time elapsed since the end of the previous large avalanche, as done on Figure \[fig:Ewtd\], for a sequence of C1-type models spanning four decades in driving rates: $\varepsilon=10^{-7}$, $10^{-6}$, and $10^{-5}$. For a simple, fully deterministic load/unload model [*e.g.* of the type originally considered by @Rosner:1978fk], avalanches are periodic and all liberate the same amount of energy; all points would then coincide. For a loading/unloading model including a stochastic trigger, one would expect a strong positive correlation between avalanche energy and wait time. For the spatially extended models considered here, individual avalanches are scattered in a cloud, whose extent in the $[E,\Delta T]$ plane increases with increasing degree of stochasticity in the avalanche model. This occurs because the unfolding of a large energy-unloading avalanche is affected by the spatial distribution of nodal variable values, itself influenced by the stochastic components introduced in the redistribution or threshold rules. Nonetheless, linear proportionality between energy and wait time is clearly apparent for the higher driving rates, but degrades towards lower $\varepsilon$-values; the linear correlation coefficients are $r=0.98$, $0.91$, and $0.89$ for $\varepsilon$ falling from $10^{-5}$ to $10^{-7}$, in decadal steps (see Figure \[fig:Ewtd\]).
![Correlation between the energy of large avalanches, and the waiting time elapsed since the end of the previous large avalanche. Scatter plots are shown for C1-type models with varying driving rates \[$\varepsilon$\], as indicated. The dashed lines are a guide to the eye corresponding to a logarithmic slope of unity, indicating linear proportionality between avalanche energy and wait time, as expected in simple energy loading/unloading models (see text).[]{data-label="fig:Ewtd"}](fig4_C1_scatter_EWTD){width="\linewidth"}
The models embedding a stochastic process in their redistribution or in their extraction rule (see Table \[tab:cons\]) all exhibit a particular population of small avalanches that is not well-fitted by a power-law (*e.g.* Figure \[fig:pdfcons\]a – b). These avalanches have an energy of the order of and lower than the unit energy \[$e_{0}$\]. In the case of the LH model, one-iteration – and one-node – avalanches always release an energy of $e_{0}$ whereas in the case of random redistribution or extraction, such avalanches can release a large range of different energies. This population of small avalanches – of few iterations – appears to deviate from the classical SOC state of the LH model. They can be easily identified in Fig. \[fig:E\_vs\_T\] where we scatter avalanches as a function of their duration \[$T$\] and released energy \[$E$\] for model C1. The alteration of very small avalanches can be different from one model to the other: model C1 shows a deviation for $E < e_{0}$ while model C3 is affected up to $E\approx 10\,e_{0}$. Despite the existence of scale-dependent populations at both ends of the avalanche distribution, the total energies of mid-size avalanches are still very well fitted by a power-law over more than four orders of magnitudes (more than two orders of magnitude for the peak energy).
![\[$E$\] *vs* \[$T$\] in model C1. The three populations of avalanches (small, mid-size power law and large quasi-periodic) are well identified in three energy regions. []{data-label="fig:E_vs_T"}](fig5_EvsT_C1){width="0.5\linewidth"}
The last three columns of Table \[tab:cons\] list the power-law indices obtained from least-square fits to the portions of the PDFs of total energy, peak energy release, and duration for the populations of mid-size avalanches in the corresponding simulations (the error bars were obtained with 10 runs using different random numbers sequences). These power-law indices are rather similar to those extracted from the 2D scalar version of the LH stochastically driven avalanche model (last line in Table \[tab:cons\]). Significant differences are observed nonetheless in going from model to model, suggesting that global statistics of event sizes are dependent to some extent on model “ingredients” and parameters in these conservative models.
We have explored a wide variety of conservative models, often combining into the same model multiple stochastic elements (see models C1 through C7 in Table \[tab:cons\]). Remarkably, none of these models succeeds in completely breaking the loading/unloading cycle. While producing scale-free avalanche size distributions spanning many decades in energy, all models also end up producing a population of very large quasi-periodic events with a more-or-less well-defined mean size. Even model C7, combining a fairly broad random stability threshold, random extraction, and random redistribution, ends up doing so. This property is truly robust, and holds over a wide range of driving rates. In contrast, the PDFs for flare sizes reconstructed from solar UV or soft X-ray data shows no such population, and is instead well represented by a single power-law [see @Dennis:1985ev; @Lu:1993cy; @Aschwanden:2000gc; @Aschwanden:2002jy and references therein]. As detailed in what immediately follows, the use of nonconservative redistribution rules can alleviate this problem over a certain range in parameter space, for all classes of deterministically driven models.
Nonconservative Redistribution\[ssec:resncons\]
-----------------------------------------------
We consider once again a representative set of deterministically driven avalanche model using non-conservative redistribution, as described in Section \[sssec:nc\], and various combination of other modeling ingredients. Table \[tab:ncons\] lists the characteristics and properties of the models that are the focus of the foregoing discussion. The general format is similar to Table \[tab:cons\], and the models were run with a non-conservation parameter $D_{nc}=0.1$. Recall that the actual fraction removed is drawn from a sequence of uniform deviates spanning $[D_{nc},1]$, so that the [*average*]{} nodal dissipation is the median of this interval. We essentially discuss in this section model NC0, the other models follow naturally the properties described in this and previous sections.
$\sigma/Z_c$ Extraction Redistribution $\alpha_E$ $\alpha_P$ $\alpha_T$
----- -------------- ------------ ---------------- ------------------ ------------------ ------------------
NC0 0.0 $Z_c$ $4\times 1/5$ $1.17 \pm 0.005$ $1.25 \pm 0.097$ $1.15 \pm 0.015$
NC1 0.0 Random $4\times 1/5$ $1.12 \pm 0.015$ $1.08 \pm 0.037$ $1.22 \pm 0.007$
NC2 0.01 $Z_c$ $4\times 1/5$ $1.20 \pm 0.015$ $1.23 \pm 0.080$ $1.16 \pm 0.066$
NC3 0.0 $Z_c$ Random $1.19 \pm 0.007$ $1.25 \pm 0.116$ $1.15 \pm 0.009$
NC4 0.01 Random $4\times 1/5$ $1.12 \pm 0.014$ $1.04 \pm 0.042$ $1.31 \pm 0.098$
NC5 0.01 $Z_c$ Random $1.21 \pm 0.026$ $1.19 \pm 0.025$ $1.19 \pm 0.022$
NC6 0.0 Random Random $1.12 \pm 0.021$ $1.08 \pm 0.036$ $1.22 \pm 0.010$
NC7 0.01 Random Random $1.12 \pm 0.015$ $1.03 \pm 0.023$ $1.29 \pm 0.037$
: Non-conservative avalanche models \[$\varepsilon =10^{-6}$, $D_{\rm nc}=0.1$\]. Error bars were obtained with ten different random-number sequences.[]{data-label="tab:ncons"}
Figure \[fig:tsncons\], similar in format to Figure \[fig:tscons\], shows a $10^6$ iterations segment of the time series for total lattice energy (a) and energy release (b) in model NC0. This is now a nonconservative model using a driving rate $\varepsilon=10^{-6}$, fixed threshold and extraction, and equal redistribution to nearest-neighbors. The only stochasticity introduced here is at the level of non-conservation, with a random fraction between 0 and 90% of the nodal variable value being lost from each unstable node during redistribution. Comparing to Figure \[fig:tscons\]a and b, one immediately notices the absence of the very large quasi-periodic avalanches characterizing models with conservative redistribution. The lattice-energy time series now has the fractal-sawtooth form characteristic of the original stochastically driven LH model. The largest avalanches still only release a small fraction of the total lattice energy (here about one percent at most, but much less on larger lattices). The peak energy release of avalanches now spans over three orders of magnitude even on the small $48^{2}$ lattice used here, indicating that the larger avalanches span the whole lattice. No hint of periodicities can be detected. These features suggest a SOC-like state with a scale-free distribution of avalanche sizes.
![Sample time series for avalanche model NC0 ($D_{nc}=0.1$). Panel (a) shows the temporal evolution of total lattice energy, and part (b) the energy released by avalanches. There is now a single population of avalanches, without a characteristic event size or quasiperiodicity in event occurrence.[]{data-label="fig:tsncons"}](fig6_NC0){width="\linewidth"}
This impression is reinforced upon computing the PDFs for the size of avalanches. Unless the conservation parameter \[$D$\] tends towards unity (more about this shortly), these PDFs now take the form of pure power laws for mid-size and large avalanches (cf. Figure \[fig:pdfncons\]). Small avalanches appear to systematically depart from a power law distribution in a similar fashion to than observed for conservative models in Section \[ssec:rescons\]. The origin of the flattening is in fact the same: small avalanches have now the possibility to release energies in a much broader range than in the original LH model. This feature notwithstanding, at small conservation parameter all of these nonconservative models show a scale-free distribution for mid-size and large avalanches in total avalanche energy, peak energy release and duration.
Models with moderate to high conservation ($D_{\rm nc}\to 1$) do exhibit deviations from power-law behavior for the largest events. The limit $D_{\rm nc}\to 1$ takes us back to conservative models, so it is not surprising to see a second population of avalanches appear. Figure \[fig:pdfncons\], shows a set of PDFs computed for a set of NC0-type model runs with driving rate fixed at $\varepsilon=10^{-6}$ and decreasing values of the dissipation parameter $D_{\rm nc}$, as labeled. Recall (Table \[tab:ncons\]) that except for the dissipation mechanism, these models are otherwise fully deterministic. The models at $D_{\rm nc}=0.1$ yield pure power laws, but at $D_{\rm nc}=0.5$ an excess of large events becomes apparent, which becomes quite pronounced at $D_{\rm nc}=0.99$.
![Probability distribution functions of total avalanche energy in a set of NC0-type model runs with fixed driving rate $\varepsilon=10^{-6}$ but decreasing conservation parameter $D_{\rm nc}$, as labeled. The layout is similar to Figure \[fig:pdfcons\]: panel a, b and c show respectively $E$, $T$ and $P$.[]{data-label="fig:pdfncons"}](fig7_NC0){width="\linewidth"}
Figure \[fig:alphancons\] shows how the power-law indices for (a) avalanche energy, (b) peak release, and (c) duration vary with driving rate \[$\varepsilon$\] and conservation parameter \[$D_{nc}$\], in the suite of NC0-type models runs. Each curve corresponds to a specific driving rate \[$\varepsilon$\], as labeled. The power-law indexes show little to no dependance to the driving rate – provided it is sufficiently low –, as expected. The general trend, namely steepening of the PDFs with conservation parameter, also characterizes the other types of nonconservative models listed in Table \[tab:ncons\]. A systematic dependence of the power-law indices with the conservation parameter is observed in Figure \[fig:alphancons\]. We fit the power law indices \[$\alpha$\] as a function of a power law of $D_{nc}$ (gray lines in Figure \[fig:alphancons\]) and obtain:
$$\begin{aligned}
\label{eq:fit_alphas}
\alpha_{E} &=& 1.18 + \left(\frac{D_{\rm nc}}{2.48}\right)^{1.60} \, , \\
\alpha_{P} &=& 1.20 + \left(\frac{D_{\rm nc}}{1.34}\right)^{2.00} \, ,\\
\alpha_{T} &=& 1.15 + \left(\frac{D_{\rm nc}}{1.35}\right)^{2.12} \, .\end{aligned}$$
![Variations of power-law indices for the PDFs of (a) total avalanche energy, (b) peak energy release and (c) duration, as a function of the conservation parameter \[$D_{\rm nc}$\] and for various driving rates \[$\varepsilon$\], as labeled. The grey thick line corresponds to the power-law fit (see text). All these runs are for NC0-type models.[]{data-label="fig:alphancons"}](fig8_alphas_NC0){width="\linewidth"}
Finally, the waiting-time distributions obtained in these nonconservative models show the same overall qualitative behavior as those characterizing the conservative models (see Figures \[fig:wtdcons\] and \[fig:Ewtd\]).
On the Breaking of Finite-Size Scaling\[ssec:fss\]
--------------------------------------------------
In statistical physics parlance, the appearance of several statistically distinct populations of avalanches (as occurs here) represents a break of [*finite-size scaling*]{}. Such breaks are known to occur in certain classes of SOC models. For instance, the @Olami:1992do model for earthquake, which is also a globally and deterministically driven model, shows a clear break in finite-size scaling when its nonconservation parameter (equivalent to $1-D$ herein) becomes too small [see @Grassberger:1994fv]. This can be traced to the effect of boundaries on internal avalanche dynamics, with lattice nodes located near the boundaries becoming favored triggering sites for large avalanches [see @Lise:2001du]. We have searched for spatial dependencies in the triggering sites of large avalanches in a few simulation runs, by constructing 2D spatial frequency distributions functions for onset locus of large and small avalanches. No obvious dependencies on the distance to boundaries could be detected.
We explored the effect of resolution on the break of finite-size scaling at both ends of our PDF (Figure \[fig:resolution\]). We observe that the population of large and quasi periodic avalanches always spans the last decade of energies (both peak and total) for the four sizes of lattice we considered. It confirms that this population consists of avalanches spanning the whole lattice and does not result from a particular scale our model could have introduced.
The small avalanches population is on the contrary sensitive to the lattice resolution. The avalanche duration PDF (c) shows no hint of resolution effect: the small avalanches population results from a shift of the typical energy an avalanche of small duration is able to release. We reiterate that the plateau of small avalanches originates from the fact that small avalanches have the ability to release energies in a significantly larger range (spanning higher energies) than the corresponding avalanches of the LH model. Calculating the energy \[$\Delta e_{NC0}$\] released by a one-node and one-iteration avalanche in model NC0 (Equation (\[eq:erel1\]), we obtain: $$\label{eq:energy_released_NC0}
\frac{\Delta e_{NC0}}{e_{0}} = \frac{\Delta e_{i,j}}{e_{0}} +
\frac{1-r_{0}}{2Z_{c}}\sum \left( A_{i\pm 1,j\pm 1}\right) +
\frac{1-r_{0}^{2}}{5}\, ,$$ where we recall that $r_{0}\in [D,1]$. If $r_0=1$, we obtain the classical LH energy release (Equation (\[eq:erel1\]). If $r_{0}\approx D$, the additional energy release is dominated by the second term – involving the sum of the nodal variable over the neighbouring node – which is of the order of the nodal variable. The worst case occurs in the middle of the lattice where the nodal variable is maximal. A larger lattice leads to larger values of the nodal variable, which explains the wider small avalanches populations observed in Figure \[fig:resolution\]a – b. For example, the nodal variable is of the order of $3\times 10^{4}$ for a lattice of $384^{2}$, which corresponds indeed to the energy limit distinguishing the small avalanches from the scale-free mid-size ones. Finally, the range of power-law mid-size avalanches also increases with resolution, as expected from classical results of the LH model.
Unlike in the seismic-fault model (which should be thought of, in some global sense, as periodic), here coronal loops do have a finite spatial extent in cross-section, and so the various “tricks” that have been developed to restore finite-size scaling to the Olami [[*et al.*]{}]{} model [*e.g.* @Manna:2006hb] cannot legitimately be used in the present context.
The model results of @Grassberger:1994fv suggested that the loss of finite-size scaling takes place quite suddenly, beyond a certain level of non-conservation. The results reported here indicate a more gradual transition, with the population of large, quasiperiodic avalanches gradually merging with the mid-size, scale-free avalanches as $D$ decreases (Figure \[fig:pdfncons\]).
![PDFs of E, T, and P for model NC0 \[$D_{nc}=0.8$, $\varepsilon
= 10^{-5}$\] for four different lattice sizes.[]{data-label="fig:resolution"}](fig9_resolution){width="\linewidth"}
Discussion and Conclusion \[sec:conclusion\]
============================================
In this article, we have described and documented the behavior of a novel and hitherto unexplored class of lattice-based avalanche models applicable to solar flares. These models, based on a spatially global and fully deterministic driving mechanism, are amenable to physical interpretation in terms of slow twisting of a small coronal loop. Stochasticity is introduced in the models at the level of the redistribution rules and stability threshold governing avalanching dynamics.
Over significant portions of model space, the simulations produce three distinct populations of avalanches. The first is composed of the smallest avalanches with a plateau-like distribution. Then, a classical scale-free distribution of avalanches takes place in the mid-size range of energies. Finally, quasi-periodic avalanches with a more-or-less well-defined mean size make up the third population. The two latter populations are observed to merge in the context of avalanche models incorporating non-conservative redistribution rules. We have argued that the large, quasiperiodic avalanches are the reflection of an energy loading/unloading cycle driven by our adopted global deterministic driving mechanism. Even though an avalanche is always triggered at a single lattice node, the effect of the global driver is that energy has been accumulating at [*all*]{} internal lattice nodes prior to avalanche onset. Once the avalanche begins, this energy must be either evacuated at the open boundaries, or dissipated locally within the lattice (in non-conservative models). In conservative models, only the former is possible, which leads inevitably to the buildup of very large avalanches that must connect to the boundaries – and thus end up spanning the whole system – and remain active until the excess energy has been drained away from the lattice – thus producing long duration avalanches. In that respect, it is therefore not surprising that a distinct population of large avalanches (and, consequently, break of finite-size scaling) is most readily observed in models based on conservative redistribution rules.
We have been careful thus far to refrain from declaring our models to be in a true self-organized critical state, using instead the term “SOC-like” to characterize their avalanching behavior. Formal demonstration of SOC typically requires the computation of critical exponents, and the demonstration that the latter satisfy mutual numerical relations that place them in a specific universality class. Although interesting in and of itself, the identification of the universality class to which the present model belongs is of limited interest from the point of view of flare modelling. We stress again that the models discussed in this article do satisfy all [*sine qua non*]{} physical conditions believed to be required for SOC [@Jensen:1998ww]: a slowly driven open system subjected to a self-limiting, local, finite-threshold instability. Moreover, in nonconservative models characterized by no quasi-periodic large scale events, the lack of dependence of the power-law indices of avalanches size and duration on the driving rate and lattice size certainly suggests that these models all belong to the same universality class.
Although non-conservative redistribution rules have received, to the best of our knowledge, little attention in the solar flare context, true SOC avalanching behavior is known to materialize in non-conservative system. The most studied such SOC system is the cellular automaton formulation of the stick–slip model for earthquake, introduced by @Olami:1992do. In that model, non-conservation is related to the fact that a moving block exerts a force not only on its neighbour blocks via the connecting springs, but also on the two plates bounding (and driving) the system from above and below; hence; when an individual block moves, a fraction of the potential energy stored in the springs is dissipated as heat via friction against the bounding plates, rather than all of it ending up in other springs by the end of the avalanche. Something similar also takes place in the non-conservative substorm models of @Liu:2006cp and @VallieresNollet:2010ed. There, a fraction of the energy released by an unstable magnetic flux tube within the central (equatorial) plasma sheet is lost via MHD waves and/or charged particles travelling away along magnetic field lines back towards Earth, eventually leading to auroral excitation in polar regions. A particularly interesting feature of the latter model is its ability to simultaneously produce “internal” avalanches with a scale-free size distribution, and quasi-periodic boundary discharge avalanches with a well-defined mean-size. This squares well with the statistically distinct properties of auroral emission on the one hand, and ring current injection events on the other, both being in principle related manifestation of energy release events in the central plasma sheet. In the model of @Liu:2006cp, this effect materialized in part because the boundaries were set up so as to respond dynamically to incoming avalanches; the results presented herein indicate that dual populations of energy release events, each with distinct statistical characteristics, would still materialize in such globally driven systems, even with more conventional “open” boundary conditions, provided conservation is high enough. [[Finally, solar flares also exhibit non-conservative properties. Flares are thought to originate from reconnection processes in coronal loops that feed on the stored magnetic energy. Along with the magnetic reorganization associated with a flare, some energy is systematically lost from the system either through radiation (direct X-ray emissions, or hard X-ray emissions triggered by reconnection-accelerated energetic particles), accelerated particles escape or even conduction along the magnetic field lines. We modeled those effects with a simple non-conservation parameter in the avalanching process. Albeit with a quite different cellular automaton model, our results are consistent with the previous finding of @Hamon:2002hs: a quasi-cyclic population of large avalanches is found in both cases when the model is close to be purely conservative.]{}]{}
Notwithstanding the presence or absence of a population of small and large avalanches, in the realm of mid-size avalanches all models are characterized by a power-law distribution of avalanche energy release spanning many orders of magnitude. In general, the corresponding power-law indices are somewhat smaller than for the LH model, implying flatter PDFs, so that all models have a power-law index smaller than the critical value $\alpha_E=2$ above which the smaller flares (*i.e.* avalanches) would become the dominant contributors to coronal heating, as in Parker’s nanoflare hypothesis.
An obvious extension of the present model is to introduce the third spatial dimension; as argued in Section \[ssec:phys\], the 2D lattices considered herein can be best regarded as a cross section at some fixed position along a coronal loop. Evidently, allowing avalanches to develop along the loop’s length would yield a more realistic model. In particular, it would become possible to map the 3D lattice to a bent coronal loop, and therefore compute fractal dimensions of flares/avalanches [see, *e.g.* @Morales:2009dc] in the same manner as carried out from observations [see, e.g. @Aschwanden:2013kb]. This would open a new comparative bridge between modelling and observation.
The broad exploration of parameter space made possible with the computationally much less demanding 2D lattices has allowed us to pinpoint the most important parameter regime, which is found for highly dissipative redistribution. Then, the way in which stochasticity is introduced in the model (stability threshold, extraction, redistribution rules) plays a comparatively lesser role in determining the size distribution of avalanches. In other words, the models are robust with respect to the details of stochastic effects.
The present model finally provides an interesting basis for a practical use of avalanches models in the context of solar flares [@Strugarek:2014we]. The deterministic driver minimizes the level of stochasticity embedded in the model, which results in very good predictive capabilities for the larger events. Interestingly, our model could provide an efficient and cheap alternative method for the prediction of large solar flares in the context of space weather.
The authors acknowledge stimulating discussions during the ISSI Workshops on Turbulence and Self-Organized Criticality (2012 – 2013) held in Bern (Switzerland); and during the “Festival de théorie” (2013) held in Aix-en-Provence (France). We acknowledge support from Canada’s Natural Sciences and Engineering Research Council.
[40]{} \#1[ISBN \#1]{}\#1[\#1]{}\#1[\#1]{}\#1[\#1]{}\#1[*\#1*]{}\#1[**\#1**]{}\#1[\#1]{}\#1[\#1]{}\#1[\#1]{}\#1[\#1]{}\#1\#1\#2\#1[\#1]{}\#1[\#1]{}\#1[*\#1*]{}\#1[\#1]{}\#1[**\#1**]{}\#1[\#1]{}\#1[*\#1*]{}\#1[\#1]{}\#1[\#1]{}\#1[\#1]{}\#1\#1[\#1]{}\#1[\#1]{} \#1[[](http://dx.doi.org/#1)]{} \#1[[](http://arxiv.org/abs/#1)]{} \#1[[](http://adsabs.harvard.edu/abs/#1)]{}\#1\#1\#1\#1\#1[\#1]{}\#1[\#1]{}\#1[\#1]{} \#1[\#1]{}
: , . (), . . .
, : , . (), . . .
, : , . (), . . .
, , : , . (), . . .
, , , , , , , : , . (), . . .
, , , : , . (), . . .
, , : , . , . . .
, , : , . , . . .
, : , , . .
: , . In: , , .
, , , : , . (), . . .
, , , , , , , , : , . (), . . .
: , . , . . .
: , . (), . . .
, , : , . , . . .
: , , ,
, , , : , . (), . . .
: , . (), . . .
, : , . (), . . .
, , , : , . (), . . .
, , , : , . (), . . .
: , . , . . .
, : , . , . . .
, , , : , . , . . .
, : , . (), . . .
: , . (), . . .
, : , . (), . . .
, : , . (), . . .
, , , : , . (), . . .
, , : , . (), . . .
: a, . , . . .
: b, , . . .
: , . , . . .
, , : , . (), . . .
: , . (), . . .
, : , . , . . .
, : 2014, [Predictive capabilites of avalanche models for solar flares]{}. *Submitted to Solar Physics*.
, , , , : , . (), . . .
: , . (), . . .
, : , . , . . .
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'In recent years cross correlation of lensing of the Cosmic Microwave Background (CMB) with other large scale structure (LSS) tracers has been used as a method to detect CMB lensing. Current experiments are also becoming sensitive enough to measure CMB lensing without the help of auxiliary tracers. As data quality improves rapidly, it has been suggested that the CMB lensing-LSS cross correlation may provide new insights into parameters describing cosmological structure growth. In this work we perform forecasts that combine the lensing potential auto power spectrum from various future CMB experiments, the galaxy power spectrum from galaxy surveys, as well as the cross power spectrum between the two, marginalizing over a number galactic and non-galactic cosmological parameters. We find that the CMB lensing-LSS cross correlation contains significant information on parameters such as the redshift distribution and bias of LSS tracers. We also find that the cross correlation information will lead to independent probes of cosmological parameters such as neutrino mass and the reionization optical depth.'
address:
- 'Department of Physics & Astronomy, University of Sussex, Brighton BN1 9QH, UK'
- 'Kavli Institute for Particle Astrophysics and Cosmology, SLAC, 2575 Sand Hill Road, Menlo Park, CA 94025, USA'
- 'Department of Physics, University of California, Berkeley, USA'
- 'Lawrence Berkeley National Laboratory, Berkeley, CA, USA'
author:
- 'R. Pearson'
- 'O. Zahn'
title: Cosmology from cross correlation of CMB lensing and galaxy surveys
---
Introduction
============
In the past few years CMB lensing science has arisen as a promising new probe of cosmology. CMB photons are deflected by matter fluctuations between the surface of last scattering and today. The effect is sensitive to the geometry of the Universe (e.g., dark energy and its evolution) and the shape of the matter power spectrum (affected by e.g., massive neutrinos). Since CMB lensing probes all matter, it should be correlated with galaxy surveys observing the large scale structure of the Universe. The cross correlation between CMB lensing and such large scale structure surveys should aid in simultaneously measuring galactic and non-galactic cosmological information. There is already a wealth of data available for such analysis, and upcoming surveys will offer even better opportunities for cross correlation science. A few examples of work to date include: CMB lensing $\times$ radio galaxies using WMAP and NVSS [@Smith:2007rg; @Feng:2012uf], CMB lensing $\times$ LRGs $\times$ quasars $\times$ radio galaxies using WMAP, NVSS and SDSS [@Hirata:2008cb], CMB lensing $\times$ optical $\times$ IR using SPT, BCS, WISE and Spitzer [@Bleem:2012gm], CMB lensing $\times$ quasars using ACT, and SDSS [@Sherwin:2012mr], CMB lensing $\times$ sub-millimetre wavelength galaxies using SPT and Herschel/SPIRE [@Holder:2013hqu].
Both CMB temperature and polarization are being measured by modern experiments. At the detector level, the polarization measurement is in terms of the Stoke’s parameters Q and U, but it is usually useful to consider CMB polarization in terms of even and odd parity modes, the E- and B-mode respectively. The temperature anisotropy has been measured well down to arc minute scales (e.g. [@Jones:2005yb; @Hinshaw:2012aka; @Planck:2013kta; @Calabrese:2013jyk; @Story:2012wx]), E-mode polarization has been detected by several groups (e.g. [@Leitch:2004gd; @Montroy:2005yx]), and the conversion of E-mode to B-mode polarization by lensing has been observed [@Hanson:2013hsb]. An observation of large scale B-modes extra to lensing and foreground components could be direct evidence of a primordial gravitational background, and used to discern inflationary theories. Even without the B-mode though, observations of E-mode polarization allow improved constraints on cosmological parameters compared to temperature information alone, since they are sensitive to the epoch of reionization as well as recombination physics. To obtain the CMB lensing signal, the quadratic estimator method of [@Hu:2001kj] is usually used. The CMB lensing potential can be reconstructed from various combinations of T-, E- and B-mode measurements. So far the lensing power spectrum has been measured using the temperature anisotropy observed by the Atacama Cosmology Telescope (ACT) [@Das:2011ak], [@Das:2013zf], the South Pole Telescope (SPT) [@vanEngelen:2012va], and most recently the Planck satellite [@Ade:2013tyw]. Lensing also affects the peaks and troughs in the angular power spectra (e.g., [@Keisler:2011aw; @Story:2012wx]). Finally, CMB lensing mixes primordial E-modes into lensing-specific B-modes. This allows reconstruction of the CMB lensing potential with much smaller variance than for the temperature modes, because no primordial B-modes are expected on small scales.
A Fisher matrix calculation marginalizing over various cosmological parameters can be used to estimate the constraining power of current and future data sets. In this work, different CMB and galaxy survey sensitivities are explored with the goal of showing how well cosmology can be constrained in the presence of uncertainties in the galaxy physics, and vice-versa. In particular, we are interested in how well the parameters are constrained using the CMB lensing - galaxy cross correlation. Seven non-galactic parameters are included in the analysis with the following fiducial values: physical baryon fraction ($\Omega_{\rm b}h^{2}$=0.02258), physical cold dark matter fraction ($\Omega_{\rm CDM}h^{2}$=0.1109), dark energy fraction ($\Omega_{\lambda}$=0.734), physical neutrino mass fraction ($\Omega_{\nu}h^{2}$=0.002), optical depth ($\tau$=0.088), spectral index ($n_{\rm s}$=0.963), scalar amplitude ($A_{\rm s}$=2.43 x10$^{-9}$); and up to three galactic parameters: the linear galaxy bias (b), the mean of the galaxy redshift distribution ($\mu$), and the standard deviation of the galaxy redshift distribution ($\sigma$).
Previous work has already shown that cross-correlation of structure and CMB lensing can constrain various biases, and thus provide useful information for cosmology[@Vallinotto:2011ge; @Vallinotto:2013eva; @Rhodes:2013fyq]. In particular, CMB lensing can break degeneracy between the matter power spectrum normalisation and systematic multiplicative biases in shear measurements [@Vallinotto:2011ge], and can constrain the linear galaxy bias [@Vallinotto:2013eva]. Our work complements the latter analysis by allowing redshift distribution parameters to also vary.
Section \[method\] introduces the Fisher methodology used and outlines the derivation of our theoretical power spectra. In section \[data\] we describe some current and future CMB and galaxy surveys as well as our data cuts. Results on cosmological parameters are given in section \[results\] and we conclude in section \[conclusion\].
Method
======
For a given cosmological model described by parameters $\lambda_i$ and future experiment with known specifications, one can approximate the likelihood in the vicinity of the best-fit (fiducial) model as a multivariate Gaussian with the Hessian given by the Fisher matrix in terms of the theoretical power spectra $C_l$ as $$\label{Fish}
F_{\alpha\beta}=\sum_{l}\frac{1}{(\delta C_{l})^2}\frac{\partial C_{l}}{\partial\lambda_{\alpha}}\frac{\partial C_{l}}{\partial\lambda_{\beta}}$$ The uncertainty $\delta C_{l}$ is given by: $$\label{error}
\delta C_{l}=\sqrt{\frac{2}{(2l+1)f_{\rm sky}}}[C_{l}+C_{l}^{\rm noise}]$$ Where the first term in the bracket accounts for cosmic variance in the signal and $C_{l}^{\rm noise}$ is the associated experimental noise. The factor $f_{\rm sky}$ accounts for the fraction of sky covered by the experiment. The error on a given parameter $\lambda_{\alpha}$ is $\sqrt{(F^{-1})_{\alpha\alpha}}$. If there is some prior constraint on the error of a parameter $\sigma_{\rm prior}$, this information is added in the Fisher matrix via $F_{\alpha\alpha}=F_{\alpha\alpha}+\frac{1}{\sigma^{2}_{\rm prior}}$.
In this analysis we form two Fisher matrices. The first includes auto and cross correlations from CMB temperature and polarization: TT, EE, TE. The second includes auto and cross correlations of the CMB lensing potential and 2-dimensional galaxy power spectrum: $\phi\phi$, GG, $\phi$G. When considering multiple power spectra (non-diagonal covariance), the Fisher matrix is formulated as shown in [@Perotto:2006rj] via: $$\label{ }
F_{\alpha\beta}=\sum^{l_{\rm max}}_{\rm l=2} \sum_{\rm PP',QQ'} \frac{\partial C_{l}^{\rm PP'}}{\partial \lambda_{\alpha}}(\mathbb{C}_{l}^{-1})_{\rm PP'QQ'}\frac{\partial C_{l}^{\rm QQ'}}{\partial \lambda_{\beta}}$$ In the first Fisher matrix $PP',QQ' \in {TT, EE, TE}$, whilst in the second Fisher matrix $PP',QQ' \in {\phi\phi, GG, \phi G}$. $\mathbb{C}_{l}$ is the 3$\times$3 covariance matrix between the different power spectra, and the elements are made up of the appropriate $(\delta C_{l})^{2}$. In the CMB temperature and polarization case, the $C_{l}^{\rm noise}$ is taken to be the standard Gaussian random detector noise as in [@PhysRevD.52.4307]. For the CMB lensing, $C_{l}^{\rm noise}$ is the N0 bias as in [@Hu:2001kj], and for the galaxy power spectrum it is simple shot noise. The two Fisher matrices are added together for the final constraint, and we neglect small correlations between CMB temperature and lensing (T$\phi$) caused by the integrated Sachs-Wolfe (ISW) effect (e.g. [@Ade:2013dsi]).
The theoretical auto and cross CMB lensing and galaxy power spectra are approximately: $$\label{ }
C_{L}^{\phi\phi}\sim\int d\chi (K^{\phi}_{L}(\chi))^{2}P(k=L/\chi,\chi)$$ $$\label{ }
C_{L}^{\rm gg}\sim\int d\chi (K^{\rm g}_{L}(\chi))^{2}P(k=L/\chi,\chi)$$ $$\label{ }
C_{L}^{\phi g}\sim\int d\chi K^{\phi}_{L}(\chi)K^{\rm g}_{L}(\chi)P(k=L/\chi,\chi)$$ Where $\chi$ is conformal distance and P(k,$\chi$) is the 3D matter power spectrum at wavenumber k and conformal look back time $\chi$. The kernels $K^{\phi}_{L}(\chi)$ and $K^{\rm g}_{L}(\chi)$ fold in information about the lensing and the galaxy dynamics respectively: $$\label{eq:Kphi}
K^{\phi}_{L}(\chi)=-\frac{3\Omega_{m}H_{0}^{2}}{L^{2}}\frac{\chi}{a}(\frac{\chi_{*}-\chi}{\chi_{*}\chi})$$ $$\label{eq:Kg}
K^{\rm g}_{L}(\chi)=\frac{dN}{dz}\frac{dz}{d\chi}\frac{b(z)}{\chi}$$
Equation \[eq:Kphi\] depends on non-galactic cosmological parameters, whereas Equation \[eq:Kg\] depends on both galactic and non-galactic cosmological parameters, since it includes the bias and redshift distribution.
Data
====
The derivatives used in the Fisher matrix are calculated using the theoretical power spectra. CMB temperature, CMB polarization, CMB lensing and galaxy power spectra were calculated for the given parameter choice using CAMB sources [@Challinor:2011bk]. To obtain the galaxy auto correlation (GG) and galaxy-CMB lensing cross correlation ($\phi$G), we used a slightly modified version of the code with source and lensing type windows.
Table \[CMBspecs\] shows the CMB experimental scenarios we consider in this work. The current type experiment is based on the [*Planck*]{} satellite, for which the blue book values agree reasonably well with the satellite’s performance so far. The 3rd generation experiment is a survey with higher angular resolution (1 arcmin FWHM) covering 10% of the sky. We verified that a three times larger beam would not qualitatively change any of the results. The 4th generation experiment has the same angular resolution as the 3rd generation experiment but covers 50% of the sky. The noise in the CMB temperature and polarization is shown in the table, and the CMB lensing noise is derived from these values. The 3rd generation is representative of surveys already taking data or soon to begin observations, such as SPTpol [@Austermann:2012ga], ACTpol [@Niemack:2010wz] and POLARBEAR [@Kermish:2012eh]. The 4th generation is representative of larger sky area surveys planned for the near future, such as the POLARBEAR extension, the Simons-array.
Table \[GALspecs\] shows experimental scenarios for 3 galaxy surveys. The Wide-field Infrared Survey Explorer (WISE) is a satellite experiment which observed the whole sky in the mid-infrared. The four observing bands are centered at 3.4, 4.6, 12 and 22 $\mu$m [@Wright:2010qw]. WISE has already been used in cross correlation with various data sets such as WMAP [@Goto:2012yc], [*Planck*]{} [@Ade:2013tyw] and SPT [@Geach13]. The Large Synoptic Survey Telescope (LSST) is a planned 8.4 meter optical ground based telescope which will observe from Cerro Pachón, Chile. The predicted redshift distribution from the LSST science book [@Abell:2009aa] is: $$\label{LSST}
P(z)=\frac{1}{2z_{0}}(\frac{z}{z_{0}})^{2}\exp (-z/z_{0})$$ With $z_{0} = 0.0417i-0.744$ (where the survey has sensitivity in the i band for magnitudes $21.5 < i < 23$). In section \[results\] we will use LSST as an example next generation survey for constraints on parameters. Choosing $i=22.25$ for a mid-range magnitude sample gives a fiducial $z_{0}=0.183$, which we use in our calculation. The LSST science book assumes a galaxy bias evolution of b=1 + 0.84z. Given a median redshift of around z=1, we will estimate the fiducial galaxy bias as b=1.84 in our simulations. Euclid is an ESA funded satellite mission due to launch in 2019. It will observe 15,000 sq deg in optical and near-infrared using a 1.2m space telescope [@Laureijs:2011gra]. Figure \[fig:dndz\] shows the redshift distributions of these galaxy surveys along with the CMB lensing kernel. Also shown is a Gaussian toy model with $\mu=1$ and $\sigma=0.5$. This Gaussian is fairly representative of the galaxy distribution redshift ranges, and all redshift distributions show an overlap with the CMB lensing kernel, suggesting that they are useful for cross correlation studies.
In table \[tab:scenarios\] we show the combinations of CMB and galaxy surveys which we consider in our Fisher calculations, labelled from A-J. To measure a cross correlation, the fractions of sky observed by CMB and galaxy surveys must overlap. If the CMB experiment has a smaller $f_{\rm sky}$ than the galaxy survey, the smaller value must also be used for the galaxy information in that case. We take a conservative approach, considering the same $f_{\rm sky}$ in the GG case as the $\phi G$ case, rather than using the full survey area available for GG. These details can be read from column 3 of the table. For the CMB information in the $TT,EE,TE$ correlations, we use the maximum $f_{\rm sky}$ possible for the deepest CMB survey (3rd or 4th generation). We then include information from the current ([*Planck*]{}) CMB survey, but only up to a combined $f_{\rm sky}=0.75$, since we do not want to double count patches of sky. This counting assumes that the remaining 25% of the sky are contaminated by foregrounds. In scenarios E, F, G, H, I and J, we add 20% priors to the galaxy parameters. This is to break degeneracies between the galactic and non-galactic parameters in order to make the Fisher matrix invertible. Adding e.g. 40% priors rather than 20% priors made negligible difference to the forecasts.
The GG and $\phi$G correlations have a sharper cut-off in multipole $l$, because non-linear clustering and scale dependence of the bias will make recovery of cosmological information from these scales challenging. The approximate onset of non-linearities at $k\sim 0.2$ h/Mpc yields a projected $l_{\rm non-linear} \simeq 300$ in GG and $\phi$G correlations for a galaxy population peaking at $z\simeq1$. The $\phi$G correlation should be less sensitive to the effect of non-linearities because it narrows in on higher redshifts. Taking this into consideration, we choose a multipole cut-off of $l_{\rm max}=500$. We also explore the cases of $l_{\rm max}=1,000$ and $l_{\rm max}=5,000$ for this observable, assuming future improvements in our understanding of the nonlinear clustering regime. Note that we have not modeled scale dependence in the bias due to non-linearities. Simulations show that non-linear bias should only affect scales $k\geq 0.7$ h/Mpc [@Zarija08], which is outside the range of modes that we use as our default $l_{\rm max}$. However when we move beyond $l_{\rm max}=1,000$ we are entering this more complicated regime.
In some cases, we show forecasted constraints for a toy model Gaussian redshift distribution where we vary redshift-integrated galaxy density and $l_{\rm max}$. Note that the [*values*]{} of number density, redshift distribution, bias, and $l_{\rm max}$ are not independent, as for example a change in the number density (by observing down to lower fluxes) would likely affect the mean galaxy bias. Since we are interested in forecasting parameter [*errors*]{} around a fiducial model, we can ignore such dependencies. Note also that we do not explicitly model a redshift dependence of the bias, b(z). This dependence is highly degenerate with the evolution of the number density N(z) when using observables integrated over all redshifts, as we do here. The degeneracy can be reduced by using redshift-binned galaxy auto and CMB lensing-galaxy cross power spectra. We defer such an analysis to a future work.
Current 3rd gen. 4th gen.
------------------------------------ --------- ---------- ----------
temperature noise ($\mu$K-arcmin) 30 2.5 2.5
polarization noise ($\mu$K-arcmin) 60 3.5 3.5
beam (arcmin) 7 1 1
$f_{\rm sky}$ 0.75 0.1 0.5
: CMB EXPERIMENTAL SCENARIOS
\[CMBspecs\]
$f_{\rm sky}$ number density (gal/deg$^2$)
-------- --------------- ------------------------------ --
WISE 0.75 10,000
LSST 0.5 198,000
EUCLID 0.4 108,000
: GALAXY EXPERIMENTAL SCENARIOS
\[GALspecs\]
![Redshift distributions for current and up coming galaxy surveys compared to the CMB lensing redshift kernel. Redshift distributions shown for the WISE, LSST and EUCLID surveys are taken from [@Geach13], [@Abell:2009aa] and [@Boldrin:2012ke] respectively. The Gaussian model has $\mu=1$, $\sigma=0.5$. []{data-label="fig:dndz"}](dndz_dists+.pdf){width="8.8cm"}
Scenario CMB surveys $GG,\phi\phi,G\phi$ surveys Parameters varied (% prior)
---------- ----------------- ------------------------------- ----------------------------------------- --
A [*Planck*]{}+4G Gaussian, 4G $\Lambda$CDM$\nu$+bias+$\mu$+$\sigma$
B [*Planck*]{}+4G LSST, 4G $\Lambda$CDM$\nu$+bias+$z_{0}$
C [*Planck*]{}+3G Gaussian 1k gal/deg$^{2}$, 3G $\Lambda$CDM$\nu$+bias+$\mu$
D [*Planck*]{}+4G Gaussian 1k gal/deg$^{2}$, 4G $\Lambda$CDM$\nu$+bias+$\mu$
E [*Planck*]{}+3G LSST, 3G $\Lambda$CDM$\nu$+bias(20%)
F [*Planck*]{}+4G LSST, 4G $\Lambda$CDM$\nu$+bias(20%)
G [*Planck*]{}+4G Gaussian, 4G $\Lambda$CDM$\nu$+bias(20%)+$\mu(20\%)$
H [*Planck*]{} LSST, [*Planck*]{} $\Lambda$CDM$\nu$+bias(20%)
I [*Planck*]{} LSST, [*Planck*]{} $\Lambda$CDM+bias(20%)
J [*Planck*]{}+4G LSST, 4G $\Lambda$CDM+bias(20%)
\[tab:scenarios\]
A description of the combinations of CMB and galaxy experimental scenarios we consider in our Fisher calculations. Column 2 lists surveys which are used in the $TT,EE,TE$ correlations. 3G denotes a third generation experiment with 10% sky coverage, 4G a fourth generation experiment with 50%. Column 3 details the surveys used for the $GG,\phi\phi,G\phi$ correlations, with the sky fraction given by that of the deeper experiment in column 2. LSST is estimated to have a number density of 198,000 gal/deg$^{2}$. For the Gaussian cases, either the number density is varied, or 1,000 gal/deg$^{2}$ is used. In column 2, surveys are separated by addition signs. This denotes that the Fisher information matrix was constructed for each of these surveys and then added together. In column 3 the surveys are separated by commas. This denotes that information from both surveys was used in one Fisher matrix to construct auto and cross correlations between the CMB lensing and galaxy power spectra. The sky areas used are always those of the the highest sensitivity CMB experiment. Column 4 shows what parameters have been varied, where $\Lambda$CDM$(\nu)$ was defined as $\Omega_{\rm b}h^{2}+\Omega_{\rm CDM}h^{2}+\Omega_{\lambda}+\tau+ n_{\rm s}+A_{\rm s} (+\Sigma m_\nu)$. Notice that priors of 20% have been added to the galaxy parameters in scenarios E, F, G, H, I, and J. These are added to break degeneracies between the galactic and non-galactic parameters in order to make the Fisher matrix invertible.
Constraints on cosmological parameters {#results}
======================================
In the following sections we refer to the [*unlensed*]{} primary CMB information using the notation CMB $ul$. This term encompasses the $TT$ primary CMB temperature power spectrum, the $EE$ primary CMB E-mode information and the $TE$ cross correlation information. Using unlensed spectra ensures that we do not over-count lensing information when the temperature and polarization power spectra are considered in conjunction with the reconstruction-based $\phi\phi$ and $\phi G$ spectra (see [@Schmittfull:2013uea; @zahn] for more detailed treatments that quantify the shared information content between these estimators). In some cases we consider the lensed theoretical power spectra for the primary CMB. This is denoted by CMB $le$.
Galaxy properties
-----------------
In this section we present constraints on parameters describing the large scale tracer structure probed by the cross correlation with CMB lensing, for both a toy model Gaussian case and for a specifically LSST parameterized distribution. In all cases, we allow the full 7 non-galactic cosmology parameters to vary. In the case of the toy model describing the populations probed by the galaxy survey, we also vary 3 parameters describing the galaxy properties: the bias (b), and the mean ($\mu$) and standard deviation ($\sigma$) of the assumed Gaussian redshift distribution, shown by the black curve of figure \[fig:dndz\].
Figure \[fig:gauss\_dndz2\] shows constraints on b, $\mu$ and $\sigma$ for a scenario-A combination of surveys (a galaxy survey with a Gaussian redshift distribution and and *Planck* + 4th. generation CMB experiments). Parameter constraints are shown over a range of galaxy survey sensitivities (galaxy number density per square degree). Different combinations of the lensing and galaxy auto correlations $\phi\phi$, GG, and the lensing galaxy cross correlation $\phi$G and the primary CMB are shown by lines on the plot. For all three parameters, the addition of $\phi$G information improves constraints significantly over using CMB + GG alone (solid line to dotted line). The constraints can be improved further by adding $\phi\phi$ information (bottom dot-dashed lines). The addition of CMB lensing information is crucial for constraining galaxy bias, because it helps break degeneracies between bias and other cosmological parameters affecting the amplitude and shape of the galaxy power spectra.
scenario-B $\delta$z$_{0}$ $\delta$b
------------------------------------ ----------------- -----------
CMB ul +GG 0.010 0.065
CMB ul + $\phi$G 0.015 0.093
CMB ul + $\phi\phi$ + $\phi$G 0.010 0.017
CMB ul + $\phi$G + GG 0.002 0.040
CMB ul + $\phi\phi$ + GG 0.007 0.060
CMB ul + $\phi\phi$ + GG + $\phi$G 0.002 0.016
\[tab:dndzLSST2\]
Constraints on the galaxy parameters b and $z_{0}$ for a scenario-B experiment combination. (LSST galaxy survey with *Planck* + 4th generation CMB) Values are shown for different combinations of GG, $\phi\phi$ and $\phi$G with the primary CMB.
Table \[tab:dndzLSST2\] shows constraints on galaxy parameters b and $z_{0}$ for an LSST type survey (described in section \[data\] and shown in equation \[LSST\]), for the combination with a 4th generation CMB experiment and [*Planck*]{}. $z_{0}$ is the single parameter describing the shape of the redshift distribution. When the cross correlation is added to the autocorrelations (CMB ul + $\phi\phi$ + GG case to CMB ul + $\phi\phi$ + GG + $\phi G$ case), the constraints on both the bias and shape parameter improve by $\sim$70%. Interestingly, the constraints on both parameters from CMB ul + GG is only 30% improved from the constraint of CMB ul + $\phi$G. This shows that the CMB lensing-galaxy correlation contains powerful information on galaxy dynamics, in particular the bias, even without using the galaxy clustering auto power spectrum. The significant improvement in the constraint on the shape parameter $z_0$ in the combination of all observables (last line) compared to CMB ul + GG or CMB ul + $\phi$G + $\phi\phi$ shows on the other hand that the galaxy clustering auto correlation becomes much more useful with the addition of CMB lensing information.
\[subfig:nu\_vs\_bias\_3g\]
Figure \[fig:astro\_ellipse\] shows the 1$\sigma$ and 2$\sigma$ ellipses for various parameters for scenario-C and scenario-D experiment combinations. (The galaxy survey is Gaussian with a number density of 1,000 gal/deg$^{2}$. The difference between scenario-C and scenario-D is that the former includes a 3rd generation CMB survey, whilst the latter includes a 4th generation CMB survey.) The top two panels show a comparison of the $\sum m_{\nu}$ and b parameters for both scenario-C (left) and scenario-D (right). The two parameters are highly degenerate when only including the galaxy auto power or galaxy-lensing cross correlation. We see again that CMB ul + $\phi$G combination does almost as well as the CMB ul + GG combination. The improvement in the CMB ul + $\phi$G cross combination is $\sim$50% for both parameters when upgrading to scenario-D from scenario-C. The lower two panels show a comparison of the redshift distribution mean ($\mu$) and bias parameter for scenario-C (left) and scenario-D (right). The improvement in the $\mu$ parameter is $\sim$40% when upgrading to scenario-D from scenario-C. In all cases we see that adding the $\phi\phi$ lensing auto correlation to the CMB ul +GG (green line to blue line) or CMB ul + $\phi$G (cyan line to purple line) improves the constraints significantly, because the lensing auto power constrains neutrino mass more directly. There is also a good improvement in the constraints when $\phi G$ is added to the CMB ul + GG (green line to yellow line), although the improvement is more dramatic with the addition of the lensing autocorrelation. In both choices of parameter pairs, when comparing scenario-C and scenario-D, the CMB ul +GG constraints improve about as much as the CMB ul + $\phi$G constraints.
Massive neutrinos and reionization optical depth
------------------------------------------------
![$\Delta\sum m_{\nu}$ for scenario-G. Constraints in the y axis are shown in electron volts, the x axis shows number density of galaxies observed per square degree. Forecasts are shown for different cutoff values of multipole $l_{\rm max}$=500, 1,000, 5,000.[]{data-label="fig:lmax_compare"}](lmax_compare_new_correctfsky.pdf){width="8.8cm"}
In this section we present forecasts on the non-galactic cosmological parameters: the neutrino mass sum $\sum m_{\nu}$, and the optical depth $\tau$ due to free electrons since the epoch of reionization.
Figure \[fig:lmax\_compare\] shows the $\sum m_{\nu}$ constraint for scenario-G for different cutoff values, $l_{\rm max}=500$, $l_{\rm max}=1,000$, $l_{\rm max}=5,000$. For each value of $l_{\rm max}$ the primary CMB is shown in combination with the $\phi G$ cross correlation (red lines), and the cross correlation plus lensing auto-correlation $\phi G+\phi\phi$ (black lines). These lines can be compared to that of the primary CMB plus only the CMB lensing auto-correlation $\phi\phi$ (blue line). (Remember that the primary CMB is unlensed and includes the TT, EE and TE correlations). In this scenario, the $\sigma$ of the redshift distribution was fixed, but the b and $\mu$ parameters were allowed to vary with a 20% prior in order to allow matrix inversions. Using a prior of e.g., 40% made negligible difference to the constraints, as did fixing the $\mu$ parameter rather than varying it. We see that adding the $\phi$G information to the CMB ul + $\phi\phi$ combination (blue line to black lines) leads to negligible improvement unless one allows $l_{\rm max}=5,000$. For a conservative value of $l_{\rm max}=500$, the combination of the unlensed CMB and $\phi$G does not give useful constraints. If we consider a less conservative $l_{\rm max}=1,000$, the CMB ul + $\phi$G combination becomes more valuable, and if we consider an optimistic $l_{\rm max}=5,000$ then the constraint becomes comparable to that of the CMB ul + $\phi\phi$ line. With optimistic $l_{\rm max}$ limits, the $\phi G$ cross correlation could provide a good independent check on the $\sum m_{\nu}$ parameter. It is worth mentioning that while the CMB ul + $\phi$G combination places better constraints than the CMB ul does alone, what we actually measure is the lensed primary CMB, and this constrains the $\sum m_{\nu}$ much better than the cross correlation and almost as well as the CMB lensing autocorrelation. (When we consider the $\phi\phi$ correlation in combination with the primary CMB, we always use the unlensed CMB to avoid over-counting the lensing effect).
[-0.5cm]{}
$\Delta \sum m_{\nu}$ (scenario-E) $\Delta \sum m_{\nu}$ (scenario-F)
----------------------------------- ------------------------------------ ------------------------------------
CMB ul only 0.328 0.201
CMB le 0.055 0.040
CMB ul + $\phi\phi$ 0.051 0.036
CMB ul + GG 0.122 0.059
CMB ul + $\phi$G 0.219 0.119
CMB ul + $\phi\phi$ + $\phi$G 0.051 0.035
CMB ul + GG + $\phi$G 0.114 0.058
CMB ul + $\phi\phi$ + GG 0.050 0.034
CMB ul + $\phi\phi$ + GG +$\phi$G 0.044 0.027
\[LSSTnu\]
Constraints in electron volts on the $\sum m_{\nu}$ parameter for an LSST type galaxy survey (equation \[LSST\]) for scenario-E (3rd generation CMB) and scenario-F (4th generation CMB). Constraints are shown for combinations of GG, $\phi\phi$ and $\phi$G with the primary CMB.
Tables \[LSSTnu\] and \[LSSTtau\] show constraints on the $\sum m_{\nu}$ and $\tau$ parameters for an LSST type galaxy survey (described in equation \[LSST\]). In this case the bias parameter was set to a fiducial value of b=1.84.
Table \[LSSTnu\] shows constraints on $\sum m_{\nu}$ for scenario-E and scenario-F. The only difference between scenario-E and scenario-F is that the former includes a 3rd generation CMB experiment, while the latter includes a 4th generation CMB experiment. In the same trend as the Gaussian redshift distribution galaxy survey, using the lensed CMB vs the unlensed CMB gives dramatic improvements, almost an order of magnitude in this case. The lensed CMB alone is much better than the CMB ul + $\phi$G, and the addition of $\phi$G to CMB ul + $\phi\phi$ gives negligible improvement. However, the improvement in the neutrino constraint from CMB ul + GG + $\phi\phi$ to CMB ul + GG + $\phi\phi$ + $\phi G$ is 10% for a 3rd generation CMB experiment and 20% for a 4th generation. This shows that when the redshift distribution is well known, the addition of $\phi G$ helps break degeneracies between cosmological and galaxy parameters in GG.
$\Lambda$CDM$\nu$ $\Delta \tau$ (scenario-H) $\Delta \tau$(scenario-F)
--------------------------- ---------------------------- ---------------------------
TT len 0.084 0.036
CMB le 0.004 0.003
TT+$\phi\phi$ 0.041 0.015
TT+GG 0.064 0.056
TT+$\phi$G 0.068 0.055
TT+$\phi\phi$+$\phi$G 0.039 0.014
TT+GG+$\phi$G 0.037 0.025
TT+GG+$\phi\phi$ 0.028 0.015
TT+GG+$\phi\phi$ +$\phi$G 0.028 0.014
$\Lambda$CDM $\Delta \tau$ (scenario-I) $\Delta \tau$(scenario-J)
TT len 0.027 0.013
CMB le 0.004 0.003
TT+$\phi\phi$ 0.020 0.010
TT+GG 0.019 0.018
TT+$\phi$G 0.047 0.033
TT+$\phi\phi$+$\phi$G 0.020 0.010
TT+GG+$\phi$G 0.019 0.016
TT+GG+$\phi\phi$ 0.017 0.010
TT+GG+$\phi\phi$ +$\phi$G 0.016 0.007
\[LSSTtau\]
Constraints on the $\tau$ parameter for an LSST type galaxy survey (equation \[LSST\]). Constraints are shown for a number of scenarios for combinations of TT, CMB, $\phi\phi$, GG and $\phi$G. Scenarios H and I are *Planck* like, scenarios F and J are 4th generation CMB like.
In table \[LSSTtau\], we show the constraint on $\tau$ for various scenarios. Scenario-H and scenario-F use the full 7 parameter non-galactic cosmology where scenario-H includes [*Planck*]{} and scenario-F includes a 4th generation CMB survey. Also shown are scenario-I and scenario-J. These two scenarios include only the $\Lambda$CDM 6 non-galactic cosmology parameters, excluding massive neutrinos. Scenario-I includes a current type CMB survey and scenario-J includes a 4th generation CMB survey. When the neutrino mass sum is fixed, the constraint on $\tau$ improves significantly for many cases. This highlights the importance of including the neutrino mass sum as an unknown parameter in the analysis.
While $\phi$G reduces degeneracies between cosmological parameters inherent in TT+GG, the improvement in the tau constraint going from TT + GG + $\phi\phi$ to TT + GG + $\phi\phi$ + $\phi G$ is small. Note that while large scale polarization information from [*Planck*]{} or future experiments will yield the tightest constraints on $\tau$, the addition of $\phi\phi$ information alone to the temperature power spectrum (TT) improves the $\tau$ constraints significantly. The constraint from a 4th generation CMB experiment [*without*]{} large scale polarization information will be comparable to that from current large scale polarization measurements from WMAP [@Hinshaw:2012aka] and will offer a good independent check on the constraint from Planck polarization, to be published next year, which could be dominated by uncertainty in the polarization foregrounds.
Figure \[fig:ellipse\_nu\_vs\_lam\] shows the 1$\sigma$ and 2$\sigma$, $\sum m_{\nu}$ vs $\Omega_{\Lambda}$ ellipses for scenario-D. (Galaxy survey with a Gaussian galaxy distribution and a number density of 1,000 gal/deg$^{2}$, *Planck* + 4th generation CMB experiments). As expected, adding the CMB lensing auto correlation to the primary CMB improves the constraint dramatically, due to the breaking of the geometric degeneracy ([@Stompor:1998zj], [@Sherwin:2011gv]). However, after the addition of the $\phi\phi$ correlation, additional information from both GG and $\phi$G make negligible improvement to the constraint on dark energy. In the absence of the $\phi\phi$ correlation, the addition of GG or $\phi$G also make negligible improvement on the constraint from the primary CMB alone.
![$\sum m_{\nu}$ vs $\Omega_{\Lambda}$ 1$\sigma$ (dashed line) and 2$\sigma$ (solid line) ellipses for scenario-D for various combination of the CMB ul, CMB lensing, galaxy and cross correlations. Galaxy survey is Gaussian with number density 1,000 gal/deg$^{2}$, and CMB experiment is 4th generation.[]{data-label="fig:ellipse_nu_vs_lam"}](fisher_2d_f_Omnuh2-OmLa.pdf){width="8.8cm"}
Conclusions {#conclusion}
===========
Our analysis is intended to begin to answer the open question of how much information about cosmology is contained in the CMB lensing - galaxy cross correlation. We have used a simple Fisher model where we included all correlations between temperature, polarization, galaxy over-density, and CMB lensing maps. We constrain both galactic and non-galactic cosmological parameters simultaneously, to explore degeneracies in these spectra between the underlying cosmology and galaxy dynamics.
In the case of the galaxy parameters, the $\phi$G cross correlation is very useful. In all cases the addition of the $\phi G$ cross correlation to the $\phi\phi$ and GG auto correlations improves the constraints significantly, and CMB ul + $\phi G$ is comparable to CMB ul + GG. We also find that the addition of CMB lensing information or any improvements in the primary CMB significantly improve the constraints on galaxy parameters. This is because the CMB lensing and primary CMB information give tighter constraints on the non-galactic cosmological parameters, allowing degeneracies with galaxy parameters to be broken.
For non-galactic cosmology, we find that in the case of the $\sum m_{\nu}$, once the $\phi\phi$ and GG auto correlations are considered, the $\phi$G cross correlation adds little improvement. We have imposed a conservative $l_{\rm max}=500$ for the $\phi$G cross correlation however, and see that when this is relaxed, the cross correlation does offer useful information, up to the same level as the auto correlations for ambitious values of $l_{\rm max}$. Although for now it may be unrealistic to extend the cross correlation information to such small scales, with some improvement in the non-linear modeling one could hope that the cross correlation provides useful independent checks of constraints derived from the GG auto-correlation. In table \[LSSTtau\] we also quantify the near future constraints available on the $\tau$ parameter when the CMB lensing information is used. We note that the TT + $\phi\phi$ correlation will constrain $\tau$ to $<$20%, which will provide a good independent check on the epoch of reionization, given galactic foreground uncertainties in the polarization constraints.
The authors thank (in alphabetical order) Carlos Cunha, Sudeep Das, Gill Holder, Antony Lewis, Adam Lidz, Blake Sherwin, Alberto Vallinotto, Kimmy Wu, and Amanda Yoho for useful discussions and comments on a draft. RP acknowledges support from the Science and Technology Facilities Council via a research studentship and thanks Prof. Chao-Lin Kuo’s group at SLAC/Stanford where they were hosted at the time of this work. OZ acknowledges support by an Inaugural Fellowship from the Berkeley Center for Cosmological Physics as well as by the National Science Foundation through grants ANT-0638937 and ANT-0130612.
[41]{} natexlab\#1[\#1]{}bibnamefont \#1[\#1]{}bibfnamefont \#1[\#1]{}citenamefont \#1[\#1]{}url \#1[`#1`]{}urlprefix\[2\][\#2]{} \[2\]\[\][[\#2](#2)]{}
, , , ****, (), .
, , , , , , ****, (), .
, , , , , ****, (), .
, , , , , (), .
, , , , , , ****, (), .
() (), .
, ****, (), .
(), ****, (), .
() (), .
, , , , , (), .
, , , , , (), .
(), .
, ****, (), .
() (), .
, ****, (), .
, , , , , , ****, (), .
, , , , , (), .
, , , , , , ****, (), .
() (), .
, , , , , , ****, (), .
, ****, (), .
(), .
, , , , , (), .
, , , , , ****, (), .
, ****, (), <http://link.aps.org/doi/10.1103/PhysRevD.52.4307>.
() (), .
, ****, (), .
, , , , , , ****, (), .
, , , , , , ****, (), .
, , , , , (), .
, , , , , , ****, (), .
, , (), .
(),
() (), .
() (), .
(), .
, , , (), .
, , , , ****, (), .
(), .
, ****, (), .
, , , , , , ****, (), .
| {
"pile_set_name": "ArXiv"
} |
---
author:
- 'Abderrahmane Maaradji, Marlon Dumas, Marcello La Rosa and Alireza Ostovar'
bibliography:
- 'IEEEabrv.bib'
- 'bibliography.bib'
title: Detecting Sudden and Gradual Drifts in Business Processes from Execution Traces
---
Acknowledgments {#acknowledgments .unnumbered}
===============
This research is funded by the Australian Research Council Discovery Project DP150103356 and the Estonian Research Council.
\[[{width="25mm" height="32mm"}]{}\] [Abderrahmane Maaradji]{} is a research fellow at the Queensland University of Technology. He obtained his PhD in computer science in 2011 at Universite Pierre and Marie Curie and Bell Laboratories, France. His research interests are in the area of process mining, with a focus on concept drift.
\[[{width="25mm" height="32mm"}]{}\] [Marlon Dumas]{} is Professor of Software Engineering and University of Tartu, Estonia and Adjunct Professor of Information Systems at Queensland University of Technology, Australia. His research focuses on combining data mining and formal methods for analysis and monitoring of business processes. He has published extensively across the fields of software engineering and information systems and has co-authored two textbooks on business process management.
\[[{width="25mm" height="32mm"}]{}\] [Marcello La Rosa]{} is Professor of Information Systems at the Information Systems school of the Queensland University of Technology, Brisbane, Australia. His research interests include process consolidation, mining and automation, in which he published over 100 papers. He is co-author of the textbook Fundamentals of Business Process Management (Springer, 2013).
\[[{width="25mm" height="32mm"}]{}\] [Alireza Ostovar]{} is a PhD student at the Queensland University of Technology. His PhD topic spans the fields of business process management and process mining, with a focus on process drift detection and characterization.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We (1) determine the number of Latin rectangles with 11 columns and each possible number of rows, including the Latin squares of order 11, (2) answer some questions of Alter by showing that the number of reduced Latin squares of order $n$ is divisible by $f!$ where $f$ is a particular integer close to $\frac12n$, (3) provide a formula for the number of Latin squares in terms of permanents of $(+1,-1)$-matrices, (4) find the extremal values for the number of 1-factorisations of $k$-regular bipartite graphs on $2n$ vertices whenever $1{\leqslant}k{\leqslant}n{\leqslant}11$, (5) show that the proportion of Latin squares with a non-trivial symmetry group tends quickly to zero as the order increases.'
author:
- |
Brendan D. McKay and Ian M. Wanless\
Department of Computer Science\
Australian National University\
Canberra, ACT 0200, Australia\
$\{$bdm,imw$\}[email protected]
title: On the number of Latin squares
---
Introduction {#intro}
============
For $1{\leqslant}k{\leqslant}n$, a $k\times n$ *Latin rectangle* is a $k\times n$ array $L=(\ell_{ij})$ with entries from $\{1,2,\ldots,n\}$ such that the entries in each row and in each column are distinct. Of course, $L$ is a *Latin square* if $k=n$. We say that $L$ is *reduced* if the first row is $(1,2,\ldots,n)$ and the first column is $(1,2,\ldots,k)^T$. If $\Red{k,n}$ is the number of reduced $k\times n$ Latin rectangles then $L_{k,n}$, the total number of $k\times n$ Latin rectangles, is $n!\,(n{-}1)!\,\Red{k,n}/(n{-}k)!$. We will sometimes write $\Red{n,n}$ as $\Red{n}$ and $L_{n,n}$ as $L_n$.
The determination of $\Red{k,n}$, especially in the case $k=n$, has been a popular pursuit for a long time. The number of reduced squares up to order 5 was known to Euler [@euler] and Cayley [@cayley]. McMahon [@macmahon] used a different method to find the same numbers, but obtained the wrong value for order 5. The number of reduced squares of order 6 was found by Frolov [@frolov] and later by Tarry [@tarry]. Frolov [@frolov] also gave an incorrect count of reduced squares of order 7. Norton [@norton] enumerated the Latin squares of order 7 but incompletely; this was completed by Sade [@sade0] and Saxena [@saxena2]. The number of reduced squares of order 8 was found by Wells [@wells], of order 9 by Bammel and Rothstein [@bammel].
The value of $\Red{10}$ was found first in 1990 by the amateur mathematician Eric Rogoyski working on his home computer and in the following year by the present first author. The resulting joint paper [@rogoyski] also presented the number of Latin rectangles with up to 10 columns. Before he died in 2002, Rogoyski worked for several years on the squares of order 11 but the computing power available to him was inadequate, despite his approach being sound. Given the advance in computers since then, we can now complete the computations moderately easily.
Several explicit formulas for general $n$ are in the literature ([@shao], for example). Saxena [@saxena2] succeeded in using such a formula to compute $\Red{7}$. We will give another very simple formula in Section \[formulas\]. At the time of writing, not even the asymptotic value of $\Red{n}$ is known. In the case of rectangles, the best asymptotic result is for $k=o(n^{6/7})$, by Godsil and McKay [@godsil].
Terminology
===========
It can be useful to think of a Latin square of order $n$ as a set of $n^2$ triples of the form (row, column, symbol). For each Latin square there are six [*conjugate*]{} squares obtained by uniformly permuting the coordinates in each of its triples. For example, the transpose of $L$ is obtained by swapping the row and column coordinates in each triple.
An [*isotopism*]{} of a Latin square $L$ is a permutation of its rows, permutation of its columns and permutation of its symbols. The resulting square is said to be [*isotopic*]{} to $L$ and the set of all squares isotopic to $L$ is called an [*isotopy class*]{}. In the special case when the same permutation is applied to the rows, columns and symbols we say that the isotopism is an [*isomorphism*]{}. An isotopism that maps $L$ to itself is called an [*autotopism*]{} of $L$ and any autotopism that is an isomorphism is called an [*automorphism*]{}. The [*main class*]{} of $L$ is the set of squares which are isotopic to some conjugate of $L$. Latin squares belonging to the same main class are said to be [*paratopic*]{} and a map which combines an isotopism with conjugation is called a [*paratopism*]{}. A paratopism which maps a Latin square to itself is called an [*autoparatopism*]{} of the square.
The number of isomorphism classes, isotopy classes and main classes has been determined by McKay, Meynert and Myrvold [@myrvold] for $n{\leqslant}10$. Our computation does not allow us to extract this information for $n=11$. However, we do show in Section \[s:nosym\] that $L_n/(6n!^3)$ provides an increasingly accurate estimate of the number of main classes as $n$ grows.
The Algorithm {#algo}
=============
Our approach is essentially that introduced by Sade [@sade0], adapted to the computer by Wells [@wells; @wellsbook], and slightly improved by Bammel and Rothstein [@bammel]. It was also used by McKay and Rogoyski [@rogoyski]. Given a $k\times n$ Latin rectangle $L$, we can define a bipartite graph $B(L)$ with vertices $C\cup S$, where $C=\{c_1,c_2,\ldots,c_n\}$ represents the columns of $L$ and $S=\{s_1,s_2,\ldots,s_n\}$ represents the symbols. There is an edge from $c_i$ to $s_j$ if and only if the symbol $j$ appears in column $i$ of $L$. Thus $B(L)$ is regular of degree $k$. Clearly $B(L)$ does not determine $L$ in general, since it does not record the order of the symbols in each column. For us this is an advantage, since it means there are many fewer graphs than there are Latin rectangles.
Given a regular bipartite graph $B$ on $C\cup S$ of degree $k$, let $m(B)$ be its number of 1-factorizations, counted without regard to the order of the factors. Obviously $m(B)$ is an invariant of the isomorphism class of $B$. In speaking of isomorphisms and automorphisms of such bipartite graphs, we will admit the possibilities that $C$ and $S$ are preserved setwise or that they are exchanged. (More complex mixings of $C$ and $S$ would, in principle, be possible in the case of disconnected graphs, but we have chosen to disallow them.) Using this convention, let $\Aut(B)$ be the automorphism group of $B$ and let $\B(k,n)$ be a set consisting of one representative of the isomorphism classes of bipartite graphs $B$ on $C\cup S$ of degree $k$.
The theoretical basis of our approach is summarized in the following theorem. Parts 1 and 3 were proved in [@rogoyski] and part 2 can be proved along similar lines.
\[main\]$ $\
1. The number of reduced $k\times n$ Latin rectangles is given by $$\Red{k,n} = 2nk!(n{-}k)! \sum_{B\in\B(k,n)} m(B)|\Aut(B)|^{-1}.$$ 2. The number of reduced Latin squares of order $n$ is given by $$\Red{n} = 2nk!(n{-}k)! \sum_{B\in\B(k,n)} m(B)m(\bar B)|\Aut(B)|^{-1},$$ where $\bar B$ is the bipartite complement (the complement in $K_{n,n}$) of $B$ and $k$ is any integer in the range $0{\leqslant}k{\leqslant}n$.\
3. Let $B\in\B(k,n)$ for $k{\geqslant}1$. Let $e$ be an arbitrary edge of $B$. Then $$m(B) = \sum_F m(B-F),$$ where the sum is over all 1-factors $F$ of $B$ that include $e$.
For each $k=1,2,\dots,11$ in turn we found $m(B)$ for all $\B(k,11)$ using (3) and were then able to deduce $\Red{k,11}$ from (1). The number of graphs in $\B(k,11)$ is 1, 14, 4196, 2806508 and 78322916, for $k=1,\ldots,5$, respectively. For $k{\geqslant}6$ the graphs in $\B(k,11)$ are the bipartite complements of those in $\B(11-k,11)$. The main practical difficulty was the efficient management of the fairly large amount of data. Two implementations were written in a way that made them independent in all substantial aspects (except for their reliance on nauty [@nauty] to recognise the isomorphism class of some graphs). For example, they used different edges $e$ in applying Theorem \[main\](3), so that generally different subgraphs were encountered. The execution time of each implementation was about 2 years (corrected to 1 GHz Pentium III), but they would have completed in under 2 months if about 3 GB memory had been available on the machines used.
We also ran the computations for $n{\leqslant}10$ and obtained the same results as reported in [@rogoyski]. We repeat those results, and include the new results, in Table \[tab1\]. It is unlikely that $\Red{12}$ will be computable by the same method for some time, since the number of regular bipartite graphs of order 24 and degree 6 is more than $10^{11}$.
Note that our value of $\Red{11}$ agrees precisely with the numerical estimate given in [@rogoyski], where estimates of $\Red{n}$ were given for $11{\leqslant}n{\leqslant}15$.
$n$ $k$ $\Red{k,n}$ $n$ $k$ $\Red{k,n}$
----- ----- -------------- ----- ----- ------------------------------------
1 1 1 9 1 1
2 1 1 2 16687
2 1 3 103443808
3 1 1 4 207624560256
2 1 5 112681643083776
3 1 6 12952605404381184
4 1 1 7 224382967916691456
2 3 8 377597570964258816
3 4 9 377597570964258816
4 4 10 1 1
5 1 1 2 148329
2 11 3 8154999232
3 46 4 147174521059584
4 56 5 746988383076286464
5 56 6 870735405591003709440
6 1 1 7 177144296983054185922560
2 53 8 4292039421591854273003520
3 1064 9 7580721483160132811489280
4 6552 10 7580721483160132811489280
5 9408 11 1 1
6 9408 2 1468457
7 1 1 3 798030483328
2 309 4 143968880078466048
3 35792 5 7533492323047902093312
4 1293216 6 96299552373292505158778880
5 11270400 7 240123216475173515502173552640
6 16942080 8 86108204357787266780858343751680
7 16942080 9 2905990310033882693113989027594240
8 1 1 10 5363937773277371298119673540771840
2 2119 11 5363937773277371298119673540771840
3 1673792
4 420909504
5 27206658048
6 335390189568
7 535281401856
8 535281401856
\[tab1\]
Some divisibility properties of $\Red{n}$ {#divprops}
=========================================
Despite obtaining the same value repeatedly for $\Red{11}$ by applying Theorem \[main\](2) for different $k$ in two independent computations, we sought to check our answer further by determining its value modulo some small prime powers. By means of the algorithms described in [@myrvold], we computed representatives $L$ of all the isotopy classes of Latin squares of order 11 for which the order of the autotopism group $\Is(L)$ is divisible by 5, 7, or 11. The numbers of such isotopy classes are listed in Table \[tab2\]. Since the number of reduced squares in the isotopy class of $L$ is $n\,n!/|\Is(L)|$, these counts imply that $\Red{11}$ equals 8515 modulo 21175, in agreement with our computations.
$|\Is(L)|$ isotopy classes
------------ -----------------
5 55621
7 8065
10 359
11 24
14 160
20 102
21 45
22 12
55 6
60 3
1210 1
\[tab2\]
We also have the following simple divisibility properties.
\[divis\] For each integer $n{\geqslant}1$,\
1. $\Red{2n+1}$ is divisible by $\gcd(n!\,(n{-}1)!\,\Red{n},(n{+}1)!)$.\
2. $\Red{2n}$ is divisible by $n!\,$.
Consider $\Red{2n+1}$ first. We define an equivalence relation on reduced Latin squares of order $2n+1$ such that each equivalence class has size either $n!\,(n{-}1)!\,\Red{n}$ or $(n{+}1)!\,$. Let $A$ be the leading principal minor of $L=(\ell_{ij})$ of order $n$.
If $A$ is a (reduced) Latin subsquare, then the squares equivalent to $L$ are those obtainable by possibly replacing $A$ by another reduced subsquare, permuting the $n$ partial rows $(\ell_{i,n+1},\ell_{i,n+2},\ldots,\ell_{i,2n+1})$ for $1{\leqslant}i{\leqslant}n$, permuting the $n-1$ partial columns $(\ell_{n+1,j},\ell_{n+2,j},\ldots,\ell_{2n+1,j})$ for $2{\leqslant}j{\leqslant}n$ then permuting columns $n+1,n+2,\ldots,2n+1$ to put the first row into natural order. These $n!\,(n{-}1)!\,\Red{n}$ operations are closed under composition and give different reduced Latin squares, so each equivalence class has size $n!\,(n{-}1)!\,\Red{n}$.
If $A$ is not a Latin subsquare, the squares equivalent to $L$ are those obtainable by applying one of the $(n{+}1)!$ isomorphisms in which the underlying permutation fixes each of the points $1,2,\dots,n$. No isomorphism of this form can be an automorphism of a square in which $A$ is not a subsquare (see [@myrvold Theorem 1]). Hence the squares obtained are different and the equivalence class has $(n{+}1)!$ elements.
The case of $\Red{2n}$ is the same except the second argument gives $n!$ instead of $(n{+}1)!\,$.
\[divcor\] If $n=2p-1$ for some prime $p$, then $\Red{n}$ is divisible by $\lfloor(n-1)/2\rfloor!\,$. Otherwise, $\Red{n}$ is divisible by $\lfloor(n+1)/2\rfloor!\,$.
This follows from Table \[tab1\] for $n{\leqslant}8$. For $n{\geqslant}9$, note that $m\mathbin{|} (m{-}2)!$ for $m>4$ unless $m$ is prime.
Note that, for $n{\geqslant}12$, the corollary gives the best divisor that can be inferred from Table \[tab1\] and Theorem \[divis\], except that $\Red{13}$ is divisible by $7!$ and not merely by $6!$.
Alter [@alter] (see also Mullen [@mullen]) asked whether an increasing power of two divides $\Red{n}$ as $n$ increases and whether $\Red{n}$ is divisible by 3 for all $n{\geqslant}6$. Theorem \[divis\] answers both these questions in the affirmative. Indeed it shows much more — that for any integer $m>1$ the power of $m$ dividing $\Red{n}$ grows at least linearly in $n$. That is, for each $m$ there exists $\lambda=\lambda(m)>0$ such that $\Red{n}$ is divisible by $m^{\lfloor\lambda n\rfloor}$ for all $n$.
Alter also asked for the highest power of two dividing $\Red{n}$, and here we must admit our ignorance. It seems from the evidence in Table \[primefact\] that the power grows faster than linearly, but we were unable to prove this.
$\begin{array}{c|c}
n&\mbox{Prime factorisation of } \Red{n}\\
\hline
2& 1\\
3& 1\\
4& 2^2\\
5& 2^3\cdot7\\
6& 2^6\cdot3\cdot7^2\\
7& 2^{10}\cdot3\cdot5\cdot1103\\
8& 2^{17}\cdot3\cdot1361291\\
9& 2^{21}\cdot3^2\cdot5231\cdot3824477\\
10& 2^{28}\cdot3^2\cdot5\cdot31\cdot37\cdot547135293937\\
11& 2^{35}\cdot3^4\cdot5\cdot2801\cdot2206499\cdot62368028479\\
\end{array}$
\[primefact\]
A formula for $\Red{n}$ {#formulas}
=======================
The literature contains quite a few exact formulas for $\Red{n}$, but none of them appear very efficient for explicit computation (though Saxena [@saxena2] managed to compute $\Red{7}$ using such a formula).
Perhaps the simplest formulas are those in [@shao], which relate $\Red{n}$ to the permanents of all 0-1 matrices of order $n$. Here we give one that is very similar but uses $\pm 1$ matrices instead. Unlike the inclusion-exclusion proof of [@shao], we give a simple analytic proof.
\[exact\] Let $p(z)$ be any monic polynomial of degree $n$ and let $\M_n$ be the family of all $n\times n$ matrices over $\{-1,+1\}$. Then $$L_n = 2^{-n^2} \sum_{X\in\M_n} p(\Per X)\,\pi(X),$$ where $\Per X$ is the permanent of $X$ and $\pi(X)$ is the product of the entries of $X$.
If $X=(x_{ij})$ is an $n\times n$ matrix of indeterminates, then by definition $\Per X = \sum_{\sigma\in S_n} T_\sigma$ where $S_n$ is the symmetric group and $T_\sigma=x_{1\sigma(1)}x_{2\sigma(2)}\cdots x_{n\sigma(n)}$. If the polynomial $p(\Per X)$ is expanded in terms of monomials in the $x_{ij}$, then the only monomial involving every $x_{ij}$ comes from products $T_{\sigma_1}T_{\sigma_2}\cdots T_{\sigma_n}$ where the permutations $\sigma_1,\sigma_2,\ldots,\sigma_n$ are the rows of a Latin square. That is, the coefficient of the only monomial with each $x_{ij}$ having odd degree is the number of Latin squares. Multiplying by $\pi(X)$ turns the required monomial into the only one that has even degree in each $x_{ij}$. Now summing over $X\in\M_n$ causes this monomial to be multiplied by $|\M_n|=2^{n^2}$ while all the other monomials cancel out.
Extremal graphs with respect to $m(B)$
======================================
In our computations we learned the values of $m(B)$ for each graph $B\in\B(k,n)$ for $n{\leqslant}11$. In Table \[tabminmax\] we record the maximum and minimum values, and the number of graphs (in the column headed “\#”) that achieve the minimum. The maximum is achieved uniquely in all cases. Of course, for $k{\leqslant}1$ the result is trivial and when $k{\geqslant}n-1$ the unique graph has $m(B)=\Red{n}$, so we omit these cases.
In most cases, the graphs maximizing $m(B)$ are the same as those with the maximum number of perfect matchings, as listed in [@wanless]. The only exceptions are as follows, where the notation is that used in [@wanless]:
- For $n=7, k=5$ the graph maximising $m(B)$ is $\overline{2J_2\oplus D_3}$;
- For $n=9, k=6$ the graph maximising $m(B)$ is $\overline{3J_3}$;
- For $n=10, k=4$ the graph maximising $m(B)$ is $J_4\oplus\overline{3J_2}$;
- For $n=11, k=4$ the graph maximising $m(B)$ is $J_4\oplus\overline{J_3\oplus D_4}$.
In the first of these cases the cited graph does, according to [@wanless], maximise the number of perfect matchings, but does not do so uniquely.
$n$ $k$ $\min m(B)$ \# $\max m(B)$
-------------- ----- ------------------------------ ---- ------------------------------
4 2 1 1 2
\[0.5ex\] 5 2 1 1 2
3 4 1 6
\[0.5ex\] 6 2 1 1 4
3 8 4 24
4 168 1 224
\[0.5ex\] 7 2 1 1 4
3 8 3 48
4 456 2 576
5 54528 1 55296
\[0.5ex\] 8 2 1 1 8
3 16 18 96
4 1120 1 13824
5 306432 1 402432
6 251894784 1 258392064
\[0.5ex\] 9 2 1 1 8
3 16 7 288
4 2720 1 32256
5 1718784 1 2312192
6 3585925120 1 3797508096
7 22606854291456 1 22710505439232
\[0.5ex\] 10 2 1 1 16
3 24 2 576
4 6992 1 129024
5 9457472 1 216760320
6 49712734208 1 71022182400
7 920073219063808 1 962525641310208
8 51072829020284387328 1 51411315765364654080
\[0.5ex\] 11 2 1 1 16
3 32 25 1152
4 17040 1 331776
5 49449728 1 1517322240
6 656992907264 1 1274550681600
7 36184087678025728 1 41312188744335360
8 6674288352734540070912 1 6904895678779049902080
9 3650989756490710602617978880 1 3665106903315598519509712896
\[tabminmax\]
Proportion of Latin squares with symmetry {#s:nosym}
=========================================
In this section we prove that the proportion of order $n$ Latin squares which have a non-trivial symmetry tends very quickly to zero as $n\rightarrow\infty$.
\[nosym\] The proportion of Latin squares of order $n$ which have a non-trivial autoparatopy group is no more than $$\label{e:nosymprob}
n^{-3n^2/8+o(n^2)}.$$
Suppose that a Latin square $L=(\ell_{ij})$ of order $n$ has a non-trivial autoparatopy group. Then by Lemma 4 in [@myrvold], $L$ has a autoparatopism $\alpha$ which fixes (pointwise) no more than one quarter of the triples of $L$.
The number of possibilities for $\alpha$ is less than $6n!^3=o(n^{3n})$. Given $\alpha$, we can construct each possible $L$ row by row. Each entry is determined either by $\alpha$ and a previous entry, or can be chosen in at most $n$ ways. The latter possibility occurs once per orbit of $\alpha$, and since $\alpha$ fixes at most $\frac14$ of the triples of $L$, the number of orbits is at most $(\frac14+\frac34\cdot\frac12)n^2=\frac58n^2$. In total we find that there are most $$o(n^{3n})n^{5n^2/8}$$ Latin squares with non-trivial autoparatopy group. Our result now follows immediately from the well known lower bound for $L_n$ (see, for example, Thm 17.2 in [@vLW]) that says that $$L_n{\geqslant}(n!)^{2n}n^{-n^2}{\geqslant}n^{n^2-o(n^2)}.$$
As a corollary to this last result, we see that the proportion of main classes or of isotopy classes or of isomorphism classes whose members have a non-trivial autoparatopy group is also bounded by . This is because each such class has somewhere between $1$ and $6n!^3=o(n^{3n})$ members.
Another corollary is that the number of isomorphism classes, isotopy classes and main classes of Latin squares of order $n$ will be asymptotic to $L_n/n!$, $L_n/n!^3$ and $L_n/(6n!^3)$, respectively.
= =
[99]{}
[0.3em ]{}
R.Alter, How many Latin squares are there? [*Amer. Math. Monthly*]{}, [**82**]{} (1975) 632–634.
S.E.Bammel and J.Rothstein, The number of $9\times9$ Latin squares, [*Discrete Math.*]{}, [**11**]{} (1975) 93–95.
J.W.Brown, Enumeration of Latin squares with application to order 8, [*J. Combinatorial Theory*]{}, [**5**]{} (1968) 177–184.
A.Cayley, On Latin squares, [*Oxford Camb. Dublin Messenger of Math.*]{}, [**19**]{} (1890) 135–137.
L.Euler, Recherches sur une nouvelle espèce de quarrés magiques, [*Verh. v. h. Zeeuwsch Genoottsch. der Wetensch., Vlissingen*]{}, [**9**]{} (1782) 85–239.
M.Frolov, Sur les permutations carrés, [*J. de Math. spéc.*]{}, [**IV**]{} (1890) 8–11, 25–30.
C.D.Godsil and B.D.McKay, Asymptotic enumeration of Latin rectangles, [*J. Combinatorial Theory, Ser. B*]{}, [**48**]{} (1990) 19–44.
P.A.MacMahon, Combinatory Analysis, Cambridge, 1915.
G.L.Mullen, How many $i$-$j$ reduced Latin squares are there? [*Amer. Math. Monthly*]{}, [**85**]{} (1978) 751–752.
B.D.McKay, `nauty` graph isomorphic software, available at\
[http://cs.anu.edu.au/bdm/nauty]{}.
B.D.McKay, A.Meynert and W.Myrvold, Small Latin squares, quasigroups and loops, [*J. Combin. Des.*]{} [**15**]{} (2007), 98–119.
B.D.McKay and E.Rogoyski, Latin squares of order 10, [*Electronic J. Combinatorics*]{}, [**2**]{} (1995) \#N3 (4 pp).
B.D.McKay and I.M.Wanless, Maximising the permanent of $(0,1)$-matrices and the number of extensions of Latin rectangles, [*Electronic J. Combinatorics*]{}, [**5**]{} (1998) \#R11 (20 pp).
H.W.Norton, The $7\times7$ squares, [*Ann. Eugenics*]{}, [**9**]{} (1939) 269–307.
A.Sade, [*Enumération des carrés latins. Application au $7^{\mbox{\`eme}}$ ordre. Conjectures pour les ordres supérieurs*]{}, privately published, Marseille, 1948, 8pp.
P.N.Saxena, A simplified method of enumerating Latin squares by MacMahon’s differential operators; II. The $7\times 7$ Latin squares, [*J. Indian Soc. Agric. Statistics*]{}, [**3**]{} (1951) 24–79.
J.Y.Shao and W.D.Wei, A formula for the number of Latin squares, [*Discrete Math.*]{}, [**110**]{} (1992) 293–296.
G.Tarry, Le problème des 36 officiers, [*Ass. Franç. Paris*]{}, (1900) 29, 170-203.
J.H.van Lint and R.M.Wilson, [*A course in combinatorics*]{}, Cambridge University Press, 1992.
M.B.Wells, The number of Latin squares of order eight, [*J. Combinatorial Theory*]{}, [**3**]{} (1967) 98–99.
M.B.Wells, Elements of combinatorial computing. Pergamon Press (Oxford-New York-Toronto, 1971).
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Let $G$ be a group and let $F$ be a field of characteristic different from $2$. Denote by $(FG)^+$ the set of symmetric elements and by $\mathcal{U}^+(FG)$ the set of symmetric units, under an oriented classical involution of the group algebra $FG$. We give some lower and upper bounds on the Lie nilpotency index of $(FG)^+$ and the nilpotency class of $\mathcal{U}^+(FG)$.'
address: 'John H. Castillo, Departamento de Matemáticas y Estadística, Universidad de Nariño'
author:
- 'John H. Castillo'
bibliography:
- 'john.bib'
title: Lie nilpotency indices of symmetric elements under oriented involutions in group algebras
---
Introduction
============
Let $FG$ denote the group algebra of a group $G$ over a field $F$ with $char(F)=p\neq 2$. A homomorphism $\sigma:G\rightarrow \{ \pm
1\}$ is called an *orientation* of the group $G$. Working in the context of $K$-theory, Novikov [@NOV], introduced an *oriented involution* $*$ of $FG$, given by $$\left(\sum_{g\in G} \alpha_gg\right)^{*}=\sum_{g\in G}\alpha_g\sigma(g)g^{-1}.$$ When $\sigma$ is trivial this involution coincides with the so called *classical involution* of $FG$.
We denote $(FG)^{+}=\{\alpha\in FG:\alpha^{*}=\alpha\}$ and $(FG)^{-}=\{\alpha\in FG:\alpha^{*}=-\alpha\}$ the set of symmetric and skew-symmetric elements of $FG$ under $*$, respectively. We denote by $N$ the kernel of $\sigma$. It is obvious that the involution $*$ coincides on the group algebra $FN$ with the classical involution. It is easy to see that, as an $F$-module, $(FG)^{+}$ is generated by the set $$\begin{aligned}
\mathcal{S}=\{g+g^{-1}:g\in N \}\cup\{g-g^{-1}:g\in G\setminus N,g^2\neq 1\}\end{aligned}$$ and $(FG)^{-}$ is generated by $$\begin{aligned}
\mathcal{L}=\{g+g^{-1}:g\in
G\setminus N\}\cup\{g-g^{-1}:g\in N,g^2\neq 1\}.\end{aligned}$$
Given $g_1, g_2\in G$, we define the commutator $(g_1,g_2)=g_1^{-1}g_2^{-1}g_1g_2$ and recursively, $(g_1,\ldots,g_{n})=((g_1,\ldots,g_{n-1}),g_{n})$ for $n$ elements $g_1,\ldots,g_n$ of $G$. By the commutator $(X,Y)$ of the subsets $X$ and $Y$ of $G$ we mean the subgroup of $G$ generated by all commutators $(x,y)$ with $x\in X$, $y\in Y$. In this way, we can define the lower central series of a nonempty subset $H$ of $G$ by: $\gamma_1(H)=H$ and $\gamma_{n+1}(H)=(\gamma_n(H),H)$, for $n\geq
1$. We say that $H$ is nilpotent if $\gamma_n(H)=1$, for some $n$. For a nilpotent subset $H\subseteq G$ the number $\cl(H)=\min\{n\in
\mathbb{N}_0: \gamma_{n+1}(H)=1\}$ is called the nilpotency class of $H$. It can be proved that $H$ is a nilpotent set if and only if $H$ satisfies the group identity $(g_1,\ldots,g_n)=1$ for some $n\geq
2$.
In an associative ring $R$, the Lie bracket on two elements $x,y
\in R$ is defined by $[x,y]=xy-yx$. This definition is extended recursively via $[x_1,\ldots,x_{n+1}]=[[x_1,\ldots,x_n],x_{n+1}]$. For $X,Y\subseteq R$ by $[X,Y]$ we denote the additive subgroup generated by all Lie commutators $[x,y]$ with $x\in X, y\in Y$. The lower Lie central series of a nonempty subset $S$ of $R$ is defined inductively by setting $\gamma^1(S)=S$ and $\gamma^{n+1}(S)=[\gamma^n(S),S]$. We say that the subset $S$ is Lie nilpotent if there exists a natural number $n$, such that $\gamma^n(S)=0$. The smallest natural number with the last property, denoted by $t(S)$, is called the Lie nilpotency index of $S$. It is possible to show that $S$ is Lie nilpotent if and only if $S$ satisfies the polynomial identity $[x_1,\ldots,x_n]=0$ for some $n\geq 2$.
Given a nonempty subset $S$ of $R$, we let $S^{(1)}=R$, and then for each $i\geq 2$, let $S^{(i)}$ be the (associative) ideal of $R$ generated by all elements of the form $[a,b]$, with $a\in
S^{(i-1)},b\in S$. We say that $S$ is strongly Lie nilpotent if $S^{(i)}=0$ for some $i$. The minimal $n$ for which $S^{(n)}=0$ is called the upper Lie nilpotency index and denoted by $\tl(S)$. Clearly, strong Lie nilpotence implies Lie nilpotence and $\t(S)\leq
\tl(S)$. Denote by $\mathcal{U}(S)$ the set of units in the subset $S$ of $R$ and suppose that it is nonempty. By the equality $(x,y)=1+x^{-1}y^{-1}[x,y]$, it is easy to see that $\gamma_n(\mathcal{U}(S))\subseteq 1+S^{(n)}$ for all $n\geq 2$. In consequence, the set of units of a strongly Lie nilpotent subset $S$ is nilpotent, and $$\label{eq1}
\cl(\mathcal{U}(S))<\tl(S).$$
In 1973, Passi, Passman and Sehgal [@PPS73] showed that the group algebra $FG$ is Lie nilpotent if and only if $G$ is nilpotent and $G'$ is a finite $p$-group, where $p$ is the characteristic of $F$. Actually, see [@S78], a group algebra is Lie nilpotent if and only if it is strongly Lie nilpotent. Next, S.K. Sehgal characterized group algebras which are Lie $n$-Engel, for some $n$.
In 1993, Giambruno and Sehgal [@SG93] began the study of Lie nilpotence of symmetric and skew-symmetric elements under the classical involution. They proved that given a group $G$ without elements of order $2$ and a field $F$ with $char(F)\neq 2$, if either $(FG)^+$ or $(FG)^-$ is Lie nilpotent, then $FG$ is Lie nilpotent. This work was completed by G.T. Lee [@Lee99], for groups in general. More specifically, he proved that the Lie nilpotence of the symmetric elements under the classical involution is equivalent to the Lie nilpotence of $FG$ when the group $G$ does not contain a copy of $Q_8$, the quaternion group of order $8$ and he also characterized the group algebras such that the set of symmetric elements is Lie nilpotent when $G$ contains a copy of $Q_8$.
Recently, Castillo and Polcino Milies, see [@CP12], studied Lie properties of the symmetric elements under an oriented classical involution. They extended some previous results from [@SG93], [@Lee99] and [@Lee00]. In particular, they gave some groups algebras such that the Lie nilpotence of the symmetric set implies the same property in the whole group algebra. Also, they obtained a complete characterization of the group algebras $FG$, such that $Q_8\subseteq G$ and $(FG)^+$ is Lie nilpotent.
Lately, Z. Balogh and T. Juhász in [@BJ11] and [@BJ12] studied the Lie nilpotency index of $(FG)^+$ and the nilpotency class of the $\mathcal{U}^+(FG)$ under the classical involution in group algebras. They gave a necessary condition to the numbers $\t((FG)^+)$ and $\cl(\mathcal{U}^+(FG))$ be maximal, as possible, in a nilpotent group algebra. Also, they studied this two numbers to group algebras such that $(FG)^+$ is Lie nilpotent but $FG$ is not.
In this article we study the Lie nilpotency index of $(FG)^+$ and the nilpotency class of $\mathcal{U}^+(FG)$ under an oriented classical involution. In the next section we give some preliminary results. In the third section we study the numbers $\t((FG)^+)$ and $\cl(\mathcal{U}^+(FG))$ in Lie nilpotent group algebras. In the fourth section we study the case when $Q_8\subseteq G$ and $(FG)^+$ is Lie nilpotent.
Throughout this paper $F$ will always denote a field of characteristic not $2$, $G$ a group and $\sigma$ a nontrivial orientation of $G$. In a number of places, all over this paper, we use arguments from [@BJ11], [@BJ12] and [@Lee10]. Some of them are reproduced here for the sake of completeness.
Preliminaries
=============
We recall the following result from [@Lee10].
\[lema2.1\] Let $R$ be a ring and $S$ a subset of $R$. Suppose, for some $i\geq
1$, that $S^{(i)}\subseteq zR$, where $z$ is central in $R$. Then for all $j>0$, we have $S^{(i+j)}\subseteq zS^{(j)}$. In particular, for any positive integer $m$, $S^{(mi)}\subseteq z^m R$.
The proof is by induction on $j$. If $j=1$, then $S^{(i+1)}\subseteq
S^{(i)}$, there is nothing to do. Assume that $S^{(i+j)}\subseteq
zS^{(j)}$. Take $a\in S^{(i+j)}, b\in S$. So $a=za_1$, for some $a_1\in S^{(j)}$. Thus, $[a,b]=[za_1,b]=z[a_1,b]\in zS^{(i+j)}$, as we want to prove.
To get the second part, notice that $$S^{(2i)}=S^{(i+i)}\subseteq zS^{(i)}\subseteq z^2R.$$ Suppose that $S^{((m-1)i)}\subseteq z^{m-1}R$. So $S^{(mi)}=S^{((m-1)i+i)}\subseteq zS^{((m-1)i)}\subseteq z^m R$.
Throughout this article we denote by $Q_8=\left<x,y:x^4=1,x^2=y^2,x^y=x^{-1}\right>$ the quaternion group of order $8$. Castillo and Polcino Milies [@CP12] characterized the group algebras of groups containing $Q_8$ and with a nontrivial orientation, such that $(FG)^+$ is Lie nilpotent. Here we prove that the conditions obtained by them are also satisfied when $(FG)^+$ is strongly Lie nilpotent.
\[teorliengelQ8+\] Let $F$ be a field of characteristic $p\neq2$, $G$ a group with a nontrivial orientation $\sigma$ and $x,y$ elements of $G$ such that $\left< x,y\right>\simeq Q_8$. Then $(FG)^{+}$ is strongly Lie nilpotent if and only if either
(i) $char(F)=0$, $N\simeq Q_8\times E$ and $G\simeq \left<Q_8,g\right> \times E$, where $E^2=1$ and $g\in
G\setminus N$ is such that $(g,x)=(g,y)=1$ and $g^2=x^2$; or,
(ii) $char(F)=p>2$, $N\simeq Q_8\times E\times P$, where $E^2=1$, $P$ is a finite $p$-group and there exists $g\in G\setminus N$ such that $G\simeq
\left<Q_8,g\right>\times E\times P$, $(g,x)=(g,y)=1$ and $g^2=x^2$.
If $(FG)^+$ is strongly Lie nilpotent, then $(FG)^+$ is Lie nilpotent and from [@CP12 Theorem 4.2] we get (i) and (ii).
Conversely, assume that $|P|=p^n$. We claim that, $((FG)^+)^{(2p^n)}=0$. The proof will be by induction on $n$. If $n=0$, then $G\simeq \left<Q_8,g\right>\times E$ and thus, from [@CP12 Lemma 4.3], $(FG)^+$ is commutative. Assume that $|P|=p^n>1$. Take $z\in \zeta(P)$ with $o(z)=p$, applying our inductive hypothesis on $\overline{G}=G/\left<z\right>$. Then, $((F\overline{G})^+)^{(2p^{n-1})}=0$. Thus $$((FG)^+)^{(2p^{n-1})}\subseteq \Delta(G,\left<z\right>)=(z-1)FG.$$ By Lema \[lema2.1\], $$((FG)^+)^{(2p^n)}\subseteq (z-1)^pFG=0,$$ as we claimed.
From the equality, $(x,y)=1+x^{-1}y^{-1}[x,y]$ we know that $\gamma_n(\mathcal{U}^+(FG))\subseteq 1+((FG)^+)^{(n)}$ and thus we get the following.
Let $F$ be a field of characteristic different from $2$. Assume that $Q_8\subseteq G$ and $(FG)^+$ is Lie nilpotent. Then, $\mathcal{U}^+(FG)$ is nilpotent.
We need the following easy observation.
\[lema5.1.1\] Let $G$ be a group, $H$ any subgroup and $A$ a normal subgroup such that $A\subseteq N$. If $(FG)^+$ is Lie nilpotent, then so are $(FH)^+$ and $(F(G/A))^+$. Furthermore, $\t((FH)^+)\leq \t((FG)^+)$ and $\t((F(G/A))^+)\leq \t((FG)^+)$.
Note that $(FH)^+$ is a subset of $(FG)^+$, and thus it has the required properties.
Since $A$ is a normal subgroup contained in the kernel of the orientation $\sigma$, we can define in $F(G/A)$ an induced oriented classical involution from $*$ in $FG$ as follows: $$\left(\sum_{\bar{g}\in G/A}\alpha_g\bar{g}\right)^{\star}= \sum_{\bar{g}\in G/A}\alpha_g\sigma(g)\bar{g}^{-1}.$$
Now, simply observe that the symmetric elements in $F(G/A)$, under $\star$, are linear combinations of terms of the form $gA+\sigma(g)g^{-1}A$, with $g\in G$. That is, every element of $(F(G/A))^+$ is the homomorphic image of an element of $(FG)^+$ under the natural map $\varepsilon_A:FG\rightarrow F(G/A)$, defined by $\varepsilon_A(\sum_{g\in G} \alpha_g g)=\sum_{g\in G} \alpha_g \bar{g}$.
So assume that $(FG)^+$ is Lie nilpotent, therefore there exists $n=\t((FG)^+)$ such that $[\alpha_1,\ldots,\alpha_n]=0$ for all $\alpha_i\in (FG)^+$. Let $\beta_1,\ldots,\beta_n\in (F(G/A))^+$. Thus $$\begin{aligned}
[\beta_1,\ldots,\beta_n]&=[\varepsilon_A(\alpha_1),\ldots,\varepsilon_A(\alpha_n)]\\
&=\varepsilon_A([\alpha_1,\ldots,\alpha_n])=\varepsilon_A(0)=0.\end{aligned}$$ Consequently, $\t((F(G/A))^+)\leq \t((FG)^+)$.
Lie nilpotent group algebras
============================
In this section we assume that $FG$ is Lie nilpotent. By [@SHB92], $\tl(FG)\leq |G'|+1$ and by [@BS04] the equality holds if and only if $G'$ is cyclic, or $G'$ is a noncentral elementary abelian group of order $4$.
Note that a group $G$ of odd finite order has trivial orientation. Indeed, let $a$ be an element of $G$. So $1=\sigma(a^{|G|})=\sigma(a)^{|G|}$ and as $|G|$ is odd we get that $\sigma(a)=1$. For the last reason when $G$ is a group of odd finite order, the involution $*$ is the classical involution. In this way, we can use the following result, that is a combination from [@BJ11 Lemma 2] and [@BJ12 Lemma 2].
\[lema2Juhasz\] Let $G$ be a finite $p$-group with a cyclic derived subgroup. Then $\t((FG)^+)\geq |G'|+1$ and $\cl(\mathcal{U}^+(FG))\geq |G'|$.
We recall that a group $G$ is called $p$-abelian if $G'$, the commutator subgroup of $G$, is a finite $p$-group and $0$-abelian means abelian.
Let $FG$ be a Lie nilpotent group algebra of odd characteristic and nontrivial orientation. Then, $\t((FG)^+)=|G'|+1$ if and only if $G'$ is cyclic. Moreover, assuming that $G$ is a torsion group, $\cl(\mathcal{U}^+(FG))=|G'|$ if and only if $G'$ is cyclic.
Assume that $\t((FG)^+)=|G'|+1$. As $G'$ is a finite $p$-group, if $G'$ is not cyclic, from [@BS04], we know that $t((FG)^+)\leq
\tl(FG)<|G'|+1$ and we get a contradiction. Thus, $G'$ is cyclic.
Conversely, suppose that $G'$ is cyclic. By the hypotheses, $G$ is a nilpotent $p$-abelian group and from [@BJ08 Lemma 1] there exists a finite $p$-group $P$ which is isomorphic to a subgroup of factor group of $G$ and $P'\simeq G'$. Actually, from the proof of [@BJ08 Lemma 1], we know that $P\simeq H/A$, where $A$ is a maximal torsion-free central subgroup of $G$.
Assume that there exists $g\in A$ such that $\sigma(g)=-1$. In this way, as $G=N\cup gN$, we get $G'=N'$. Using in $FP$ the classical involution, by lemmas \[lema2Juhasz\] and \[lema5.1.1\], we obtain that $$|G'|+1=|N'|+1=|P'|+1\leq \t((FP)^+)\leq \t((FN)^+)\leq \t((FG)^+).$$ In the other hand, suppose that $A\subseteq N$. Then we can define an induced oriented classical involution in $P\simeq H/A$, from that one in $FG$. Consequently, $$|G'|+1=|P'|+1\leq \t((FP)^+)\leq \t((FG)^+).$$
The proof of the second part is similar.
Groups that contain a copy of $Q_8$
===================================
We assume that $Q_8\subseteq G$ and $(FG)^+$ is Lie nilpotent. This means that the group algebra $FG$ is not Lie nilpotent. Recently, this kind of group algebras was characterized by Castillo and Polcino Milies [@CP12]. This characterization is the same as in Theorem \[teorliengelQ8+\], so during this section we assume that $G$ is as in that result. In this section, we will study the Lie nilpotency index of the symmetric elements under oriented classical involutions.
It is easy to show that $$\label{eq5.5}
g^m-1\equiv m(g-1)\pmod {\Delta(G)^2}.$$ for every $g\in G$ and any integer $m$.
We begin with the following result.
\[lema5Juhasz\] Consider $FG$ with an oriented classical involution. Then $$((FG)^+)^{(n)}\subseteq FG\Delta(P)^n$$ for all $n\geq 2$
Recall that the symmetric elements are spanned as an $F$-module by the set $$\mathcal{S}=\{z+z^{-1}: z\in N\}\cup \{z-z^{-1}:z\in
G\setminus N\}.$$ If $z\in N$, then $z=ah$ with $a\in Q_8\times E$ and $h\in P$. Note that if $a^2h=1$, then $h=1$ and $a^2=1$. Thus, $a\in \zeta(Q_8\times E)$. Assuming $a^2h\neq 1$, follows that $z+z^{-1}=ah+a^{-1}h^{-1}=ah+a^3h^{-1}=a(h+a^2h^{-1})$.
Also, if $z\in G\setminus N$; we can write $z=gah$ with $a\in
Q_8\times E$ and $h\in P$. If $a^2h=1$, then $a^2=h=1$. Again, $a\in
\zeta(Q_8\times E)$ and thus $z-z^{-1}=gah-g^{-1}a^{-1}h^{-1}=ga-g^{-1}a=ga(1-g^2)\in
\zeta(Q_8\times E)$. Now we suppose that $a^2h\neq 1$ and we get the following cases:
1. If $a^2=1$ and $h\neq 1$, then $z-z^{-1}=gah-g^{-1}a^{-1}h^{-1}=ag(h-g^2h^{-1})$.
2. If $a^2\neq 1$ and $h=1$, then $z-z^{-1}=gah-g^{-1}a^{-1}h^{-1}=ga-g^3a^3=ga-ga=0$.
3. If $a^2\neq 1$ and $h\neq 1$, then $z-z^{-1}=gah-g^{-1}a^{-1}h^{-1}=agh-a^3g^3h^{-1}=ag(h-h^{-1})$, because $a^3g^3=ag$.
From the above considerations, we obtain that $$\begin{aligned}
\mathcal{S}=\mathcal{A}\cup \mathcal{B}\cup\mathcal{C} \cup\zeta(Q_8\times
E),
\end{aligned}$$ where $$\begin{aligned}
\mathcal{A}&=\{a(h+a^2h^{-1}): a\in Q_8\times E, h\in P \text{ and }
a^2h\neq 1 \},\\
\mathcal{B}&=\{ag(h-g^2h^{-1}):a\in Q_8\times E , h\in P \text{ and
}
(a^2=1 \text{ and } h\neq 1) \},\\
\mathcal{C}&=\{ag(h-h^{-1}):a\in Q_8\times E , h\in P \text{ and }
(a^2\neq 1 \text{ and } h\neq 1)\}.\end{aligned}$$
Given $a\in Q_8\times E$, such that $a^2\neq 1$ we know that $1+a^2$ is symmetric and $a^2\in \zeta(Q_8\times E)$. In this way, $$a(h+a^2h^{-1})+1+a^2=a(h-1)+a^3(h^{-1}-1)+1+a+a^2+a^3,$$ where $1+a+a^2+a^3$ is a central element in $FG$ and $a(h-1)+a^3(h^{-1}-1)\in FG\Delta(P)$. It is clear that, $ag(h-h^{-1})\in FG\Delta(P)$. Furthermore, if $a^2=1$ and $h\neq
1$, then $ag(h-g^2h^{-1})=ag(h-1)-ag^3(h^{-1}-1)+a(g-g^{-1})\in
FG\Delta(P)+\zeta(FG)$.
So $$\begin{aligned}
\mathcal{\widetilde{S}}=\mathcal{A}'\cup \mathcal{B}\cup \mathcal{C} \cup\zeta(Q_8\times
E),\end{aligned}$$ also spans $(FG)^+$ as an $F$-module, where $$\mathcal{A}'=\{a(h+a^2h^{-1})+1+a^2: a\in Q_8\times E, h\in P
\text{ and } a^2h\neq 1 \}$$ and $\mathcal{B}, \mathcal{C}$ are as above.
In consequence, $$\label{contencion1} (FG)^+\subseteq
FG\Delta(P)+\zeta(FG).$$ The proof follows by induction on $n$. Indeed, if $n=2$ $$\begin{aligned}
[(FG)^+,(FG)^+]\subseteq [FG\Delta(P),FG\Delta(P)]\subseteq
FG\Delta(P)^2.\end{aligned}$$ Suppose that the lemma is true for some $n\geq 2$. Take $\alpha\in
((FG)^+)^{(n)}$ and $\beta \in (FG)^+$. So $$\begin{aligned}
[\alpha,\beta] \in [FG\Delta(P)^n,FG\Delta(P)]\subseteq
FG\Delta(P)^{n+1}.\end{aligned}$$ and we get that $((FG)^+)^{(n+1)}\subseteq FG\Delta(P)^{n+1}$ as required.
Denote by $c$ the central element of $Q_8\times E$, such that $(Q_8\times E)^2=\left<c\right>$. Given $n\geq 2$, we denote with $M_n$ the $F$-subspace of the vector space $FG$ generates by the set $$\{(h_1-h_1^{-1})\cdots (h_n-h_n^{-1})(1-c)a:h_1,\ldots,h_n\in P, a\in (Q_8\times E)\setminus \zeta(Q_8\times E)\}.$$ To simplify, we write $f_{1,\ldots,n}$ instead of $(h_1-h_1^{-1})\cdots (h_n-h_n^{-1})$.
Let $S_n$ be the symmetric group of degree $n$ and $FS_n$ its group algebra over the field $F$. It is possible to define a group action of $S_n$ on $M_n$ via: for a $\sigma\in S_n$ and a generator element $f_{1,\ldots,n}(1-c)a$ of $M_n$ let $$\sigma\cdot
f_{1,\ldots,n}(1-c)a=f_{\sigma(1),\ldots,\sigma(n)}(1-c)a.$$ Naturally, this group action on a generator set of $M_n$ can be extended linearly to the whole $M_n$. We extend this group action to a group algebra action: for $x=\sum_{\sigma\in
S_n}\alpha_{\sigma}\sigma \in FS_n$ and $z\in M_n$, let $$x\cdot z=\sum_{\sigma\in S_n}\alpha_{\sigma}(\sigma\cdot z).$$ For $n\geq 2$ we define the elements $x_{2,n},x_{3,n},\ldots,x_{n,n}$ of $FS_n$ recursively as: $$\begin{aligned}
&x_{2,n}=1+(2,1),\\
&x_{i,n}=x_{i-1,n}+x_{i-1,n}(i,i-1,\ldots,1); \text{ for $3\leq
i\leq n$.}\end{aligned}$$
Since $(FN)^+\subseteq (FG)^+$, from Lemma 4 and Lemma 5 in [@BJ12], we get the following results.
$x_{n,n}M_n\in \gamma^n((FG)^+)(1-c)$ for all $n\geq 2$.
\[lema6Juhasz\] If $|P|=p^k$, then $\widehat{P}(1-c)a\in \gamma^{k(p-1)}((FG)^+)$ for some $a\in Q_8\times E$
We recall that the augmentation ideal $\Delta(P)$ of a finite $p$-group $P$ is a nilpotent ideal, see [@PS02 Theorem 6.3.1], we will denote by $\tnil(P)$ its nilpotency index. Also, we remind that a finite $p$-group $P$, is called *powerful* if $P'\subseteq P ^p$. Let $P$ be a powerful group. We denote with $D_i=D_i(FP)$ the $i$-th dimensional subgroup. By Theorem 5.5 in [@J41], $D_1=P$ and for $n>1$, $$D_n=\left<(D_{n-1},P),(D_{\lceil \frac{n}{p}\rceil})^p\right>.$$ It can be showed that, $(P^{p^i})^{p^j}=P^{p^{i+j}}$ and $(P^{p^i},P)\subseteq P^{p^{i+1}}$ for every pair $i,j$. So, if $p^{i-1}<n\leq p^i$ then $D_n=P^{p^i}$.
\[lema5.3.4\] Let $P$ be a powerful group and $h_i-1\in\Delta(P)^{k_i}$ and $h_j-1\in \Delta(P)^{k_j}$, where $k_i$ and $k_j$ are positive integers. Then $$\label{eq5.24}
(h_i-1)(h_j-1)\equiv (h_j-1)(h_i-1)\pmod{\Delta(P)^{k_i+k_j+1}}.$$
First, we prove that $(D_i,D_j)\subseteq D_{i+j+1}$, for every $i,j$. Take $h_i\in D_i$ and $h_j\in D_j$. We get the following equation $$\label{eq5.25}
(h_i,h_j)-1=h_i^{-1}h_j^{-1}((h_i-1)(h_j-1)-(h_j-1)(h_i-1)).$$
If either $i$ or $j$, say $i$, is not a power of $p$, then $h_i\in
D_i=D_{i+1}$, so by , $(h_i,h_j)-1\in
\Delta(P)^{i+j+1}$; thus $(h_i,h_j)\in D_{i+j+1}$. If both $i$ and $j$ are powers of $p$, then $i+j$ cannot be a power of $p$ and consequently $D_{i+j}=D_{i+j+1}$. By follows $(h_i,h_j)\in D_{i+j+1}$; therefore our claim is proved.
Let $h_i-1\in \Delta(P)^{k_i}$ and $h_j-1\in \Delta(P)^{k_j}$ for some positive integers $k_i,k_j$. Then $$(h_i-1)(h_j-1)=(h_j-1)(h_i-1)+h_jh_i((h_i,h_j)-1),$$ and as $(h_i,h_j)\in D_{k_i+k_j+1}$, the result follows.
Now we can prove our main result in this section.
Let $F$ be a field of characteristic $p>2$. Consider the group algebra $FG$ with an oriented classical involution. Assume that $Q_8\subseteq G$, $(FG)^+$ is Lie nilpotent and the Sylow $p$-group $P$ of $G$ is of order $p^m$, with $m\geq 1$. Then
(i) $1+m(p-1)\leq \t((FG)^+)\leq \tl((FG)^+)\leq \tnil(P)$ and $\cl(\mathcal{U}^+(FG))\leq \tnil(P)-1$.
(ii) If $\t((FG)^+)=\tnil(P)$, then $\cl(\mathcal{U}^+(FG))+1=\t((FG)^+)$.
(iii) If $P$ is powerful, then $\t((FG)^+)=\tnil(P)$.
(iv) If $P$ is abelian, then, for all $k\geq 2$, the $F$-space $\gamma^k((FG)^+)$ is generated by the set $$\begin{split}
\mathcal{M}_k=&\{(h_1-h_1^{-1})\cdots (h_k-h_k^{-1})(1-a^2)a:h_i\in
P, a\in (Q_8\times E)\setminus \zeta(Q_8\times E)\}\cup \\
&\{g(h_1-h_1^{-1})\cdots (h_k-h_k^{-1})(1-a^2)a:h_i\in P, a\in
(Q_8\times E)\setminus \zeta(Q_8\times E)\}.
\end{split}$$
From Theorem \[teorliengelQ8+\], we know that $N\simeq Q_8\times
E\times P$, where $E^2=1$, $P$ is a finite $p$-group and there exists $g\in G\setminus N$ such that $G\simeq
\left<Q_8,g\right>\times E\times P$, $(g,x)=(g,y)=1$ and $g^2=x^2$. By Lemma \[lema6Juhasz\], there exists $0 \neq
\widehat{P}(1-c)a\in \gamma^{m(p-1)}((FG)^+)$ for some $a\in
Q_8\times E$. In this way, $1+m(p-1)\leq \t((FG)^+)$. Furthermore, Lemma \[lema5Juhasz\] implies that $\tl((FG)^+)\leq \tnil(P)$.
To show (ii), consider the symmetric elements $u_i=1-a_i(1+a_i^2)+x_i$, where $x_i=a_i(h_i+a_i^2h_i^{-1})\in
\mathcal S$, $a_i\in Q_8\times E$ and $h_i\in P$. Thus, $u_i=1+a_i(h_i-1)+a_i^3(h_i^{-1}-1)\in 1+FG\Delta(P)$. Since $FG\Delta(P)$ is a nilpotent ideal, we get that $1+FG\Delta(P)$ is a normal subgroup of $\mathcal{U}(FG)$ and in consequence $u_i$ is a unit in $FG$. We will prove, by induction, that $$\label{eq5.24}
(u_1,u_2,\ldots,u_n)\equiv 1+[x_1,x_2,\ldots,x_n] \pmod{
FG\Delta(P)^{n+1}}.$$
Since $u_1^{-1}u_2^{-1}\equiv 1 \pmod{FG\Delta(P)}$, Lemma \[lema5Juhasz\] implies that $$\begin{aligned}
(u_1,u_2)&=1+u_1^{-1}u_2^{-1}[u_1,u_2]=1+(u_1^{-1}u_2^{-1}-1)[u_1,u_2]+[u_1,u_2]\\
&\equiv 1+[u_1,u_2]\pmod{FG\Delta(P)^3}.\end{aligned}$$ We recall that $\widehat{a}_i=1+a_i+a_i^2+a_i^3$ and $1+a_i^2$, for each $a_i\in Q_8\times E$, are central elements of $FG$. So $$\begin{aligned}
[u_1,u_2]&=[1-a_1(1+a_1^2)+x_1,1-a_2(1+a_2^2)+x_2]\\
&=[x_1,x_2]+[a_1(1+a_1^2),a_2(1+a_2^2)]-[x_1,a_2(1+a_2^2)]-[a_1(1+a_1^2),x_2]\\
&=[x_1,x_2]+[\widehat{a}_1,\widehat{a}_2]-[x_1,\widehat{a}_2]-[\widehat{a}_1,x_2]\\
&=[x_1,x_2],\end{aligned}$$ which proves the congruence when $n=2$.
Suppose that , is true to $n-1$; that is $$(u_1,u_2,\ldots,u_{n-1})\equiv
1+[x_1,x_2,\ldots,x_{n-1}]\pmod{FG\Delta(P)^n}.$$ Then, Lemma \[lema5Juhasz\] and as $(u_1,u_2,\ldots,u_{n-1})^{-1}u_n^{-1}-1\in FG\Delta(P)$ imply $$\begin{aligned}
(u_1,u_2,&\ldots,u_{n})\\&=1+((u_1,u_2,\ldots,u_{n-1})^{-1}u_n^{-1}-1)[(u_1,u_2,\ldots,u_{n-1}),u_n]+[(u_1,u_2,\ldots,u_{n-1}),u_n]\\
&\equiv 1+[(u_1,u_2,\ldots,u_{n-1}),u_n]\pmod{FG\Delta(P)^{n+1}}\\
&\equiv
1+[[x_1,x_2,\ldots,x_{n-1}],1-a_n(1+a_n^2)+x_n]\pmod{FG\Delta(P)^{n+1}}\\
&\equiv
1+[x_1,x_2,\ldots,x_n]-[[x_1,x_2,\ldots,x_{n-1}],\widehat{a}_n]\pmod{FG\Delta(P)^{n+1}}\\
&\equiv 1+[x_1,x_2,\ldots,x_n] \pmod{FG\Delta(P)^{n+1}},\end{aligned}$$ and the statement is true for all $n\geq 2$.
Let $n=\tnil(P)-1$. If $\t((FG)^+)=\tnil(P)$, then there are $x_1,\ldots,x_n\in \mathcal{S}$ such that $[x_1,\ldots,x_n]\neq 0$. Thus, by the congruence , $\gamma_n(\mathcal{U}^+(FG))\neq 1$. So $n\leq
\cl(\mathcal{U}^+(FG))$. Moreover, we know that $\cl(\mathcal{U}^+(FG))< \tl((FG)^+)\leq \tnil(P)=n+1$ and we get (ii).
Assume that $P$ is powerful. Then, by Lemma \[lema5.3.4\], we obtain $$x_{n,n}f_{1,\ldots,n}(1-c)a\equiv
2^nf_{1,\ldots,n}(1-c)a\pmod{FG\Delta(P)^{n+1}}.$$ Furthermore, if $h_i-1\in \Delta(P)^{k_i}$, then by $$h_i-h_i^{-1}=(h_i-1)-(h_i^{-1}-1)\equiv 2(h_i-1)
\pmod{\Delta(P)^{k_i+1}},$$ thus $$\begin{aligned}
x_{n,n}f_{1,\ldots,n}(1-c)a&\equiv 2^n
(h_1-h_1^{-1})\cdots(h_n-h_n^{-1})(1-c)a\\
&\equiv 2^{2n}(h_1-1)\cdots(h_n-1)(1-c)a \pmod{FG\Delta(P)^{n+1}}.\end{aligned}$$ It is clear that, if $n<\tnil(P)$, there exist $h_1,\ldots,h_n\in P$ such that $\prod_{i=1}^n(h_i-1)\neq 0$, and then $x_{n,n}M_n\neq 0$. Thus, $\tnil(P)\leq\t((FG)^+)$ and (iii) follows.
Finally, assume that $P$ is abelian. Let $a,b\in (Q_8\times
E)\setminus\zeta(Q_8\times E)$, and $h_1,h_2\in P$, such that $(a,b)\neq 1$. Then $$\label{eq5.21}
\begin{split}
[a(h_1+a^2h_1^{-1}),b(h_2+b^2h_2^{-1})]&=(h_2+b^2h_2^{-1})(h_1+a^2h_1^{-1})[a,b]\\
&=(h_2-h_2^{-1})(h_1-h_1^{-1})(1-c)ab.
\end{split}$$ If $\alpha\in FP$, $h\in P$, then $$\label{eq5.22}
\begin{split}
[\alpha(1-c)a,b(h+b^2h^{-1})]&=\alpha(1-c)(h+b^2h^{-1})[a,b]\\
&=\alpha(h-h^{-1})(1-c)^2ab=2\alpha(h-h^{-1})(1-c)ab.
\end{split}$$
$$\label{eq5.28}
\begin{split}
[a(h_1+a^2h_1^{-1}),bg(h_2-h_2^{-1})]&=(h_2-h_2^{-1})(h_1+a^2h_1^{-1})[a,bg]\\
&=g(h_2-h_2^{-1})(h_1+ch_1^{-1})(1-c)ab\\
&=g(h_2-h_2^{-1})(h_1-h_1^{-1})(1-c)ab,
\end{split}$$
and $$\label{eq5.29}
\begin{split}
[ag(h_1-h_1^{-1}),bg(h_2-h_2^{-1})]&=(h_2-h_2^{-1})(h_1-h_1^{-1})[ag,bg]\\
&=(h_2-h_2^{-1})(h_1-h_1^{-1})g^2(1-c)ab\\
&=-(h_2-h_2^{-1})(h_1-h_1^{-1})(1-c)ab.
\end{split}$$ The equations , and imply that $\gamma^2((FG)^+)=\mathcal{M}_2$. Suppose that $\gamma^{n-1}((FG)^+)=\mathcal{M}_{n-1}$ for some $n\geq 3$. Take $\alpha \in FP$, $h\in P$ and $a,b\in (Q_8\times E)\setminus
\zeta(Q_8\times E)$, such that $(a,b)\neq 1$. We get the following equalities: $$\label{eq5.30}
\begin{split}
[\alpha(1-c)a,gb(h-h^{-1})]&=\alpha(h-h^{-1})(1-c)[a,gb]\\
&=g\alpha(h-h^{-1})(1-c)[a,b]\\
&=g\alpha(h-h^{-1})(1-c)^2ab\\
&=2g\alpha(h-h^{-1})(1-c)ab,
\end{split}$$ and $$\label{eq5.31}
\begin{split}
[g\alpha(1-c)a,gb(h-h^{-1})]&=g^2\alpha(h-h^{-1})(1-c)[a,b]\\
&=c\alpha(h-h^{-1})(1-c)[a,b]\\
&=c\alpha(h-h^{-1})(1-c)^2ab\\
&=-2\alpha(h-h^{-1})(1-c)ab.
\end{split}$$ By substituting $f_{1,\ldots,n-1}$ for $\alpha$ in , and , we get that $[\mathcal{M}_{n-1},(FG)^+]=\mathcal{M}_n$ and therefore $$\gamma^n((FG)^+)=[\gamma^{n-1}((FG)^+),(FG)^+]=[\mathcal{M}_{n-1},(FG)^+]=\mathcal{M}_n,$$ as we wanted to prove.
Acknowledgements {#acknowledgements .unnumbered}
================
The results of this paper are part of the author’s Ph.D. thesis, at the Instituto de Matemática e Estatística of the Universidade de São Paulo, under the guidance of Prof. César Polcino Milies. This work was partially supported by CAPES and CNPq. proc. 141857/2011-0 of Brazil.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Currently, accurate detection of natural phenomena, such as red tide, that adversely affect wildlife and human, using satellite images has been increasingly utilized. However, red tide detection on satellite images still remains a very hard task due to unpredictable nature of red tide occurrence, extreme sparsity of red tide samples, difficulties in accurate groundtruthing, etc. In this paper, we aim to tackle both the data sparsity and groundtruthing issues by primarily addressing two challenges: i) significant lack of hard examples of non-red tide that can enhance detection performance and ii) extreme data imbalance between red tide and non-red tide examples. In the proposed work, we devise a 9-layer fully convolutional network jointly optimized with two plug-in modules tailored to overcoming the two challenges: i) a hard negative example generator (HNG) to supplement the hard negative (non-red tide) examples and ii) cascaded online hard example mining (cOHEM) to ease the data imbalance. Our proposed network jointly trained with HNG and cOHEM provides state-of-the-art red tide detection accuracy on GOCI satellite images.'
author:
- |
Hyungtae Lee$^{\star\dagger}$ Heesung Kwon$^{\star}$ Wonkook Kim$^{\ddagger}$\
\
$^{\star}$US Army Research Lab $^{\dagger}$Booz Allen Hamilton $^{\ddagger}$Pusan National University
bibliography:
- 'egbib.bib'
title: Discovering and Generating Hard Examples for Training a Red Tide Detector
---
Introduction
============
Accurate and timely detection of short-term and long-term variations in naturally occurring phenomena (e.g. red tide, sea fog, yellow dust, etc.) that adversely affect wildlife as well as humans is highly critical. For instance, red tide is a toxic microscopic organism that inflicts serious damages to not only near-shore fishery but also large marine ecosystems in general. To investigate how, when, where these harmful natural phenomena occur and spread, many countries launched geostationary satellites closely observing areas of interest surrounding their territory. Accordingly, there have been a number of attempts to detect the harmful natural phenomena by analyzing remotely sensed images [@LDongTGARS2017; @AGermanIGARSS2017; @BPanJSTARS2017; @ARahmanIGARSS2016; @HSuIGARSS2016; @JZhaoIGARSS2014] from the geostationary satellites. In this paper, we propose a convolutional neural network (CNN)-based approach that can detect red tide embedded into a large scale image dataset.
![[ In the above figure, red tide appears as elongated red bands. The images are a false color image by combining the 6th, 4th and 1st band of the GOCI multi-spectral image representing the red, green and blue colors, respectively.]{}[]{data-label="fig:red_tide"}](red_tide_examples.png){width="\linewidth"}
To develop a CNN-based red tide detection approach, we have used the large-scale multi-spectral image dataset obtained from GOCI (Geostationary Ocean Color Imager) [@SChoKJRS2010] on a geostationary satellite. Several red tide examples on GOCI multi-spectral images are shown in Figure \[fig:red\_tide\]. Since the characteristics of biological properties of red tide do not clearly appear in the image, we used the information on real-world red tide occurrences reported by NIFS (National Institute of Fisheries Science) [@NIFS] of South Korea. However, NIFS manually examined red tide occurrence only at a limited number of locations along the southern seashore of South Korea, certainly not being able to cover the entire area infested by red tide. Therefore, in training, we end up with having only a small number of spectral samples from a fraction of areas where red tide actually occur. In our work, we use the images taken in December where red tide do not occur due to the low water temperature as negative examples[^1]. Figure \[fig:goci\_img\] shows the GOCI images used for the positive (red tide) and negative (non-red tide) training examples, and the red tide region annotation of the positive image.
![[ In the middle figure, the red tide label is marked as a white pixel.]{}[]{data-label="fig:goci_img"}](goci_images.pdf){width="\linewidth"}
There are two challenges to use GOCI images and their ground truth labels for training the red tide detection. First, the spectral characteristics of the images taken in December are very different from those of the images taken in the summer when red tide mostly occurs because the marine environment in summer and winter is very different. Therefore, the negative examples from the December images do not normally represent the non-red tide area from the images collected in summer months. Hence, using the negative examples from the winter season in training may fail to accurately separate the red tide area from non-red tide area in summer images generating a high level of false positives. Second, the imbalance between the number of positive examples and the number of negative examples is quite significant; the number of positive examples is in the order of one hundred pixels per image, and the number of negative examples is about 31M pixels per image as all the pixels of the GOCI image (5567$\times$5685) taken in winter are used as negative examples. Since class distributions are significantly unbalanced, the resulting detection model trained on the unbalanced data would be also significantly biased and inaccurate.
To overcome the first challenge, we implemented a generator based on CNN which artificially creates spectral signals with non-red tide property whose spectral characteristics are somewhat close to those of red tide examples. We train the generator in such a way that the artificially generated negative examples have non-red tide property but are recognized by the red tide detection model as red tide (i.e. false positive) via the generative adversarial learning (GAN) [@IGoodfellowNIPS2014]. To prevent the generated examples from losing typical non-red tide characteristics, we trained a network to distinguish the real negative examples from the artificially generated non-red tide examples, which serves as *discriminator* in the GAN framework [@IGoodfellowNIPS2014]. Accordingly, the red tide detection network and the discriminator become the counterparts that the generator should deceive. When the generator takes non-red tide examples as input, its output examples are used as hard negative examples–non-red tide samples that lie close to the boundaries of red-tide regions–in re-training the red tide detection model. Therefore, this generator is called a *hard negative generator* (HNG).
In order to cope with a large imbalance between positive and negative examples (the second challenge), we redesigned the online hard example mining (OHEM) [@AShrivastavaCVPR2016] in training the network. Shrivastava et al. introduced OHEM which, for every iteration, sampled a small number of examples with a high loss from the overwhelming examples and use them in training CNN. We used OHEM in a cascaded way to mine a huge volume of negative examples. First, we have taken all positive examples and a randomly chosen subset of negative examples and then apply OHEM to the selected examples to choose hard examples. We called this learning strategy as a *cascaded online hard example mining* (cOHEM).
For the red tide detection model, we implemented a 9-layer fully convolutional network (FCN) inspired by [@HLeeTIP2017]. The FCN architecture proved to be suitable for pixel-wise classification where classification is performed for each pixel in an image by using the network trained over only a few pixels [@HLeeIGARSS2018; @HLeeIGARSS2016; @HLeeTIP2017]. Red tide detection is a pixel-wise classification problem as opposed to image segmentation because each pixel region is classified in test based on only a small number of red tide labels available in training. We conducted extensive experiments to find out how HNG and cOHEM jointly address the problems that arise in training the proposed network on GOCI images. Experiments have confirmed that both modules used together significantly enhance red tide detection performance. To show that the proposed network can be easily extended to other tasks, we also applied the network to pixel-wise hyperspectral image classification.
In this paper, we made the following contributions:
- Augmented the training dataset by generating hard negative examples via HNG to reduce false alarms.
- Used cOHEM to cope with significantly unbalanced class distribution of GOCI images.
- Jointly optimized a 9 layer FCN with cOHEM and HNG.
- Provided the state-of-the-art accuracy on red tide detection and hyperspectral image classification.
Related Works {#sec:related_work}
=============
[**CNN used for Detecting Natural Phenomena in Marine Environment.**]{} Since CNN have provided promising performance in image classification, there have been several attempts to use it in the marine environment. CNNs have been effectively used for detection of coral reefs [@AKingCVPRW2018; @AMahmoodICIP2016], classification of fish [@GDingOCEANS2017; @ZGeICIP2015; @RMandalIJCNN2018], detection of oil from shipwreck [@MNietoHidalgoTGARS2018], and so on. However, applying deep neural network to detect objects-of-interest in the marine environment has been quite limited due mainly to difficulties in acquiring large amounts of annotated data unlike general object detection applications. In this paper, we devise a CNN training strategy coupled with advanced network architecture tailored to red tide detection while minimizing human labeling efforts.\
[**Training Generator via Adversarial Learning.**]{} Szegedy et al. [@CSzegedyICLR2014] introduced a method to generate an adversarial image by adding perturbation so that it is misclassified by a CNN-based recognition approach. These perturbed images become adversarial images to the recognition approach. Goodfellow et al. [@IGoodfellowNIPS2014] introduce two models: a generator that captures the data distribution and a discriminator that estimates the probability that a example came from the training data rather than the generator. A generator and a discriminator are trained at the same time in a direction to interfere with each other. This is called an adversarial learning framework.
Radford et al. [@ARadfordICLR2016] devised an image generation approach based on CNN by adopting this adversarial learning framework. Wang et al. [@XWangCVPR2017] used the adversarial learning framework to train a network that creates artificial occlusion and deformation on images. An object detection network is trained against this adversary to improve performance. We also use adversarial learning for training a hard negative generation network.\
[**Hard Example Mining.**]{} Sung and Poggio [@KSungMITAIMemo1994] firstly introduced hard negative mining (also known as bootstrapping) that trains the initial model with randomly chosen negatives and adapts the model to a hard negatives that consist of false positives of the initial model. Thereafter, hard example mining has been widely used in various applications such as pedestrian detection [@LBourdevECCV2010; @NDalalCVPR2005], human pose estimation [@LBourdevICCV2009; @HLeeACCV2012], action recognition [@SMajiCVPR2011], event recognition [@HLeeWACV2015], object detection [@PFelzenszwalbTPAMI2010; @RGirshickTPAMI2016; @KHeTPAMI2015; @TMalisiewiczICCV2011], and so on. There are alternative ways to find hard examples using heuristic [@RGirshickICCV2015] or other hard example selection algorithms [@SRenTPAMI2017; @JUijlingsIJCV2013], which avoid training multiple times.
Shrivastava et al. [@AShrivastavaCVPR2016] introduced online hard example mining which, for every training iteration, carries out hard example mining that chooses examples with high training loss. However, it is too exhaustive to evaluate all examples on each iteration. Hence, when using a very large example set like our problem, it is impractical to examine all examples for each iteration. Therefore, we use OHEM in a cascaded fashion to randomly select a subset of examples and then perform efficient mining on it.
Red Tide Detection Approach {#sec:pixelCNN}
===========================
Red Tide Detection {#ssec:detection}
------------------
In this section, we describe the proposed CNN-based red tide detection algorithm. This approach takes the GOCI image as input and evaluates whether each pixel in the image belongs to a red tide area or not. Therefore, red tide detection can be considered as pixel-wise classification. The architecture of the proposed approach is built on a network introduced by [@HLeeTIP2017], which is known to be suitable for pixel-wise classification. We apply a sliding window method to deal with limited GPU memory when processing GOCI images.\
![[ The training and test architectures have slightly different initial multi-scale filter banks tailored to pixel-wise classification. The weights of the convolutional layers do not change between training and test.]{}[]{data-label="fig:architecture"}](architecture2.pdf){width="\linewidth"}
[**Pixel-wise Classification.**]{} Pixel-wise classification is similar to image segmentation in that it clusters pixels with similar characteristics within each category. While the objective loss of image segmentation is calculated by comparing pixel-level labels of the entire image with the output of the network in training, pixel-wise classification is trained to minimize the difference between the labels of several pixels and their network outputs. In GOCI images, a red tide label is not available for all the red tide pixels, so red tide detection can be treated as a pixel-wise classification problem.\
{width="\linewidth"}
[**Architecture.**]{} The architecture of the proposed red tide detection model is shown in Figure \[fig:architecture\]. In order to cope with the pixel-wise classification of red tide detection, we use a 9-layer fully convolutional network (FCN) which intakes an image of arbitrary size. The network takes image patches of 25$\times$25 as input during training while an image patch of a certain dimension determined by the maximum size of GPU memory is fed into the network in test time.
The initial module of the network is a multi-scale filter bank which consists of convolutional filters with four different sizes (1$\times$1, 5$\times$5, 9$\times$9, and 13$\times$13). The architecture of the multi-scale filter bank is slightly different between training and testing. Given an image patch of 25$\times$25 in training, each of $k\times k$ filters is convolved with a patch of $(2k-1)\times(2k-1)$ centered on the 25$\times$25 patch. The size of the smaller patch, $(2k-1)\times(2k-1)$, is determined in such a way that each convolution always includes the center of the larger 25$\times$25 patch. For example, when applying a 5$\times$5 filter to a 9$\times$9 patch, the 5$\times$5 window always contains the center pixel of the 25$\times$25 patch being evaluated. After the initial convolution, a max pooling is applied to the outputs of the convolutional filters so that those pooled feature maps have a size of 1$\times$1 except for the 1$\times$1 convolution. In the test, four filters are applied to the same patch of the same size. These convolutions use appropriate padding to ensure that the four pooled feature maps have the same size. For both training and testing, four output feature maps are concatenated and then fed to the second convolutional layer. Accordingly, due to the multi-scale filter bank architecture, the receptive field of our network becomes 25$\times$25 and the network uses spatial information based on this receptive field when evaluating each pixel.
In training, dropout layers commonly used to solve the overfitting issue to some extent are added at the end of the $7^{th}$ and $8^{th}$ layers. The rest of the network is same in training and testing. Specifically, the binary sigmoid classifier which is useful for either single-label or multi-label classification is used for the output layer so that we can perform the tasks of identifying other natural phenomena (e.g. sea fog, yellow dust, etc.) from the GOCI images later using the same architecture.\
[**Sliding Window Strategy.**]{} The proposed method cannot process a very large GOCI image at once because of limitations of GPU memory, where the size of the image is 5567$\times$5685. To overcome this issue, we use a sliding window based strategy, where each window size is $H\times W$. In our experiment, we set $H$ and $W$ to 600. Considering our network’s receptive field, only the scores corresponding to the central $(H-24)\times(W-24)$ region are used as the final output.
Training: Adopting HNG and cOHEM {#ssec:training}
--------------------------------
[**Hard Negative Generator (HNG).**]{} We introduce a HNG network that generates hard negatives that are likely classified by the red tide detection network as false positives. The HNG architecture is designed as a 10-layered conv-deconv network consisting of eight convolutional layers and two deconvolutional layers, which is inspired by U-Net with high image generation capability [@HCaiCVPR2018; @DFourureBMVC2017; @WLiuCVPR2018; @HNohICCV2015; @ORonnebergerMICCAI2015], as shown in Figure \[fig:hng\]. We aim to achieve two goals in the training of the HNG network. First, HNG must be able to fool the red tide detection network so that the generated examples are incorrectly classified as red tides. Second, examples generated by HNG should have typical non-red-tide spectral characteristics. To do this, we introduce a network that distinguishes real examples from artificially generated ones, and HNG is trained to deceive the network referred to as the *discriminator* in the GAN framework, Accordingly, examples generated by HNG become hard examples that are designed to maximize the losses of the red tide detection network and the discriminator, which conflicts with the two networks’ objectives. Training process of HNG is shown in Figure \[fig:hng\].
To mathematically formulate the process of generating hard negative examples, the red tide detection network and its loss are represented by $\CMcal{F}_{rtd}$ and $\CMcal{L}_{rtd}$, respectively. The red tide detection network is trained by minimizing its loss expressed as: $$ \CMcal{L}_{rtd}(E,L_{rtd}) = \CMcal{H}(\CMcal{F}_{rtd}{(E)},L_{rtd}),$$ where $E$ and $L_{rtd}$ are training examples and their associated labels, respectively. For each example $e\in E$, its red tide labels $l_{rtd} \in L_{rtd}$ can be either 1 (red tide) or 0 (non-red tide). $\CMcal{H}(p,q)$ is the cross-entropy for the distributions $p$ and $q$.
The discriminator and its loss are denoted as $\CMcal{F}_{d}$ and $\CMcal{L}_{d}$, respectively, in Equation \[eq:discr\_train\]. The discriminator architecture consists of four convolutional layers and one fully-connected layer, as shown in Figure \[fig:hng\]. The discriminator is optimized by minimizing its loss which is expressed as: $$\CMcal{L}_{d}(N) = \CMcal{H}{(\CMcal{F}_{d}{(N)},\bf{1})} + \CMcal{H}{(\CMcal{F}_{d}{(\CMcal{F}_{hng}{(N)})},\bf{0})},
\label{eq:discr_train}$$ where $N$ is a set of real negative examples. The discriminator’s labels can be either 1 (real example) or 0 (artificially generated example). $\CMcal{F}_{hng}$ denotes the HNG network. The randomly selected 256 negative examples and the generated examples are used for every iteration when training the discriminator. The generated examples are the output of HNG taking the randomly selected real negative examples as input.
The objective of HNG is to generate negative examples incorrectly classified as red tide by the red tide detection network and as real negative examples by the discriminator. Accordingly, HNG loss ($\CMcal{L}_{hng}$) can be expressed as: $$\CMcal{L}_{hng}(N) = \CMcal{H}(\CMcal{F}_{rtd}{(\CMcal{F}_{hng}{(N)})},{\bf 1})+\CMcal{H}{(\CMcal{F}_{d}{(\CMcal{F}_{hng}{(N)})},{\bf 1})},
\label{eq:hng_train}$$ where ${\bf 1}$ indicates that labels associated with the generated negative examples are red tide for $\CMcal{F}_{rtd}$ or real negatives for $\CMcal{F}_{d}$. HNG can be trained by minimizing $\CMcal{L}_{hng}(N)$, where $\CMcal{F}_{hng}{(N)}$ becomes adversarial examples for the red tide detection network and the discriminator.\
[**Cascaded Online Hard Example Mining (cOHEM).**]{} OHEM introduced by Shrivastava et al. [@AShrivastavaCVPR2016] builds batches by collecting hard examples to collectively minimize objective loss throughout the entire examples. However, in problems like ours with a very large number of examples, it takes quite long time to find hard examples, so we used OHEM in a cascaded fashion. First, we build a large pool of randomly chosen examples and then carry out OHEM on the pool. This method is called cascaded OHEM (cOHEM).
![[ cOHEM finds hard examples ($E_{hard}$) through two-stage sampling (*Negative Random Sampler* and *Hard Example Sampler*) for every iteration. (RT Detect Network: the proposed red tide detection network, $|P|$: the size of the positive example set.)]{}[]{data-label="fig:ohem"}](cohem.pdf){width="\linewidth"}
Figure \[fig:ohem\] shows the training protocol of the network using cOHEM. For every pixel labeled with red tide, a 25$\times$25 area around each pixel is collected and used as a positive example. In *Negative Random Sampler*, we randomly select 100 regions of a size 37$\times$37 from a negative image, which is equivalent to 16,900 examples of size 25$\times$25. In our experiments, we found that it is important to randomly select multiple regions instead of one wide region for negative examples, in order to effectively improve the accuracy of red tide detection using cOHEM. For all selected examples, their classification losses are calculated by feeding the examples to the network. Note that this loss represents the extent to which the current network correctly classifies each example. *Hard Example Sampler* randomly selects the high loss examples with predetermined batch size. Then the proposed network is trained with these batches consisting of hard examples.\
[**Training Red Tide Network with HNG and cOHEM.**]{} For training the red tide detection network using hard negatives generated by HNG and hard examples mined by cOHEM, we adopt a 3-stage training algorithm. The first stage is to train the red tide detection network using cOHEM. In the second stage, HNG and the discriminator are trained, as shown in Figure \[fig:hng\]. In this stage, the discriminator is trained first with HNG weights unchanged and then HNG is trained while keeping the discriminator and the red tide detection network fixed. In the last stage, the red tide detection network is updated by using examples augmented by HNG. The real negative examples and artificially generated negative examples are combined into a negative set, and then hard negatives are discovered via cOHEM. In the third stage, HNG weights are fixed.
For the first and third stage, all the 25$\times$25 positive regions and 100 of the 37$\times$37 negative regions are used for training. For the second stage, 256 of the 25$\times$25 negative regions are used. Note that our FCN can intake an input image of arbitrary size.
The proposed networks are optimized by using a mini-batch Stochastic Gradient Descent (SGD) approach with a batch size of 256 examples, momentum of 0.9, and weight decay of 0.0005. For the red tide detection network, its training objective is to minimize the cross entropy losses between the red tide labels and the final output scores. Each batch consists of examples extracted from one positive image with red tide occurrence and one negative image with no red tide occurrence. The positive-to-negative ratio in each batch is set to 1:3.
To reduce overfitting in training, data augmentation is carried out. Since a GOCI image is taken from a top view, training examples are augmented by mirroring across the horizontal, vertical, and diagonal axes. This mirroring can be performed in one direction or in multiple directions. This will increase the number of examples by 8 times.
When training the red tide detection network, all learnable layers except for the layers of residual modules ($3^{rd}$, $4^{th}$, $5^{th}$, and $6^{th}$ layers) are initialized according to Gaussian distribution with zero mean and 0.01 standard deviation. The layers of the residual modules are initialized according to Gaussian distribution with a mean of zero and a standard deviation of 0.005. All layers of HNG except for the last layer are initialized according to Gaussian distribution with a mean of zero and a standard deviation of 0.02. The last layer is initialized according to Gaussian distribution with a mean of zero and a standard deviation of 50.
{width="\linewidth"} \[fig:roc\]
GOCI Satellite Images {#sec:goci_image}
=====================
GOCI (Geostationary Ocean Color Imager) acquires multispectral images from a large area surrounding the Korean peninsula. The GOCI image [@SChoKJRS2010] has 8 channels consisting of 6 visible and 2 near infrared (NIR) frequency bands and 500 m spatial resolution. The size of the GOCI image is 5567$\times$5685. Some examples of GOCI images are shown in Figure \[fig:goci\_img\].
In this paper, we use GOCI images taken in July, August, and December of 2013 to evaluate our red tide detection network. Images from July and August where red tide occurred are used as positive images, and images from December are used as negative images. Based on some conditions such as the atmosphere, we chose 8 images in July and August and 4 images in December. Half of them were used for training and the other half for testing.
In order to label red tide pixels, we used the red tide information reported by NIFS (National Institute of Fisheries) of South Korea which directly tested seawater from a ship. NIFS examined red tide occurrence only at a limited number of locations; so it is not possible to cover the entire red tide areas. Furthermore, the red tide positions indicated in the reports were not very accurate due to the error-prone manual process that included mapping geo-coordinates of red tide locations onto GOCI images. Hence, we have extended potential red tide regions up to 25 km (50 pixel distance) in all directions from the red tide location indicated in the report and then labeled red tide with the help of experts. Approximately 100 pixels from each training image were sparsely labeled as a red tide area. We used pixels labeled as red tide as positive examples and all the pixels of the December images as negative examples.
Experiments {#sec:experiments}
===========
Evaluation Settings {#ssec:setting}
-------------------
[**Evaluation Metrics.**]{} We used two different metrics for the evaluation of the proposed network: the receiver operating characteristic (ROC) curve and the ROC variation curve. The ROC variation curve describes changes in the detection rate based on varying numbers of (true or false) detections per image (NDPI) instead of the false positive rate. This metric is especially useful when there are numerous unlabeled examples whose identity is unknown. Note that in a GOCI image only a fraction of red tide pixels are labeled and the rest of the image remains unlabeled. For quantitative analysis, we calculate the AUC (the area under the ROC curve) and ndpi@dr=0.25, ndpi@dr=0.5 and ndpi@dr=0.75 indicating the NDPI values when the detection rate reaches 0.25, 0.5, and 0.75, respectively.\
[**Network Training.**]{} The proposed networks are trained from scratch. When both HNG and cOHEM are used for training, we used a three-stage training strategy and trained the network with 1250 iterations for each stage. A base learning rate is 0.01 for the red tide detection network and HNG and 0.0001 for the discriminator. The base learning rate drops to a factor of 10 for every 500 iterations. When the three stage training strategy is not used to train the network (i.e. HNG is not adopted for training), we trained the network with 2500 iterations. A base learning rate is 0.01 and dropps by a factor of 10 for every 1K iterations.
Experimental Results {#ssec:performance}
--------------------
[**Baselines.**]{} We implement two baselines of SVM and CNN-based hyperspectral image classification approach [@HLeeTIP2017] by which our network is inspired. In SVM, a 25$\times$25 region centered on the pixel in test is used as a feature representing the center pixel. In order to know the advantages of adopting hard example mining, we applied conventional hard negative mining [@KSungMITAIMemo1994] to SVM training.\
[**Performance Comparison.**]{} Table \[tab:performance\] shows that our network trained by using both hard examples (via cOHEM) and generated hard negatives (via HNG) provides the highest accuracy in all four metrics. Using the HNG network was effective in improving the performance of our network and [@HLeeTIP2017]. It is also observed that adopting a hard example mining approach consistently improves the accuracy of all three methods as it efficiently eases the significant imbalance between red tide and non-red tide examples.
![[ The top images and the bottom images are input images and the corresponding red tide detection results from our approach, respectively. The first and second input images have red tide in them and the last image is a negative image without red tide.]{}[]{data-label="fig:results"}](results_horizontal2.jpg){width="\linewidth"}
Figure \[fig:roc\] shows the ROC variation curves for our network and baselines. From Figure \[fig:roc\], we can confirm that our network provides significantly enhanced detection performance over other techniques over the entire range of NDPI. Figure \[fig:results\] shows red tide detection results from our approach.
Analyzing HNG Network
---------------------
![[ Losses changed during training HNG and discriminator in 2nd training stage are shown.]{}[]{data-label="fig:training_loss"}](training_loss.pdf){width="0.85\linewidth"}
In Figure \[fig:training\_loss\], we observe that the HNG and discriminator converge successfully in the $2^{nd}$ training stage. This shows that HNG optimization overcomes the interference of the red tide detection network and the discriminator.
Some examples of generated hard negatives, the output of the successfully trained HNG, are presented in Figure \[fig:gen\_examples\]. In Figure \[fig:gen\_examples\], it can be observed that activated pixel regions (the $3^{rd}$ column of each set) tend to have different colors–i.e., different intensity values for some spectral bands–for certain areas than the ones covered by real non-red tide examples. Note that the activated regions in Figure \[fig:gen\_examples\] do not show typical characteristics of red tide regions–narrow elongate bands with sharp boundaries as shown in Figure \[fig:red\_tide\]. Furthermore, the activated regions by artificial non-red examples appear to be pink while the real red tide are in general red.
{width="\linewidth"}
\[fig:gen\_examples\]
{width="\linewidth"}
\[fig:training\_example\_portion\]
Figure \[fig:training\_example\_portion\] shows an evolution of the change of the ratio of training examples between real negative examples and generated negative examples within one batch as the $3^{rd}$ training stage progresses. Overall, the proportion of generated negative examples within a batch gradually decreases as the training iteration increases. To analyze the relationship among three different types of training examples–i.e. positive, real negative, and artificially generated negative examples– after training HNG at the 2nd training stage, we use a visualization technique called t-distributed stochastic neighborhood embedding (t-SNE) [@LMaatenJMLR2008], as shown in Figure \[fig:tsne\_analysis\]. For the visualization, 500 examples are randomly chosen separately from real positive and negative examples and the artificially generated examples are the output of HNG, which takes real negative examples as input. We use the output of $8^{th}$ convolutional layer (after ReLU layer) as the features of the selected examples. As shown on the left side of Figure \[fig:tsne\_analysis\], the artificially generated negative examples are placed between the positive and real negative examples. This shows that the generated negative examples become hard examples when retraining the red tide detection network. When training HNG without using the discriminator, the artificially generated negative examples are no longer hard examples as they are placed far to the right of the positive and real negative examples, as shown on the right side of Figure \[fig:tsne\_analysis\].
Finding Sampling Strategy of cOHEM {#ssec:sampling_of_cOHEM}
----------------------------------
We compare various negative example sampling strategies of cOHEM with respect to detection accuracy and training time. Negative example sampling can change based on two factors: a window size and the number of windows. In Table \[tab:cOHEM\_strategy\], we compare four different sampling strategies with different factors. The first three strategies are chosen to maintain a similar number of negative examples as used in training. The last strategy (192 windows of 25$\times$25) is training without using cOHEM. Accordingly, its training time is the shortest.
![[ Data points close to each other in the embedded space are likely to have similar spectral characteristics.]{}[]{data-label="fig:tsne_analysis"}](vis_w_discr.pdf "fig:"){width="0.4\linewidth"} ![[ Data points close to each other in the embedded space are likely to have similar spectral characteristics.]{}[]{data-label="fig:tsne_analysis"}](vis_wo_discr.pdf "fig:"){width="0.4\linewidth"}
In Table \[tab:cOHEM\_strategy\], the third strategy gives the best performance in terms of AUC and training time. This observation indicates that increasing spatial diversity of sampling is important in providing competitive performance. Therefore, even though the third strategy requires a large amount of memory due to the large pixels, it is adopted in our training approach.
Hyperspectral Image Classification
==================================
Our approach can also be easily generalized to hyperspectral image (HSI) classification, which also requires pixel-wise classification. For the HSI classification, we have used three benchmark datasets: Indian pines, Salinas, and PaviaU. For each HSI dataset, 200 pixels randomly selected from each category are used for training and all the remaining pixels are used for testing.\
[**Hard Example Generator (HEG).**]{} When applying the proposed training strategy to hyperspectral image classification, we replace HNG with a hard example generator (HEG) that generates hard examples of individual material categories instead of hard negatives. The HEG is trained to generate examples misclassified by the HSI classification network. For each example $e$, its adversarial training label ($l^a_{hsi}$) is the category with the largest loss among all categories except for the class the example belongs to: $$l^{a}_{hsi} = \operatorname*{arg\,max}_{l\in C,~l\neq l_{hsi}}{\CMcal{L}_{hsi}(e,l)},$$ where $l_{hsi}$ is a label associated with the example $e$ and $C$ is the set of all categories. HEG optimization is carried out with Equation \[eq:hng\_train\] by replacing $L_{rtd}$ (denoted as [**1**]{} in the Equation) with $L^{a}_{hsi}=\{l^{a}_{hsi}\}$.\
[**Results.**]{} Table \[tab:HICarrucary\] shows the classification accuracy for the three datasets. For all the three datasets, our 9-layer CNN outperforms baseline introduced in [@HLeeTIP2017]. Our strategy to discover and generate hard examples further improves performance by at least 0.44%.
-------------- ------------------ ------------------ -------------------------------------------------------------------------------
Dataset [@HLeeTIP2017] Ours + cOHEM & HNG
Indian pines 93.61 $\pm$ 0.56 95.17 $\pm$ 0.48 **[96.43]{} $\pm$ **[0.42]{}\
Salinas & 95.07 $\pm$ 0.23 & 96.09 $\pm$ 0.40 & **[96.53]{} $\pm$ **[0.48]{}\
PaviaU & 95.97 $\pm$ 0.46 & 96.27 $\pm$ 0.51 & **[97.33]{} $\pm$ **[0.37]{}\
************
-------------- ------------------ ------------------ -------------------------------------------------------------------------------
: [ When calculating the accuracy for our network, training and evaluation protocols follow those in [@HLeeTIP2017].]{}
\[tab:HICarrucary\]
Note that the problems we encountered with GOCI images (i.e. extreme sparsity of red tide samples, difficulties in accurate groundtruthing, etc.) are not normally observed on hyperspectral images. Enhanced performance for HSI classification verifies that our strategy of discovering and artificially generating hard examples is also applicable to other related problem domains, such as HSI analysis.\
[**Acknowledgement.**]{} This research was supported by the “Development of the integrated data processing system for GOCI-II” funded by the Ministry of Ocean and Fisheries, Korea.
[^1]: In South Korea, summer is in July and August and winter in December. Red tide occurs mainly in summer when the water temperature is high.
| {
"pile_set_name": "ArXiv"
} |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.